Upload 104 files

audiosr/__init__.py

@@ -1,2 +1,2 @@
-from .utils import seed_everything, save_wave, get_time, get_duration, read_list
-from .pipeline import *
+from .utils import seed_everything, save_wave, get_time, get_duration, read_list
+from .pipeline import *

audiosr/__main__.py

@@ -1,123 +1,123 @@
-#!/usr/bin/python3
-import os
-import torch
-import logging
-from audiosr import super_resolution, build_model, save_wave, get_time, read_list
-import argparse
-
-os.environ["TOKENIZERS_PARALLELISM"] = "true"
-matplotlib_logger = logging.getLogger('matplotlib')
-matplotlib_logger.setLevel(logging.WARNING)
-
-parser = argparse.ArgumentParser()
-
-parser.add_argument(
-    "-i",
-    "--input_audio_file",
-    type=str,
-    required=False,
-    help="Input audio file for audio super resolution",
-)
-
-parser.add_argument(
-    "-il",
-    "--input_file_list",
-    type=str,
-    required=False,
-    default="",
-    help="A file that contains all audio files that need to perform audio super resolution",
-)
-
-parser.add_argument(
-    "-s",
-    "--save_path",
-    type=str,
-    required=False,
-    help="The path to save model output",
-    default="./output",
-)
-
-parser.add_argument(
-    "--model_name",
-    type=str,
-    required=False,
-    help="The checkpoint you gonna use",
-    default="basic",
-    choices=["basic","speech"]
-)
-
-parser.add_argument(
-    "-d",
-    "--device",
-    type=str,
-    required=False,
-    help="The device for computation. If not specified, the script will automatically choose the device based on your environment.",
-    default="auto",
-)
-
-parser.add_argument(
-    "--ddim_steps",
-    type=int,
-    required=False,
-    default=50,
-    help="The sampling step for DDIM",
-)
-
-parser.add_argument(
-    "-gs",
-    "--guidance_scale",
-    type=float,
-    required=False,
-    default=3.5,
-    help="Guidance scale (Large => better quality and relavancy to text; Small => better diversity)",
-)
-
-parser.add_argument(
-    "--seed",
-    type=int,
-    required=False,
-    default=42,
-    help="Changing this value (any integer number) will lead to a different generation result.",
-)
-
-parser.add_argument(
-    "--suffix",
-    type=str,
-    required=False,
-    help="Suffix for the output file",
-    default="_AudioSR_Processed_48K",
-)
-
-args = parser.parse_args()
-torch.set_float32_matmul_precision("high")
-save_path = os.path.join(args.save_path, get_time())
-
-assert args.input_file_list is not None or args.input_audio_file is not None,"Please provide either a list of audio files or a single audio file"
-
-input_file = args.input_audio_file
-random_seed = args.seed
-sample_rate=48000
-latent_t_per_second=12.8
-guidance_scale = args.guidance_scale
-
-os.makedirs(save_path, exist_ok=True)
-audiosr = build_model(model_name=args.model_name, device=args.device)
-
-if(args.input_file_list):
-    print("Generate audio based on the text prompts in %s" % args.input_file_list)
-    files_todo = read_list(args.input_file_list)
-else: 
-    files_todo = [input_file]
-    
-for input_file in files_todo:
-    name = os.path.splitext(os.path.basename(input_file))[0] + args.suffix
-    
-    waveform = super_resolution(
-        audiosr,
-        input_file,
-        seed=random_seed,
-        guidance_scale=guidance_scale,
-        ddim_steps=args.ddim_steps,
-        latent_t_per_second=latent_t_per_second
-    )
-    save_wave(waveform, inputpath=input_file, savepath=save_path, name=name, samplerate=sample_rate)
+#!/usr/bin/python3
+import os
+import torch
+import logging
+from audiosr import super_resolution, build_model, save_wave, get_time, read_list
+import argparse
+
+os.environ["TOKENIZERS_PARALLELISM"] = "true"
+matplotlib_logger = logging.getLogger('matplotlib')
+matplotlib_logger.setLevel(logging.WARNING)
+
+parser = argparse.ArgumentParser()
+
+parser.add_argument(
+    "-i",
+    "--input_audio_file",
+    type=str,
+    required=False,
+    help="Input audio file for audio super resolution",
+)
+
+parser.add_argument(
+    "-il",
+    "--input_file_list",
+    type=str,
+    required=False,
+    default="",
+    help="A file that contains all audio files that need to perform audio super resolution",
+)
+
+parser.add_argument(
+    "-s",
+    "--save_path",
+    type=str,
+    required=False,
+    help="The path to save model output",
+    default="./output",
+)
+
+parser.add_argument(
+    "--model_name",
+    type=str,
+    required=False,
+    help="The checkpoint you gonna use",
+    default="basic",
+    choices=["basic","speech"]
+)
+
+parser.add_argument(
+    "-d",
+    "--device",
+    type=str,
+    required=False,
+    help="The device for computation. If not specified, the script will automatically choose the device based on your environment.",
+    default="auto",
+)
+
+parser.add_argument(
+    "--ddim_steps",
+    type=int,
+    required=False,
+    default=50,
+    help="The sampling step for DDIM",
+)
+
+parser.add_argument(
+    "-gs",
+    "--guidance_scale",
+    type=float,
+    required=False,
+    default=3.5,
+    help="Guidance scale (Large => better quality and relavancy to text; Small => better diversity)",
+)
+
+parser.add_argument(
+    "--seed",
+    type=int,
+    required=False,
+    default=42,
+    help="Change this value (any integer number) will lead to a different generation result.",
+)
+
+parser.add_argument(
+    "--suffix",
+    type=str,
+    required=False,
+    help="Suffix for the output file",
+    default="_AudioSR_Processed_48K",
+)
+
+args = parser.parse_args()
+torch.set_float32_matmul_precision("high")
+save_path = os.path.join(args.save_path, get_time())
+
+assert args.input_file_list is not None or args.input_audio_file is not None,"Please provide either a list of audio files or a single audio file"
+
+input_file = args.input_audio_file
+random_seed = args.seed
+sample_rate=48000
+latent_t_per_second=12.8
+guidance_scale = args.guidance_scale
+
+os.makedirs(save_path, exist_ok=True)
+audiosr = build_model(model_name=args.model_name, device=args.device)
+
+if(args.input_file_list):
+    print("Generate audio based on the text prompts in %s" % args.input_file_list)
+    files_todo = read_list(args.input_file_list)
+else: 
+    files_todo = [input_file]
+    
+for input_file in files_todo:
+    name = os.path.splitext(os.path.basename(input_file))[0] + args.suffix
+
+    waveform = super_resolution(
+        audiosr,
+        input_file,
+        seed=random_seed,
+        guidance_scale=guidance_scale,
+        ddim_steps=args.ddim_steps,
+        latent_t_per_second=latent_t_per_second
+    )
+    save_wave(waveform, inputpath=input_file, savepath=save_path, name=name, samplerate=sample_rate)

audiosr/clap/open_clip/__init__.py

@@ -1,25 +1,25 @@
-from .factory import (
-    list_models,
-    create_model,
-    create_model_and_transforms,
-    add_model_config,
-)
-from .loss import ClipLoss, gather_features, LPLoss, lp_gather_features, LPMetrics
-from .model import (
-    CLAP,
-    CLAPTextCfg,
-    CLAPVisionCfg,
-    CLAPAudioCfp,
-    convert_weights_to_fp16,
-    trace_model,
-)
-from .openai import load_openai_model, list_openai_models
-from .pretrained import (
-    list_pretrained,
-    list_pretrained_tag_models,
-    list_pretrained_model_tags,
-    get_pretrained_url,
-    download_pretrained,
-)
-from .tokenizer import SimpleTokenizer, tokenize
-from .transform import image_transform
+from .factory import (
+    list_models,
+    create_model,
+    create_model_and_transforms,
+    add_model_config,
+)
+from .loss import ClipLoss, gather_features, LPLoss, lp_gather_features, LPMetrics
+from .model import (
+    CLAP,
+    CLAPTextCfg,
+    CLAPVisionCfg,
+    CLAPAudioCfp,
+    convert_weights_to_fp16,
+    trace_model,
+)
+from .openai import load_openai_model, list_openai_models
+from .pretrained import (
+    list_pretrained,
+    list_pretrained_tag_models,
+    list_pretrained_model_tags,
+    get_pretrained_url,
+    download_pretrained,
+)
+from .tokenizer import SimpleTokenizer, tokenize
+from .transform import image_transform

audiosr/clap/open_clip/factory.py

@@ -1,276 +1,276 @@
-import json
-import logging
-import os
-import re
-from copy import deepcopy
-from pathlib import Path
-
-import torch
-
-from .model import CLAP, convert_weights_to_fp16
-from .openai import load_openai_model
-from .pretrained import get_pretrained_url, download_pretrained
-from .transform import image_transform
-
-_MODEL_CONFIG_PATHS = [Path(__file__).parent / f"model_configs/"]
-_MODEL_CONFIGS = {}  # directory (model_name: config) of model architecture configs
-
-
-def _natural_key(string_):
-    return [int(s) if s.isdigit() else s for s in re.split(r"(\d+)", string_.lower())]
-
-
-def _rescan_model_configs():
-    global _MODEL_CONFIGS
-
-    config_ext = (".json",)
-    config_files = []
-    for config_path in _MODEL_CONFIG_PATHS:
-        if config_path.is_file() and config_path.suffix in config_ext:
-            config_files.append(config_path)
-        elif config_path.is_dir():
-            for ext in config_ext:
-                config_files.extend(config_path.glob(f"*{ext}"))
-
-    for cf in config_files:
-        if os.path.basename(cf)[0] == ".":
-            continue  # Ignore hidden files
-
-        with open(cf, "r") as f:
-            model_cfg = json.load(f)
-            if all(a in model_cfg for a in ("embed_dim", "audio_cfg", "text_cfg")):
-                _MODEL_CONFIGS[cf.stem] = model_cfg
-
-    _MODEL_CONFIGS = {
-        k: v
-        for k, v in sorted(_MODEL_CONFIGS.items(), key=lambda x: _natural_key(x[0]))
-    }
-
-
-_rescan_model_configs()  # initial populate of model config registry
-
-
-def load_state_dict(checkpoint_path: str, map_location="cpu", skip_params=True):
-    checkpoint = torch.load(checkpoint_path, map_location=map_location)
-    if isinstance(checkpoint, dict) and "state_dict" in checkpoint:
-        state_dict = checkpoint["state_dict"]
-    else:
-        state_dict = checkpoint
-    if skip_params:
-        if next(iter(state_dict.items()))[0].startswith("module"):
-            state_dict = {k[7:]: v for k, v in state_dict.items()}
-    # for k in state_dict:
-    #     if k.startswith('transformer'):
-    #         v = state_dict.pop(k)
-    #         state_dict['text_branch.' + k[12:]] = v
-    return state_dict
-
-
-def create_model(
-    amodel_name: str,
-    tmodel_name: str,
-    pretrained: str = "",
-    precision: str = "fp32",
-    device: torch.device = torch.device("cpu"),
-    jit: bool = False,
-    force_quick_gelu: bool = False,
-    openai_model_cache_dir: str = os.path.expanduser("~/.cache/clip"),
-    skip_params=True,
-    pretrained_audio: str = "",
-    pretrained_text: str = "",
-    enable_fusion: bool = False,
-    fusion_type: str = "None"
-    # pretrained_image: bool = False,
-):
-    amodel_name = amodel_name.replace(
-        "/", "-"
-    )  # for callers using old naming with / in ViT names
-    pretrained_orig = pretrained
-    pretrained = pretrained.lower()
-    if pretrained == "openai":
-        if amodel_name in _MODEL_CONFIGS:
-            logging.info(f"Loading {amodel_name} model config.")
-            model_cfg = deepcopy(_MODEL_CONFIGS[amodel_name])
-        else:
-            logging.error(
-                f"Model config for {amodel_name} not found; available models {list_models()}."
-            )
-            raise RuntimeError(f"Model config for {amodel_name} not found.")
-
-        logging.info(f"Loading pretrained ViT-B-16 text encoder from OpenAI.")
-        # Hard Code in model name
-        model_cfg["text_cfg"]["model_type"] = tmodel_name
-        model = load_openai_model(
-            "ViT-B-16",
-            model_cfg,
-            device=device,
-            jit=jit,
-            cache_dir=openai_model_cache_dir,
-            enable_fusion=enable_fusion,
-            fusion_type=fusion_type,
-        )
-        # See https://discuss.pytorch.org/t/valueerror-attemting-to-unscale-fp16-gradients/81372
-        if precision == "amp" or precision == "fp32":
-            model = model.float()
-    else:
-        if amodel_name in _MODEL_CONFIGS:
-            logging.info(f"Loading {amodel_name} model config.")
-            model_cfg = deepcopy(_MODEL_CONFIGS[amodel_name])
-        else:
-            logging.error(
-                f"Model config for {amodel_name} not found; available models {list_models()}."
-            )
-            raise RuntimeError(f"Model config for {amodel_name} not found.")
-
-        if force_quick_gelu:
-            # override for use of QuickGELU on non-OpenAI transformer models
-            model_cfg["quick_gelu"] = True
-
-        # if pretrained_image:
-        #     if 'timm_amodel_name' in model_cfg.get('vision_cfg', {}):
-        #         # pretrained weight loading for timm models set via vision_cfg
-        #         model_cfg['vision_cfg']['timm_model_pretrained'] = True
-        #     else:
-        #         assert False, 'pretrained image towers currently only supported for timm models'
-        model_cfg["text_cfg"]["model_type"] = tmodel_name
-        model_cfg["enable_fusion"] = enable_fusion
-        model_cfg["fusion_type"] = fusion_type
-        model = CLAP(**model_cfg)
-
-        if pretrained:
-            checkpoint_path = ""
-            url = get_pretrained_url(amodel_name, pretrained)
-            if url:
-                checkpoint_path = download_pretrained(url, root=openai_model_cache_dir)
-            elif os.path.exists(pretrained_orig):
-                checkpoint_path = pretrained_orig
-            if checkpoint_path:
-                logging.info(
-                    f"Loading pretrained {amodel_name}-{tmodel_name} weights ({pretrained})."
-                )
-                ckpt = load_state_dict(checkpoint_path, skip_params=True)
-                model.load_state_dict(ckpt)
-                param_names = [n for n, p in model.named_parameters()]
-                # for n in param_names:
-                #     print(n, "\t", "Loaded" if n in ckpt else "Unloaded")
-            else:
-                logging.warning(
-                    f"Pretrained weights ({pretrained}) not found for model {amodel_name}."
-                )
-                raise RuntimeError(
-                    f"Pretrained weights ({pretrained}) not found for model {amodel_name}."
-                )
-
-        if pretrained_audio:
-            if amodel_name.startswith("PANN"):
-                if "Cnn14_mAP" in pretrained_audio:  # official checkpoint
-                    audio_ckpt = torch.load(pretrained_audio, map_location="cpu")
-                    audio_ckpt = audio_ckpt["model"]
-                    keys = list(audio_ckpt.keys())
-                    for key in keys:
-                        if (
-                            "spectrogram_extractor" not in key
-                            and "logmel_extractor" not in key
-                        ):
-                            v = audio_ckpt.pop(key)
-                            audio_ckpt["audio_branch." + key] = v
-                elif os.path.basename(pretrained_audio).startswith(
-                    "PANN"
-                ):  # checkpoint trained via HTSAT codebase
-                    audio_ckpt = torch.load(pretrained_audio, map_location="cpu")
-                    audio_ckpt = audio_ckpt["state_dict"]
-                    keys = list(audio_ckpt.keys())
-                    for key in keys:
-                        if key.startswith("sed_model"):
-                            v = audio_ckpt.pop(key)
-                            audio_ckpt["audio_branch." + key[10:]] = v
-                elif os.path.basename(pretrained_audio).startswith(
-                    "finetuned"
-                ):  # checkpoint trained via linear probe codebase
-                    audio_ckpt = torch.load(pretrained_audio, map_location="cpu")
-                else:
-                    raise ValueError("Unknown audio checkpoint")
-            elif amodel_name.startswith("HTSAT"):
-                if "HTSAT_AudioSet_Saved" in pretrained_audio:  # official checkpoint
-                    audio_ckpt = torch.load(pretrained_audio, map_location="cpu")
-                    audio_ckpt = audio_ckpt["state_dict"]
-                    keys = list(audio_ckpt.keys())
-                    for key in keys:
-                        if key.startswith("sed_model") and (
-                            "spectrogram_extractor" not in key
-                            and "logmel_extractor" not in key
-                        ):
-                            v = audio_ckpt.pop(key)
-                            audio_ckpt["audio_branch." + key[10:]] = v
-                elif os.path.basename(pretrained_audio).startswith(
-                    "HTSAT"
-                ):  # checkpoint trained via HTSAT codebase
-                    audio_ckpt = torch.load(pretrained_audio, map_location="cpu")
-                    audio_ckpt = audio_ckpt["state_dict"]
-                    keys = list(audio_ckpt.keys())
-                    for key in keys:
-                        if key.startswith("sed_model"):
-                            v = audio_ckpt.pop(key)
-                            audio_ckpt["audio_branch." + key[10:]] = v
-                elif os.path.basename(pretrained_audio).startswith(
-                    "finetuned"
-                ):  # checkpoint trained via linear probe codebase
-                    audio_ckpt = torch.load(pretrained_audio, map_location="cpu")
-                else:
-                    raise ValueError("Unknown audio checkpoint")
-            else:
-                raise f"this audio encoder pretrained checkpoint is not support"
-
-            model.load_state_dict(audio_ckpt, strict=False)
-            logging.info(
-                f"Loading pretrained {amodel_name} weights ({pretrained_audio})."
-            )
-            param_names = [n for n, p in model.named_parameters()]
-            for n in param_names:
-                print(n, "\t", "Loaded" if n in audio_ckpt else "Unloaded")
-
-        model.to(device=device)
-        if precision == "fp16":
-            assert device.type != "cpu"
-            convert_weights_to_fp16(model)
-
-        if jit:
-            model = torch.jit.script(model)
-
-    return model, model_cfg
-
-
-def create_model_and_transforms(
-    model_name: str,
-    pretrained: str = "",
-    precision: str = "fp32",
-    device: torch.device = torch.device("cpu"),
-    jit: bool = False,
-    force_quick_gelu: bool = False,
-    # pretrained_image: bool = False,
-):
-    model = create_model(
-        model_name,
-        pretrained,
-        precision,
-        device,
-        jit,
-        force_quick_gelu=force_quick_gelu,
-        # pretrained_image=pretrained_image
-    )
-    preprocess_train = image_transform(model.visual.image_size, is_train=True)
-    preprocess_val = image_transform(model.visual.image_size, is_train=False)
-    return model, preprocess_train, preprocess_val
-
-
-def list_models():
-    """enumerate available model architectures based on config files"""
-    return list(_MODEL_CONFIGS.keys())
-
-
-def add_model_config(path):
-    """add model config path or file and update registry"""
-    if not isinstance(path, Path):
-        path = Path(path)
-    _MODEL_CONFIG_PATHS.append(path)
-    _rescan_model_configs()
+import json
+import logging
+import os
+import re
+from copy import deepcopy
+from pathlib import Path
+
+import torch
+
+from .model import CLAP, convert_weights_to_fp16
+from .openai import load_openai_model
+from .pretrained import get_pretrained_url, download_pretrained
+from .transform import image_transform
+
+_MODEL_CONFIG_PATHS = [Path(__file__).parent / f"model_configs/"]
+_MODEL_CONFIGS = {}  # directory (model_name: config) of model architecture configs
+
+
+def _natural_key(string_):
+    return [int(s) if s.isdigit() else s for s in re.split(r"(\d+)", string_.lower())]
+
+
+def _rescan_model_configs():
+    global _MODEL_CONFIGS
+
+    config_ext = (".json",)
+    config_files = []
+    for config_path in _MODEL_CONFIG_PATHS:
+        if config_path.is_file() and config_path.suffix in config_ext:
+            config_files.append(config_path)
+        elif config_path.is_dir():
+            for ext in config_ext:
+                config_files.extend(config_path.glob(f"*{ext}"))
+
+    for cf in config_files:
+        if os.path.basename(cf)[0] == ".":
+            continue  # Ignore hidden files
+
+        with open(cf, "r") as f:
+            model_cfg = json.load(f)
+            if all(a in model_cfg for a in ("embed_dim", "audio_cfg", "text_cfg")):
+                _MODEL_CONFIGS[cf.stem] = model_cfg
+
+    _MODEL_CONFIGS = {
+        k: v
+        for k, v in sorted(_MODEL_CONFIGS.items(), key=lambda x: _natural_key(x[0]))
+    }
+
+
+_rescan_model_configs()  # initial populate of model config registry
+
+
+def load_state_dict(checkpoint_path: str, map_location="cpu", skip_params=True):
+    checkpoint = torch.load(checkpoint_path, map_location=map_location)
+    if isinstance(checkpoint, dict) and "state_dict" in checkpoint:
+        state_dict = checkpoint["state_dict"]
+    else:
+        state_dict = checkpoint
+    if skip_params:
+        if next(iter(state_dict.items()))[0].startswith("module"):
+            state_dict = {k[7:]: v for k, v in state_dict.items()}
+    # for k in state_dict:
+    #     if k.startswith('transformer'):
+    #         v = state_dict.pop(k)
+    #         state_dict['text_branch.' + k[12:]] = v
+    return state_dict
+
+
+def create_model(
+    amodel_name: str,
+    tmodel_name: str,
+    pretrained: str = "",
+    precision: str = "fp32",
+    device: torch.device = torch.device("cpu"),
+    jit: bool = False,
+    force_quick_gelu: bool = False,
+    openai_model_cache_dir: str = os.path.expanduser("~/.cache/clip"),
+    skip_params=True,
+    pretrained_audio: str = "",
+    pretrained_text: str = "",
+    enable_fusion: bool = False,
+    fusion_type: str = "None"
+    # pretrained_image: bool = False,
+):
+    amodel_name = amodel_name.replace(
+        "/", "-"
+    )  # for callers using old naming with / in ViT names
+    pretrained_orig = pretrained
+    pretrained = pretrained.lower()
+    if pretrained == "openai":
+        if amodel_name in _MODEL_CONFIGS:
+            logging.info(f"Loading {amodel_name} model config.")
+            model_cfg = deepcopy(_MODEL_CONFIGS[amodel_name])
+        else:
+            logging.error(
+                f"Model config for {amodel_name} not found; available models {list_models()}."
+            )
+            raise RuntimeError(f"Model config for {amodel_name} not found.")
+
+        logging.info(f"Loading pretrained ViT-B-16 text encoder from OpenAI.")
+        # Hard Code in model name
+        model_cfg["text_cfg"]["model_type"] = tmodel_name
+        model = load_openai_model(
+            "ViT-B-16",
+            model_cfg,
+            device=device,
+            jit=jit,
+            cache_dir=openai_model_cache_dir,
+            enable_fusion=enable_fusion,
+            fusion_type=fusion_type,
+        )
+        # See https://discuss.pytorch.org/t/valueerror-attemting-to-unscale-fp16-gradients/81372
+        if precision == "amp" or precision == "fp32":
+            model = model.float()
+    else:
+        if amodel_name in _MODEL_CONFIGS:
+            logging.info(f"Loading {amodel_name} model config.")
+            model_cfg = deepcopy(_MODEL_CONFIGS[amodel_name])
+        else:
+            logging.error(
+                f"Model config for {amodel_name} not found; available models {list_models()}."
+            )
+            raise RuntimeError(f"Model config for {amodel_name} not found.")
+
+        if force_quick_gelu:
+            # override for use of QuickGELU on non-OpenAI transformer models
+            model_cfg["quick_gelu"] = True
+
+        # if pretrained_image:
+        #     if 'timm_amodel_name' in model_cfg.get('vision_cfg', {}):
+        #         # pretrained weight loading for timm models set via vision_cfg
+        #         model_cfg['vision_cfg']['timm_model_pretrained'] = True
+        #     else:
+        #         assert False, 'pretrained image towers currently only supported for timm models'
+        model_cfg["text_cfg"]["model_type"] = tmodel_name
+        model_cfg["enable_fusion"] = enable_fusion
+        model_cfg["fusion_type"] = fusion_type
+        model = CLAP(**model_cfg)
+
+        if pretrained:
+            checkpoint_path = ""
+            url = get_pretrained_url(amodel_name, pretrained)
+            if url:
+                checkpoint_path = download_pretrained(url, root=openai_model_cache_dir)
+            elif os.path.exists(pretrained_orig):
+                checkpoint_path = pretrained_orig
+            if checkpoint_path:
+                logging.info(
+                    f"Loading pretrained {amodel_name}-{tmodel_name} weights ({pretrained})."
+                )
+                ckpt = load_state_dict(checkpoint_path, skip_params=True)
+                model.load_state_dict(ckpt)
+                param_names = [n for n, p in model.named_parameters()]
+                # for n in param_names:
+                #     print(n, "\t", "Loaded" if n in ckpt else "Unloaded")
+            else:
+                logging.warning(
+                    f"Pretrained weights ({pretrained}) not found for model {amodel_name}."
+                )
+                raise RuntimeError(
+                    f"Pretrained weights ({pretrained}) not found for model {amodel_name}."
+                )
+
+        if pretrained_audio:
+            if amodel_name.startswith("PANN"):
+                if "Cnn14_mAP" in pretrained_audio:  # official checkpoint
+                    audio_ckpt = torch.load(pretrained_audio, map_location="cpu")
+                    audio_ckpt = audio_ckpt["model"]
+                    keys = list(audio_ckpt.keys())
+                    for key in keys:
+                        if (
+                            "spectrogram_extractor" not in key
+                            and "logmel_extractor" not in key
+                        ):
+                            v = audio_ckpt.pop(key)
+                            audio_ckpt["audio_branch." + key] = v
+                elif os.path.basename(pretrained_audio).startswith(
+                    "PANN"
+                ):  # checkpoint trained via HTSAT codebase
+                    audio_ckpt = torch.load(pretrained_audio, map_location="cpu")
+                    audio_ckpt = audio_ckpt["state_dict"]
+                    keys = list(audio_ckpt.keys())
+                    for key in keys:
+                        if key.startswith("sed_model"):
+                            v = audio_ckpt.pop(key)
+                            audio_ckpt["audio_branch." + key[10:]] = v
+                elif os.path.basename(pretrained_audio).startswith(
+                    "finetuned"
+                ):  # checkpoint trained via linear probe codebase
+                    audio_ckpt = torch.load(pretrained_audio, map_location="cpu")
+                else:
+                    raise ValueError("Unknown audio checkpoint")
+            elif amodel_name.startswith("HTSAT"):
+                if "HTSAT_AudioSet_Saved" in pretrained_audio:  # official checkpoint
+                    audio_ckpt = torch.load(pretrained_audio, map_location="cpu")
+                    audio_ckpt = audio_ckpt["state_dict"]
+                    keys = list(audio_ckpt.keys())
+                    for key in keys:
+                        if key.startswith("sed_model") and (
+                            "spectrogram_extractor" not in key
+                            and "logmel_extractor" not in key
+                        ):
+                            v = audio_ckpt.pop(key)
+                            audio_ckpt["audio_branch." + key[10:]] = v
+                elif os.path.basename(pretrained_audio).startswith(
+                    "HTSAT"
+                ):  # checkpoint trained via HTSAT codebase
+                    audio_ckpt = torch.load(pretrained_audio, map_location="cpu")
+                    audio_ckpt = audio_ckpt["state_dict"]
+                    keys = list(audio_ckpt.keys())
+                    for key in keys:
+                        if key.startswith("sed_model"):
+                            v = audio_ckpt.pop(key)
+                            audio_ckpt["audio_branch." + key[10:]] = v
+                elif os.path.basename(pretrained_audio).startswith(
+                    "finetuned"
+                ):  # checkpoint trained via linear probe codebase
+                    audio_ckpt = torch.load(pretrained_audio, map_location="cpu")
+                else:
+                    raise ValueError("Unknown audio checkpoint")
+            else:
+                raise f"this audio encoder pretrained checkpoint is not support"
+
+            model.load_state_dict(audio_ckpt, strict=False)
+            logging.info(
+                f"Loading pretrained {amodel_name} weights ({pretrained_audio})."
+            )
+            param_names = [n for n, p in model.named_parameters()]
+            for n in param_names:
+                print(n, "\t", "Loaded" if n in audio_ckpt else "Unloaded")
+
+        model.to(device=device)
+        if precision == "fp16":
+            assert device.type != "cpu"
+            convert_weights_to_fp16(model)
+
+        if jit:
+            model = torch.jit.script(model)
+
+    return model, model_cfg
+
+
+def create_model_and_transforms(
+    model_name: str,
+    pretrained: str = "",
+    precision: str = "fp32",
+    device: torch.device = torch.device("cpu"),
+    jit: bool = False,
+    force_quick_gelu: bool = False,
+    # pretrained_image: bool = False,
+):
+    model = create_model(
+        model_name,
+        pretrained,
+        precision,
+        device,
+        jit,
+        force_quick_gelu=force_quick_gelu,
+        # pretrained_image=pretrained_image
+    )
+    preprocess_train = image_transform(model.visual.image_size, is_train=True)
+    preprocess_val = image_transform(model.visual.image_size, is_train=False)
+    return model, preprocess_train, preprocess_val
+
+
+def list_models():
+    """enumerate available model architectures based on config files"""
+    return list(_MODEL_CONFIGS.keys())
+
+
+def add_model_config(path):
+    """add model config path or file and update registry"""
+    if not isinstance(path, Path):
+        path = Path(path)
+    _MODEL_CONFIG_PATHS.append(path)
+    _rescan_model_configs()

audiosr/clap/open_clip/feature_fusion.py

@@ -1,192 +1,192 @@
-"""
-Feature Fusion for Varible-Length Data Processing
-AFF/iAFF is referred and modified from https://github.com/YimianDai/open-aff/blob/master/aff_pytorch/aff_net/fusion.py
-According to the paper: Yimian Dai et al, Attentional Feature Fusion, IEEE Winter Conference on Applications of Computer Vision, WACV 2021
-"""
-
-import torch
-import torch.nn as nn
-
-
-class DAF(nn.Module):
-    """
-    直接相加 DirectAddFuse
-    """
-
-    def __init__(self):
-        super(DAF, self).__init__()
-
-    def forward(self, x, residual):
-        return x + residual
-
-
-class iAFF(nn.Module):
-    """
-    多特征融合 iAFF
-    """
-
-    def __init__(self, channels=64, r=4, type="2D"):
-        super(iAFF, self).__init__()
-        inter_channels = int(channels // r)
-
-        if type == "1D":
-            # 本地注意力
-            self.local_att = nn.Sequential(
-                nn.Conv1d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
-                nn.BatchNorm1d(inter_channels),
-                nn.ReLU(inplace=True),
-                nn.Conv1d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
-                nn.BatchNorm1d(channels),
-            )
-
-            # 全局注意力
-            self.global_att = nn.Sequential(
-                nn.AdaptiveAvgPool1d(1),
-                nn.Conv1d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
-                nn.BatchNorm1d(inter_channels),
-                nn.ReLU(inplace=True),
-                nn.Conv1d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
-                nn.BatchNorm1d(channels),
-            )
-
-            # 第二次本地注意力
-            self.local_att2 = nn.Sequential(
-                nn.Conv1d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
-                nn.BatchNorm1d(inter_channels),
-                nn.ReLU(inplace=True),
-                nn.Conv1d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
-                nn.BatchNorm1d(channels),
-            )
-            # 第二次全局注意力
-            self.global_att2 = nn.Sequential(
-                nn.AdaptiveAvgPool1d(1),
-                nn.Conv1d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
-                nn.BatchNorm1d(inter_channels),
-                nn.ReLU(inplace=True),
-                nn.Conv1d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
-                nn.BatchNorm1d(channels),
-            )
-        elif type == "2D":
-            # 本地注意力
-            self.local_att = nn.Sequential(
-                nn.Conv2d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
-                nn.BatchNorm2d(inter_channels),
-                nn.ReLU(inplace=True),
-                nn.Conv2d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
-                nn.BatchNorm2d(channels),
-            )
-
-            # 全局注意力
-            self.global_att = nn.Sequential(
-                nn.AdaptiveAvgPool2d(1),
-                nn.Conv2d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
-                nn.BatchNorm2d(inter_channels),
-                nn.ReLU(inplace=True),
-                nn.Conv2d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
-                nn.BatchNorm2d(channels),
-            )
-
-            # 第二次本地注意力
-            self.local_att2 = nn.Sequential(
-                nn.Conv2d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
-                nn.BatchNorm2d(inter_channels),
-                nn.ReLU(inplace=True),
-                nn.Conv2d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
-                nn.BatchNorm2d(channels),
-            )
-            # 第二次全局注意力
-            self.global_att2 = nn.Sequential(
-                nn.AdaptiveAvgPool2d(1),
-                nn.Conv2d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
-                nn.BatchNorm2d(inter_channels),
-                nn.ReLU(inplace=True),
-                nn.Conv2d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
-                nn.BatchNorm2d(channels),
-            )
-        else:
-            raise f"the type is not supported"
-
-        self.sigmoid = nn.Sigmoid()
-
-    def forward(self, x, residual):
-        flag = False
-        xa = x + residual
-        if xa.size(0) == 1:
-            xa = torch.cat([xa, xa], dim=0)
-            flag = True
-        xl = self.local_att(xa)
-        xg = self.global_att(xa)
-        xlg = xl + xg
-        wei = self.sigmoid(xlg)
-        xi = x * wei + residual * (1 - wei)
-
-        xl2 = self.local_att2(xi)
-        xg2 = self.global_att(xi)
-        xlg2 = xl2 + xg2
-        wei2 = self.sigmoid(xlg2)
-        xo = x * wei2 + residual * (1 - wei2)
-        if flag:
-            xo = xo[0].unsqueeze(0)
-        return xo
-
-
-class AFF(nn.Module):
-    """
-    多特征融合 AFF
-    """
-
-    def __init__(self, channels=64, r=4, type="2D"):
-        super(AFF, self).__init__()
-        inter_channels = int(channels // r)
-
-        if type == "1D":
-            self.local_att = nn.Sequential(
-                nn.Conv1d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
-                nn.BatchNorm1d(inter_channels),
-                nn.ReLU(inplace=True),
-                nn.Conv1d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
-                nn.BatchNorm1d(channels),
-            )
-            self.global_att = nn.Sequential(
-                nn.AdaptiveAvgPool1d(1),
-                nn.Conv1d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
-                nn.BatchNorm1d(inter_channels),
-                nn.ReLU(inplace=True),
-                nn.Conv1d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
-                nn.BatchNorm1d(channels),
-            )
-        elif type == "2D":
-            self.local_att = nn.Sequential(
-                nn.Conv2d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
-                nn.BatchNorm2d(inter_channels),
-                nn.ReLU(inplace=True),
-                nn.Conv2d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
-                nn.BatchNorm2d(channels),
-            )
-            self.global_att = nn.Sequential(
-                nn.AdaptiveAvgPool2d(1),
-                nn.Conv2d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
-                nn.BatchNorm2d(inter_channels),
-                nn.ReLU(inplace=True),
-                nn.Conv2d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
-                nn.BatchNorm2d(channels),
-            )
-        else:
-            raise f"the type is not supported."
-
-        self.sigmoid = nn.Sigmoid()
-
-    def forward(self, x, residual):
-        flag = False
-        xa = x + residual
-        if xa.size(0) == 1:
-            xa = torch.cat([xa, xa], dim=0)
-            flag = True
-        xl = self.local_att(xa)
-        xg = self.global_att(xa)
-        xlg = xl + xg
-        wei = self.sigmoid(xlg)
-        xo = 2 * x * wei + 2 * residual * (1 - wei)
-        if flag:
-            xo = xo[0].unsqueeze(0)
-        return xo
+"""
+Feature Fusion for Varible-Length Data Processing
+AFF/iAFF is referred and modified from https://github.com/YimianDai/open-aff/blob/master/aff_pytorch/aff_net/fusion.py
+According to the paper: Yimian Dai et al, Attentional Feature Fusion, IEEE Winter Conference on Applications of Computer Vision, WACV 2021
+"""
+
+import torch
+import torch.nn as nn
+
+
+class DAF(nn.Module):
+    """
+    直接相加 DirectAddFuse
+    """
+
+    def __init__(self):
+        super(DAF, self).__init__()
+
+    def forward(self, x, residual):
+        return x + residual
+
+
+class iAFF(nn.Module):
+    """
+    多特征融合 iAFF
+    """
+
+    def __init__(self, channels=64, r=4, type="2D"):
+        super(iAFF, self).__init__()
+        inter_channels = int(channels // r)
+
+        if type == "1D":
+            # 本地注意力
+            self.local_att = nn.Sequential(
+                nn.Conv1d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm1d(inter_channels),
+                nn.ReLU(inplace=True),
+                nn.Conv1d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm1d(channels),
+            )
+
+            # 全局注意力
+            self.global_att = nn.Sequential(
+                nn.AdaptiveAvgPool1d(1),
+                nn.Conv1d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm1d(inter_channels),
+                nn.ReLU(inplace=True),
+                nn.Conv1d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm1d(channels),
+            )
+
+            # 第二次本地注意力
+            self.local_att2 = nn.Sequential(
+                nn.Conv1d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm1d(inter_channels),
+                nn.ReLU(inplace=True),
+                nn.Conv1d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm1d(channels),
+            )
+            # 第二次全局注意力
+            self.global_att2 = nn.Sequential(
+                nn.AdaptiveAvgPool1d(1),
+                nn.Conv1d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm1d(inter_channels),
+                nn.ReLU(inplace=True),
+                nn.Conv1d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm1d(channels),
+            )
+        elif type == "2D":
+            # 本地注意力
+            self.local_att = nn.Sequential(
+                nn.Conv2d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm2d(inter_channels),
+                nn.ReLU(inplace=True),
+                nn.Conv2d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm2d(channels),
+            )
+
+            # 全局注意力
+            self.global_att = nn.Sequential(
+                nn.AdaptiveAvgPool2d(1),
+                nn.Conv2d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm2d(inter_channels),
+                nn.ReLU(inplace=True),
+                nn.Conv2d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm2d(channels),
+            )
+
+            # 第二次本地注意力
+            self.local_att2 = nn.Sequential(
+                nn.Conv2d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm2d(inter_channels),
+                nn.ReLU(inplace=True),
+                nn.Conv2d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm2d(channels),
+            )
+            # 第二次全局注意力
+            self.global_att2 = nn.Sequential(
+                nn.AdaptiveAvgPool2d(1),
+                nn.Conv2d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm2d(inter_channels),
+                nn.ReLU(inplace=True),
+                nn.Conv2d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm2d(channels),
+            )
+        else:
+            raise f"the type is not supported"
+
+        self.sigmoid = nn.Sigmoid()
+
+    def forward(self, x, residual):
+        flag = False
+        xa = x + residual
+        if xa.size(0) == 1:
+            xa = torch.cat([xa, xa], dim=0)
+            flag = True
+        xl = self.local_att(xa)
+        xg = self.global_att(xa)
+        xlg = xl + xg
+        wei = self.sigmoid(xlg)
+        xi = x * wei + residual * (1 - wei)
+
+        xl2 = self.local_att2(xi)
+        xg2 = self.global_att(xi)
+        xlg2 = xl2 + xg2
+        wei2 = self.sigmoid(xlg2)
+        xo = x * wei2 + residual * (1 - wei2)
+        if flag:
+            xo = xo[0].unsqueeze(0)
+        return xo
+
+
+class AFF(nn.Module):
+    """
+    多特征融合 AFF
+    """
+
+    def __init__(self, channels=64, r=4, type="2D"):
+        super(AFF, self).__init__()
+        inter_channels = int(channels // r)
+
+        if type == "1D":
+            self.local_att = nn.Sequential(
+                nn.Conv1d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm1d(inter_channels),
+                nn.ReLU(inplace=True),
+                nn.Conv1d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm1d(channels),
+            )
+            self.global_att = nn.Sequential(
+                nn.AdaptiveAvgPool1d(1),
+                nn.Conv1d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm1d(inter_channels),
+                nn.ReLU(inplace=True),
+                nn.Conv1d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm1d(channels),
+            )
+        elif type == "2D":
+            self.local_att = nn.Sequential(
+                nn.Conv2d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm2d(inter_channels),
+                nn.ReLU(inplace=True),
+                nn.Conv2d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm2d(channels),
+            )
+            self.global_att = nn.Sequential(
+                nn.AdaptiveAvgPool2d(1),
+                nn.Conv2d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm2d(inter_channels),
+                nn.ReLU(inplace=True),
+                nn.Conv2d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm2d(channels),
+            )
+        else:
+            raise f"the type is not supported."
+
+        self.sigmoid = nn.Sigmoid()
+
+    def forward(self, x, residual):
+        flag = False
+        xa = x + residual
+        if xa.size(0) == 1:
+            xa = torch.cat([xa, xa], dim=0)
+            flag = True
+        xl = self.local_att(xa)
+        xg = self.global_att(xa)
+        xlg = xl + xg
+        wei = self.sigmoid(xlg)
+        xo = 2 * x * wei + 2 * residual * (1 - wei)
+        if flag:
+            xo = xo[0].unsqueeze(0)
+        return xo

audiosr/clap/open_clip/loss.py

@@ -1,397 +1,397 @@
-import torch
-import torch.distributed.nn
-from torch import distributed as dist, nn as nn
-from torch.nn import functional as F
-import numpy as np
-from sklearn.metrics import average_precision_score, roc_auc_score, accuracy_score
-
-try:
-    import horovod.torch as hvd
-except ImportError:
-    hvd = None
-
-
-def gather_features(
-    audio_features,
-    text_features,
-    audio_features_mlp=None,
-    text_features_mlp=None,
-    local_loss=False,
-    gather_with_grad=False,
-    rank=0,
-    world_size=1,
-    use_horovod=False,
-    mlp_loss=False,
-):
-    if use_horovod:
-        assert hvd is not None, "Please install horovod"
-        if gather_with_grad:
-            all_audio_features = hvd.allgather(audio_features)
-            all_text_features = hvd.allgather(text_features)
-            if mlp_loss:
-                all_audio_features_mlp = hvd.allgather(audio_features_mlp)
-                all_text_features_mlp = hvd.allgather(text_features_mlp)
-        else:
-            with torch.no_grad():
-                all_audio_features = hvd.allgather(audio_features)
-                all_text_features = hvd.allgather(text_features)
-                if mlp_loss:
-                    all_audio_features_mlp = hvd.allgather(audio_features_mlp)
-                    all_text_features_mlp = hvd.allgather(text_features_mlp)
-            if not local_loss:
-                # ensure grads for local rank when all_* features don't have a gradient
-                gathered_audio_features = list(
-                    all_audio_features.chunk(world_size, dim=0)
-                )
-                gathered_text_features = list(
-                    all_text_features.chunk(world_size, dim=0)
-                )
-                gathered_audio_features[rank] = audio_features
-                gathered_text_features[rank] = text_features
-                all_audio_features = torch.cat(gathered_audio_features, dim=0)
-                all_text_features = torch.cat(gathered_text_features, dim=0)
-                if mlp_loss:
-                    gathered_audio_features_mlp = list(
-                        all_audio_features_mlp.chunk(world_size, dim=0)
-                    )
-                    gathered_text_features_mlp = list(
-                        all_text_features_mlp.chunk(world_size, dim=0)
-                    )
-                    gathered_audio_features_mlp[rank] = audio_features_mlp
-                    gathered_text_features_mlp[rank] = text_features_mlp
-                    all_audio_features_mlp = torch.cat(
-                        gathered_audio_features_mlp, dim=0
-                    )
-                    all_text_features_mlp = torch.cat(gathered_text_features_mlp, dim=0)
-    else:
-        # We gather tensors from all gpus
-        if gather_with_grad:
-            all_audio_features = torch.cat(
-                torch.distributed.nn.all_gather(audio_features), dim=0
-            )
-            all_text_features = torch.cat(
-                torch.distributed.nn.all_gather(text_features), dim=0
-            )
-            if mlp_loss:
-                all_audio_features_mlp = torch.cat(
-                    torch.distributed.nn.all_gather(audio_features_mlp), dim=0
-                )
-                all_text_features_mlp = torch.cat(
-                    torch.distributed.nn.all_gather(text_features_mlp), dim=0
-                )
-        else:
-            gathered_audio_features = [
-                torch.zeros_like(audio_features) for _ in range(world_size)
-            ]
-            gathered_text_features = [
-                torch.zeros_like(text_features) for _ in range(world_size)
-            ]
-            dist.all_gather(gathered_audio_features, audio_features)
-            dist.all_gather(gathered_text_features, text_features)
-            if mlp_loss:
-                gathered_audio_features_mlp = [
-                    torch.zeros_like(audio_features_mlp) for _ in range(world_size)
-                ]
-                gathered_text_features_mlp = [
-                    torch.zeros_like(text_features_mlp) for _ in range(world_size)
-                ]
-                dist.all_gather(gathered_audio_features_mlp, audio_features_mlp)
-                dist.all_gather(gathered_text_features_mlp, text_features_mlp)
-            if not local_loss:
-                # ensure grads for local rank when all_* features don't have a gradient
-                gathered_audio_features[rank] = audio_features
-                gathered_text_features[rank] = text_features
-                if mlp_loss:
-                    gathered_audio_features_mlp[rank] = audio_features_mlp
-                    gathered_text_features_mlp[rank] = text_features_mlp
-
-            all_audio_features = torch.cat(gathered_audio_features, dim=0)
-            all_text_features = torch.cat(gathered_text_features, dim=0)
-            if mlp_loss:
-                all_audio_features_mlp = torch.cat(gathered_audio_features_mlp, dim=0)
-                all_text_features_mlp = torch.cat(gathered_text_features_mlp, dim=0)
-    if mlp_loss:
-        return (
-            all_audio_features,
-            all_text_features,
-            all_audio_features_mlp,
-            all_text_features_mlp,
-        )
-    else:
-        return all_audio_features, all_text_features
-
-
-class ClipLoss(nn.Module):
-    def __init__(
-        self,
-        local_loss=False,
-        gather_with_grad=False,
-        cache_labels=False,
-        rank=0,
-        world_size=1,
-        use_horovod=False,
-        mlp_loss=False,
-        weight_loss_kappa=0,
-    ):
-        super().__init__()
-        self.local_loss = local_loss
-        self.gather_with_grad = gather_with_grad
-        self.cache_labels = cache_labels
-        self.rank = rank
-        self.world_size = world_size
-        self.use_horovod = use_horovod
-        self.mlp_loss = mlp_loss
-        self.weighted_loss = bool(weight_loss_kappa != 0)
-        self.weight_loss_kappa = weight_loss_kappa
-        # cache state
-        self.prev_num_logits = 0
-        self.labels = {}
-
-    def forward(
-        self,
-        audio_features,
-        text_features,
-        logit_scale_a,
-        logit_scale_t=None,
-        audio_features_mlp=None,
-        text_features_mlp=None,
-    ):
-        device = audio_features.device
-        if self.mlp_loss:
-            if self.world_size > 1:
-                (
-                    all_audio_features,
-                    all_text_features,
-                    all_audio_features_mlp,
-                    all_text_features_mlp,
-                ) = gather_features(
-                    audio_features=audio_features,
-                    text_features=text_features,
-                    audio_features_mlp=audio_features_mlp,
-                    text_features_mlp=text_features_mlp,
-                    local_loss=self.local_loss,
-                    gather_with_grad=self.gather_with_grad,
-                    rank=self.rank,
-                    world_size=self.world_size,
-                    use_horovod=self.use_horovod,
-                    mlp_loss=self.mlp_loss,
-                )
-                if self.local_loss:
-                    a_logits_per_audio = (
-                        logit_scale_a * audio_features @ all_text_features_mlp.T
-                    )
-                    a_logits_per_text = (
-                        logit_scale_a * text_features_mlp @ all_audio_features.T
-                    )
-                    t_logits_per_audio = (
-                        logit_scale_t * audio_features_mlp @ all_text_features.T
-                    )
-                    t_logits_per_text = (
-                        logit_scale_t * text_features @ all_audio_features_mlp.T
-                    )
-                else:
-                    a_logits_per_audio = (
-                        logit_scale_a * all_audio_features @ all_text_features_mlp.T
-                    )
-                    a_logits_per_text = a_logits_per_audio.T
-                    t_logits_per_audio = (
-                        logit_scale_t * all_audio_features_mlp @ all_text_features.T
-                    )
-                    t_logits_per_text = t_logits_per_audio.T
-            else:
-                a_logits_per_audio = (
-                    logit_scale_a * audio_features @ text_features_mlp.T
-                )
-                a_logits_per_text = logit_scale_a * text_features_mlp @ audio_features.T
-                t_logits_per_audio = (
-                    logit_scale_t * audio_features_mlp @ text_features.T
-                )
-                t_logits_per_text = logit_scale_t * text_features @ audio_features_mlp.T
-
-            # calculated ground-truth and cache if enabled
-            num_logits = a_logits_per_audio.shape[0]
-            if self.prev_num_logits != num_logits or device not in self.labels:
-                labels = torch.arange(num_logits, device=device, dtype=torch.long)
-                if self.world_size > 1 and self.local_loss:
-                    labels = labels + num_logits * self.rank
-                if self.cache_labels:
-                    self.labels[device] = labels
-                    self.prev_num_logits = num_logits
-            else:
-                labels = self.labels[device]
-
-            if not self.weighted_loss:
-                total_loss = (
-                    F.cross_entropy(a_logits_per_audio, labels)
-                    + F.cross_entropy(a_logits_per_text, labels)
-                    + F.cross_entropy(t_logits_per_audio, labels)
-                    + F.cross_entropy(t_logits_per_text, labels)
-                ) / 4
-            else:
-                audio_weight = (audio_features @ audio_features.T).detach()
-                audio_weight = (
-                    torch.exp(
-                        torch.sum(audio_weight, axis=1)
-                        / (self.weight_loss_kappa * len(audio_weight))
-                    )
-                ).detach()
-                text_weight = (text_features @ text_features.T).detach()
-                text_weight = (
-                    torch.exp(
-                        torch.sum(text_weight, axis=1)
-                        / (self.weight_loss_kappa * len(text_features))
-                    )
-                ).detach()
-                total_loss = (
-                    F.cross_entropy(a_logits_per_audio, labels, weight=audio_weight)
-                    + F.cross_entropy(a_logits_per_text, labels, weight=audio_weight)
-                    + F.cross_entropy(t_logits_per_audio, labels, weight=text_weight)
-                    + F.cross_entropy(t_logits_per_text, labels, weight=text_weight)
-                ) / 4
-        else:
-            if self.world_size > 1:
-                all_audio_features, all_text_features = gather_features(
-                    audio_features=audio_features,
-                    text_features=text_features,
-                    local_loss=self.local_loss,
-                    gather_with_grad=self.gather_with_grad,
-                    rank=self.rank,
-                    world_size=self.world_size,
-                    use_horovod=self.use_horovod,
-                    mlp_loss=self.mlp_loss,
-                )
-
-                if self.local_loss:
-                    logits_per_audio = (
-                        logit_scale_a * audio_features @ all_text_features.T
-                    )
-                    logits_per_text = (
-                        logit_scale_a * text_features @ all_audio_features.T
-                    )
-                else:
-                    logits_per_audio = (
-                        logit_scale_a * all_audio_features @ all_text_features.T
-                    )
-                    logits_per_text = logits_per_audio.T
-            else:
-                logits_per_audio = logit_scale_a * audio_features @ text_features.T
-                logits_per_text = logit_scale_a * text_features @ audio_features.T
-
-            # calculated ground-truth and cache if enabled
-            num_logits = logits_per_audio.shape[0]
-            if self.prev_num_logits != num_logits or device not in self.labels:
-                labels = torch.arange(num_logits, device=device, dtype=torch.long)
-                if self.world_size > 1 and self.local_loss:
-                    labels = labels + num_logits * self.rank
-                if self.cache_labels:
-                    self.labels[device] = labels
-                    self.prev_num_logits = num_logits
-            else:
-                labels = self.labels[device]
-            if not self.weighted_loss:
-                total_loss = (
-                    F.cross_entropy(logits_per_audio, labels)
-                    + F.cross_entropy(logits_per_text, labels)
-                ) / 2
-            else:
-                audio_weight = (all_audio_features @ all_audio_features.T).detach()
-                audio_weight = (
-                    torch.exp(
-                        torch.sum(audio_weight, axis=1)
-                        / (self.weight_loss_kappa * len(all_audio_features))
-                    )
-                ).detach()
-                text_weight = (all_text_features @ all_text_features.T).detach()
-                text_weight = (
-                    torch.exp(
-                        torch.sum(text_weight, axis=1)
-                        / (self.weight_loss_kappa * len(all_text_features))
-                    )
-                ).detach()
-                total_loss = (
-                    F.cross_entropy(logits_per_audio, labels, weight=text_weight)
-                    + F.cross_entropy(logits_per_text, labels, weight=audio_weight)
-                ) / 2
-        return total_loss
-
-
-def lp_gather_features(pred, target, world_size=1, use_horovod=False):
-    if use_horovod:
-        assert hvd is not None, "Please install horovod"
-        with torch.no_grad():
-            all_preds = hvd.allgather(pred)
-            all_targets = hvd.allgath(target)
-    else:
-        gathered_preds = [torch.zeros_like(pred) for _ in range(world_size)]
-        gathered_targets = [torch.zeros_like(target) for _ in range(world_size)]
-
-        dist.all_gather(gathered_preds, pred)
-        dist.all_gather(gathered_targets, target)
-        all_preds = torch.cat(gathered_preds, dim=0)
-        all_targets = torch.cat(gathered_targets, dim=0)
-
-    return all_preds, all_targets
-
-
-def get_map(pred, target):
-    pred = torch.sigmoid(pred).numpy()
-    target = target.numpy()
-    return np.mean(average_precision_score(target, pred, average=None))
-
-
-def get_acc(pred, target):
-    pred = torch.argmax(pred, 1).numpy()
-    target = torch.argmax(target, 1).numpy()
-    return accuracy_score(target, pred)
-
-
-def get_mauc(pred, target):
-    pred = torch.sigmoid(pred).numpy()
-    target = target.numpy()
-    return np.mean(roc_auc_score(target, pred, average=None))
-
-
-class LPMetrics(object):
-    def __init__(self, metric_names=["map", "acc", "mauc"]):
-        self.metrics = []
-        for name in metric_names:
-            self.metrics.append(self.get_metric(name))
-        self.metric_names = metric_names
-
-    def get_metric(self, name):
-        if name == "map":
-            return get_map
-        elif name == "acc":
-            return get_acc
-        elif name == "mauc":
-            return get_mauc
-        else:
-            raise ValueError(f"the metric should be at least one of [map, acc, mauc]")
-
-    def evaluate_mertics(self, pred, target):
-        metric_dict = {}
-        for i in range(len(self.metric_names)):
-            metric_dict[self.metric_names[i]] = self.metrics[i](pred, target)
-        return metric_dict
-
-
-def calc_celoss(pred, target):
-    target = torch.argmax(target, 1).long()
-    return nn.CrossEntropyLoss()(pred, target)
-
-
-class LPLoss(nn.Module):
-    def __init__(self, loss_name):
-        super().__init__()
-        if loss_name == "bce":
-            self.loss_func = nn.BCEWithLogitsLoss()
-        elif loss_name == "ce":
-            self.loss_func = calc_celoss
-        elif loss_name == "mse":
-            self.loss_func = nn.MSELoss()
-        else:
-            raise ValueError(f"the loss func should be at least one of [bce, ce, mse]")
-
-    def forward(self, pred, target):
-        loss = self.loss_func(pred, target)
-        return loss
+import torch
+import torch.distributed.nn
+from torch import distributed as dist, nn as nn
+from torch.nn import functional as F
+import numpy as np
+from sklearn.metrics import average_precision_score, roc_auc_score, accuracy_score
+
+try:
+    import horovod.torch as hvd
+except ImportError:
+    hvd = None
+
+
+def gather_features(
+    audio_features,
+    text_features,
+    audio_features_mlp=None,
+    text_features_mlp=None,
+    local_loss=False,
+    gather_with_grad=False,
+    rank=0,
+    world_size=1,
+    use_horovod=False,
+    mlp_loss=False,
+):
+    if use_horovod:
+        assert hvd is not None, "Please install horovod"
+        if gather_with_grad:
+            all_audio_features = hvd.allgather(audio_features)
+            all_text_features = hvd.allgather(text_features)
+            if mlp_loss:
+                all_audio_features_mlp = hvd.allgather(audio_features_mlp)
+                all_text_features_mlp = hvd.allgather(text_features_mlp)
+        else:
+            with torch.no_grad():
+                all_audio_features = hvd.allgather(audio_features)
+                all_text_features = hvd.allgather(text_features)
+                if mlp_loss:
+                    all_audio_features_mlp = hvd.allgather(audio_features_mlp)
+                    all_text_features_mlp = hvd.allgather(text_features_mlp)
+            if not local_loss:
+                # ensure grads for local rank when all_* features don't have a gradient
+                gathered_audio_features = list(
+                    all_audio_features.chunk(world_size, dim=0)
+                )
+                gathered_text_features = list(
+                    all_text_features.chunk(world_size, dim=0)
+                )
+                gathered_audio_features[rank] = audio_features
+                gathered_text_features[rank] = text_features
+                all_audio_features = torch.cat(gathered_audio_features, dim=0)
+                all_text_features = torch.cat(gathered_text_features, dim=0)
+                if mlp_loss:
+                    gathered_audio_features_mlp = list(
+                        all_audio_features_mlp.chunk(world_size, dim=0)
+                    )
+                    gathered_text_features_mlp = list(
+                        all_text_features_mlp.chunk(world_size, dim=0)
+                    )
+                    gathered_audio_features_mlp[rank] = audio_features_mlp
+                    gathered_text_features_mlp[rank] = text_features_mlp
+                    all_audio_features_mlp = torch.cat(
+                        gathered_audio_features_mlp, dim=0
+                    )
+                    all_text_features_mlp = torch.cat(gathered_text_features_mlp, dim=0)
+    else:
+        # We gather tensors from all gpus
+        if gather_with_grad:
+            all_audio_features = torch.cat(
+                torch.distributed.nn.all_gather(audio_features), dim=0
+            )
+            all_text_features = torch.cat(
+                torch.distributed.nn.all_gather(text_features), dim=0
+            )
+            if mlp_loss:
+                all_audio_features_mlp = torch.cat(
+                    torch.distributed.nn.all_gather(audio_features_mlp), dim=0
+                )
+                all_text_features_mlp = torch.cat(
+                    torch.distributed.nn.all_gather(text_features_mlp), dim=0
+                )
+        else:
+            gathered_audio_features = [
+                torch.zeros_like(audio_features) for _ in range(world_size)
+            ]
+            gathered_text_features = [
+                torch.zeros_like(text_features) for _ in range(world_size)
+            ]
+            dist.all_gather(gathered_audio_features, audio_features)
+            dist.all_gather(gathered_text_features, text_features)
+            if mlp_loss:
+                gathered_audio_features_mlp = [
+                    torch.zeros_like(audio_features_mlp) for _ in range(world_size)
+                ]
+                gathered_text_features_mlp = [
+                    torch.zeros_like(text_features_mlp) for _ in range(world_size)
+                ]
+                dist.all_gather(gathered_audio_features_mlp, audio_features_mlp)
+                dist.all_gather(gathered_text_features_mlp, text_features_mlp)
+            if not local_loss:
+                # ensure grads for local rank when all_* features don't have a gradient
+                gathered_audio_features[rank] = audio_features
+                gathered_text_features[rank] = text_features
+                if mlp_loss:
+                    gathered_audio_features_mlp[rank] = audio_features_mlp
+                    gathered_text_features_mlp[rank] = text_features_mlp
+
+            all_audio_features = torch.cat(gathered_audio_features, dim=0)
+            all_text_features = torch.cat(gathered_text_features, dim=0)
+            if mlp_loss:
+                all_audio_features_mlp = torch.cat(gathered_audio_features_mlp, dim=0)
+                all_text_features_mlp = torch.cat(gathered_text_features_mlp, dim=0)
+    if mlp_loss:
+        return (
+            all_audio_features,
+            all_text_features,
+            all_audio_features_mlp,
+            all_text_features_mlp,
+        )
+    else:
+        return all_audio_features, all_text_features
+
+
+class ClipLoss(nn.Module):
+    def __init__(
+        self,
+        local_loss=False,
+        gather_with_grad=False,
+        cache_labels=False,
+        rank=0,
+        world_size=1,
+        use_horovod=False,
+        mlp_loss=False,
+        weight_loss_kappa=0,
+    ):
+        super().__init__()
+        self.local_loss = local_loss
+        self.gather_with_grad = gather_with_grad
+        self.cache_labels = cache_labels
+        self.rank = rank
+        self.world_size = world_size
+        self.use_horovod = use_horovod
+        self.mlp_loss = mlp_loss
+        self.weighted_loss = bool(weight_loss_kappa != 0)
+        self.weight_loss_kappa = weight_loss_kappa
+        # cache state
+        self.prev_num_logits = 0
+        self.labels = {}
+
+    def forward(
+        self,
+        audio_features,
+        text_features,
+        logit_scale_a,
+        logit_scale_t=None,
+        audio_features_mlp=None,
+        text_features_mlp=None,
+    ):
+        device = audio_features.device
+        if self.mlp_loss:
+            if self.world_size > 1:
+                (
+                    all_audio_features,
+                    all_text_features,
+                    all_audio_features_mlp,
+                    all_text_features_mlp,
+                ) = gather_features(
+                    audio_features=audio_features,
+                    text_features=text_features,
+                    audio_features_mlp=audio_features_mlp,
+                    text_features_mlp=text_features_mlp,
+                    local_loss=self.local_loss,
+                    gather_with_grad=self.gather_with_grad,
+                    rank=self.rank,
+                    world_size=self.world_size,
+                    use_horovod=self.use_horovod,
+                    mlp_loss=self.mlp_loss,
+                )
+                if self.local_loss:
+                    a_logits_per_audio = (
+                        logit_scale_a * audio_features @ all_text_features_mlp.T
+                    )
+                    a_logits_per_text = (
+                        logit_scale_a * text_features_mlp @ all_audio_features.T
+                    )
+                    t_logits_per_audio = (
+                        logit_scale_t * audio_features_mlp @ all_text_features.T
+                    )
+                    t_logits_per_text = (
+                        logit_scale_t * text_features @ all_audio_features_mlp.T
+                    )
+                else:
+                    a_logits_per_audio = (
+                        logit_scale_a * all_audio_features @ all_text_features_mlp.T
+                    )
+                    a_logits_per_text = a_logits_per_audio.T
+                    t_logits_per_audio = (
+                        logit_scale_t * all_audio_features_mlp @ all_text_features.T
+                    )
+                    t_logits_per_text = t_logits_per_audio.T
+            else:
+                a_logits_per_audio = (
+                    logit_scale_a * audio_features @ text_features_mlp.T
+                )
+                a_logits_per_text = logit_scale_a * text_features_mlp @ audio_features.T
+                t_logits_per_audio = (
+                    logit_scale_t * audio_features_mlp @ text_features.T
+                )
+                t_logits_per_text = logit_scale_t * text_features @ audio_features_mlp.T
+
+            # calculated ground-truth and cache if enabled
+            num_logits = a_logits_per_audio.shape[0]
+            if self.prev_num_logits != num_logits or device not in self.labels:
+                labels = torch.arange(num_logits, device=device, dtype=torch.long)
+                if self.world_size > 1 and self.local_loss:
+                    labels = labels + num_logits * self.rank
+                if self.cache_labels:
+                    self.labels[device] = labels
+                    self.prev_num_logits = num_logits
+            else:
+                labels = self.labels[device]
+
+            if not self.weighted_loss:
+                total_loss = (
+                    F.cross_entropy(a_logits_per_audio, labels)
+                    + F.cross_entropy(a_logits_per_text, labels)
+                    + F.cross_entropy(t_logits_per_audio, labels)
+                    + F.cross_entropy(t_logits_per_text, labels)
+                ) / 4
+            else:
+                audio_weight = (audio_features @ audio_features.T).detach()
+                audio_weight = (
+                    torch.exp(
+                        torch.sum(audio_weight, axis=1)
+                        / (self.weight_loss_kappa * len(audio_weight))
+                    )
+                ).detach()
+                text_weight = (text_features @ text_features.T).detach()
+                text_weight = (
+                    torch.exp(
+                        torch.sum(text_weight, axis=1)
+                        / (self.weight_loss_kappa * len(text_features))
+                    )
+                ).detach()
+                total_loss = (
+                    F.cross_entropy(a_logits_per_audio, labels, weight=audio_weight)
+                    + F.cross_entropy(a_logits_per_text, labels, weight=audio_weight)
+                    + F.cross_entropy(t_logits_per_audio, labels, weight=text_weight)
+                    + F.cross_entropy(t_logits_per_text, labels, weight=text_weight)
+                ) / 4
+        else:
+            if self.world_size > 1:
+                all_audio_features, all_text_features = gather_features(
+                    audio_features=audio_features,
+                    text_features=text_features,
+                    local_loss=self.local_loss,
+                    gather_with_grad=self.gather_with_grad,
+                    rank=self.rank,
+                    world_size=self.world_size,
+                    use_horovod=self.use_horovod,
+                    mlp_loss=self.mlp_loss,
+                )
+
+                if self.local_loss:
+                    logits_per_audio = (
+                        logit_scale_a * audio_features @ all_text_features.T
+                    )
+                    logits_per_text = (
+                        logit_scale_a * text_features @ all_audio_features.T
+                    )
+                else:
+                    logits_per_audio = (
+                        logit_scale_a * all_audio_features @ all_text_features.T
+                    )
+                    logits_per_text = logits_per_audio.T
+            else:
+                logits_per_audio = logit_scale_a * audio_features @ text_features.T
+                logits_per_text = logit_scale_a * text_features @ audio_features.T
+
+            # calculated ground-truth and cache if enabled
+            num_logits = logits_per_audio.shape[0]
+            if self.prev_num_logits != num_logits or device not in self.labels:
+                labels = torch.arange(num_logits, device=device, dtype=torch.long)
+                if self.world_size > 1 and self.local_loss:
+                    labels = labels + num_logits * self.rank
+                if self.cache_labels:
+                    self.labels[device] = labels
+                    self.prev_num_logits = num_logits
+            else:
+                labels = self.labels[device]
+            if not self.weighted_loss:
+                total_loss = (
+                    F.cross_entropy(logits_per_audio, labels)
+                    + F.cross_entropy(logits_per_text, labels)
+                ) / 2
+            else:
+                audio_weight = (all_audio_features @ all_audio_features.T).detach()
+                audio_weight = (
+                    torch.exp(
+                        torch.sum(audio_weight, axis=1)
+                        / (self.weight_loss_kappa * len(all_audio_features))
+                    )
+                ).detach()
+                text_weight = (all_text_features @ all_text_features.T).detach()
+                text_weight = (
+                    torch.exp(
+                        torch.sum(text_weight, axis=1)
+                        / (self.weight_loss_kappa * len(all_text_features))
+                    )
+                ).detach()
+                total_loss = (
+                    F.cross_entropy(logits_per_audio, labels, weight=text_weight)
+                    + F.cross_entropy(logits_per_text, labels, weight=audio_weight)
+                ) / 2
+        return total_loss
+
+
+def lp_gather_features(pred, target, world_size=1, use_horovod=False):
+    if use_horovod:
+        assert hvd is not None, "Please install horovod"
+        with torch.no_grad():
+            all_preds = hvd.allgather(pred)
+            all_targets = hvd.allgath(target)
+    else:
+        gathered_preds = [torch.zeros_like(pred) for _ in range(world_size)]
+        gathered_targets = [torch.zeros_like(target) for _ in range(world_size)]
+
+        dist.all_gather(gathered_preds, pred)
+        dist.all_gather(gathered_targets, target)
+        all_preds = torch.cat(gathered_preds, dim=0)
+        all_targets = torch.cat(gathered_targets, dim=0)
+
+    return all_preds, all_targets
+
+
+def get_map(pred, target):
+    pred = torch.sigmoid(pred).numpy()
+    target = target.numpy()
+    return np.mean(average_precision_score(target, pred, average=None))
+
+
+def get_acc(pred, target):
+    pred = torch.argmax(pred, 1).numpy()
+    target = torch.argmax(target, 1).numpy()
+    return accuracy_score(target, pred)
+
+
+def get_mauc(pred, target):
+    pred = torch.sigmoid(pred).numpy()
+    target = target.numpy()
+    return np.mean(roc_auc_score(target, pred, average=None))
+
+
+class LPMetrics(object):
+    def __init__(self, metric_names=["map", "acc", "mauc"]):
+        self.metrics = []
+        for name in metric_names:
+            self.metrics.append(self.get_metric(name))
+        self.metric_names = metric_names
+
+    def get_metric(self, name):
+        if name == "map":
+            return get_map
+        elif name == "acc":
+            return get_acc
+        elif name == "mauc":
+            return get_mauc
+        else:
+            raise ValueError(f"the metric should be at least one of [map, acc, mauc]")
+
+    def evaluate_mertics(self, pred, target):
+        metric_dict = {}
+        for i in range(len(self.metric_names)):
+            metric_dict[self.metric_names[i]] = self.metrics[i](pred, target)
+        return metric_dict
+
+
+def calc_celoss(pred, target):
+    target = torch.argmax(target, 1).long()
+    return nn.CrossEntropyLoss()(pred, target)
+
+
+class LPLoss(nn.Module):
+    def __init__(self, loss_name):
+        super().__init__()
+        if loss_name == "bce":
+            self.loss_func = nn.BCEWithLogitsLoss()
+        elif loss_name == "ce":
+            self.loss_func = calc_celoss
+        elif loss_name == "mse":
+            self.loss_func = nn.MSELoss()
+        else:
+            raise ValueError(f"the loss func should be at least one of [bce, ce, mse]")
+
+    def forward(self, pred, target):
+        loss = self.loss_func(pred, target)
+        return loss

audiosr/clap/open_clip/model.py

@@ -1,931 +1,931 @@
-""" CLAP Model
-
-Adapted from CLIP: https://github.com/openai/CLIP. Originally MIT License, Copyright (c) 2021 OpenAI.
-Adapted to the Audio Task.
-"""
-
-from collections import OrderedDict
-from dataclasses import dataclass
-from typing import Tuple, Union, Callable, Optional
-
-import numpy as np
-import torch
-import torch.nn.functional as F
-from torch import nn
-
-import logging
-from .utils import freeze_batch_norm_2d
-
-from .pann_model import create_pann_model
-from .htsat import create_htsat_model
-from transformers import BertModel, RobertaModel, BartModel, RobertaConfig
-
-
-class MLPLayers(nn.Module):
-    def __init__(self, units=[512, 512, 512], nonlin=nn.ReLU(), dropout=0.1):
-        super(MLPLayers, self).__init__()
-        self.nonlin = nonlin
-        self.dropout = dropout
-
-        sequence = []
-        for u0, u1 in zip(units[:-1], units[1:]):
-            sequence.append(nn.Linear(u0, u1))
-            sequence.append(self.nonlin)
-            sequence.append(nn.Dropout(self.dropout))
-        sequence = sequence[:-2]
-
-        self.sequential = nn.Sequential(*sequence)
-
-    def forward(self, X):
-        X = self.sequential(X)
-        return X
-
-
-class Bottleneck(nn.Module):
-    expansion = 4
-
-    def __init__(self, inplanes, planes, stride=1):
-        super().__init__()
-
-        # all conv layers have stride 1. an avgpool is performed after the second convolution when stride > 1
-        self.conv1 = nn.Conv2d(inplanes, planes, 1, bias=False)
-        self.bn1 = nn.BatchNorm2d(planes)
-
-        self.conv2 = nn.Conv2d(planes, planes, 3, padding=1, bias=False)
-        self.bn2 = nn.BatchNorm2d(planes)
-
-        self.avgpool = nn.AvgPool2d(stride) if stride > 1 else nn.Identity()
-
-        self.conv3 = nn.Conv2d(planes, planes * self.expansion, 1, bias=False)
-        self.bn3 = nn.BatchNorm2d(planes * self.expansion)
-
-        self.relu = nn.ReLU(inplace=True)
-        self.downsample = None
-        self.stride = stride
-
-        if stride > 1 or inplanes != planes * Bottleneck.expansion:
-            # downsampling layer is prepended with an avgpool, and the subsequent convolution has stride 1
-            self.downsample = nn.Sequential(
-                OrderedDict(
-                    [
-                        ("-1", nn.AvgPool2d(stride)),
-                        (
-                            "0",
-                            nn.Conv2d(
-                                inplanes,
-                                planes * self.expansion,
-                                1,
-                                stride=1,
-                                bias=False,
-                            ),
-                        ),
-                        ("1", nn.BatchNorm2d(planes * self.expansion)),
-                    ]
-                )
-            )
-
-    def forward(self, x: torch.Tensor):
-        identity = x
-
-        out = self.relu(self.bn1(self.conv1(x)))
-        out = self.relu(self.bn2(self.conv2(out)))
-        out = self.avgpool(out)
-        out = self.bn3(self.conv3(out))
-
-        if self.downsample is not None:
-            identity = self.downsample(x)
-
-        out += identity
-        out = self.relu(out)
-        return out
-
-
-class AttentionPool2d(nn.Module):
-    def __init__(
-        self, spacial_dim: int, embed_dim: int, num_heads: int, output_dim: int = None
-    ):
-        super().__init__()
-        self.positional_embedding = nn.Parameter(
-            torch.randn(spacial_dim**2 + 1, embed_dim) / embed_dim**0.5
-        )
-        self.k_proj = nn.Linear(embed_dim, embed_dim)
-        self.q_proj = nn.Linear(embed_dim, embed_dim)
-        self.v_proj = nn.Linear(embed_dim, embed_dim)
-        self.c_proj = nn.Linear(embed_dim, output_dim or embed_dim)
-        self.num_heads = num_heads
-
-    def forward(self, x):
-        x = x.reshape(x.shape[0], x.shape[1], x.shape[2] * x.shape[3]).permute(
-            2, 0, 1
-        )  # NCHW -> (HW)NC
-        x = torch.cat([x.mean(dim=0, keepdim=True), x], dim=0)  # (HW+1)NC
-        x = x + self.positional_embedding[:, None, :].to(x.dtype)  # (HW+1)NC
-        x, _ = F.multi_head_attention_forward(
-            query=x,
-            key=x,
-            value=x,
-            embed_dim_to_check=x.shape[-1],
-            num_heads=self.num_heads,
-            q_proj_weight=self.q_proj.weight,
-            k_proj_weight=self.k_proj.weight,
-            v_proj_weight=self.v_proj.weight,
-            in_proj_weight=None,
-            in_proj_bias=torch.cat(
-                [self.q_proj.bias, self.k_proj.bias, self.v_proj.bias]
-            ),
-            bias_k=None,
-            bias_v=None,
-            add_zero_attn=False,
-            dropout_p=0,
-            out_proj_weight=self.c_proj.weight,
-            out_proj_bias=self.c_proj.bias,
-            use_separate_proj_weight=True,
-            training=self.training,
-            need_weights=False,
-        )
-
-        return x[0]
-
-
-class ModifiedResNet(nn.Module):
-    """
-    A ResNet class that is similar to torchvision's but contains the following changes:
-    - There are now 3 "stem" convolutions as opposed to 1, with an average pool instead of a max pool.
-    - Performs anti-aliasing strided convolutions, where an avgpool is prepended to convolutions with stride > 1
-    - The final pooling layer is a QKV attention instead of an average pool
-    """
-
-    def __init__(self, layers, output_dim, heads, image_size=224, width=64):
-        super().__init__()
-        self.output_dim = output_dim
-        self.image_size = image_size
-
-        # the 3-layer stem
-        self.conv1 = nn.Conv2d(
-            3, width // 2, kernel_size=3, stride=2, padding=1, bias=False
-        )
-        self.bn1 = nn.BatchNorm2d(width // 2)
-        self.conv2 = nn.Conv2d(
-            width // 2, width // 2, kernel_size=3, padding=1, bias=False
-        )
-        self.bn2 = nn.BatchNorm2d(width // 2)
-        self.conv3 = nn.Conv2d(width // 2, width, kernel_size=3, padding=1, bias=False)
-        self.bn3 = nn.BatchNorm2d(width)
-        self.avgpool = nn.AvgPool2d(2)
-        self.relu = nn.ReLU(inplace=True)
-
-        # residual layers
-        self._inplanes = width  # this is a *mutable* variable used during construction
-        self.layer1 = self._make_layer(width, layers[0])
-        self.layer2 = self._make_layer(width * 2, layers[1], stride=2)
-        self.layer3 = self._make_layer(width * 4, layers[2], stride=2)
-        self.layer4 = self._make_layer(width * 8, layers[3], stride=2)
-
-        embed_dim = width * 32  # the ResNet feature dimension
-        self.attnpool = AttentionPool2d(image_size // 32, embed_dim, heads, output_dim)
-
-        self.init_parameters()
-
-    def _make_layer(self, planes, blocks, stride=1):
-        layers = [Bottleneck(self._inplanes, planes, stride)]
-
-        self._inplanes = planes * Bottleneck.expansion
-        for _ in range(1, blocks):
-            layers.append(Bottleneck(self._inplanes, planes))
-
-        return nn.Sequential(*layers)
-
-    def init_parameters(self):
-        if self.attnpool is not None:
-            std = self.attnpool.c_proj.in_features**-0.5
-            nn.init.normal_(self.attnpool.q_proj.weight, std=std)
-            nn.init.normal_(self.attnpool.k_proj.weight, std=std)
-            nn.init.normal_(self.attnpool.v_proj.weight, std=std)
-            nn.init.normal_(self.attnpool.c_proj.weight, std=std)
-
-        for resnet_block in [self.layer1, self.layer2, self.layer3, self.layer4]:
-            for name, param in resnet_block.named_parameters():
-                if name.endswith("bn3.weight"):
-                    nn.init.zeros_(param)
-
-    def lock(self, unlocked_groups=0, freeze_bn_stats=False):
-        assert (
-            unlocked_groups == 0
-        ), "partial locking not currently supported for this model"
-        for param in self.parameters():
-            param.requires_grad = False
-        if freeze_bn_stats:
-            freeze_batch_norm_2d(self)
-
-    def stem(self, x):
-        for conv, bn in [
-            (self.conv1, self.bn1),
-            (self.conv2, self.bn2),
-            (self.conv3, self.bn3),
-        ]:
-            x = self.relu(bn(conv(x)))
-        x = self.avgpool(x)
-        return x
-
-    def forward(self, x):
-        x = self.stem(x)
-        x = self.layer1(x)
-        x = self.layer2(x)
-        x = self.layer3(x)
-        x = self.layer4(x)
-        x = self.attnpool(x)
-
-        return x
-
-
-class LayerNorm(nn.LayerNorm):
-    """Subclass torch's LayerNorm to handle fp16."""
-
-    def forward(self, x: torch.Tensor):
-        orig_type = x.dtype
-        x = F.layer_norm(x, self.normalized_shape, self.weight, self.bias, self.eps)
-        return x.to(orig_type)
-
-
-class QuickGELU(nn.Module):
-    # NOTE This is slower than nn.GELU or nn.SiLU and uses more GPU memory
-    def forward(self, x: torch.Tensor):
-        return x * torch.sigmoid(1.702 * x)
-
-
-class ResidualAttentionBlock(nn.Module):
-    def __init__(self, d_model: int, n_head: int, act_layer: Callable = nn.GELU):
-        super().__init__()
-
-        self.attn = nn.MultiheadAttention(d_model, n_head)
-        self.ln_1 = LayerNorm(d_model)
-        self.mlp = nn.Sequential(
-            OrderedDict(
-                [
-                    ("c_fc", nn.Linear(d_model, d_model * 4)),
-                    ("gelu", act_layer()),
-                    ("c_proj", nn.Linear(d_model * 4, d_model)),
-                ]
-            )
-        )
-        self.ln_2 = LayerNorm(d_model)
-
-    def attention(self, x: torch.Tensor, attn_mask: Optional[torch.Tensor] = None):
-        return self.attn(x, x, x, need_weights=False, attn_mask=attn_mask)[0]
-
-    def forward(self, x: torch.Tensor, attn_mask: Optional[torch.Tensor] = None):
-        x = x + self.attention(self.ln_1(x), attn_mask=attn_mask)
-        x = x + self.mlp(self.ln_2(x))
-        return x
-
-
-class Transformer(nn.Module):
-    def __init__(
-        self, width: int, layers: int, heads: int, act_layer: Callable = nn.GELU
-    ):
-        super().__init__()
-        self.width = width
-        self.layers = layers
-        self.resblocks = nn.ModuleList(
-            [
-                ResidualAttentionBlock(width, heads, act_layer=act_layer)
-                for _ in range(layers)
-            ]
-        )
-
-    def forward(self, x: torch.Tensor, attn_mask: Optional[torch.Tensor] = None):
-        for r in self.resblocks:
-            x = r(x, attn_mask=attn_mask)
-        return x
-
-
-class VisualTransformer(nn.Module):
-    def __init__(
-        self,
-        image_size: int,
-        patch_size: int,
-        width: int,
-        layers: int,
-        heads: int,
-        output_dim: int,
-        act_layer: Callable = nn.GELU,
-    ):
-        super().__init__()
-        self.image_size = image_size
-        self.output_dim = output_dim
-        self.conv1 = nn.Conv2d(
-            in_channels=3,
-            out_channels=width,
-            kernel_size=patch_size,
-            stride=patch_size,
-            bias=False,
-        )
-
-        scale = width**-0.5
-        self.class_embedding = nn.Parameter(scale * torch.randn(width))
-        self.positional_embedding = nn.Parameter(
-            scale * torch.randn((image_size // patch_size) ** 2 + 1, width)
-        )
-        self.ln_pre = LayerNorm(width)
-
-        self.text_branch = Transformer(width, layers, heads, act_layer=act_layer)
-
-        self.ln_post = LayerNorm(width)
-        self.proj = nn.Parameter(scale * torch.randn(width, output_dim))
-
-    def lock(self, unlocked_groups=0, freeze_bn_stats=False):
-        assert (
-            unlocked_groups == 0
-        ), "partial locking not currently supported for this model"
-        for param in self.parameters():
-            param.requires_grad = False
-
-    def forward(self, x: torch.Tensor):
-        x = self.conv1(x)  # shape = [*, width, grid, grid]
-        x = x.reshape(x.shape[0], x.shape[1], -1)  # shape = [*, width, grid ** 2]
-        x = x.permute(0, 2, 1)  # shape = [*, grid ** 2, width]
-        x = torch.cat(
-            [
-                self.class_embedding.to(x.dtype)
-                + torch.zeros(
-                    x.shape[0], 1, x.shape[-1], dtype=x.dtype, device=x.device
-                ),
-                x,
-            ],
-            dim=1,
-        )  # shape = [*, grid ** 2 + 1, width]
-        x = x + self.positional_embedding.to(x.dtype)
-        x = self.ln_pre(x)
-
-        x = x.permute(1, 0, 2)  # NLD -> LND
-        x = self.text_branch(x)
-        x = x.permute(1, 0, 2)  # LND -> NLD
-
-        x = self.ln_post(x[:, 0, :])
-
-        if self.proj is not None:
-            x = x @ self.proj
-
-        return x
-
-
-@dataclass
-class CLAPVisionCfg:
-    layers: Union[Tuple[int, int, int, int], int] = 12
-    width: int = 768
-    patch_size: int = 16
-    image_size: Union[Tuple[int, int], int] = 224
-    timm_model_name: str = (
-        None  # a valid model name overrides layers, width, patch_size
-    )
-    timm_model_pretrained: bool = (
-        False  # use (imagenet) pretrained weights for named model
-    )
-    timm_pool: str = (
-        "avg"  # feature pooling for timm model ('abs_attn', 'rot_attn', 'avg', '')
-    )
-    timm_proj: str = (
-        "linear"  # linear projection for timm model output ('linear', 'mlp', '')
-    )
-
-
-# Audio Config Class
-@dataclass
-class CLAPAudioCfp:
-    model_type: str = "PANN"
-    model_name: str = "Cnn14"
-    sample_rate: int = 48000
-    # Param
-    audio_length: int = 1024
-    window_size: int = 1024
-    hop_size: int = 1024
-    fmin: int = 50
-    fmax: int = 14000
-    class_num: int = 527
-    mel_bins: int = 64
-    clip_samples: int = 480000
-
-
-@dataclass
-class CLAPTextCfg:
-    context_length: int
-    vocab_size: int
-    width: int
-    heads: int
-    layers: int
-    model_type: str
-
-
-class CLAP(nn.Module):
-    def __init__(
-        self,
-        embed_dim: int,
-        audio_cfg: CLAPAudioCfp,
-        text_cfg: CLAPTextCfg,
-        quick_gelu: bool = False,
-        enable_fusion: bool = False,
-        fusion_type: str = "None",
-        joint_embed_shape: int = 512,
-        mlp_act: str = "relu",
-    ):
-        super().__init__()
-        if isinstance(audio_cfg, dict):
-            audio_cfg = CLAPAudioCfp(**audio_cfg)
-        if isinstance(text_cfg, dict):
-            text_cfg = CLAPTextCfg(**text_cfg)
-
-        self.audio_cfg = audio_cfg
-        self.text_cfg = text_cfg
-        self.enable_fusion = enable_fusion
-        self.fusion_type = fusion_type
-        self.joint_embed_shape = joint_embed_shape
-        self.mlp_act = mlp_act
-
-        self.context_length = text_cfg.context_length
-
-        # OpenAI models are pretrained w/ QuickGELU but native nn.GELU is both faster and more
-        # memory efficient in recent PyTorch releases (>= 1.10).
-        # NOTE: timm models always use native GELU regardless of quick_gelu flag.
-        act_layer = QuickGELU if quick_gelu else nn.GELU
-
-        if mlp_act == "relu":
-            mlp_act_layer = nn.ReLU()
-        elif mlp_act == "gelu":
-            mlp_act_layer = nn.GELU()
-        else:
-            raise NotImplementedError
-
-        # audio branch
-        # audio branch parameters
-        if audio_cfg.model_type == "PANN":
-            self.audio_branch = create_pann_model(audio_cfg, enable_fusion, fusion_type)
-        elif audio_cfg.model_type == "HTSAT":
-            self.audio_branch = create_htsat_model(
-                audio_cfg, enable_fusion, fusion_type
-            )
-        else:
-            logging.error(f"Model config for {audio_cfg.model_type} not found")
-            raise RuntimeError(f"Model config for {audio_cfg.model_type} not found.")
-
-        # text branch
-        # text branch parameters
-        if text_cfg.model_type == "transformer":
-            self.text_branch = Transformer(
-                width=text_cfg.width,
-                layers=text_cfg.layers,
-                heads=text_cfg.heads,
-                act_layer=act_layer,
-            )
-            self.vocab_size = text_cfg.vocab_size
-            self.token_embedding = nn.Embedding(text_cfg.vocab_size, text_cfg.width)
-            self.positional_embedding = nn.Parameter(
-                torch.empty(self.context_length, text_cfg.width)
-            )
-            self.ln_final = LayerNorm(text_cfg.width)
-            self.text_transform = MLPLayers(
-                units=[
-                    self.joint_embed_shape,
-                    self.joint_embed_shape,
-                    self.joint_embed_shape,
-                ],
-                dropout=0.1,
-            )
-            self.text_projection = nn.Sequential(
-                nn.Linear(text_cfg.width, self.joint_embed_shape),
-                mlp_act_layer,
-                nn.Linear(self.joint_embed_shape, self.joint_embed_shape),
-            )
-        elif text_cfg.model_type == "bert":
-            self.text_branch = BertModel.from_pretrained("bert-base-uncased")
-            self.text_transform = MLPLayers(
-                units=[
-                    self.joint_embed_shape,
-                    self.joint_embed_shape,
-                    self.joint_embed_shape,
-                ],
-                dropout=0.1,
-            )
-            self.text_projection = nn.Sequential(
-                nn.Linear(768, self.joint_embed_shape),
-                mlp_act_layer,
-                nn.Linear(self.joint_embed_shape, self.joint_embed_shape),
-            )
-        elif text_cfg.model_type == "roberta":
-            self.text_branch = RobertaModel(
-                RobertaConfig.from_pretrained("roberta-base")
-            )
-            self.text_transform = MLPLayers(
-                units=[
-                    self.joint_embed_shape,
-                    self.joint_embed_shape,
-                    self.joint_embed_shape,
-                ],
-                dropout=0.1,
-            )
-            self.text_projection = nn.Sequential(
-                nn.Linear(768, self.joint_embed_shape),
-                mlp_act_layer,
-                nn.Linear(self.joint_embed_shape, self.joint_embed_shape),
-            )
-        elif text_cfg.model_type == "bart":
-            self.text_branch = BartModel.from_pretrained("facebook/bart-base")
-            self.text_transform = MLPLayers(
-                units=[
-                    self.joint_embed_shape,
-                    self.joint_embed_shape,
-                    self.joint_embed_shape,
-                ],
-                dropout=0.1,
-            )
-            self.text_projection = nn.Sequential(
-                nn.Linear(768, self.joint_embed_shape),
-                mlp_act_layer,
-                nn.Linear(self.joint_embed_shape, self.joint_embed_shape),
-            )
-        else:
-            logging.error(f"Model config for {text_cfg.model_type} not found")
-            raise RuntimeError(f"Model config for {text_cfg.model_type} not found.")
-        self.text_branch_type = text_cfg.model_type
-        # text branch parameters
-
-        # audio branch parameters
-        self.audio_transform = MLPLayers(
-            units=[
-                self.joint_embed_shape,
-                self.joint_embed_shape,
-                self.joint_embed_shape,
-            ],
-            dropout=0.1,
-        )
-
-        # below here is text branch parameters
-
-        # ============================================================================================================
-        self.audio_projection = nn.Sequential(
-            nn.Linear(embed_dim, self.joint_embed_shape),
-            mlp_act_layer,
-            nn.Linear(self.joint_embed_shape, self.joint_embed_shape),
-        )
-
-        self.logit_scale_a = nn.Parameter(torch.ones([]) * np.log(1 / 0.07))
-        self.logit_scale_t = nn.Parameter(torch.ones([]) * np.log(1 / 0.07))
-        self.register_buffer("attn_mask", self.build_attention_mask(), persistent=False)
-
-        self.init_text_branch_parameters()
-
-    def init_text_branch_parameters(self):
-        if self.text_branch_type == "transformer":
-            nn.init.normal_(self.token_embedding.weight, std=0.02)
-            nn.init.normal_(self.positional_embedding, std=0.01)
-            proj_std = (self.text_branch.width**-0.5) * (
-                (2 * self.text_branch.layers) ** -0.5
-            )
-            attn_std = self.text_branch.width**-0.5
-            fc_std = (2 * self.text_branch.width) ** -0.5
-            for block in self.text_branch.resblocks:
-                nn.init.normal_(block.attn.in_proj_weight, std=attn_std)
-                nn.init.normal_(block.attn.out_proj.weight, std=proj_std)
-                nn.init.normal_(block.mlp.c_fc.weight, std=fc_std)
-                nn.init.normal_(block.mlp.c_proj.weight, std=proj_std)
-        if self.text_branch_type == "bert" or self.text_branch_type == "roberta":
-            self.text_branch.embeddings.word_embeddings.weight.shape[-1]
-        elif self.text_branch_type == "bart":
-            self.text_branch.shared.weight.shape[-1]
-        else:
-            self.text_branch.width
-        nn.init.constant_(self.logit_scale_a, np.log(1 / 0.07))
-        nn.init.constant_(self.logit_scale_t, np.log(1 / 0.07))
-
-        # deprecated
-        # if hasattr(self.visual, 'init_parameters'):
-        # self.visual.init_parameters()
-
-        # if self.text_projection is not None:
-        #     nn.init.normal_(self.text_projection, std=width**-0.5)
-
-    def build_attention_mask(self):
-        # lazily create causal attention mask, with full attention between the vision tokens
-        # pytorch uses additive attention mask; fill with -inf
-        mask = torch.empty(self.context_length, self.context_length)
-        mask.fill_(float("-inf"))
-        mask.triu_(1)  # zero out the lower diagonal
-        return mask
-
-    def encode_audio(self, audio, device):
-        return self.audio_branch(
-            audio, mixup_lambda=None, device=device
-        )  # mix lambda needs to add
-
-    # def list_of_dict_of_tensor2dict_of_tensor(self, x, device):
-    #     tmp = {}
-    #     for k in x[0].keys():
-    #         tmp[k] = []
-    #         for i in range(len(x)):
-    #             tmp[k].append(x[i][k][:77])
-    #     for k in x[0].keys():
-    #         tmp[k] = torch.tensor(tmp[k]).to(device=device, non_blocking=True)
-    #     return tmp
-
-    def encode_text(self, text, device):
-        if self.text_branch_type == "transformer":
-            text = text.to(device=device, non_blocking=True)
-            x = self.token_embedding(text)  # [batch_size, n_ctx, d_model]
-
-            x = x + self.positional_embedding
-            x = x.permute(1, 0, 2)  # NLD -> LND
-            x = self.text_branch(x, attn_mask=self.attn_mask)
-            x = x.permute(1, 0, 2)  # LND -> NLD
-            x = self.ln_final(x)
-
-            # x.shape = [batch_size, n_ctx, transformer.width]
-            # take features from the eot embedding (eot_token is the highest number in each sequence)
-            x = self.text_projection(x[torch.arange(x.shape[0]), text.argmax(dim=-1)])
-        elif self.text_branch_type == "bert":
-            # text = self.list_of_dict_of_tensor2dict_of_tensor(text, device)
-            # text = BatchEncoding(text)
-            x = self.text_branch(
-                input_ids=text["input_ids"].to(device=device, non_blocking=True),
-                attention_mask=text["attention_mask"].to(
-                    device=device, non_blocking=True
-                ),
-                token_type_ids=text["token_type_ids"].to(
-                    device=device, non_blocking=True
-                ),
-            )["pooler_output"]
-            x = self.text_projection(x)
-        elif self.text_branch_type == "roberta":
-            x = self.text_branch(
-                input_ids=text["input_ids"].to(device=device, non_blocking=True),
-                attention_mask=text["attention_mask"].to(
-                    device=device, non_blocking=True
-                ),
-            )["pooler_output"]
-            x = self.text_projection(x)
-        elif self.text_branch_type == "bart":
-            x = torch.mean(
-                self.text_branch(
-                    input_ids=text["input_ids"].to(device=device, non_blocking=True),
-                    attention_mask=text["attention_mask"].to(
-                        device=device, non_blocking=True
-                    ),
-                )["encoder_last_hidden_state"],
-                axis=1,
-            )
-            x = self.text_projection(x)
-        else:
-            logging.error(f"Model type {self.text_branch_type} not found")
-            raise RuntimeError(f"Model type {self.text_branch_type} not found.")
-        return x
-
-    def forward(self, audio, text, device=None):
-        """Forward audio and text into the CLAP
-
-        Parameters
-        ----------
-        audio: torch.Tensor (batch_size, audio_length)
-            the time-domain audio input / the batch of mel_spec and longer list.
-        text: torch.Tensor () // need to add
-            the text token input
-        """
-        if device is None:
-            if audio is not None:
-                device = audio.device
-            elif text is not None:
-                device = text.device
-        if audio is None and text is None:
-            # a hack to get the logit scale
-            return self.logit_scale_a.exp(), self.logit_scale_t.exp()
-        elif audio is None:
-            return self.encode_text(text, device=device)
-        elif text is None:
-            return self.audio_projection(
-                self.encode_audio(audio, device=device)["embedding"]
-            )
-        audio_features = self.audio_projection(
-            self.encode_audio(audio, device=device)["embedding"]
-        )
-        audio_features = F.normalize(audio_features, dim=-1)
-
-        text_features = self.encode_text(text, device=device)
-        # print("text_features", text_features)
-        # print("text_features.shape", text_features.shape)
-        # print("text_features.type", type(text_features))
-        text_features = F.normalize(text_features, dim=-1)
-
-        audio_features_mlp = self.audio_transform(audio_features)
-        text_features_mlp = self.text_transform(text_features)
-        # Four outputs: audio features (basic & MLP), text features (basic & MLP)
-        return (
-            audio_features,
-            text_features,
-            audio_features_mlp,
-            text_features_mlp,
-            self.logit_scale_a.exp(),
-            self.logit_scale_t.exp(),
-        )
-
-    def get_logit_scale(self):
-        return self.logit_scale_a.exp(), self.logit_scale_t.exp()
-
-    def get_text_embedding(self, data):
-        """Get the text embedding from the model
-
-        Parameters
-        ----------
-        data: torch.Tensor
-            a tensor of text embedding
-
-        Returns
-        ----------
-        text_embed: torch.Tensor
-            a tensor of text_embeds (N, D)
-
-        """
-        device = next(self.parameters()).device
-        for k in data:
-            data[k] = data[k].to(device)
-        text_embeds = self.encode_text(data, device=device)
-        text_embeds = F.normalize(text_embeds, dim=-1)
-
-        return text_embeds
-
-    def get_audio_embedding(self, data):
-        """Get the audio embedding from the model
-
-        Parameters
-        ----------
-        data: a list of dict
-            the audio input dict list from 'get_audio_feature' method
-
-        Returns
-        ----------
-        audio_embed: torch.Tensor
-            a tensor of audio_embeds (N, D)
-
-        """
-        device = next(self.parameters()).device
-        # input_dict = {}
-        # keys = data[0].keys()
-        # for k in keys:
-        #     input_dict[k] = torch.cat([d[k].unsqueeze(0) for d in data], dim=0).to(
-        #         device
-        #     )
-        audio_embeds = self.audio_projection(
-            self.encode_audio(data, device=device)["embedding"]
-        )
-        audio_embeds = F.normalize(audio_embeds, dim=-1)
-
-        return audio_embeds
-
-    def audio_infer(self, audio, hopsize=None, device=None):
-        """Forward one audio and produce the audio embedding
-
-        Parameters
-        ----------
-        audio:  (audio_length)
-            the time-domain audio input, notice that it must be only one input
-        hopsize: int
-            the overlap hopsize as the sliding window
-
-        Returns
-        ----------
-        output_dict: {
-            key: [n, (embedding_shape)] if "HTS-AT"
-            or
-            key: [(embedding_shape)] if "PANN"
-        }
-            the list of key values of the audio branch
-
-        """
-
-        assert not self.training, "the inference mode must be run at eval stage"
-        output_dict = {}
-        # PANN
-        if self.audio_cfg.model_type == "PANN":
-            audio_input = audio.unsqueeze(dim=0)
-            output_dict[key] = self.encode_audio(audio_input, device=device)[
-                key
-            ].squeeze(dim=0)
-        elif self.audio_cfg.model_type == "HTSAT":
-            # repeat
-            audio_len = len(audio)
-            k = self.audio_cfg.clip_samples // audio_len
-            if k > 1:
-                audio = audio.repeat(k)
-                audio_len = len(audio)
-
-            if hopsize is None:
-                hopsize = min(hopsize, audio_len)
-
-            if audio_len > self.audio_cfg.clip_samples:
-                audio_input = [
-                    audio[pos : pos + self.audio_cfg.clip_samples].clone()
-                    for pos in range(
-                        0, audio_len - self.audio_cfg.clip_samples, hopsize
-                    )
-                ]
-                audio_input.append(audio[-self.audio_cfg.clip_samples :].clone())
-                audio_input = torch.stack(audio_input)
-                output_dict[key] = self.encode_audio(audio_input, device=device)[key]
-            else:
-                audio_input = audio.unsqueeze(dim=0)
-                output_dict[key] = self.encode_audio(audio_input, device=device)[
-                    key
-                ].squeeze(dim=0)
-
-        return output_dict
-
-
-def convert_weights_to_fp16(model: nn.Module):
-    """Convert applicable model parameters to fp16"""
-
-    def _convert_weights_to_fp16(l):
-        if isinstance(l, (nn.Conv1d, nn.Conv2d, nn.Linear)):
-            l.weight.data = l.weight.data.half()
-            if l.bias is not None:
-                l.bias.data = l.bias.data.half()
-
-        if isinstance(l, nn.MultiheadAttention):
-            for attr in [
-                *[f"{s}_proj_weight" for s in ["in", "q", "k", "v"]],
-                "in_proj_bias",
-                "bias_k",
-                "bias_v",
-            ]:
-                tensor = getattr(l, attr)
-                if tensor is not None:
-                    tensor.data = tensor.data.half()
-
-        for name in ["text_projection", "proj"]:
-            if hasattr(l, name):
-                attr = getattr(l, name)
-                if attr is not None:
-                    attr.data = attr.data.half()
-
-    model.apply(_convert_weights_to_fp16)
-
-
-# Ignore the state dict of the vision part
-def build_model_from_openai_state_dict(
-    state_dict: dict, model_cfg, enable_fusion: bool = False, fusion_type: str = "None"
-):
-    embed_dim = model_cfg["embed_dim"]
-    audio_cfg = model_cfg["audio_cfg"]
-    text_cfg = model_cfg["text_cfg"]
-    state_dict["positional_embedding"].shape[0]
-    state_dict["token_embedding.weight"].shape[0]
-    transformer_width = state_dict["ln_final.weight"].shape[0]
-    transformer_width // 64
-    transformer_layers = len(
-        set(
-            k.split(".")[2]
-            for k in state_dict
-            if k.startswith(f"transformer.resblocks")
-        )
-    )
-
-    audio_cfg = CLAPAudioCfp(**audio_cfg)
-    text_cfg = CLAPTextCfg(**text_cfg)
-
-    model = CLAP(
-        embed_dim,
-        audio_cfg=audio_cfg,
-        text_cfg=text_cfg,
-        quick_gelu=True,  # OpenAI models were trained with QuickGELU
-        enable_fusion=enable_fusion,
-        fusion_type=fusion_type,
-    )
-    state_dict["logit_scale_a"] = state_dict["logit_scale"]
-    state_dict["logit_scale_t"] = state_dict["logit_scale"]
-    pop_keys = list(state_dict.keys())[::]
-    # pop the visual branch saved weights
-    for key in pop_keys:
-        if key.startswith("visual."):
-            state_dict.pop(key, None)
-
-    for key in ["logit_scale", "input_resolution", "context_length", "vocab_size"]:
-        state_dict.pop(key, None)
-
-    # not use fp16
-    # convert_weights_to_fp16(model)
-    model.load_state_dict(state_dict, strict=False)
-    return model.eval()
-
-
-def trace_model(model, batch_size=256, device=torch.device("cpu")):
-    model.eval()
-    audio_length = model.audio_cfg.audio_length
-    example_audio = torch.ones((batch_size, audio_length), device=device)
-    example_text = torch.zeros(
-        (batch_size, model.context_length), dtype=torch.int, device=device
-    )
-    model = torch.jit.trace_module(
-        model,
-        inputs=dict(
-            forward=(example_audio, example_text),
-            encode_text=(example_text,),
-            encode_image=(example_audio,),
-        ),
-    )
-    model.audio_cfg.audio_length = audio_length  # Question: what does this do?
-    return model
+""" CLAP Model
+
+Adapted from CLIP: https://github.com/openai/CLIP. Originally MIT License, Copyright (c) 2021 OpenAI.
+Adapted to the Audio Task.
+"""
+
+from collections import OrderedDict
+from dataclasses import dataclass
+from typing import Tuple, Union, Callable, Optional
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+import logging
+from .utils import freeze_batch_norm_2d
+
+from .pann_model import create_pann_model
+from .htsat import create_htsat_model
+from transformers import BertModel, RobertaModel, BartModel, RobertaConfig
+
+
+class MLPLayers(nn.Module):
+    def __init__(self, units=[512, 512, 512], nonlin=nn.ReLU(), dropout=0.1):
+        super(MLPLayers, self).__init__()
+        self.nonlin = nonlin
+        self.dropout = dropout
+
+        sequence = []
+        for u0, u1 in zip(units[:-1], units[1:]):
+            sequence.append(nn.Linear(u0, u1))
+            sequence.append(self.nonlin)
+            sequence.append(nn.Dropout(self.dropout))
+        sequence = sequence[:-2]
+
+        self.sequential = nn.Sequential(*sequence)
+
+    def forward(self, X):
+        X = self.sequential(X)
+        return X
+
+
+class Bottleneck(nn.Module):
+    expansion = 4
+
+    def __init__(self, inplanes, planes, stride=1):
+        super().__init__()
+
+        # all conv layers have stride 1. an avgpool is performed after the second convolution when stride > 1
+        self.conv1 = nn.Conv2d(inplanes, planes, 1, bias=False)
+        self.bn1 = nn.BatchNorm2d(planes)
+
+        self.conv2 = nn.Conv2d(planes, planes, 3, padding=1, bias=False)
+        self.bn2 = nn.BatchNorm2d(planes)
+
+        self.avgpool = nn.AvgPool2d(stride) if stride > 1 else nn.Identity()
+
+        self.conv3 = nn.Conv2d(planes, planes * self.expansion, 1, bias=False)
+        self.bn3 = nn.BatchNorm2d(planes * self.expansion)
+
+        self.relu = nn.ReLU(inplace=True)
+        self.downsample = None
+        self.stride = stride
+
+        if stride > 1 or inplanes != planes * Bottleneck.expansion:
+            # downsampling layer is prepended with an avgpool, and the subsequent convolution has stride 1
+            self.downsample = nn.Sequential(
+                OrderedDict(
+                    [
+                        ("-1", nn.AvgPool2d(stride)),
+                        (
+                            "0",
+                            nn.Conv2d(
+                                inplanes,
+                                planes * self.expansion,
+                                1,
+                                stride=1,
+                                bias=False,
+                            ),
+                        ),
+                        ("1", nn.BatchNorm2d(planes * self.expansion)),
+                    ]
+                )
+            )
+
+    def forward(self, x: torch.Tensor):
+        identity = x
+
+        out = self.relu(self.bn1(self.conv1(x)))
+        out = self.relu(self.bn2(self.conv2(out)))
+        out = self.avgpool(out)
+        out = self.bn3(self.conv3(out))
+
+        if self.downsample is not None:
+            identity = self.downsample(x)
+
+        out += identity
+        out = self.relu(out)
+        return out
+
+
+class AttentionPool2d(nn.Module):
+    def __init__(
+        self, spacial_dim: int, embed_dim: int, num_heads: int, output_dim: int = None
+    ):
+        super().__init__()
+        self.positional_embedding = nn.Parameter(
+            torch.randn(spacial_dim**2 + 1, embed_dim) / embed_dim**0.5
+        )
+        self.k_proj = nn.Linear(embed_dim, embed_dim)
+        self.q_proj = nn.Linear(embed_dim, embed_dim)
+        self.v_proj = nn.Linear(embed_dim, embed_dim)
+        self.c_proj = nn.Linear(embed_dim, output_dim or embed_dim)
+        self.num_heads = num_heads
+
+    def forward(self, x):
+        x = x.reshape(x.shape[0], x.shape[1], x.shape[2] * x.shape[3]).permute(
+            2, 0, 1
+        )  # NCHW -> (HW)NC
+        x = torch.cat([x.mean(dim=0, keepdim=True), x], dim=0)  # (HW+1)NC
+        x = x + self.positional_embedding[:, None, :].to(x.dtype)  # (HW+1)NC
+        x, _ = F.multi_head_attention_forward(
+            query=x,
+            key=x,
+            value=x,
+            embed_dim_to_check=x.shape[-1],
+            num_heads=self.num_heads,
+            q_proj_weight=self.q_proj.weight,
+            k_proj_weight=self.k_proj.weight,
+            v_proj_weight=self.v_proj.weight,
+            in_proj_weight=None,
+            in_proj_bias=torch.cat(
+                [self.q_proj.bias, self.k_proj.bias, self.v_proj.bias]
+            ),
+            bias_k=None,
+            bias_v=None,
+            add_zero_attn=False,
+            dropout_p=0,
+            out_proj_weight=self.c_proj.weight,
+            out_proj_bias=self.c_proj.bias,
+            use_separate_proj_weight=True,
+            training=self.training,
+            need_weights=False,
+        )
+
+        return x[0]
+
+
+class ModifiedResNet(nn.Module):
+    """
+    A ResNet class that is similar to torchvision's but contains the following changes:
+    - There are now 3 "stem" convolutions as opposed to 1, with an average pool instead of a max pool.
+    - Performs anti-aliasing strided convolutions, where an avgpool is prepended to convolutions with stride > 1
+    - The final pooling layer is a QKV attention instead of an average pool
+    """
+
+    def __init__(self, layers, output_dim, heads, image_size=224, width=64):
+        super().__init__()
+        self.output_dim = output_dim
+        self.image_size = image_size
+
+        # the 3-layer stem
+        self.conv1 = nn.Conv2d(
+            3, width // 2, kernel_size=3, stride=2, padding=1, bias=False
+        )
+        self.bn1 = nn.BatchNorm2d(width // 2)
+        self.conv2 = nn.Conv2d(
+            width // 2, width // 2, kernel_size=3, padding=1, bias=False
+        )
+        self.bn2 = nn.BatchNorm2d(width // 2)
+        self.conv3 = nn.Conv2d(width // 2, width, kernel_size=3, padding=1, bias=False)
+        self.bn3 = nn.BatchNorm2d(width)
+        self.avgpool = nn.AvgPool2d(2)
+        self.relu = nn.ReLU(inplace=True)
+
+        # residual layers
+        self._inplanes = width  # this is a *mutable* variable used during construction
+        self.layer1 = self._make_layer(width, layers[0])
+        self.layer2 = self._make_layer(width * 2, layers[1], stride=2)
+        self.layer3 = self._make_layer(width * 4, layers[2], stride=2)
+        self.layer4 = self._make_layer(width * 8, layers[3], stride=2)
+
+        embed_dim = width * 32  # the ResNet feature dimension
+        self.attnpool = AttentionPool2d(image_size // 32, embed_dim, heads, output_dim)
+
+        self.init_parameters()
+
+    def _make_layer(self, planes, blocks, stride=1):
+        layers = [Bottleneck(self._inplanes, planes, stride)]
+
+        self._inplanes = planes * Bottleneck.expansion
+        for _ in range(1, blocks):
+            layers.append(Bottleneck(self._inplanes, planes))
+
+        return nn.Sequential(*layers)
+
+    def init_parameters(self):
+        if self.attnpool is not None:
+            std = self.attnpool.c_proj.in_features**-0.5
+            nn.init.normal_(self.attnpool.q_proj.weight, std=std)
+            nn.init.normal_(self.attnpool.k_proj.weight, std=std)
+            nn.init.normal_(self.attnpool.v_proj.weight, std=std)
+            nn.init.normal_(self.attnpool.c_proj.weight, std=std)
+
+        for resnet_block in [self.layer1, self.layer2, self.layer3, self.layer4]:
+            for name, param in resnet_block.named_parameters():
+                if name.endswith("bn3.weight"):
+                    nn.init.zeros_(param)
+
+    def lock(self, unlocked_groups=0, freeze_bn_stats=False):
+        assert (
+            unlocked_groups == 0
+        ), "partial locking not currently supported for this model"
+        for param in self.parameters():
+            param.requires_grad = False
+        if freeze_bn_stats:
+            freeze_batch_norm_2d(self)
+
+    def stem(self, x):
+        for conv, bn in [
+            (self.conv1, self.bn1),
+            (self.conv2, self.bn2),
+            (self.conv3, self.bn3),
+        ]:
+            x = self.relu(bn(conv(x)))
+        x = self.avgpool(x)
+        return x
+
+    def forward(self, x):
+        x = self.stem(x)
+        x = self.layer1(x)
+        x = self.layer2(x)
+        x = self.layer3(x)
+        x = self.layer4(x)
+        x = self.attnpool(x)
+
+        return x
+
+
+class LayerNorm(nn.LayerNorm):
+    """Subclass torch's LayerNorm to handle fp16."""
+
+    def forward(self, x: torch.Tensor):
+        orig_type = x.dtype
+        x = F.layer_norm(x, self.normalized_shape, self.weight, self.bias, self.eps)
+        return x.to(orig_type)
+
+
+class QuickGELU(nn.Module):
+    # NOTE This is slower than nn.GELU or nn.SiLU and uses more GPU memory
+    def forward(self, x: torch.Tensor):
+        return x * torch.sigmoid(1.702 * x)
+
+
+class ResidualAttentionBlock(nn.Module):
+    def __init__(self, d_model: int, n_head: int, act_layer: Callable = nn.GELU):
+        super().__init__()
+
+        self.attn = nn.MultiheadAttention(d_model, n_head)
+        self.ln_1 = LayerNorm(d_model)
+        self.mlp = nn.Sequential(
+            OrderedDict(
+                [
+                    ("c_fc", nn.Linear(d_model, d_model * 4)),
+                    ("gelu", act_layer()),
+                    ("c_proj", nn.Linear(d_model * 4, d_model)),
+                ]
+            )
+        )
+        self.ln_2 = LayerNorm(d_model)
+
+    def attention(self, x: torch.Tensor, attn_mask: Optional[torch.Tensor] = None):
+        return self.attn(x, x, x, need_weights=False, attn_mask=attn_mask)[0]
+
+    def forward(self, x: torch.Tensor, attn_mask: Optional[torch.Tensor] = None):
+        x = x + self.attention(self.ln_1(x), attn_mask=attn_mask)
+        x = x + self.mlp(self.ln_2(x))
+        return x
+
+
+class Transformer(nn.Module):
+    def __init__(
+        self, width: int, layers: int, heads: int, act_layer: Callable = nn.GELU
+    ):
+        super().__init__()
+        self.width = width
+        self.layers = layers
+        self.resblocks = nn.ModuleList(
+            [
+                ResidualAttentionBlock(width, heads, act_layer=act_layer)
+                for _ in range(layers)
+            ]
+        )
+
+    def forward(self, x: torch.Tensor, attn_mask: Optional[torch.Tensor] = None):
+        for r in self.resblocks:
+            x = r(x, attn_mask=attn_mask)
+        return x
+
+
+class VisualTransformer(nn.Module):
+    def __init__(
+        self,
+        image_size: int,
+        patch_size: int,
+        width: int,
+        layers: int,
+        heads: int,
+        output_dim: int,
+        act_layer: Callable = nn.GELU,
+    ):
+        super().__init__()
+        self.image_size = image_size
+        self.output_dim = output_dim
+        self.conv1 = nn.Conv2d(
+            in_channels=3,
+            out_channels=width,
+            kernel_size=patch_size,
+            stride=patch_size,
+            bias=False,
+        )
+
+        scale = width**-0.5
+        self.class_embedding = nn.Parameter(scale * torch.randn(width))
+        self.positional_embedding = nn.Parameter(
+            scale * torch.randn((image_size // patch_size) ** 2 + 1, width)
+        )
+        self.ln_pre = LayerNorm(width)
+
+        self.text_branch = Transformer(width, layers, heads, act_layer=act_layer)
+
+        self.ln_post = LayerNorm(width)
+        self.proj = nn.Parameter(scale * torch.randn(width, output_dim))
+
+    def lock(self, unlocked_groups=0, freeze_bn_stats=False):
+        assert (
+            unlocked_groups == 0
+        ), "partial locking not currently supported for this model"
+        for param in self.parameters():
+            param.requires_grad = False
+
+    def forward(self, x: torch.Tensor):
+        x = self.conv1(x)  # shape = [*, width, grid, grid]
+        x = x.reshape(x.shape[0], x.shape[1], -1)  # shape = [*, width, grid ** 2]
+        x = x.permute(0, 2, 1)  # shape = [*, grid ** 2, width]
+        x = torch.cat(
+            [
+                self.class_embedding.to(x.dtype)
+                + torch.zeros(
+                    x.shape[0], 1, x.shape[-1], dtype=x.dtype, device=x.device
+                ),
+                x,
+            ],
+            dim=1,
+        )  # shape = [*, grid ** 2 + 1, width]
+        x = x + self.positional_embedding.to(x.dtype)
+        x = self.ln_pre(x)
+
+        x = x.permute(1, 0, 2)  # NLD -> LND
+        x = self.text_branch(x)
+        x = x.permute(1, 0, 2)  # LND -> NLD
+
+        x = self.ln_post(x[:, 0, :])
+
+        if self.proj is not None:
+            x = x @ self.proj
+
+        return x
+
+
+@dataclass
+class CLAPVisionCfg:
+    layers: Union[Tuple[int, int, int, int], int] = 12
+    width: int = 768
+    patch_size: int = 16
+    image_size: Union[Tuple[int, int], int] = 224
+    timm_model_name: str = (
+        None  # a valid model name overrides layers, width, patch_size
+    )
+    timm_model_pretrained: bool = (
+        False  # use (imagenet) pretrained weights for named model
+    )
+    timm_pool: str = (
+        "avg"  # feature pooling for timm model ('abs_attn', 'rot_attn', 'avg', '')
+    )
+    timm_proj: str = (
+        "linear"  # linear projection for timm model output ('linear', 'mlp', '')
+    )
+
+
+# Audio Config Class
+@dataclass
+class CLAPAudioCfp:
+    model_type: str = "PANN"
+    model_name: str = "Cnn14"
+    sample_rate: int = 48000
+    # Param
+    audio_length: int = 1024
+    window_size: int = 1024
+    hop_size: int = 1024
+    fmin: int = 50
+    fmax: int = 14000
+    class_num: int = 527
+    mel_bins: int = 64
+    clip_samples: int = 480000
+
+
+@dataclass
+class CLAPTextCfg:
+    context_length: int
+    vocab_size: int
+    width: int
+    heads: int
+    layers: int
+    model_type: str
+
+
+class CLAP(nn.Module):
+    def __init__(
+        self,
+        embed_dim: int,
+        audio_cfg: CLAPAudioCfp,
+        text_cfg: CLAPTextCfg,
+        quick_gelu: bool = False,
+        enable_fusion: bool = False,
+        fusion_type: str = "None",
+        joint_embed_shape: int = 512,
+        mlp_act: str = "relu",
+    ):
+        super().__init__()
+        if isinstance(audio_cfg, dict):
+            audio_cfg = CLAPAudioCfp(**audio_cfg)
+        if isinstance(text_cfg, dict):
+            text_cfg = CLAPTextCfg(**text_cfg)
+
+        self.audio_cfg = audio_cfg
+        self.text_cfg = text_cfg
+        self.enable_fusion = enable_fusion
+        self.fusion_type = fusion_type
+        self.joint_embed_shape = joint_embed_shape
+        self.mlp_act = mlp_act
+
+        self.context_length = text_cfg.context_length
+
+        # OpenAI models are pretrained w/ QuickGELU but native nn.GELU is both faster and more
+        # memory efficient in recent PyTorch releases (>= 1.10).
+        # NOTE: timm models always use native GELU regardless of quick_gelu flag.
+        act_layer = QuickGELU if quick_gelu else nn.GELU
+
+        if mlp_act == "relu":
+            mlp_act_layer = nn.ReLU()
+        elif mlp_act == "gelu":
+            mlp_act_layer = nn.GELU()
+        else:
+            raise NotImplementedError
+
+        # audio branch
+        # audio branch parameters
+        if audio_cfg.model_type == "PANN":
+            self.audio_branch = create_pann_model(audio_cfg, enable_fusion, fusion_type)
+        elif audio_cfg.model_type == "HTSAT":
+            self.audio_branch = create_htsat_model(
+                audio_cfg, enable_fusion, fusion_type
+            )
+        else:
+            logging.error(f"Model config for {audio_cfg.model_type} not found")
+            raise RuntimeError(f"Model config for {audio_cfg.model_type} not found.")
+
+        # text branch
+        # text branch parameters
+        if text_cfg.model_type == "transformer":
+            self.text_branch = Transformer(
+                width=text_cfg.width,
+                layers=text_cfg.layers,
+                heads=text_cfg.heads,
+                act_layer=act_layer,
+            )
+            self.vocab_size = text_cfg.vocab_size
+            self.token_embedding = nn.Embedding(text_cfg.vocab_size, text_cfg.width)
+            self.positional_embedding = nn.Parameter(
+                torch.empty(self.context_length, text_cfg.width)
+            )
+            self.ln_final = LayerNorm(text_cfg.width)
+            self.text_transform = MLPLayers(
+                units=[
+                    self.joint_embed_shape,
+                    self.joint_embed_shape,
+                    self.joint_embed_shape,
+                ],
+                dropout=0.1,
+            )
+            self.text_projection = nn.Sequential(
+                nn.Linear(text_cfg.width, self.joint_embed_shape),
+                mlp_act_layer,
+                nn.Linear(self.joint_embed_shape, self.joint_embed_shape),
+            )
+        elif text_cfg.model_type == "bert":
+            self.text_branch = BertModel.from_pretrained("bert-base-uncased")
+            self.text_transform = MLPLayers(
+                units=[
+                    self.joint_embed_shape,
+                    self.joint_embed_shape,
+                    self.joint_embed_shape,
+                ],
+                dropout=0.1,
+            )
+            self.text_projection = nn.Sequential(
+                nn.Linear(768, self.joint_embed_shape),
+                mlp_act_layer,
+                nn.Linear(self.joint_embed_shape, self.joint_embed_shape),
+            )
+        elif text_cfg.model_type == "roberta":
+            self.text_branch = RobertaModel(
+                RobertaConfig.from_pretrained("roberta-base")
+            )
+            self.text_transform = MLPLayers(
+                units=[
+                    self.joint_embed_shape,
+                    self.joint_embed_shape,
+                    self.joint_embed_shape,
+                ],
+                dropout=0.1,
+            )
+            self.text_projection = nn.Sequential(
+                nn.Linear(768, self.joint_embed_shape),
+                mlp_act_layer,
+                nn.Linear(self.joint_embed_shape, self.joint_embed_shape),
+            )
+        elif text_cfg.model_type == "bart":
+            self.text_branch = BartModel.from_pretrained("facebook/bart-base")
+            self.text_transform = MLPLayers(
+                units=[
+                    self.joint_embed_shape,
+                    self.joint_embed_shape,
+                    self.joint_embed_shape,
+                ],
+                dropout=0.1,
+            )
+            self.text_projection = nn.Sequential(
+                nn.Linear(768, self.joint_embed_shape),
+                mlp_act_layer,
+                nn.Linear(self.joint_embed_shape, self.joint_embed_shape),
+            )
+        else:
+            logging.error(f"Model config for {text_cfg.model_type} not found")
+            raise RuntimeError(f"Model config for {text_cfg.model_type} not found.")
+        self.text_branch_type = text_cfg.model_type
+        # text branch parameters
+
+        # audio branch parameters
+        self.audio_transform = MLPLayers(
+            units=[
+                self.joint_embed_shape,
+                self.joint_embed_shape,
+                self.joint_embed_shape,
+            ],
+            dropout=0.1,
+        )
+
+        # below here is text branch parameters
+
+        # ============================================================================================================
+        self.audio_projection = nn.Sequential(
+            nn.Linear(embed_dim, self.joint_embed_shape),
+            mlp_act_layer,
+            nn.Linear(self.joint_embed_shape, self.joint_embed_shape),
+        )
+
+        self.logit_scale_a = nn.Parameter(torch.ones([]) * np.log(1 / 0.07))
+        self.logit_scale_t = nn.Parameter(torch.ones([]) * np.log(1 / 0.07))
+        self.register_buffer("attn_mask", self.build_attention_mask(), persistent=False)
+
+        self.init_text_branch_parameters()
+
+    def init_text_branch_parameters(self):
+        if self.text_branch_type == "transformer":
+            nn.init.normal_(self.token_embedding.weight, std=0.02)
+            nn.init.normal_(self.positional_embedding, std=0.01)
+            proj_std = (self.text_branch.width**-0.5) * (
+                (2 * self.text_branch.layers) ** -0.5
+            )
+            attn_std = self.text_branch.width**-0.5
+            fc_std = (2 * self.text_branch.width) ** -0.5
+            for block in self.text_branch.resblocks:
+                nn.init.normal_(block.attn.in_proj_weight, std=attn_std)
+                nn.init.normal_(block.attn.out_proj.weight, std=proj_std)
+                nn.init.normal_(block.mlp.c_fc.weight, std=fc_std)
+                nn.init.normal_(block.mlp.c_proj.weight, std=proj_std)
+        if self.text_branch_type == "bert" or self.text_branch_type == "roberta":
+            self.text_branch.embeddings.word_embeddings.weight.shape[-1]
+        elif self.text_branch_type == "bart":
+            self.text_branch.shared.weight.shape[-1]
+        else:
+            self.text_branch.width
+        nn.init.constant_(self.logit_scale_a, np.log(1 / 0.07))
+        nn.init.constant_(self.logit_scale_t, np.log(1 / 0.07))
+
+        # deprecated
+        # if hasattr(self.visual, 'init_parameters'):
+        # self.visual.init_parameters()
+
+        # if self.text_projection is not None:
+        #     nn.init.normal_(self.text_projection, std=width**-0.5)
+
+    def build_attention_mask(self):
+        # lazily create causal attention mask, with full attention between the vision tokens
+        # pytorch uses additive attention mask; fill with -inf
+        mask = torch.empty(self.context_length, self.context_length)
+        mask.fill_(float("-inf"))
+        mask.triu_(1)  # zero out the lower diagonal
+        return mask
+
+    def encode_audio(self, audio, device):
+        return self.audio_branch(
+            audio, mixup_lambda=None, device=device
+        )  # mix lambda needs to add
+
+    # def list_of_dict_of_tensor2dict_of_tensor(self, x, device):
+    #     tmp = {}
+    #     for k in x[0].keys():
+    #         tmp[k] = []
+    #         for i in range(len(x)):
+    #             tmp[k].append(x[i][k][:77])
+    #     for k in x[0].keys():
+    #         tmp[k] = torch.tensor(tmp[k]).to(device=device, non_blocking=True)
+    #     return tmp
+
+    def encode_text(self, text, device):
+        if self.text_branch_type == "transformer":
+            text = text.to(device=device, non_blocking=True)
+            x = self.token_embedding(text)  # [batch_size, n_ctx, d_model]
+
+            x = x + self.positional_embedding
+            x = x.permute(1, 0, 2)  # NLD -> LND
+            x = self.text_branch(x, attn_mask=self.attn_mask)
+            x = x.permute(1, 0, 2)  # LND -> NLD
+            x = self.ln_final(x)
+
+            # x.shape = [batch_size, n_ctx, transformer.width]
+            # take features from the eot embedding (eot_token is the highest number in each sequence)
+            x = self.text_projection(x[torch.arange(x.shape[0]), text.argmax(dim=-1)])
+        elif self.text_branch_type == "bert":
+            # text = self.list_of_dict_of_tensor2dict_of_tensor(text, device)
+            # text = BatchEncoding(text)
+            x = self.text_branch(
+                input_ids=text["input_ids"].to(device=device, non_blocking=True),
+                attention_mask=text["attention_mask"].to(
+                    device=device, non_blocking=True
+                ),
+                token_type_ids=text["token_type_ids"].to(
+                    device=device, non_blocking=True
+                ),
+            )["pooler_output"]
+            x = self.text_projection(x)
+        elif self.text_branch_type == "roberta":
+            x = self.text_branch(
+                input_ids=text["input_ids"].to(device=device, non_blocking=True),
+                attention_mask=text["attention_mask"].to(
+                    device=device, non_blocking=True
+                ),
+            )["pooler_output"]
+            x = self.text_projection(x)
+        elif self.text_branch_type == "bart":
+            x = torch.mean(
+                self.text_branch(
+                    input_ids=text["input_ids"].to(device=device, non_blocking=True),
+                    attention_mask=text["attention_mask"].to(
+                        device=device, non_blocking=True
+                    ),
+                )["encoder_last_hidden_state"],
+                axis=1,
+            )
+            x = self.text_projection(x)
+        else:
+            logging.error(f"Model type {self.text_branch_type} not found")
+            raise RuntimeError(f"Model type {self.text_branch_type} not found.")
+        return x
+
+    def forward(self, audio, text, device=None):
+        """Forward audio and text into the CLAP
+
+        Parameters
+        ----------
+        audio: torch.Tensor (batch_size, audio_length)
+            the time-domain audio input / the batch of mel_spec and longer list.
+        text: torch.Tensor () // need to add
+            the text token input
+        """
+        if device is None:
+            if audio is not None:
+                device = audio.device
+            elif text is not None:
+                device = text.device
+        if audio is None and text is None:
+            # a hack to get the logit scale
+            return self.logit_scale_a.exp(), self.logit_scale_t.exp()
+        elif audio is None:
+            return self.encode_text(text, device=device)
+        elif text is None:
+            return self.audio_projection(
+                self.encode_audio(audio, device=device)["embedding"]
+            )
+        audio_features = self.audio_projection(
+            self.encode_audio(audio, device=device)["embedding"]
+        )
+        audio_features = F.normalize(audio_features, dim=-1)
+
+        text_features = self.encode_text(text, device=device)
+        # print("text_features", text_features)
+        # print("text_features.shape", text_features.shape)
+        # print("text_features.type", type(text_features))
+        text_features = F.normalize(text_features, dim=-1)
+
+        audio_features_mlp = self.audio_transform(audio_features)
+        text_features_mlp = self.text_transform(text_features)
+        # Four outputs: audio features (basic & MLP), text features (basic & MLP)
+        return (
+            audio_features,
+            text_features,
+            audio_features_mlp,
+            text_features_mlp,
+            self.logit_scale_a.exp(),
+            self.logit_scale_t.exp(),
+        )
+
+    def get_logit_scale(self):
+        return self.logit_scale_a.exp(), self.logit_scale_t.exp()
+
+    def get_text_embedding(self, data):
+        """Get the text embedding from the model
+
+        Parameters
+        ----------
+        data: torch.Tensor
+            a tensor of text embedding
+
+        Returns
+        ----------
+        text_embed: torch.Tensor
+            a tensor of text_embeds (N, D)
+
+        """
+        device = next(self.parameters()).device
+        for k in data:
+            data[k] = data[k].to(device)
+        text_embeds = self.encode_text(data, device=device)
+        text_embeds = F.normalize(text_embeds, dim=-1)
+
+        return text_embeds
+
+    def get_audio_embedding(self, data):
+        """Get the audio embedding from the model
+
+        Parameters
+        ----------
+        data: a list of dict
+            the audio input dict list from 'get_audio_feature' method
+
+        Returns
+        ----------
+        audio_embed: torch.Tensor
+            a tensor of audio_embeds (N, D)
+
+        """
+        device = next(self.parameters()).device
+        # input_dict = {}
+        # keys = data[0].keys()
+        # for k in keys:
+        #     input_dict[k] = torch.cat([d[k].unsqueeze(0) for d in data], dim=0).to(
+        #         device
+        #     )
+        audio_embeds = self.audio_projection(
+            self.encode_audio(data, device=device)["embedding"]
+        )
+        audio_embeds = F.normalize(audio_embeds, dim=-1)
+
+        return audio_embeds
+
+    def audio_infer(self, audio, hopsize=None, device=None):
+        """Forward one audio and produce the audio embedding
+
+        Parameters
+        ----------
+        audio:  (audio_length)
+            the time-domain audio input, notice that it must be only one input
+        hopsize: int
+            the overlap hopsize as the sliding window
+
+        Returns
+        ----------
+        output_dict: {
+            key: [n, (embedding_shape)] if "HTS-AT"
+            or
+            key: [(embedding_shape)] if "PANN"
+        }
+            the list of key values of the audio branch
+
+        """
+
+        assert not self.training, "the inference mode must be run at eval stage"
+        output_dict = {}
+        # PANN
+        if self.audio_cfg.model_type == "PANN":
+            audio_input = audio.unsqueeze(dim=0)
+            output_dict[key] = self.encode_audio(audio_input, device=device)[
+                key
+            ].squeeze(dim=0)
+        elif self.audio_cfg.model_type == "HTSAT":
+            # repeat
+            audio_len = len(audio)
+            k = self.audio_cfg.clip_samples // audio_len
+            if k > 1:
+                audio = audio.repeat(k)
+                audio_len = len(audio)
+
+            if hopsize is None:
+                hopsize = min(hopsize, audio_len)
+
+            if audio_len > self.audio_cfg.clip_samples:
+                audio_input = [
+                    audio[pos : pos + self.audio_cfg.clip_samples].clone()
+                    for pos in range(
+                        0, audio_len - self.audio_cfg.clip_samples, hopsize
+                    )
+                ]
+                audio_input.append(audio[-self.audio_cfg.clip_samples :].clone())
+                audio_input = torch.stack(audio_input)
+                output_dict[key] = self.encode_audio(audio_input, device=device)[key]
+            else:
+                audio_input = audio.unsqueeze(dim=0)
+                output_dict[key] = self.encode_audio(audio_input, device=device)[
+                    key
+                ].squeeze(dim=0)
+
+        return output_dict
+
+
+def convert_weights_to_fp16(model: nn.Module):
+    """Convert applicable model parameters to fp16"""
+
+    def _convert_weights_to_fp16(l):
+        if isinstance(l, (nn.Conv1d, nn.Conv2d, nn.Linear)):
+            l.weight.data = l.weight.data.half()
+            if l.bias is not None:
+                l.bias.data = l.bias.data.half()
+
+        if isinstance(l, nn.MultiheadAttention):
+            for attr in [
+                *[f"{s}_proj_weight" for s in ["in", "q", "k", "v"]],
+                "in_proj_bias",
+                "bias_k",
+                "bias_v",
+            ]:
+                tensor = getattr(l, attr)
+                if tensor is not None:
+                    tensor.data = tensor.data.half()
+
+        for name in ["text_projection", "proj"]:
+            if hasattr(l, name):
+                attr = getattr(l, name)
+                if attr is not None:
+                    attr.data = attr.data.half()
+
+    model.apply(_convert_weights_to_fp16)
+
+
+# Ignore the state dict of the vision part
+def build_model_from_openai_state_dict(
+    state_dict: dict, model_cfg, enable_fusion: bool = False, fusion_type: str = "None"
+):
+    embed_dim = model_cfg["embed_dim"]
+    audio_cfg = model_cfg["audio_cfg"]
+    text_cfg = model_cfg["text_cfg"]
+    state_dict["positional_embedding"].shape[0]
+    state_dict["token_embedding.weight"].shape[0]
+    transformer_width = state_dict["ln_final.weight"].shape[0]
+    transformer_width // 64
+    transformer_layers = len(
+        set(
+            k.split(".")[2]
+            for k in state_dict
+            if k.startswith(f"transformer.resblocks")
+        )
+    )
+
+    audio_cfg = CLAPAudioCfp(**audio_cfg)
+    text_cfg = CLAPTextCfg(**text_cfg)
+
+    model = CLAP(
+        embed_dim,
+        audio_cfg=audio_cfg,
+        text_cfg=text_cfg,
+        quick_gelu=True,  # OpenAI models were trained with QuickGELU
+        enable_fusion=enable_fusion,
+        fusion_type=fusion_type,
+    )
+    state_dict["logit_scale_a"] = state_dict["logit_scale"]
+    state_dict["logit_scale_t"] = state_dict["logit_scale"]
+    pop_keys = list(state_dict.keys())[::]
+    # pop the visual branch saved weights
+    for key in pop_keys:
+        if key.startswith("visual."):
+            state_dict.pop(key, None)
+
+    for key in ["logit_scale", "input_resolution", "context_length", "vocab_size"]:
+        state_dict.pop(key, None)
+
+    # not use fp16
+    # convert_weights_to_fp16(model)
+    model.load_state_dict(state_dict, strict=False)
+    return model.eval()
+
+
+def trace_model(model, batch_size=256, device=torch.device("cpu")):
+    model.eval()
+    audio_length = model.audio_cfg.audio_length
+    example_audio = torch.ones((batch_size, audio_length), device=device)
+    example_text = torch.zeros(
+        (batch_size, model.context_length), dtype=torch.int, device=device
+    )
+    model = torch.jit.trace_module(
+        model,
+        inputs=dict(
+            forward=(example_audio, example_text),
+            encode_text=(example_text,),
+            encode_image=(example_audio,),
+        ),
+    )
+    model.audio_cfg.audio_length = audio_length  # Question: what does this do?
+    return model

audiosr/clap/open_clip/model_configs/HTSAT-base.json

@@ -1,23 +1,23 @@
-{
-    "embed_dim": 1024,
-    "audio_cfg": {
-        "audio_length": 1024,
-        "clip_samples": 480000,
-        "mel_bins": 64,
-        "sample_rate": 48000,
-        "window_size": 1024,
-        "hop_size": 480,
-        "fmin": 50,
-        "fmax": 14000,
-        "class_num": 527,
-        "model_type": "HTSAT",
-        "model_name": "base"
-    },
-    "text_cfg": {
-        "context_length": 77,
-        "vocab_size": 49408,
-        "width": 512,
-        "heads": 8,
-        "layers": 12
-    }
+{
+    "embed_dim": 1024,
+    "audio_cfg": {
+        "audio_length": 1024,
+        "clip_samples": 480000,
+        "mel_bins": 64,
+        "sample_rate": 48000,
+        "window_size": 1024,
+        "hop_size": 480,
+        "fmin": 50,
+        "fmax": 14000,
+        "class_num": 527,
+        "model_type": "HTSAT",
+        "model_name": "base"
+    },
+    "text_cfg": {
+        "context_length": 77,
+        "vocab_size": 49408,
+        "width": 512,
+        "heads": 8,
+        "layers": 12
+    }
 }

audiosr/clap/open_clip/model_configs/HTSAT-large.json

@@ -1,23 +1,23 @@
-{
-    "embed_dim": 2048,
-    "audio_cfg": {
-        "audio_length": 1024,
-        "clip_samples": 480000,
-        "mel_bins": 64,
-        "sample_rate": 48000,
-        "window_size": 1024,
-        "hop_size": 480,
-        "fmin": 50,
-        "fmax": 14000,
-        "class_num": 527,
-        "model_type": "HTSAT",
-        "model_name": "large"
-    },
-    "text_cfg": {
-        "context_length": 77,
-        "vocab_size": 49408,
-        "width": 512,
-        "heads": 8,
-        "layers": 12
-    }
+{
+    "embed_dim": 2048,
+    "audio_cfg": {
+        "audio_length": 1024,
+        "clip_samples": 480000,
+        "mel_bins": 64,
+        "sample_rate": 48000,
+        "window_size": 1024,
+        "hop_size": 480,
+        "fmin": 50,
+        "fmax": 14000,
+        "class_num": 527,
+        "model_type": "HTSAT",
+        "model_name": "large"
+    },
+    "text_cfg": {
+        "context_length": 77,
+        "vocab_size": 49408,
+        "width": 512,
+        "heads": 8,
+        "layers": 12
+    }
 }

audiosr/clap/open_clip/model_configs/HTSAT-tiny-win-1536.json

@@ -1,23 +1,23 @@
-{
-    "embed_dim": 768,
-    "audio_cfg": {
-        "audio_length": 1024,
-        "clip_samples": 480000,
-        "mel_bins": 64,
-        "sample_rate": 48000,
-        "window_size": 1536,
-        "hop_size": 480,
-        "fmin": 50,
-        "fmax": 14000,
-        "class_num": 527,
-        "model_type": "HTSAT",
-        "model_name": "tiny"
-    },
-    "text_cfg": {
-        "context_length": 77,
-        "vocab_size": 49408,
-        "width": 512,
-        "heads": 8,
-        "layers": 12
-    }
+{
+    "embed_dim": 768,
+    "audio_cfg": {
+        "audio_length": 1024,
+        "clip_samples": 480000,
+        "mel_bins": 64,
+        "sample_rate": 48000,
+        "window_size": 1536,
+        "hop_size": 480,
+        "fmin": 50,
+        "fmax": 14000,
+        "class_num": 527,
+        "model_type": "HTSAT",
+        "model_name": "tiny"
+    },
+    "text_cfg": {
+        "context_length": 77,
+        "vocab_size": 49408,
+        "width": 512,
+        "heads": 8,
+        "layers": 12
+    }
 }

audiosr/clap/open_clip/model_configs/HTSAT-tiny.json

@@ -1,23 +1,23 @@
-{
-    "embed_dim": 768,
-    "audio_cfg": {
-        "audio_length": 1024,
-        "clip_samples": 480000,
-        "mel_bins": 64,
-        "sample_rate": 48000,
-        "window_size": 1024,
-        "hop_size": 480,
-        "fmin": 50,
-        "fmax": 14000,
-        "class_num": 527,
-        "model_type": "HTSAT",
-        "model_name": "tiny"
-    },
-    "text_cfg": {
-        "context_length": 77,
-        "vocab_size": 49408,
-        "width": 512,
-        "heads": 8,
-        "layers": 12
-    }
+{
+    "embed_dim": 768,
+    "audio_cfg": {
+        "audio_length": 1024,
+        "clip_samples": 480000,
+        "mel_bins": 64,
+        "sample_rate": 48000,
+        "window_size": 1024,
+        "hop_size": 480,
+        "fmin": 50,
+        "fmax": 14000,
+        "class_num": 527,
+        "model_type": "HTSAT",
+        "model_name": "tiny"
+    },
+    "text_cfg": {
+        "context_length": 77,
+        "vocab_size": 49408,
+        "width": 512,
+        "heads": 8,
+        "layers": 12
+    }
 }

audiosr/clap/open_clip/model_configs/PANN-10.json

@@ -1,23 +1,23 @@
-{
-    "embed_dim": 1024,
-    "audio_cfg": {
-        "audio_length": 1024,
-        "clip_samples": 480000,
-        "mel_bins": 64,
-        "sample_rate": 48000,
-        "window_size": 1024,
-        "hop_size": 480,
-        "fmin": 50,
-        "fmax": 14000,
-        "class_num": 527,
-        "model_type": "PANN",
-        "model_name": "Cnn10"
-    },
-    "text_cfg": {
-        "context_length": 77,
-        "vocab_size": 49408,
-        "width": 512,
-        "heads": 8,
-        "layers": 12
-    }
+{
+    "embed_dim": 1024,
+    "audio_cfg": {
+        "audio_length": 1024,
+        "clip_samples": 480000,
+        "mel_bins": 64,
+        "sample_rate": 48000,
+        "window_size": 1024,
+        "hop_size": 480,
+        "fmin": 50,
+        "fmax": 14000,
+        "class_num": 527,
+        "model_type": "PANN",
+        "model_name": "Cnn10"
+    },
+    "text_cfg": {
+        "context_length": 77,
+        "vocab_size": 49408,
+        "width": 512,
+        "heads": 8,
+        "layers": 12
+    }
 }

audiosr/clap/open_clip/model_configs/PANN-14-fmax-18k.json

@@ -1,23 +1,23 @@
-{
-    "embed_dim": 2048,
-    "audio_cfg": {
-        "audio_length": 1024,
-        "clip_samples": 480000,
-        "mel_bins": 64,
-        "sample_rate": 48000,
-        "window_size": 1024,
-        "hop_size": 480,
-        "fmin": 50,
-        "fmax": 18000,
-        "class_num": 527,
-        "model_type": "PANN",
-        "model_name": "Cnn14"
-    },
-    "text_cfg": {
-        "context_length": 77,
-        "vocab_size": 49408,
-        "width": 512,
-        "heads": 8,
-        "layers": 12
-    }
+{
+    "embed_dim": 2048,
+    "audio_cfg": {
+        "audio_length": 1024,
+        "clip_samples": 480000,
+        "mel_bins": 64,
+        "sample_rate": 48000,
+        "window_size": 1024,
+        "hop_size": 480,
+        "fmin": 50,
+        "fmax": 18000,
+        "class_num": 527,
+        "model_type": "PANN",
+        "model_name": "Cnn14"
+    },
+    "text_cfg": {
+        "context_length": 77,
+        "vocab_size": 49408,
+        "width": 512,
+        "heads": 8,
+        "layers": 12
+    }
 }

audiosr/clap/open_clip/model_configs/PANN-14-fmax-8k-20s.json

@@ -1,23 +1,23 @@
-{
-    "embed_dim": 2048,
-    "audio_cfg": {
-        "audio_length": 1024,
-        "clip_samples": 960000,
-        "mel_bins": 64,
-        "sample_rate": 48000,
-        "window_size": 1024,
-        "hop_size": 360,
-        "fmin": 50,
-        "fmax": 8000,
-        "class_num": 527,
-        "model_type": "PANN",
-        "model_name": "Cnn14"
-    },
-    "text_cfg": {
-        "context_length": 77,
-        "vocab_size": 49408,
-        "width": 512,
-        "heads": 8,
-        "layers": 12
-    }
+{
+    "embed_dim": 2048,
+    "audio_cfg": {
+        "audio_length": 1024,
+        "clip_samples": 960000,
+        "mel_bins": 64,
+        "sample_rate": 48000,
+        "window_size": 1024,
+        "hop_size": 360,
+        "fmin": 50,
+        "fmax": 8000,
+        "class_num": 527,
+        "model_type": "PANN",
+        "model_name": "Cnn14"
+    },
+    "text_cfg": {
+        "context_length": 77,
+        "vocab_size": 49408,
+        "width": 512,
+        "heads": 8,
+        "layers": 12
+    }
 }

audiosr/clap/open_clip/model_configs/PANN-14-tiny-transformer.json

@@ -1,23 +1,23 @@
-{
-    "embed_dim": 2048,
-    "audio_cfg": {
-        "audio_length": 1024,
-        "clip_samples": 480000,
-        "mel_bins": 64,
-        "sample_rate": 48000,
-        "window_size": 1024,
-        "hop_size": 480,
-        "fmin": 50,
-        "fmax": 14000,
-        "class_num": 527,
-        "model_type": "PANN",
-        "model_name": "Cnn14"
-    },
-    "text_cfg": {
-        "context_length": 77,
-        "vocab_size": 49408,
-        "width": 512,
-        "heads": 8,
-        "layers": 4
-    }
+{
+    "embed_dim": 2048,
+    "audio_cfg": {
+        "audio_length": 1024,
+        "clip_samples": 480000,
+        "mel_bins": 64,
+        "sample_rate": 48000,
+        "window_size": 1024,
+        "hop_size": 480,
+        "fmin": 50,
+        "fmax": 14000,
+        "class_num": 527,
+        "model_type": "PANN",
+        "model_name": "Cnn14"
+    },
+    "text_cfg": {
+        "context_length": 77,
+        "vocab_size": 49408,
+        "width": 512,
+        "heads": 8,
+        "layers": 4
+    }
 }

audiosr/clap/open_clip/model_configs/PANN-14-win-1536.json

@@ -1,23 +1,23 @@
-{
-    "embed_dim": 2048,
-    "audio_cfg": {
-        "audio_length": 1024,
-        "clip_samples": 480000,
-        "mel_bins": 64,
-        "sample_rate": 48000,
-        "window_size": 1536,
-        "hop_size": 480,
-        "fmin": 50,
-        "fmax": 14000,
-        "class_num": 527,
-        "model_type": "PANN",
-        "model_name": "Cnn14"
-    },
-    "text_cfg": {
-        "context_length": 77,
-        "vocab_size": 49408,
-        "width": 512,
-        "heads": 8,
-        "layers": 12
-    }
+{
+    "embed_dim": 2048,
+    "audio_cfg": {
+        "audio_length": 1024,
+        "clip_samples": 480000,
+        "mel_bins": 64,
+        "sample_rate": 48000,
+        "window_size": 1536,
+        "hop_size": 480,
+        "fmin": 50,
+        "fmax": 14000,
+        "class_num": 527,
+        "model_type": "PANN",
+        "model_name": "Cnn14"
+    },
+    "text_cfg": {
+        "context_length": 77,
+        "vocab_size": 49408,
+        "width": 512,
+        "heads": 8,
+        "layers": 12
+    }
 }

audiosr/clap/open_clip/model_configs/PANN-14.json

@@ -1,23 +1,23 @@
-{
-    "embed_dim": 2048,
-    "audio_cfg": {
-        "audio_length": 1024,
-        "clip_samples": 480000,
-        "mel_bins": 64,
-        "sample_rate": 48000,
-        "window_size": 1024,
-        "hop_size": 480,
-        "fmin": 50,
-        "fmax": 14000,
-        "class_num": 527,
-        "model_type": "PANN",
-        "model_name": "Cnn14"
-    },
-    "text_cfg": {
-        "context_length": 77,
-        "vocab_size": 49408,
-        "width": 512,
-        "heads": 8,
-        "layers": 12
-    }
+{
+    "embed_dim": 2048,
+    "audio_cfg": {
+        "audio_length": 1024,
+        "clip_samples": 480000,
+        "mel_bins": 64,
+        "sample_rate": 48000,
+        "window_size": 1024,
+        "hop_size": 480,
+        "fmin": 50,
+        "fmax": 14000,
+        "class_num": 527,
+        "model_type": "PANN",
+        "model_name": "Cnn14"
+    },
+    "text_cfg": {
+        "context_length": 77,
+        "vocab_size": 49408,
+        "width": 512,
+        "heads": 8,
+        "layers": 12
+    }
 }

audiosr/clap/open_clip/model_configs/PANN-6.json

@@ -1,23 +1,23 @@
-{
-    "embed_dim": 512,
-    "audio_cfg": {
-        "audio_length": 1024,
-        "clip_samples": 480000,
-        "mel_bins": 64,
-        "sample_rate": 48000,
-        "window_size": 1024,
-        "hop_size": 480,
-        "fmin": 50,
-        "fmax": 14000,
-        "class_num": 527,
-        "model_type": "PANN",
-        "model_name": "Cnn6"
-    },
-    "text_cfg": {
-        "context_length": 77,
-        "vocab_size": 49408,
-        "width": 512,
-        "heads": 8,
-        "layers": 12
-    }
+{
+    "embed_dim": 512,
+    "audio_cfg": {
+        "audio_length": 1024,
+        "clip_samples": 480000,
+        "mel_bins": 64,
+        "sample_rate": 48000,
+        "window_size": 1024,
+        "hop_size": 480,
+        "fmin": 50,
+        "fmax": 14000,
+        "class_num": 527,
+        "model_type": "PANN",
+        "model_name": "Cnn6"
+    },
+    "text_cfg": {
+        "context_length": 77,
+        "vocab_size": 49408,
+        "width": 512,
+        "heads": 8,
+        "layers": 12
+    }
 }

audiosr/clap/open_clip/model_configs/RN101-quickgelu.json

@@ -1,22 +1,22 @@
-{
-    "embed_dim": 512,
-    "quick_gelu": true,
-    "vision_cfg": {
-        "image_size": 224,
-        "layers": [
-            3,
-            4,
-            23,
-            3
-        ],
-        "width": 64,
-        "patch_size": null
-    },
-    "text_cfg": {
-        "context_length": 77,
-        "vocab_size": 49408,
-        "width": 512,
-        "heads": 8,
-        "layers": 12
-    }
+{
+    "embed_dim": 512,
+    "quick_gelu": true,
+    "vision_cfg": {
+        "image_size": 224,
+        "layers": [
+            3,
+            4,
+            23,
+            3
+        ],
+        "width": 64,
+        "patch_size": null
+    },
+    "text_cfg": {
+        "context_length": 77,
+        "vocab_size": 49408,
+        "width": 512,
+        "heads": 8,
+        "layers": 12
+    }
 }

audiosr/clap/open_clip/model_configs/RN101.json

@@ -1,21 +1,21 @@
-{
-    "embed_dim": 512,
-    "vision_cfg": {
-        "image_size": 224,
-        "layers": [
-            3,
-            4,
-            23,
-            3
-        ],
-        "width": 64,
-        "patch_size": null
-    },
-    "text_cfg": {
-        "context_length": 77,
-        "vocab_size": 49408,
-        "width": 512,
-        "heads": 8,
-        "layers": 12
-    }
+{
+    "embed_dim": 512,
+    "vision_cfg": {
+        "image_size": 224,
+        "layers": [
+            3,
+            4,
+            23,
+            3
+        ],
+        "width": 64,
+        "patch_size": null
+    },
+    "text_cfg": {
+        "context_length": 77,
+        "vocab_size": 49408,
+        "width": 512,
+        "heads": 8,
+        "layers": 12
+    }
 }

audiosr/clap/open_clip/model_configs/RN50-quickgelu.json

@@ -1,22 +1,22 @@
-{
-    "embed_dim": 1024,
-    "quick_gelu": true,
-    "vision_cfg": {
-        "image_size": 224,
-        "layers": [
-            3,
-            4,
-            6,
-            3
-        ],
-        "width": 64,
-        "patch_size": null
-    },
-    "text_cfg": {
-        "context_length": 77,
-        "vocab_size": 49408,
-        "width": 512,
-        "heads": 8,
-        "layers": 12
-    }
-}
+{
+    "embed_dim": 1024,
+    "quick_gelu": true,
+    "vision_cfg": {
+        "image_size": 224,
+        "layers": [
+            3,
+            4,
+            6,
+            3
+        ],
+        "width": 64,
+        "patch_size": null
+    },
+    "text_cfg": {
+        "context_length": 77,
+        "vocab_size": 49408,
+        "width": 512,
+        "heads": 8,
+        "layers": 12
+    }
+}

audiosr/clap/open_clip/model_configs/RN50.json

@@ -1,21 +1,21 @@
-{
-    "embed_dim": 1024,
-    "vision_cfg": {
-        "image_size": 224,
-        "layers": [
-            3,
-            4,
-            6,
-            3
-        ],
-        "width": 64,
-        "patch_size": null
-    },
-    "text_cfg": {
-        "context_length": 77,
-        "vocab_size": 49408,
-        "width": 512,
-        "heads": 8,
-        "layers": 12
-    }
+{
+    "embed_dim": 1024,
+    "vision_cfg": {
+        "image_size": 224,
+        "layers": [
+            3,
+            4,
+            6,
+            3
+        ],
+        "width": 64,
+        "patch_size": null
+    },
+    "text_cfg": {
+        "context_length": 77,
+        "vocab_size": 49408,
+        "width": 512,
+        "heads": 8,
+        "layers": 12
+    }
 }

audiosr/clap/open_clip/model_configs/RN50x16.json

@@ -1,21 +1,21 @@
-{
-    "embed_dim": 768,
-    "vision_cfg": {
-        "image_size": 384,
-        "layers": [
-            6,
-            8,
-            18,
-            8
-        ],
-        "width": 96,
-        "patch_size": null
-    },
-    "text_cfg": {
-        "context_length": 77,
-        "vocab_size": 49408,
-        "width": 768,
-        "heads": 12,
-        "layers": 12
-    }
+{
+    "embed_dim": 768,
+    "vision_cfg": {
+        "image_size": 384,
+        "layers": [
+            6,
+            8,
+            18,
+            8
+        ],
+        "width": 96,
+        "patch_size": null
+    },
+    "text_cfg": {
+        "context_length": 77,
+        "vocab_size": 49408,
+        "width": 768,
+        "heads": 12,
+        "layers": 12
+    }
 }

audiosr/clap/open_clip/model_configs/RN50x4.json

@@ -1,21 +1,21 @@
-{
-    "embed_dim": 640,
-    "vision_cfg": {
-        "image_size": 288,
-        "layers": [
-            4,
-            6,
-            10,
-            6
-        ],
-        "width": 80,
-        "patch_size": null
-    },
-    "text_cfg": {
-        "context_length": 77,
-        "vocab_size": 49408,
-        "width": 640,
-        "heads": 10,
-        "layers": 12
-    }
+{
+    "embed_dim": 640,
+    "vision_cfg": {
+        "image_size": 288,
+        "layers": [
+            4,
+            6,
+            10,
+            6
+        ],
+        "width": 80,
+        "patch_size": null
+    },
+    "text_cfg": {
+        "context_length": 77,
+        "vocab_size": 49408,
+        "width": 640,
+        "heads": 10,
+        "layers": 12
+    }
 }

audiosr/clap/open_clip/model_configs/ViT-B-16.json

@@ -1,16 +1,16 @@
-{
-    "embed_dim": 512,
-    "vision_cfg": {
-        "image_size": 224,
-        "layers": 12,
-        "width": 768,
-        "patch_size": 16
-    },
-    "text_cfg": {
-        "context_length": 77,
-        "vocab_size": 49408,
-        "width": 512,
-        "heads": 8,
-        "layers": 12
-    }
+{
+    "embed_dim": 512,
+    "vision_cfg": {
+        "image_size": 224,
+        "layers": 12,
+        "width": 768,
+        "patch_size": 16
+    },
+    "text_cfg": {
+        "context_length": 77,
+        "vocab_size": 49408,
+        "width": 512,
+        "heads": 8,
+        "layers": 12
+    }
 }

audiosr/clap/open_clip/model_configs/ViT-B-32-quickgelu.json

@@ -1,17 +1,17 @@
-{
-    "embed_dim": 512,
-    "quick_gelu": true,
-    "vision_cfg": {
-        "image_size": 224,
-        "layers": 12,
-        "width": 768,
-        "patch_size": 32
-    },
-    "text_cfg": {
-        "context_length": 77,
-        "vocab_size": 49408,
-        "width": 512,
-        "heads": 8,
-        "layers": 12
-    }
+{
+    "embed_dim": 512,
+    "quick_gelu": true,
+    "vision_cfg": {
+        "image_size": 224,
+        "layers": 12,
+        "width": 768,
+        "patch_size": 32
+    },
+    "text_cfg": {
+        "context_length": 77,
+        "vocab_size": 49408,
+        "width": 512,
+        "heads": 8,
+        "layers": 12
+    }
 }

audiosr/clap/open_clip/model_configs/ViT-B-32.json

@@ -1,16 +1,16 @@
-{
-    "embed_dim": 512,
-    "vision_cfg": {
-        "image_size": 224,
-        "layers": 12,
-        "width": 768,
-        "patch_size": 32
-    },
-    "text_cfg": {
-        "context_length": 77,
-        "vocab_size": 49408,
-        "width": 512,
-        "heads": 8,
-        "layers": 12
-    }
+{
+    "embed_dim": 512,
+    "vision_cfg": {
+        "image_size": 224,
+        "layers": 12,
+        "width": 768,
+        "patch_size": 32
+    },
+    "text_cfg": {
+        "context_length": 77,
+        "vocab_size": 49408,
+        "width": 512,
+        "heads": 8,
+        "layers": 12
+    }
 }

audiosr/clap/open_clip/model_configs/ViT-L-14.json

@@ -1,16 +1,16 @@
-{
-    "embed_dim": 768,
-    "vision_cfg": {
-        "image_size": 224,
-        "layers": 24,
-        "width": 1024,
-        "patch_size": 14
-    },
-    "text_cfg": {
-        "context_length": 77,
-        "vocab_size": 49408,
-        "width": 768,
-        "heads": 12,
-        "layers": 12
-    }
+{
+    "embed_dim": 768,
+    "vision_cfg": {
+        "image_size": 224,
+        "layers": 24,
+        "width": 1024,
+        "patch_size": 14
+    },
+    "text_cfg": {
+        "context_length": 77,
+        "vocab_size": 49408,
+        "width": 768,
+        "heads": 12,
+        "layers": 12
+    }
 }

audiosr/clap/open_clip/openai.py

@@ -1,156 +1,156 @@
-""" OpenAI pretrained model functions
-
-Adapted from https://github.com/openai/CLIP. Originally MIT License, Copyright (c) 2021 OpenAI.
-"""
-
-import os
-import warnings
-from typing import Union, List
-
-import torch
-
-from .model import build_model_from_openai_state_dict
-from .pretrained import (
-    get_pretrained_url,
-    list_pretrained_tag_models,
-    download_pretrained,
-)
-
-__all__ = ["list_openai_models", "load_openai_model"]
-
-
-def list_openai_models() -> List[str]:
-    """Returns the names of available CLIP models"""
-    return list_pretrained_tag_models("openai")
-
-
-def load_openai_model(
-    name: str,
-    model_cfg,
-    device: Union[str, torch.device] = "cuda" if torch.cuda.is_available() else "cpu",
-    jit=True,
-    cache_dir=os.path.expanduser("~/.cache/clip"),
-    enable_fusion: bool = False,
-    fusion_type: str = "None",
-):
-    """Load a CLIP model, preserve its text pretrained part, and set in the CLAP model
-
-    Parameters
-    ----------
-    name : str
-        A model name listed by `clip.available_models()`, or the path to a model checkpoint containing the state_dict
-    device : Union[str, torch.device]
-        The device to put the loaded model
-    jit : bool
-        Whether to load the optimized JIT model (default) or more hackable non-JIT model.
-
-    Returns
-    -------
-    model : torch.nn.Module
-        The CLAP model
-    preprocess : Callable[[PIL.Image], torch.Tensor]
-        A torchvision transform that converts a PIL image into a tensor that the returned model can take as its input
-    """
-    if get_pretrained_url(name, "openai"):
-        model_path = download_pretrained(
-            get_pretrained_url(name, "openai"), root=cache_dir
-        )
-    elif os.path.isfile(name):
-        model_path = name
-    else:
-        raise RuntimeError(
-            f"Model {name} not found; available models = {list_openai_models()}"
-        )
-
-    try:
-        # loading JIT archive
-        model = torch.jit.load(model_path, map_location=device if jit else "cpu").eval()
-        state_dict = None
-    except RuntimeError:
-        # loading saved state dict
-        if jit:
-            warnings.warn(
-                f"File {model_path} is not a JIT archive. Loading as a state dict instead"
-            )
-            jit = False
-        state_dict = torch.load(model_path, map_location="cpu")
-
-    if not jit:
-        try:
-            model = build_model_from_openai_state_dict(
-                state_dict or model.state_dict(), model_cfg, enable_fusion, fusion_type
-            ).to(device)
-        except KeyError:
-            sd = {k[7:]: v for k, v in state_dict["state_dict"].items()}
-            model = build_model_from_openai_state_dict(
-                sd, model_cfg, enable_fusion, fusion_type
-            ).to(device)
-
-        if str(device) == "cpu":
-            model.float()
-        return model
-
-    # patch the device names
-    device_holder = torch.jit.trace(
-        lambda: torch.ones([]).to(torch.device(device)), example_inputs=[]
-    )
-    device_node = [
-        n
-        for n in device_holder.graph.findAllNodes("prim::Constant")
-        if "Device" in repr(n)
-    ][-1]
-
-    def patch_device(module):
-        try:
-            graphs = [module.graph] if hasattr(module, "graph") else []
-        except RuntimeError:
-            graphs = []
-
-        if hasattr(module, "forward1"):
-            graphs.append(module.forward1.graph)
-
-        for graph in graphs:
-            for node in graph.findAllNodes("prim::Constant"):
-                if "value" in node.attributeNames() and str(node["value"]).startswith(
-                    "cuda"
-                ):
-                    node.copyAttributes(device_node)
-
-    model.apply(patch_device)
-    patch_device(model.encode_audio)
-    patch_device(model.encode_text)
-
-    # patch dtype to float32 on CPU
-    if str(device) == "cpu":
-        float_holder = torch.jit.trace(
-            lambda: torch.ones([]).float(), example_inputs=[]
-        )
-        float_input = list(float_holder.graph.findNode("aten::to").inputs())[1]
-        float_node = float_input.node()
-
-        def patch_float(module):
-            try:
-                graphs = [module.graph] if hasattr(module, "graph") else []
-            except RuntimeError:
-                graphs = []
-
-            if hasattr(module, "forward1"):
-                graphs.append(module.forward1.graph)
-
-            for graph in graphs:
-                for node in graph.findAllNodes("aten::to"):
-                    inputs = list(node.inputs())
-                    for i in [
-                        1,
-                        2,
-                    ]:  # dtype can be the second or third argument to aten::to()
-                        if inputs[i].node()["value"] == 5:
-                            inputs[i].node().copyAttributes(float_node)
-
-        model.apply(patch_float)
-        patch_float(model.encode_audio)
-        patch_float(model.encode_text)
-        model.float()
-
-    model.audio_branch.audio_length = model.audio_cfg.audio_length
-    return model
+""" OpenAI pretrained model functions
+
+Adapted from https://github.com/openai/CLIP. Originally MIT License, Copyright (c) 2021 OpenAI.
+"""
+
+import os
+import warnings
+from typing import Union, List
+
+import torch
+
+from .model import build_model_from_openai_state_dict
+from .pretrained import (
+    get_pretrained_url,
+    list_pretrained_tag_models,
+    download_pretrained,
+)
+
+__all__ = ["list_openai_models", "load_openai_model"]
+
+
+def list_openai_models() -> List[str]:
+    """Returns the names of available CLIP models"""
+    return list_pretrained_tag_models("openai")
+
+
+def load_openai_model(
+    name: str,
+    model_cfg,
+    device: Union[str, torch.device] = "cuda" if torch.cuda.is_available() else "cpu",
+    jit=True,
+    cache_dir=os.path.expanduser("~/.cache/clip"),
+    enable_fusion: bool = False,
+    fusion_type: str = "None",
+):
+    """Load a CLIP model, preserve its text pretrained part, and set in the CLAP model
+
+    Parameters
+    ----------
+    name : str
+        A model name listed by `clip.available_models()`, or the path to a model checkpoint containing the state_dict
+    device : Union[str, torch.device]
+        The device to put the loaded model
+    jit : bool
+        Whether to load the optimized JIT model (default) or more hackable non-JIT model.
+
+    Returns
+    -------
+    model : torch.nn.Module
+        The CLAP model
+    preprocess : Callable[[PIL.Image], torch.Tensor]
+        A torchvision transform that converts a PIL image into a tensor that the returned model can take as its input
+    """
+    if get_pretrained_url(name, "openai"):
+        model_path = download_pretrained(
+            get_pretrained_url(name, "openai"), root=cache_dir
+        )
+    elif os.path.isfile(name):
+        model_path = name
+    else:
+        raise RuntimeError(
+            f"Model {name} not found; available models = {list_openai_models()}"
+        )
+
+    try:
+        # loading JIT archive
+        model = torch.jit.load(model_path, map_location=device if jit else "cpu").eval()
+        state_dict = None
+    except RuntimeError:
+        # loading saved state dict
+        if jit:
+            warnings.warn(
+                f"File {model_path} is not a JIT archive. Loading as a state dict instead"
+            )
+            jit = False
+        state_dict = torch.load(model_path, map_location="cpu")
+
+    if not jit:
+        try:
+            model = build_model_from_openai_state_dict(
+                state_dict or model.state_dict(), model_cfg, enable_fusion, fusion_type
+            ).to(device)
+        except KeyError:
+            sd = {k[7:]: v for k, v in state_dict["state_dict"].items()}
+            model = build_model_from_openai_state_dict(
+                sd, model_cfg, enable_fusion, fusion_type
+            ).to(device)
+
+        if str(device) == "cpu":
+            model.float()
+        return model
+
+    # patch the device names
+    device_holder = torch.jit.trace(
+        lambda: torch.ones([]).to(torch.device(device)), example_inputs=[]
+    )
+    device_node = [
+        n
+        for n in device_holder.graph.findAllNodes("prim::Constant")
+        if "Device" in repr(n)
+    ][-1]
+
+    def patch_device(module):
+        try:
+            graphs = [module.graph] if hasattr(module, "graph") else []
+        except RuntimeError:
+            graphs = []
+
+        if hasattr(module, "forward1"):
+            graphs.append(module.forward1.graph)
+
+        for graph in graphs:
+            for node in graph.findAllNodes("prim::Constant"):
+                if "value" in node.attributeNames() and str(node["value"]).startswith(
+                    "cuda"
+                ):
+                    node.copyAttributes(device_node)
+
+    model.apply(patch_device)
+    patch_device(model.encode_audio)
+    patch_device(model.encode_text)
+
+    # patch dtype to float32 on CPU
+    if str(device) == "cpu":
+        float_holder = torch.jit.trace(
+            lambda: torch.ones([]).float(), example_inputs=[]
+        )
+        float_input = list(float_holder.graph.findNode("aten::to").inputs())[1]
+        float_node = float_input.node()
+
+        def patch_float(module):
+            try:
+                graphs = [module.graph] if hasattr(module, "graph") else []
+            except RuntimeError:
+                graphs = []
+
+            if hasattr(module, "forward1"):
+                graphs.append(module.forward1.graph)
+
+            for graph in graphs:
+                for node in graph.findAllNodes("aten::to"):
+                    inputs = list(node.inputs())
+                    for i in [
+                        1,
+                        2,
+                    ]:  # dtype can be the second or third argument to aten::to()
+                        if inputs[i].node()["value"] == 5:
+                            inputs[i].node().copyAttributes(float_node)
+
+        model.apply(patch_float)
+        patch_float(model.encode_audio)
+        patch_float(model.encode_text)
+        model.float()
+
+    model.audio_branch.audio_length = model.audio_cfg.audio_length
+    return model

audiosr/clap/open_clip/pann_model.py

@@ -1,697 +1,697 @@
-# PANNs: Large-Scale Pretrained Audio Neural Networks for Audio Pattern Recognition
-# Reference from https://github.com/qiuqiangkong/audioset_tagging_cnn
-# Some layers are re-designed for CLAP
-import os
-
-os.environ["NUMBA_CACHE_DIR"] = "/tmp/"
-
-import torch
-import torch.nn as nn
-import torch.nn.functional as F
-from torchlibrosa.stft import Spectrogram, LogmelFilterBank
-from torchlibrosa.augmentation import SpecAugmentation
-
-from .utils import do_mixup, interpolate
-from .feature_fusion import iAFF, AFF, DAF
-
-
-def init_layer(layer):
-    """Initialize a Linear or Convolutional layer."""
-    nn.init.xavier_uniform_(layer.weight)
-
-    if hasattr(layer, "bias"):
-        if layer.bias is not None:
-            layer.bias.data.fill_(0.0)
-
-
-def init_bn(bn):
-    """Initialize a Batchnorm layer."""
-    bn.bias.data.fill_(0.0)
-    bn.weight.data.fill_(1.0)
-
-
-class ConvBlock(nn.Module):
-    def __init__(self, in_channels, out_channels):
-        super(ConvBlock, self).__init__()
-
-        self.conv1 = nn.Conv2d(
-            in_channels=in_channels,
-            out_channels=out_channels,
-            kernel_size=(3, 3),
-            stride=(1, 1),
-            padding=(1, 1),
-            bias=False,
-        )
-
-        self.conv2 = nn.Conv2d(
-            in_channels=out_channels,
-            out_channels=out_channels,
-            kernel_size=(3, 3),
-            stride=(1, 1),
-            padding=(1, 1),
-            bias=False,
-        )
-
-        self.bn1 = nn.BatchNorm2d(out_channels)
-        self.bn2 = nn.BatchNorm2d(out_channels)
-
-        self.init_weight()
-
-    def init_weight(self):
-        init_layer(self.conv1)
-        init_layer(self.conv2)
-        init_bn(self.bn1)
-        init_bn(self.bn2)
-
-    def forward(self, input, pool_size=(2, 2), pool_type="avg"):
-        x = input
-        x = F.relu_(self.bn1(self.conv1(x)))
-        x = F.relu_(self.bn2(self.conv2(x)))
-        if pool_type == "max":
-            x = F.max_pool2d(x, kernel_size=pool_size)
-        elif pool_type == "avg":
-            x = F.avg_pool2d(x, kernel_size=pool_size)
-        elif pool_type == "avg+max":
-            x1 = F.avg_pool2d(x, kernel_size=pool_size)
-            x2 = F.max_pool2d(x, kernel_size=pool_size)
-            x = x1 + x2
-        else:
-            raise Exception("Incorrect argument!")
-
-        return x
-
-
-class ConvBlock5x5(nn.Module):
-    def __init__(self, in_channels, out_channels):
-        super(ConvBlock5x5, self).__init__()
-
-        self.conv1 = nn.Conv2d(
-            in_channels=in_channels,
-            out_channels=out_channels,
-            kernel_size=(5, 5),
-            stride=(1, 1),
-            padding=(2, 2),
-            bias=False,
-        )
-
-        self.bn1 = nn.BatchNorm2d(out_channels)
-
-        self.init_weight()
-
-    def init_weight(self):
-        init_layer(self.conv1)
-        init_bn(self.bn1)
-
-    def forward(self, input, pool_size=(2, 2), pool_type="avg"):
-        x = input
-        x = F.relu_(self.bn1(self.conv1(x)))
-        if pool_type == "max":
-            x = F.max_pool2d(x, kernel_size=pool_size)
-        elif pool_type == "avg":
-            x = F.avg_pool2d(x, kernel_size=pool_size)
-        elif pool_type == "avg+max":
-            x1 = F.avg_pool2d(x, kernel_size=pool_size)
-            x2 = F.max_pool2d(x, kernel_size=pool_size)
-            x = x1 + x2
-        else:
-            raise Exception("Incorrect argument!")
-
-        return x
-
-
-class AttBlock(nn.Module):
-    def __init__(self, n_in, n_out, activation="linear", temperature=1.0):
-        super(AttBlock, self).__init__()
-
-        self.activation = activation
-        self.temperature = temperature
-        self.att = nn.Conv1d(
-            in_channels=n_in,
-            out_channels=n_out,
-            kernel_size=1,
-            stride=1,
-            padding=0,
-            bias=True,
-        )
-        self.cla = nn.Conv1d(
-            in_channels=n_in,
-            out_channels=n_out,
-            kernel_size=1,
-            stride=1,
-            padding=0,
-            bias=True,
-        )
-
-        self.bn_att = nn.BatchNorm1d(n_out)
-        self.init_weights()
-
-    def init_weights(self):
-        init_layer(self.att)
-        init_layer(self.cla)
-        init_bn(self.bn_att)
-
-    def forward(self, x):
-        # x: (n_samples, n_in, n_time)
-        norm_att = torch.softmax(torch.clamp(self.att(x), -10, 10), dim=-1)
-        cla = self.nonlinear_transform(self.cla(x))
-        x = torch.sum(norm_att * cla, dim=2)
-        return x, norm_att, cla
-
-    def nonlinear_transform(self, x):
-        if self.activation == "linear":
-            return x
-        elif self.activation == "sigmoid":
-            return torch.sigmoid(x)
-
-
-class Cnn14(nn.Module):
-    def __init__(
-        self,
-        sample_rate,
-        window_size,
-        hop_size,
-        mel_bins,
-        fmin,
-        fmax,
-        classes_num,
-        enable_fusion=False,
-        fusion_type="None",
-    ):
-        super(Cnn14, self).__init__()
-
-        window = "hann"
-        center = True
-        pad_mode = "reflect"
-        ref = 1.0
-        amin = 1e-10
-        top_db = None
-
-        self.enable_fusion = enable_fusion
-        self.fusion_type = fusion_type
-
-        # Spectrogram extractor
-        self.spectrogram_extractor = Spectrogram(
-            n_fft=window_size,
-            hop_length=hop_size,
-            win_length=window_size,
-            window=window,
-            center=center,
-            pad_mode=pad_mode,
-            freeze_parameters=True,
-        )
-
-        # Logmel feature extractor
-        self.logmel_extractor = LogmelFilterBank(
-            sr=sample_rate,
-            n_fft=window_size,
-            n_mels=mel_bins,
-            fmin=fmin,
-            fmax=fmax,
-            ref=ref,
-            amin=amin,
-            top_db=top_db,
-            freeze_parameters=True,
-        )
-
-        # Spec augmenter
-        self.spec_augmenter = SpecAugmentation(
-            time_drop_width=64,
-            time_stripes_num=2,
-            freq_drop_width=8,
-            freq_stripes_num=2,
-        )
-
-        self.bn0 = nn.BatchNorm2d(64)
-
-        if (self.enable_fusion) and (self.fusion_type == "channel_map"):
-            self.conv_block1 = ConvBlock(in_channels=4, out_channels=64)
-        else:
-            self.conv_block1 = ConvBlock(in_channels=1, out_channels=64)
-        self.conv_block2 = ConvBlock(in_channels=64, out_channels=128)
-        self.conv_block3 = ConvBlock(in_channels=128, out_channels=256)
-        self.conv_block4 = ConvBlock(in_channels=256, out_channels=512)
-        self.conv_block5 = ConvBlock(in_channels=512, out_channels=1024)
-        self.conv_block6 = ConvBlock(in_channels=1024, out_channels=2048)
-
-        self.fc1 = nn.Linear(2048, 2048, bias=True)
-        self.fc_audioset = nn.Linear(2048, classes_num, bias=True)
-
-        if (self.enable_fusion) and (
-            self.fusion_type in ["daf_1d", "aff_1d", "iaff_1d"]
-        ):
-            self.mel_conv1d = nn.Sequential(
-                nn.Conv1d(64, 64, kernel_size=5, stride=3, padding=2),
-                nn.BatchNorm1d(64),  # No Relu
-            )
-            if self.fusion_type == "daf_1d":
-                self.fusion_model = DAF()
-            elif self.fusion_type == "aff_1d":
-                self.fusion_model = AFF(channels=64, type="1D")
-            elif self.fusion_type == "iaff_1d":
-                self.fusion_model = iAFF(channels=64, type="1D")
-
-        if (self.enable_fusion) and (
-            self.fusion_type in ["daf_2d", "aff_2d", "iaff_2d"]
-        ):
-            self.mel_conv2d = nn.Sequential(
-                nn.Conv2d(1, 64, kernel_size=(5, 5), stride=(6, 2), padding=(2, 2)),
-                nn.BatchNorm2d(64),
-                nn.ReLU(inplace=True),
-            )
-
-            if self.fusion_type == "daf_2d":
-                self.fusion_model = DAF()
-            elif self.fusion_type == "aff_2d":
-                self.fusion_model = AFF(channels=64, type="2D")
-            elif self.fusion_type == "iaff_2d":
-                self.fusion_model = iAFF(channels=64, type="2D")
-        self.init_weight()
-
-    def init_weight(self):
-        init_bn(self.bn0)
-        init_layer(self.fc1)
-        init_layer(self.fc_audioset)
-
-    def forward(self, input, mixup_lambda=None, device=None):
-        """
-        Input: (batch_size, data_length)"""
-
-        if self.enable_fusion and input["longer"].sum() == 0:
-            # if no audio is longer than 10s, then randomly select one audio to be longer
-            input["longer"][torch.randint(0, input["longer"].shape[0], (1,))] = True
-
-        if not self.enable_fusion:
-            x = self.spectrogram_extractor(
-                input["waveform"].to(device=device, non_blocking=True)
-            )  # (batch_size, 1, time_steps, freq_bins)
-            x = self.logmel_extractor(x)  # (batch_size, 1, time_steps, mel_bins)
-
-            x = x.transpose(1, 3)
-            x = self.bn0(x)
-            x = x.transpose(1, 3)
-        else:
-            longer_list = input["longer"].to(device=device, non_blocking=True)
-            x = input["mel_fusion"].to(device=device, non_blocking=True)
-            longer_list_idx = torch.where(longer_list)[0]
-            x = x.transpose(1, 3)
-            x = self.bn0(x)
-            x = x.transpose(1, 3)
-            if self.fusion_type in ["daf_1d", "aff_1d", "iaff_1d"]:
-                new_x = x[:, 0:1, :, :].clone().contiguous()
-                # local processing
-                if len(longer_list_idx) > 0:
-                    fusion_x_local = x[longer_list_idx, 1:, :, :].clone().contiguous()
-                    FB, FC, FT, FF = fusion_x_local.size()
-                    fusion_x_local = fusion_x_local.view(FB * FC, FT, FF)
-                    fusion_x_local = torch.permute(
-                        fusion_x_local, (0, 2, 1)
-                    ).contiguous()
-                    fusion_x_local = self.mel_conv1d(fusion_x_local)
-                    fusion_x_local = fusion_x_local.view(
-                        FB, FC, FF, fusion_x_local.size(-1)
-                    )
-                    fusion_x_local = (
-                        torch.permute(fusion_x_local, (0, 2, 1, 3))
-                        .contiguous()
-                        .flatten(2)
-                    )
-                    if fusion_x_local.size(-1) < FT:
-                        fusion_x_local = torch.cat(
-                            [
-                                fusion_x_local,
-                                torch.zeros(
-                                    (FB, FF, FT - fusion_x_local.size(-1)),
-                                    device=device,
-                                ),
-                            ],
-                            dim=-1,
-                        )
-                    else:
-                        fusion_x_local = fusion_x_local[:, :, :FT]
-                    # 1D fusion
-                    new_x = new_x.squeeze(1).permute((0, 2, 1)).contiguous()
-                    new_x[longer_list_idx] = self.fusion_model(
-                        new_x[longer_list_idx], fusion_x_local
-                    )
-                    x = new_x.permute((0, 2, 1)).contiguous()[:, None, :, :]
-                else:
-                    x = new_x
-            elif self.fusion_type in ["daf_2d", "aff_2d", "iaff_2d", "channel_map"]:
-                x = x  # no change
-
-        if self.training:
-            x = self.spec_augmenter(x)
-        # Mixup on spectrogram
-        if self.training and mixup_lambda is not None:
-            x = do_mixup(x, mixup_lambda)
-        if (self.enable_fusion) and (
-            self.fusion_type in ["daf_2d", "aff_2d", "iaff_2d"]
-        ):
-            global_x = x[:, 0:1, :, :]
-
-            # global processing
-            B, C, H, W = global_x.shape
-            global_x = self.conv_block1(global_x, pool_size=(2, 2), pool_type="avg")
-            if len(longer_list_idx) > 0:
-                local_x = x[longer_list_idx, 1:, :, :].contiguous()
-                TH = global_x.size(-2)
-                # local processing
-                B, C, H, W = local_x.shape
-                local_x = local_x.view(B * C, 1, H, W)
-                local_x = self.mel_conv2d(local_x)
-                local_x = local_x.view(
-                    B, C, local_x.size(1), local_x.size(2), local_x.size(3)
-                )
-                local_x = local_x.permute((0, 2, 1, 3, 4)).contiguous().flatten(2, 3)
-                TB, TC, _, TW = local_x.size()
-                if local_x.size(-2) < TH:
-                    local_x = torch.cat(
-                        [
-                            local_x,
-                            torch.zeros(
-                                (TB, TC, TH - local_x.size(-2), TW),
-                                device=global_x.device,
-                            ),
-                        ],
-                        dim=-2,
-                    )
-                else:
-                    local_x = local_x[:, :, :TH, :]
-
-                global_x[longer_list_idx] = self.fusion_model(
-                    global_x[longer_list_idx], local_x
-                )
-            x = global_x
-        else:
-            x = self.conv_block1(x, pool_size=(2, 2), pool_type="avg")
-
-        x = F.dropout(x, p=0.2, training=self.training)
-        x = self.conv_block2(x, pool_size=(2, 2), pool_type="avg")
-        x = F.dropout(x, p=0.2, training=self.training)
-        x = self.conv_block3(x, pool_size=(2, 2), pool_type="avg")
-        x = F.dropout(x, p=0.2, training=self.training)
-        x = self.conv_block4(x, pool_size=(2, 2), pool_type="avg")
-        x = F.dropout(x, p=0.2, training=self.training)
-        x = self.conv_block5(x, pool_size=(2, 2), pool_type="avg")
-        x = F.dropout(x, p=0.2, training=self.training)
-        x = self.conv_block6(x, pool_size=(1, 1), pool_type="avg")
-        x = F.dropout(x, p=0.2, training=self.training)
-        x = torch.mean(x, dim=3)
-
-        latent_x1 = F.max_pool1d(x, kernel_size=3, stride=1, padding=1)
-        latent_x2 = F.avg_pool1d(x, kernel_size=3, stride=1, padding=1)
-        latent_x = latent_x1 + latent_x2
-        latent_x = latent_x.transpose(1, 2)
-        latent_x = F.relu_(self.fc1(latent_x))
-        latent_output = interpolate(latent_x, 32)
-
-        (x1, _) = torch.max(x, dim=2)
-        x2 = torch.mean(x, dim=2)
-        x = x1 + x2
-        x = F.dropout(x, p=0.5, training=self.training)
-        x = F.relu_(self.fc1(x))
-        embedding = F.dropout(x, p=0.5, training=self.training)
-        clipwise_output = torch.sigmoid(self.fc_audioset(x))
-
-        output_dict = {
-            "clipwise_output": clipwise_output,
-            "embedding": embedding,
-            "fine_grained_embedding": latent_output,
-        }
-        return output_dict
-
-
-class Cnn6(nn.Module):
-    def __init__(
-        self,
-        sample_rate,
-        window_size,
-        hop_size,
-        mel_bins,
-        fmin,
-        fmax,
-        classes_num,
-        enable_fusion=False,
-        fusion_type="None",
-    ):
-        super(Cnn6, self).__init__()
-
-        window = "hann"
-        center = True
-        pad_mode = "reflect"
-        ref = 1.0
-        amin = 1e-10
-        top_db = None
-
-        self.enable_fusion = enable_fusion
-        self.fusion_type = fusion_type
-
-        # Spectrogram extractor
-        self.spectrogram_extractor = Spectrogram(
-            n_fft=window_size,
-            hop_length=hop_size,
-            win_length=window_size,
-            window=window,
-            center=center,
-            pad_mode=pad_mode,
-            freeze_parameters=True,
-        )
-
-        # Logmel feature extractor
-        self.logmel_extractor = LogmelFilterBank(
-            sr=sample_rate,
-            n_fft=window_size,
-            n_mels=mel_bins,
-            fmin=fmin,
-            fmax=fmax,
-            ref=ref,
-            amin=amin,
-            top_db=top_db,
-            freeze_parameters=True,
-        )
-
-        # Spec augmenter
-        self.spec_augmenter = SpecAugmentation(
-            time_drop_width=64,
-            time_stripes_num=2,
-            freq_drop_width=8,
-            freq_stripes_num=2,
-        )
-
-        self.bn0 = nn.BatchNorm2d(64)
-
-        self.conv_block1 = ConvBlock5x5(in_channels=1, out_channels=64)
-        self.conv_block2 = ConvBlock5x5(in_channels=64, out_channels=128)
-        self.conv_block3 = ConvBlock5x5(in_channels=128, out_channels=256)
-        self.conv_block4 = ConvBlock5x5(in_channels=256, out_channels=512)
-
-        self.fc1 = nn.Linear(512, 512, bias=True)
-        self.fc_audioset = nn.Linear(512, classes_num, bias=True)
-
-        self.init_weight()
-
-    def init_weight(self):
-        init_bn(self.bn0)
-        init_layer(self.fc1)
-        init_layer(self.fc_audioset)
-
-    def forward(self, input, mixup_lambda=None, device=None):
-        """
-        Input: (batch_size, data_length)"""
-
-        x = self.spectrogram_extractor(input)  # (batch_size, 1, time_steps, freq_bins)
-        x = self.logmel_extractor(x)  # (batch_size, 1, time_steps, mel_bins)
-
-        x = x.transpose(1, 3)
-        x = self.bn0(x)
-        x = x.transpose(1, 3)
-
-        if self.training:
-            x = self.spec_augmenter(x)
-
-        # Mixup on spectrogram
-        if self.training and mixup_lambda is not None:
-            x = do_mixup(x, mixup_lambda)
-
-        x = self.conv_block1(x, pool_size=(2, 2), pool_type="avg")
-        x = F.dropout(x, p=0.2, training=self.training)
-        x = self.conv_block2(x, pool_size=(2, 2), pool_type="avg")
-        x = F.dropout(x, p=0.2, training=self.training)
-        x = self.conv_block3(x, pool_size=(2, 2), pool_type="avg")
-        x = F.dropout(x, p=0.2, training=self.training)
-        x = self.conv_block4(x, pool_size=(2, 2), pool_type="avg")
-        x = F.dropout(x, p=0.2, training=self.training)
-        x = torch.mean(x, dim=3)
-
-        latent_x1 = F.max_pool1d(x, kernel_size=3, stride=1, padding=1)
-        latent_x2 = F.avg_pool1d(x, kernel_size=3, stride=1, padding=1)
-        latent_x = latent_x1 + latent_x2
-        latent_x = latent_x.transpose(1, 2)
-        latent_x = F.relu_(self.fc1(latent_x))
-        latent_output = interpolate(latent_x, 16)
-
-        (x1, _) = torch.max(x, dim=2)
-        x2 = torch.mean(x, dim=2)
-        x = x1 + x2
-        x = F.dropout(x, p=0.5, training=self.training)
-        x = F.relu_(self.fc1(x))
-        embedding = F.dropout(x, p=0.5, training=self.training)
-        clipwise_output = torch.sigmoid(self.fc_audioset(x))
-
-        output_dict = {
-            "clipwise_output": clipwise_output,
-            "embedding": embedding,
-            "fine_grained_embedding": latent_output,
-        }
-
-        return output_dict
-
-
-class Cnn10(nn.Module):
-    def __init__(
-        self,
-        sample_rate,
-        window_size,
-        hop_size,
-        mel_bins,
-        fmin,
-        fmax,
-        classes_num,
-        enable_fusion=False,
-        fusion_type="None",
-    ):
-        super(Cnn10, self).__init__()
-
-        window = "hann"
-        center = True
-        pad_mode = "reflect"
-        ref = 1.0
-        amin = 1e-10
-        top_db = None
-
-        self.enable_fusion = enable_fusion
-        self.fusion_type = fusion_type
-
-        # Spectrogram extractor
-        self.spectrogram_extractor = Spectrogram(
-            n_fft=window_size,
-            hop_length=hop_size,
-            win_length=window_size,
-            window=window,
-            center=center,
-            pad_mode=pad_mode,
-            freeze_parameters=True,
-        )
-
-        # Logmel feature extractor
-        self.logmel_extractor = LogmelFilterBank(
-            sr=sample_rate,
-            n_fft=window_size,
-            n_mels=mel_bins,
-            fmin=fmin,
-            fmax=fmax,
-            ref=ref,
-            amin=amin,
-            top_db=top_db,
-            freeze_parameters=True,
-        )
-
-        # Spec augmenter
-        self.spec_augmenter = SpecAugmentation(
-            time_drop_width=64,
-            time_stripes_num=2,
-            freq_drop_width=8,
-            freq_stripes_num=2,
-        )
-
-        self.bn0 = nn.BatchNorm2d(64)
-
-        self.conv_block1 = ConvBlock(in_channels=1, out_channels=64)
-        self.conv_block2 = ConvBlock(in_channels=64, out_channels=128)
-        self.conv_block3 = ConvBlock(in_channels=128, out_channels=256)
-        self.conv_block4 = ConvBlock(in_channels=256, out_channels=512)
-        self.conv_block5 = ConvBlock(in_channels=512, out_channels=1024)
-
-        self.fc1 = nn.Linear(1024, 1024, bias=True)
-        self.fc_audioset = nn.Linear(1024, classes_num, bias=True)
-
-        self.init_weight()
-
-    def init_weight(self):
-        init_bn(self.bn0)
-        init_layer(self.fc1)
-        init_layer(self.fc_audioset)
-
-    def forward(self, input, mixup_lambda=None, device=None):
-        """
-        Input: (batch_size, data_length)"""
-
-        x = self.spectrogram_extractor(input)  # (batch_size, 1, time_steps, freq_bins)
-        x = self.logmel_extractor(x)  # (batch_size, 1, time_steps, mel_bins)
-
-        x = x.transpose(1, 3)
-        x = self.bn0(x)
-        x = x.transpose(1, 3)
-
-        if self.training:
-            x = self.spec_augmenter(x)
-
-        # Mixup on spectrogram
-        if self.training and mixup_lambda is not None:
-            x = do_mixup(x, mixup_lambda)
-
-        x = self.conv_block1(x, pool_size=(2, 2), pool_type="avg")
-        x = F.dropout(x, p=0.2, training=self.training)
-        x = self.conv_block2(x, pool_size=(2, 2), pool_type="avg")
-        x = F.dropout(x, p=0.2, training=self.training)
-        x = self.conv_block3(x, pool_size=(2, 2), pool_type="avg")
-        x = F.dropout(x, p=0.2, training=self.training)
-        x = self.conv_block4(x, pool_size=(2, 2), pool_type="avg")
-        x = F.dropout(x, p=0.2, training=self.training)
-        x = self.conv_block5(x, pool_size=(2, 2), pool_type="avg")
-        x = F.dropout(x, p=0.2, training=self.training)
-        x = torch.mean(x, dim=3)
-
-        latent_x1 = F.max_pool1d(x, kernel_size=3, stride=1, padding=1)
-        latent_x2 = F.avg_pool1d(x, kernel_size=3, stride=1, padding=1)
-        latent_x = latent_x1 + latent_x2
-        latent_x = latent_x.transpose(1, 2)
-        latent_x = F.relu_(self.fc1(latent_x))
-        latent_output = interpolate(latent_x, 32)
-
-        (x1, _) = torch.max(x, dim=2)
-        x2 = torch.mean(x, dim=2)
-        x = x1 + x2
-        x = F.dropout(x, p=0.5, training=self.training)
-        x = F.relu_(self.fc1(x))
-        embedding = F.dropout(x, p=0.5, training=self.training)
-        clipwise_output = torch.sigmoid(self.fc_audioset(x))
-
-        output_dict = {
-            "clipwise_output": clipwise_output,
-            "embedding": embedding,
-            "fine_grained_embedding": latent_output,
-        }
-
-        return output_dict
-
-
-def create_pann_model(audio_cfg, enable_fusion=False, fusion_type="None"):
-    try:
-        ModelProto = eval(audio_cfg.model_name)
-        model = ModelProto(
-            sample_rate=audio_cfg.sample_rate,
-            window_size=audio_cfg.window_size,
-            hop_size=audio_cfg.hop_size,
-            mel_bins=audio_cfg.mel_bins,
-            fmin=audio_cfg.fmin,
-            fmax=audio_cfg.fmax,
-            classes_num=audio_cfg.class_num,
-            enable_fusion=enable_fusion,
-            fusion_type=fusion_type,
-        )
-        return model
-    except:
-        raise RuntimeError(
-            f"Import Model for {audio_cfg.model_name} not found, or the audio cfg parameters are not enough."
-        )
+# PANNs: Large-Scale Pretrained Audio Neural Networks for Audio Pattern Recognition
+# Reference from https://github.com/qiuqiangkong/audioset_tagging_cnn
+# Some layers are re-designed for CLAP
+import os
+
+os.environ["NUMBA_CACHE_DIR"] = "/tmp/"
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torchlibrosa.stft import Spectrogram, LogmelFilterBank
+from torchlibrosa.augmentation import SpecAugmentation
+
+from .utils import do_mixup, interpolate
+from .feature_fusion import iAFF, AFF, DAF
+
+
+def init_layer(layer):
+    """Initialize a Linear or Convolutional layer."""
+    nn.init.xavier_uniform_(layer.weight)
+
+    if hasattr(layer, "bias"):
+        if layer.bias is not None:
+            layer.bias.data.fill_(0.0)
+
+
+def init_bn(bn):
+    """Initialize a Batchnorm layer."""
+    bn.bias.data.fill_(0.0)
+    bn.weight.data.fill_(1.0)
+
+
+class ConvBlock(nn.Module):
+    def __init__(self, in_channels, out_channels):
+        super(ConvBlock, self).__init__()
+
+        self.conv1 = nn.Conv2d(
+            in_channels=in_channels,
+            out_channels=out_channels,
+            kernel_size=(3, 3),
+            stride=(1, 1),
+            padding=(1, 1),
+            bias=False,
+        )
+
+        self.conv2 = nn.Conv2d(
+            in_channels=out_channels,
+            out_channels=out_channels,
+            kernel_size=(3, 3),
+            stride=(1, 1),
+            padding=(1, 1),
+            bias=False,
+        )
+
+        self.bn1 = nn.BatchNorm2d(out_channels)
+        self.bn2 = nn.BatchNorm2d(out_channels)
+
+        self.init_weight()
+
+    def init_weight(self):
+        init_layer(self.conv1)
+        init_layer(self.conv2)
+        init_bn(self.bn1)
+        init_bn(self.bn2)
+
+    def forward(self, input, pool_size=(2, 2), pool_type="avg"):
+        x = input
+        x = F.relu_(self.bn1(self.conv1(x)))
+        x = F.relu_(self.bn2(self.conv2(x)))
+        if pool_type == "max":
+            x = F.max_pool2d(x, kernel_size=pool_size)
+        elif pool_type == "avg":
+            x = F.avg_pool2d(x, kernel_size=pool_size)
+        elif pool_type == "avg+max":
+            x1 = F.avg_pool2d(x, kernel_size=pool_size)
+            x2 = F.max_pool2d(x, kernel_size=pool_size)
+            x = x1 + x2
+        else:
+            raise Exception("Incorrect argument!")
+
+        return x
+
+
+class ConvBlock5x5(nn.Module):
+    def __init__(self, in_channels, out_channels):
+        super(ConvBlock5x5, self).__init__()
+
+        self.conv1 = nn.Conv2d(
+            in_channels=in_channels,
+            out_channels=out_channels,
+            kernel_size=(5, 5),
+            stride=(1, 1),
+            padding=(2, 2),
+            bias=False,
+        )
+
+        self.bn1 = nn.BatchNorm2d(out_channels)
+
+        self.init_weight()
+
+    def init_weight(self):
+        init_layer(self.conv1)
+        init_bn(self.bn1)
+
+    def forward(self, input, pool_size=(2, 2), pool_type="avg"):
+        x = input
+        x = F.relu_(self.bn1(self.conv1(x)))
+        if pool_type == "max":
+            x = F.max_pool2d(x, kernel_size=pool_size)
+        elif pool_type == "avg":
+            x = F.avg_pool2d(x, kernel_size=pool_size)
+        elif pool_type == "avg+max":
+            x1 = F.avg_pool2d(x, kernel_size=pool_size)
+            x2 = F.max_pool2d(x, kernel_size=pool_size)
+            x = x1 + x2
+        else:
+            raise Exception("Incorrect argument!")
+
+        return x
+
+
+class AttBlock(nn.Module):
+    def __init__(self, n_in, n_out, activation="linear", temperature=1.0):
+        super(AttBlock, self).__init__()
+
+        self.activation = activation
+        self.temperature = temperature
+        self.att = nn.Conv1d(
+            in_channels=n_in,
+            out_channels=n_out,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=True,
+        )
+        self.cla = nn.Conv1d(
+            in_channels=n_in,
+            out_channels=n_out,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=True,
+        )
+
+        self.bn_att = nn.BatchNorm1d(n_out)
+        self.init_weights()
+
+    def init_weights(self):
+        init_layer(self.att)
+        init_layer(self.cla)
+        init_bn(self.bn_att)
+
+    def forward(self, x):
+        # x: (n_samples, n_in, n_time)
+        norm_att = torch.softmax(torch.clamp(self.att(x), -10, 10), dim=-1)
+        cla = self.nonlinear_transform(self.cla(x))
+        x = torch.sum(norm_att * cla, dim=2)
+        return x, norm_att, cla
+
+    def nonlinear_transform(self, x):
+        if self.activation == "linear":
+            return x
+        elif self.activation == "sigmoid":
+            return torch.sigmoid(x)
+
+
+class Cnn14(nn.Module):
+    def __init__(
+        self,
+        sample_rate,
+        window_size,
+        hop_size,
+        mel_bins,
+        fmin,
+        fmax,
+        classes_num,
+        enable_fusion=False,
+        fusion_type="None",
+    ):
+        super(Cnn14, self).__init__()
+
+        window = "hann"
+        center = True
+        pad_mode = "reflect"
+        ref = 1.0
+        amin = 1e-10
+        top_db = None
+
+        self.enable_fusion = enable_fusion
+        self.fusion_type = fusion_type
+
+        # Spectrogram extractor
+        self.spectrogram_extractor = Spectrogram(
+            n_fft=window_size,
+            hop_length=hop_size,
+            win_length=window_size,
+            window=window,
+            center=center,
+            pad_mode=pad_mode,
+            freeze_parameters=True,
+        )
+
+        # Logmel feature extractor
+        self.logmel_extractor = LogmelFilterBank(
+            sr=sample_rate,
+            n_fft=window_size,
+            n_mels=mel_bins,
+            fmin=fmin,
+            fmax=fmax,
+            ref=ref,
+            amin=amin,
+            top_db=top_db,
+            freeze_parameters=True,
+        )
+
+        # Spec augmenter
+        self.spec_augmenter = SpecAugmentation(
+            time_drop_width=64,
+            time_stripes_num=2,
+            freq_drop_width=8,
+            freq_stripes_num=2,
+        )
+
+        self.bn0 = nn.BatchNorm2d(64)
+
+        if (self.enable_fusion) and (self.fusion_type == "channel_map"):
+            self.conv_block1 = ConvBlock(in_channels=4, out_channels=64)
+        else:
+            self.conv_block1 = ConvBlock(in_channels=1, out_channels=64)
+        self.conv_block2 = ConvBlock(in_channels=64, out_channels=128)
+        self.conv_block3 = ConvBlock(in_channels=128, out_channels=256)
+        self.conv_block4 = ConvBlock(in_channels=256, out_channels=512)
+        self.conv_block5 = ConvBlock(in_channels=512, out_channels=1024)
+        self.conv_block6 = ConvBlock(in_channels=1024, out_channels=2048)
+
+        self.fc1 = nn.Linear(2048, 2048, bias=True)
+        self.fc_audioset = nn.Linear(2048, classes_num, bias=True)
+
+        if (self.enable_fusion) and (
+            self.fusion_type in ["daf_1d", "aff_1d", "iaff_1d"]
+        ):
+            self.mel_conv1d = nn.Sequential(
+                nn.Conv1d(64, 64, kernel_size=5, stride=3, padding=2),
+                nn.BatchNorm1d(64),  # No Relu
+            )
+            if self.fusion_type == "daf_1d":
+                self.fusion_model = DAF()
+            elif self.fusion_type == "aff_1d":
+                self.fusion_model = AFF(channels=64, type="1D")
+            elif self.fusion_type == "iaff_1d":
+                self.fusion_model = iAFF(channels=64, type="1D")
+
+        if (self.enable_fusion) and (
+            self.fusion_type in ["daf_2d", "aff_2d", "iaff_2d"]
+        ):
+            self.mel_conv2d = nn.Sequential(
+                nn.Conv2d(1, 64, kernel_size=(5, 5), stride=(6, 2), padding=(2, 2)),
+                nn.BatchNorm2d(64),
+                nn.ReLU(inplace=True),
+            )
+
+            if self.fusion_type == "daf_2d":
+                self.fusion_model = DAF()
+            elif self.fusion_type == "aff_2d":
+                self.fusion_model = AFF(channels=64, type="2D")
+            elif self.fusion_type == "iaff_2d":
+                self.fusion_model = iAFF(channels=64, type="2D")
+        self.init_weight()
+
+    def init_weight(self):
+        init_bn(self.bn0)
+        init_layer(self.fc1)
+        init_layer(self.fc_audioset)
+
+    def forward(self, input, mixup_lambda=None, device=None):
+        """
+        Input: (batch_size, data_length)"""
+
+        if self.enable_fusion and input["longer"].sum() == 0:
+            # if no audio is longer than 10s, then randomly select one audio to be longer
+            input["longer"][torch.randint(0, input["longer"].shape[0], (1,))] = True
+
+        if not self.enable_fusion:
+            x = self.spectrogram_extractor(
+                input["waveform"].to(device=device, non_blocking=True)
+            )  # (batch_size, 1, time_steps, freq_bins)
+            x = self.logmel_extractor(x)  # (batch_size, 1, time_steps, mel_bins)
+
+            x = x.transpose(1, 3)
+            x = self.bn0(x)
+            x = x.transpose(1, 3)
+        else:
+            longer_list = input["longer"].to(device=device, non_blocking=True)
+            x = input["mel_fusion"].to(device=device, non_blocking=True)
+            longer_list_idx = torch.where(longer_list)[0]
+            x = x.transpose(1, 3)
+            x = self.bn0(x)
+            x = x.transpose(1, 3)
+            if self.fusion_type in ["daf_1d", "aff_1d", "iaff_1d"]:
+                new_x = x[:, 0:1, :, :].clone().contiguous()
+                # local processing
+                if len(longer_list_idx) > 0:
+                    fusion_x_local = x[longer_list_idx, 1:, :, :].clone().contiguous()
+                    FB, FC, FT, FF = fusion_x_local.size()
+                    fusion_x_local = fusion_x_local.view(FB * FC, FT, FF)
+                    fusion_x_local = torch.permute(
+                        fusion_x_local, (0, 2, 1)
+                    ).contiguous()
+                    fusion_x_local = self.mel_conv1d(fusion_x_local)
+                    fusion_x_local = fusion_x_local.view(
+                        FB, FC, FF, fusion_x_local.size(-1)
+                    )
+                    fusion_x_local = (
+                        torch.permute(fusion_x_local, (0, 2, 1, 3))
+                        .contiguous()
+                        .flatten(2)
+                    )
+                    if fusion_x_local.size(-1) < FT:
+                        fusion_x_local = torch.cat(
+                            [
+                                fusion_x_local,
+                                torch.zeros(
+                                    (FB, FF, FT - fusion_x_local.size(-1)),
+                                    device=device,
+                                ),
+                            ],
+                            dim=-1,
+                        )
+                    else:
+                        fusion_x_local = fusion_x_local[:, :, :FT]
+                    # 1D fusion
+                    new_x = new_x.squeeze(1).permute((0, 2, 1)).contiguous()
+                    new_x[longer_list_idx] = self.fusion_model(
+                        new_x[longer_list_idx], fusion_x_local
+                    )
+                    x = new_x.permute((0, 2, 1)).contiguous()[:, None, :, :]
+                else:
+                    x = new_x
+            elif self.fusion_type in ["daf_2d", "aff_2d", "iaff_2d", "channel_map"]:
+                x = x  # no change
+
+        if self.training:
+            x = self.spec_augmenter(x)
+        # Mixup on spectrogram
+        if self.training and mixup_lambda is not None:
+            x = do_mixup(x, mixup_lambda)
+        if (self.enable_fusion) and (
+            self.fusion_type in ["daf_2d", "aff_2d", "iaff_2d"]
+        ):
+            global_x = x[:, 0:1, :, :]
+
+            # global processing
+            B, C, H, W = global_x.shape
+            global_x = self.conv_block1(global_x, pool_size=(2, 2), pool_type="avg")
+            if len(longer_list_idx) > 0:
+                local_x = x[longer_list_idx, 1:, :, :].contiguous()
+                TH = global_x.size(-2)
+                # local processing
+                B, C, H, W = local_x.shape
+                local_x = local_x.view(B * C, 1, H, W)
+                local_x = self.mel_conv2d(local_x)
+                local_x = local_x.view(
+                    B, C, local_x.size(1), local_x.size(2), local_x.size(3)
+                )
+                local_x = local_x.permute((0, 2, 1, 3, 4)).contiguous().flatten(2, 3)
+                TB, TC, _, TW = local_x.size()
+                if local_x.size(-2) < TH:
+                    local_x = torch.cat(
+                        [
+                            local_x,
+                            torch.zeros(
+                                (TB, TC, TH - local_x.size(-2), TW),
+                                device=global_x.device,
+                            ),
+                        ],
+                        dim=-2,
+                    )
+                else:
+                    local_x = local_x[:, :, :TH, :]
+
+                global_x[longer_list_idx] = self.fusion_model(
+                    global_x[longer_list_idx], local_x
+                )
+            x = global_x
+        else:
+            x = self.conv_block1(x, pool_size=(2, 2), pool_type="avg")
+
+        x = F.dropout(x, p=0.2, training=self.training)
+        x = self.conv_block2(x, pool_size=(2, 2), pool_type="avg")
+        x = F.dropout(x, p=0.2, training=self.training)
+        x = self.conv_block3(x, pool_size=(2, 2), pool_type="avg")
+        x = F.dropout(x, p=0.2, training=self.training)
+        x = self.conv_block4(x, pool_size=(2, 2), pool_type="avg")
+        x = F.dropout(x, p=0.2, training=self.training)
+        x = self.conv_block5(x, pool_size=(2, 2), pool_type="avg")
+        x = F.dropout(x, p=0.2, training=self.training)
+        x = self.conv_block6(x, pool_size=(1, 1), pool_type="avg")
+        x = F.dropout(x, p=0.2, training=self.training)
+        x = torch.mean(x, dim=3)
+
+        latent_x1 = F.max_pool1d(x, kernel_size=3, stride=1, padding=1)
+        latent_x2 = F.avg_pool1d(x, kernel_size=3, stride=1, padding=1)
+        latent_x = latent_x1 + latent_x2
+        latent_x = latent_x.transpose(1, 2)
+        latent_x = F.relu_(self.fc1(latent_x))
+        latent_output = interpolate(latent_x, 32)
+
+        (x1, _) = torch.max(x, dim=2)
+        x2 = torch.mean(x, dim=2)
+        x = x1 + x2
+        x = F.dropout(x, p=0.5, training=self.training)
+        x = F.relu_(self.fc1(x))
+        embedding = F.dropout(x, p=0.5, training=self.training)
+        clipwise_output = torch.sigmoid(self.fc_audioset(x))
+
+        output_dict = {
+            "clipwise_output": clipwise_output,
+            "embedding": embedding,
+            "fine_grained_embedding": latent_output,
+        }
+        return output_dict
+
+
+class Cnn6(nn.Module):
+    def __init__(
+        self,
+        sample_rate,
+        window_size,
+        hop_size,
+        mel_bins,
+        fmin,
+        fmax,
+        classes_num,
+        enable_fusion=False,
+        fusion_type="None",
+    ):
+        super(Cnn6, self).__init__()
+
+        window = "hann"
+        center = True
+        pad_mode = "reflect"
+        ref = 1.0
+        amin = 1e-10
+        top_db = None
+
+        self.enable_fusion = enable_fusion
+        self.fusion_type = fusion_type
+
+        # Spectrogram extractor
+        self.spectrogram_extractor = Spectrogram(
+            n_fft=window_size,
+            hop_length=hop_size,
+            win_length=window_size,
+            window=window,
+            center=center,
+            pad_mode=pad_mode,
+            freeze_parameters=True,
+        )
+
+        # Logmel feature extractor
+        self.logmel_extractor = LogmelFilterBank(
+            sr=sample_rate,
+            n_fft=window_size,
+            n_mels=mel_bins,
+            fmin=fmin,
+            fmax=fmax,
+            ref=ref,
+            amin=amin,
+            top_db=top_db,
+            freeze_parameters=True,
+        )
+
+        # Spec augmenter
+        self.spec_augmenter = SpecAugmentation(
+            time_drop_width=64,
+            time_stripes_num=2,
+            freq_drop_width=8,
+            freq_stripes_num=2,
+        )
+
+        self.bn0 = nn.BatchNorm2d(64)
+
+        self.conv_block1 = ConvBlock5x5(in_channels=1, out_channels=64)
+        self.conv_block2 = ConvBlock5x5(in_channels=64, out_channels=128)
+        self.conv_block3 = ConvBlock5x5(in_channels=128, out_channels=256)
+        self.conv_block4 = ConvBlock5x5(in_channels=256, out_channels=512)
+
+        self.fc1 = nn.Linear(512, 512, bias=True)
+        self.fc_audioset = nn.Linear(512, classes_num, bias=True)
+
+        self.init_weight()
+
+    def init_weight(self):
+        init_bn(self.bn0)
+        init_layer(self.fc1)
+        init_layer(self.fc_audioset)
+
+    def forward(self, input, mixup_lambda=None, device=None):
+        """
+        Input: (batch_size, data_length)"""
+
+        x = self.spectrogram_extractor(input)  # (batch_size, 1, time_steps, freq_bins)
+        x = self.logmel_extractor(x)  # (batch_size, 1, time_steps, mel_bins)
+
+        x = x.transpose(1, 3)
+        x = self.bn0(x)
+        x = x.transpose(1, 3)
+
+        if self.training:
+            x = self.spec_augmenter(x)
+
+        # Mixup on spectrogram
+        if self.training and mixup_lambda is not None:
+            x = do_mixup(x, mixup_lambda)
+
+        x = self.conv_block1(x, pool_size=(2, 2), pool_type="avg")
+        x = F.dropout(x, p=0.2, training=self.training)
+        x = self.conv_block2(x, pool_size=(2, 2), pool_type="avg")
+        x = F.dropout(x, p=0.2, training=self.training)
+        x = self.conv_block3(x, pool_size=(2, 2), pool_type="avg")
+        x = F.dropout(x, p=0.2, training=self.training)
+        x = self.conv_block4(x, pool_size=(2, 2), pool_type="avg")
+        x = F.dropout(x, p=0.2, training=self.training)
+        x = torch.mean(x, dim=3)
+
+        latent_x1 = F.max_pool1d(x, kernel_size=3, stride=1, padding=1)
+        latent_x2 = F.avg_pool1d(x, kernel_size=3, stride=1, padding=1)
+        latent_x = latent_x1 + latent_x2
+        latent_x = latent_x.transpose(1, 2)
+        latent_x = F.relu_(self.fc1(latent_x))
+        latent_output = interpolate(latent_x, 16)
+
+        (x1, _) = torch.max(x, dim=2)
+        x2 = torch.mean(x, dim=2)
+        x = x1 + x2
+        x = F.dropout(x, p=0.5, training=self.training)
+        x = F.relu_(self.fc1(x))
+        embedding = F.dropout(x, p=0.5, training=self.training)
+        clipwise_output = torch.sigmoid(self.fc_audioset(x))
+
+        output_dict = {
+            "clipwise_output": clipwise_output,
+            "embedding": embedding,
+            "fine_grained_embedding": latent_output,
+        }
+
+        return output_dict
+
+
+class Cnn10(nn.Module):
+    def __init__(
+        self,
+        sample_rate,
+        window_size,
+        hop_size,
+        mel_bins,
+        fmin,
+        fmax,
+        classes_num,
+        enable_fusion=False,
+        fusion_type="None",
+    ):
+        super(Cnn10, self).__init__()
+
+        window = "hann"
+        center = True
+        pad_mode = "reflect"
+        ref = 1.0
+        amin = 1e-10
+        top_db = None
+
+        self.enable_fusion = enable_fusion
+        self.fusion_type = fusion_type
+
+        # Spectrogram extractor
+        self.spectrogram_extractor = Spectrogram(
+            n_fft=window_size,
+            hop_length=hop_size,
+            win_length=window_size,
+            window=window,
+            center=center,
+            pad_mode=pad_mode,
+            freeze_parameters=True,
+        )
+
+        # Logmel feature extractor
+        self.logmel_extractor = LogmelFilterBank(
+            sr=sample_rate,
+            n_fft=window_size,
+            n_mels=mel_bins,
+            fmin=fmin,
+            fmax=fmax,
+            ref=ref,
+            amin=amin,
+            top_db=top_db,
+            freeze_parameters=True,
+        )
+
+        # Spec augmenter
+        self.spec_augmenter = SpecAugmentation(
+            time_drop_width=64,
+            time_stripes_num=2,
+            freq_drop_width=8,
+            freq_stripes_num=2,
+        )
+
+        self.bn0 = nn.BatchNorm2d(64)
+
+        self.conv_block1 = ConvBlock(in_channels=1, out_channels=64)
+        self.conv_block2 = ConvBlock(in_channels=64, out_channels=128)
+        self.conv_block3 = ConvBlock(in_channels=128, out_channels=256)
+        self.conv_block4 = ConvBlock(in_channels=256, out_channels=512)
+        self.conv_block5 = ConvBlock(in_channels=512, out_channels=1024)
+
+        self.fc1 = nn.Linear(1024, 1024, bias=True)
+        self.fc_audioset = nn.Linear(1024, classes_num, bias=True)
+
+        self.init_weight()
+
+    def init_weight(self):
+        init_bn(self.bn0)
+        init_layer(self.fc1)
+        init_layer(self.fc_audioset)
+
+    def forward(self, input, mixup_lambda=None, device=None):
+        """
+        Input: (batch_size, data_length)"""
+
+        x = self.spectrogram_extractor(input)  # (batch_size, 1, time_steps, freq_bins)
+        x = self.logmel_extractor(x)  # (batch_size, 1, time_steps, mel_bins)
+
+        x = x.transpose(1, 3)
+        x = self.bn0(x)
+        x = x.transpose(1, 3)
+
+        if self.training:
+            x = self.spec_augmenter(x)
+
+        # Mixup on spectrogram
+        if self.training and mixup_lambda is not None:
+            x = do_mixup(x, mixup_lambda)
+
+        x = self.conv_block1(x, pool_size=(2, 2), pool_type="avg")
+        x = F.dropout(x, p=0.2, training=self.training)
+        x = self.conv_block2(x, pool_size=(2, 2), pool_type="avg")
+        x = F.dropout(x, p=0.2, training=self.training)
+        x = self.conv_block3(x, pool_size=(2, 2), pool_type="avg")
+        x = F.dropout(x, p=0.2, training=self.training)
+        x = self.conv_block4(x, pool_size=(2, 2), pool_type="avg")
+        x = F.dropout(x, p=0.2, training=self.training)
+        x = self.conv_block5(x, pool_size=(2, 2), pool_type="avg")
+        x = F.dropout(x, p=0.2, training=self.training)
+        x = torch.mean(x, dim=3)
+
+        latent_x1 = F.max_pool1d(x, kernel_size=3, stride=1, padding=1)
+        latent_x2 = F.avg_pool1d(x, kernel_size=3, stride=1, padding=1)
+        latent_x = latent_x1 + latent_x2
+        latent_x = latent_x.transpose(1, 2)
+        latent_x = F.relu_(self.fc1(latent_x))
+        latent_output = interpolate(latent_x, 32)
+
+        (x1, _) = torch.max(x, dim=2)
+        x2 = torch.mean(x, dim=2)
+        x = x1 + x2
+        x = F.dropout(x, p=0.5, training=self.training)
+        x = F.relu_(self.fc1(x))
+        embedding = F.dropout(x, p=0.5, training=self.training)
+        clipwise_output = torch.sigmoid(self.fc_audioset(x))
+
+        output_dict = {
+            "clipwise_output": clipwise_output,
+            "embedding": embedding,
+            "fine_grained_embedding": latent_output,
+        }
+
+        return output_dict
+
+
+def create_pann_model(audio_cfg, enable_fusion=False, fusion_type="None"):
+    try:
+        ModelProto = eval(audio_cfg.model_name)
+        model = ModelProto(
+            sample_rate=audio_cfg.sample_rate,
+            window_size=audio_cfg.window_size,
+            hop_size=audio_cfg.hop_size,
+            mel_bins=audio_cfg.mel_bins,
+            fmin=audio_cfg.fmin,
+            fmax=audio_cfg.fmax,
+            classes_num=audio_cfg.class_num,
+            enable_fusion=enable_fusion,
+            fusion_type=fusion_type,
+        )
+        return model
+    except:
+        raise RuntimeError(
+            f"Import Model for {audio_cfg.model_name} not found, or the audio cfg parameters are not enough."
+        )

audiosr/clap/open_clip/pretrained.py

@@ -1,167 +1,167 @@
-import hashlib
-import os
-import urllib
-import warnings
-
-from tqdm import tqdm
-
-_RN50 = dict(
-    openai="https://openaipublic.azureedge.net/clip/models/afeb0e10f9e5a86da6080e35cf09123aca3b358a0c3e3b6c78a7b63bc04b6762/RN50.pt",
-    yfcc15m="https://github.com/mlfoundations/open_clip/releases/download/v0.2-weights/rn50-quickgelu-yfcc15m-455df137.pt",
-    cc12m="https://github.com/mlfoundations/open_clip/releases/download/v0.2-weights/rn50-quickgelu-cc12m-f000538c.pt",
-)
-
-_RN50_quickgelu = dict(
-    openai="https://openaipublic.azureedge.net/clip/models/afeb0e10f9e5a86da6080e35cf09123aca3b358a0c3e3b6c78a7b63bc04b6762/RN50.pt",
-    yfcc15m="https://github.com/mlfoundations/open_clip/releases/download/v0.2-weights/rn50-quickgelu-yfcc15m-455df137.pt",
-    cc12m="https://github.com/mlfoundations/open_clip/releases/download/v0.2-weights/rn50-quickgelu-cc12m-f000538c.pt",
-)
-
-_RN101 = dict(
-    openai="https://openaipublic.azureedge.net/clip/models/8fa8567bab74a42d41c5915025a8e4538c3bdbe8804a470a72f30b0d94fab599/RN101.pt",
-    yfcc15m="https://github.com/mlfoundations/open_clip/releases/download/v0.2-weights/rn101-quickgelu-yfcc15m-3e04b30e.pt",
-)
-
-_RN101_quickgelu = dict(
-    openai="https://openaipublic.azureedge.net/clip/models/8fa8567bab74a42d41c5915025a8e4538c3bdbe8804a470a72f30b0d94fab599/RN101.pt",
-    yfcc15m="https://github.com/mlfoundations/open_clip/releases/download/v0.2-weights/rn101-quickgelu-yfcc15m-3e04b30e.pt",
-)
-
-_RN50x4 = dict(
-    openai="https://openaipublic.azureedge.net/clip/models/7e526bd135e493cef0776de27d5f42653e6b4c8bf9e0f653bb11773263205fdd/RN50x4.pt",
-)
-
-_RN50x16 = dict(
-    openai="https://openaipublic.azureedge.net/clip/models/52378b407f34354e150460fe41077663dd5b39c54cd0bfd2b27167a4a06ec9aa/RN50x16.pt",
-)
-
-_RN50x64 = dict(
-    openai="https://openaipublic.azureedge.net/clip/models/be1cfb55d75a9666199fb2206c106743da0f6468c9d327f3e0d0a543a9919d9c/RN50x64.pt",
-)
-
-_VITB32 = dict(
-    openai="https://openaipublic.azureedge.net/clip/models/40d365715913c9da98579312b702a82c18be219cc2a73407c4526f58eba950af/ViT-B-32.pt",
-    laion400m_e31="https://github.com/mlfoundations/open_clip/releases/download/v0.2-weights/vit_b_32-quickgelu-laion400m_e31-d867053b.pt",
-    laion400m_e32="https://github.com/mlfoundations/open_clip/releases/download/v0.2-weights/vit_b_32-quickgelu-laion400m_e32-46683a32.pt",
-    laion400m_avg="https://github.com/mlfoundations/open_clip/releases/download/v0.2-weights/vit_b_32-quickgelu-laion400m_avg-8a00ab3c.pt",
-)
-
-_VITB32_quickgelu = dict(
-    openai="https://openaipublic.azureedge.net/clip/models/40d365715913c9da98579312b702a82c18be219cc2a73407c4526f58eba950af/ViT-B-32.pt",
-    laion400m_e31="https://github.com/mlfoundations/open_clip/releases/download/v0.2-weights/vit_b_32-quickgelu-laion400m_e31-d867053b.pt",
-    laion400m_e32="https://github.com/mlfoundations/open_clip/releases/download/v0.2-weights/vit_b_32-quickgelu-laion400m_e32-46683a32.pt",
-    laion400m_avg="https://github.com/mlfoundations/open_clip/releases/download/v0.2-weights/vit_b_32-quickgelu-laion400m_avg-8a00ab3c.pt",
-)
-
-_VITB16 = dict(
-    openai="https://openaipublic.azureedge.net/clip/models/5806e77cd80f8b59890b7e101eabd078d9fb84e6937f9e85e4ecb61988df416f/ViT-B-16.pt",
-)
-
-_VITL14 = dict(
-    openai="https://openaipublic.azureedge.net/clip/models/b8cca3fd41ae0c99ba7e8951adf17d267cdb84cd88be6f7c2e0eca1737a03836/ViT-L-14.pt",
-)
-
-_PRETRAINED = {
-    "RN50": _RN50,
-    "RN50-quickgelu": _RN50_quickgelu,
-    "RN101": _RN101,
-    "RN101-quickgelu": _RN101_quickgelu,
-    "RN50x4": _RN50x4,
-    "RN50x16": _RN50x16,
-    "ViT-B-32": _VITB32,
-    "ViT-B-32-quickgelu": _VITB32_quickgelu,
-    "ViT-B-16": _VITB16,
-    "ViT-L-14": _VITL14,
-}
-
-
-def list_pretrained(as_str: bool = False):
-    """returns list of pretrained models
-    Returns a tuple (model_name, pretrain_tag) by default or 'name:tag' if as_str == True
-    """
-    return [
-        ":".join([k, t]) if as_str else (k, t)
-        for k in _PRETRAINED.keys()
-        for t in _PRETRAINED[k].keys()
-    ]
-
-
-def list_pretrained_tag_models(tag: str):
-    """return all models having the specified pretrain tag"""
-    models = []
-    for k in _PRETRAINED.keys():
-        if tag in _PRETRAINED[k]:
-            models.append(k)
-    return models
-
-
-def list_pretrained_model_tags(model: str):
-    """return all pretrain tags for the specified model architecture"""
-    tags = []
-    if model in _PRETRAINED:
-        tags.extend(_PRETRAINED[model].keys())
-    return tags
-
-
-def get_pretrained_url(model: str, tag: str):
-    if model not in _PRETRAINED:
-        return ""
-    model_pretrained = _PRETRAINED[model]
-    if tag not in model_pretrained:
-        return ""
-    return model_pretrained[tag]
-
-
-def download_pretrained(url: str, root: str = os.path.expanduser("~/.cache/clip")):
-    os.makedirs(root, exist_ok=True)
-    filename = os.path.basename(url)
-
-    if "openaipublic" in url:
-        expected_sha256 = url.split("/")[-2]
-    else:
-        expected_sha256 = ""
-
-    download_target = os.path.join(root, filename)
-
-    if os.path.exists(download_target) and not os.path.isfile(download_target):
-        raise RuntimeError(f"{download_target} exists and is not a regular file")
-
-    if os.path.isfile(download_target):
-        if expected_sha256:
-            if (
-                hashlib.sha256(open(download_target, "rb").read()).hexdigest()
-                == expected_sha256
-            ):
-                return download_target
-            else:
-                warnings.warn(
-                    f"{download_target} exists, but the SHA256 checksum does not match; re-downloading the file"
-                )
-        else:
-            return download_target
-
-    with urllib.request.urlopen(url) as source, open(download_target, "wb") as output:
-        with tqdm(
-            total=int(source.info().get("Content-Length")),
-            ncols=80,
-            unit="iB",
-            unit_scale=True,
-        ) as loop:
-            while True:
-                buffer = source.read(8192)
-                if not buffer:
-                    break
-
-                output.write(buffer)
-                loop.update(len(buffer))
-
-    if (
-        expected_sha256
-        and hashlib.sha256(open(download_target, "rb").read()).hexdigest()
-        != expected_sha256
-    ):
-        raise RuntimeError(
-            f"Model has been downloaded but the SHA256 checksum does not not match"
-        )
-
-    return download_target
+import hashlib
+import os
+import urllib
+import warnings
+
+from tqdm import tqdm
+
+_RN50 = dict(
+    openai="https://openaipublic.azureedge.net/clip/models/afeb0e10f9e5a86da6080e35cf09123aca3b358a0c3e3b6c78a7b63bc04b6762/RN50.pt",
+    yfcc15m="https://github.com/mlfoundations/open_clip/releases/download/v0.2-weights/rn50-quickgelu-yfcc15m-455df137.pt",
+    cc12m="https://github.com/mlfoundations/open_clip/releases/download/v0.2-weights/rn50-quickgelu-cc12m-f000538c.pt",
+)
+
+_RN50_quickgelu = dict(
+    openai="https://openaipublic.azureedge.net/clip/models/afeb0e10f9e5a86da6080e35cf09123aca3b358a0c3e3b6c78a7b63bc04b6762/RN50.pt",
+    yfcc15m="https://github.com/mlfoundations/open_clip/releases/download/v0.2-weights/rn50-quickgelu-yfcc15m-455df137.pt",
+    cc12m="https://github.com/mlfoundations/open_clip/releases/download/v0.2-weights/rn50-quickgelu-cc12m-f000538c.pt",
+)
+
+_RN101 = dict(
+    openai="https://openaipublic.azureedge.net/clip/models/8fa8567bab74a42d41c5915025a8e4538c3bdbe8804a470a72f30b0d94fab599/RN101.pt",
+    yfcc15m="https://github.com/mlfoundations/open_clip/releases/download/v0.2-weights/rn101-quickgelu-yfcc15m-3e04b30e.pt",
+)
+
+_RN101_quickgelu = dict(
+    openai="https://openaipublic.azureedge.net/clip/models/8fa8567bab74a42d41c5915025a8e4538c3bdbe8804a470a72f30b0d94fab599/RN101.pt",
+    yfcc15m="https://github.com/mlfoundations/open_clip/releases/download/v0.2-weights/rn101-quickgelu-yfcc15m-3e04b30e.pt",
+)
+
+_RN50x4 = dict(
+    openai="https://openaipublic.azureedge.net/clip/models/7e526bd135e493cef0776de27d5f42653e6b4c8bf9e0f653bb11773263205fdd/RN50x4.pt",
+)
+
+_RN50x16 = dict(
+    openai="https://openaipublic.azureedge.net/clip/models/52378b407f34354e150460fe41077663dd5b39c54cd0bfd2b27167a4a06ec9aa/RN50x16.pt",
+)
+
+_RN50x64 = dict(
+    openai="https://openaipublic.azureedge.net/clip/models/be1cfb55d75a9666199fb2206c106743da0f6468c9d327f3e0d0a543a9919d9c/RN50x64.pt",
+)
+
+_VITB32 = dict(
+    openai="https://openaipublic.azureedge.net/clip/models/40d365715913c9da98579312b702a82c18be219cc2a73407c4526f58eba950af/ViT-B-32.pt",
+    laion400m_e31="https://github.com/mlfoundations/open_clip/releases/download/v0.2-weights/vit_b_32-quickgelu-laion400m_e31-d867053b.pt",
+    laion400m_e32="https://github.com/mlfoundations/open_clip/releases/download/v0.2-weights/vit_b_32-quickgelu-laion400m_e32-46683a32.pt",
+    laion400m_avg="https://github.com/mlfoundations/open_clip/releases/download/v0.2-weights/vit_b_32-quickgelu-laion400m_avg-8a00ab3c.pt",
+)
+
+_VITB32_quickgelu = dict(
+    openai="https://openaipublic.azureedge.net/clip/models/40d365715913c9da98579312b702a82c18be219cc2a73407c4526f58eba950af/ViT-B-32.pt",
+    laion400m_e31="https://github.com/mlfoundations/open_clip/releases/download/v0.2-weights/vit_b_32-quickgelu-laion400m_e31-d867053b.pt",
+    laion400m_e32="https://github.com/mlfoundations/open_clip/releases/download/v0.2-weights/vit_b_32-quickgelu-laion400m_e32-46683a32.pt",
+    laion400m_avg="https://github.com/mlfoundations/open_clip/releases/download/v0.2-weights/vit_b_32-quickgelu-laion400m_avg-8a00ab3c.pt",
+)
+
+_VITB16 = dict(
+    openai="https://openaipublic.azureedge.net/clip/models/5806e77cd80f8b59890b7e101eabd078d9fb84e6937f9e85e4ecb61988df416f/ViT-B-16.pt",
+)
+
+_VITL14 = dict(
+    openai="https://openaipublic.azureedge.net/clip/models/b8cca3fd41ae0c99ba7e8951adf17d267cdb84cd88be6f7c2e0eca1737a03836/ViT-L-14.pt",
+)
+
+_PRETRAINED = {
+    "RN50": _RN50,
+    "RN50-quickgelu": _RN50_quickgelu,
+    "RN101": _RN101,
+    "RN101-quickgelu": _RN101_quickgelu,
+    "RN50x4": _RN50x4,
+    "RN50x16": _RN50x16,
+    "ViT-B-32": _VITB32,
+    "ViT-B-32-quickgelu": _VITB32_quickgelu,
+    "ViT-B-16": _VITB16,
+    "ViT-L-14": _VITL14,
+}
+
+
+def list_pretrained(as_str: bool = False):
+    """returns list of pretrained models
+    Returns a tuple (model_name, pretrain_tag) by default or 'name:tag' if as_str == True
+    """
+    return [
+        ":".join([k, t]) if as_str else (k, t)
+        for k in _PRETRAINED.keys()
+        for t in _PRETRAINED[k].keys()
+    ]
+
+
+def list_pretrained_tag_models(tag: str):
+    """return all models having the specified pretrain tag"""
+    models = []
+    for k in _PRETRAINED.keys():
+        if tag in _PRETRAINED[k]:
+            models.append(k)
+    return models
+
+
+def list_pretrained_model_tags(model: str):
+    """return all pretrain tags for the specified model architecture"""
+    tags = []
+    if model in _PRETRAINED:
+        tags.extend(_PRETRAINED[model].keys())
+    return tags
+
+
+def get_pretrained_url(model: str, tag: str):
+    if model not in _PRETRAINED:
+        return ""
+    model_pretrained = _PRETRAINED[model]
+    if tag not in model_pretrained:
+        return ""
+    return model_pretrained[tag]
+
+
+def download_pretrained(url: str, root: str = os.path.expanduser("~/.cache/clip")):
+    os.makedirs(root, exist_ok=True)
+    filename = os.path.basename(url)
+
+    if "openaipublic" in url:
+        expected_sha256 = url.split("/")[-2]
+    else:
+        expected_sha256 = ""
+
+    download_target = os.path.join(root, filename)
+
+    if os.path.exists(download_target) and not os.path.isfile(download_target):
+        raise RuntimeError(f"{download_target} exists and is not a regular file")
+
+    if os.path.isfile(download_target):
+        if expected_sha256:
+            if (
+                hashlib.sha256(open(download_target, "rb").read()).hexdigest()
+                == expected_sha256
+            ):
+                return download_target
+            else:
+                warnings.warn(
+                    f"{download_target} exists, but the SHA256 checksum does not match; re-downloading the file"
+                )
+        else:
+            return download_target
+
+    with urllib.request.urlopen(url) as source, open(download_target, "wb") as output:
+        with tqdm(
+            total=int(source.info().get("Content-Length")),
+            ncols=80,
+            unit="iB",
+            unit_scale=True,
+        ) as loop:
+            while True:
+                buffer = source.read(8192)
+                if not buffer:
+                    break
+
+                output.write(buffer)
+                loop.update(len(buffer))
+
+    if (
+        expected_sha256
+        and hashlib.sha256(open(download_target, "rb").read()).hexdigest()
+        != expected_sha256
+    ):
+        raise RuntimeError(
+            f"Model has been downloaded but the SHA256 checksum does not not match"
+        )
+
+    return download_target

audiosr/clap/open_clip/timm_model.py

@@ -1,112 +1,112 @@
-""" timm model adapter
-
-Wraps timm (https://github.com/rwightman/pytorch-image-models) models for use as a vision tower in CLIP model.
-"""
-from collections import OrderedDict
-
-import torch.nn as nn
-
-try:
-    import timm
-    from timm.models.layers import Mlp, to_2tuple
-    from timm.models.layers.attention_pool2d import RotAttentionPool2d
-    from timm.models.layers.attention_pool2d import (
-        AttentionPool2d as AbsAttentionPool2d,
-    )
-except ImportError:
-    timm = None
-
-from .utils import freeze_batch_norm_2d
-
-
-class TimmModel(nn.Module):
-    """timm model adapter
-    # FIXME this adapter is a work in progress, may change in ways that break weight compat
-    """
-
-    def __init__(
-        self,
-        model_name,
-        embed_dim,
-        image_size=224,
-        pool="avg",
-        proj="linear",
-        drop=0.0,
-        pretrained=False,
-    ):
-        super().__init__()
-        if timm is None:
-            raise RuntimeError("Please `pip install timm` to use timm models.")
-
-        self.image_size = to_2tuple(image_size)
-        self.trunk = timm.create_model(model_name, pretrained=pretrained)
-        feat_size = self.trunk.default_cfg.get("pool_size", None)
-        feature_ndim = 1 if not feat_size else 2
-        if pool in ("abs_attn", "rot_attn"):
-            assert feature_ndim == 2
-            # if attn pooling used, remove both classifier and default pool
-            self.trunk.reset_classifier(0, global_pool="")
-        else:
-            # reset global pool if pool config set, otherwise leave as network default
-            reset_kwargs = dict(global_pool=pool) if pool else {}
-            self.trunk.reset_classifier(0, **reset_kwargs)
-        prev_chs = self.trunk.num_features
-
-        head_layers = OrderedDict()
-        if pool == "abs_attn":
-            head_layers["pool"] = AbsAttentionPool2d(
-                prev_chs, feat_size=feat_size, out_features=embed_dim
-            )
-            prev_chs = embed_dim
-        elif pool == "rot_attn":
-            head_layers["pool"] = RotAttentionPool2d(prev_chs, out_features=embed_dim)
-            prev_chs = embed_dim
-        else:
-            assert proj, "projection layer needed if non-attention pooling is used."
-
-        # NOTE attention pool ends with a projection layer, so proj should usually be set to '' if such pooling is used
-        if proj == "linear":
-            head_layers["drop"] = nn.Dropout(drop)
-            head_layers["proj"] = nn.Linear(prev_chs, embed_dim)
-        elif proj == "mlp":
-            head_layers["mlp"] = Mlp(prev_chs, 2 * embed_dim, embed_dim, drop=drop)
-
-        self.head = nn.Sequential(head_layers)
-
-    def lock(self, unlocked_groups=0, freeze_bn_stats=False):
-        """lock modules
-        Args:
-            unlocked_groups (int): leave last n layer groups unlocked (default: 0)
-        """
-        if not unlocked_groups:
-            # lock full model
-            for param in self.trunk.parameters():
-                param.requires_grad = False
-            if freeze_bn_stats:
-                freeze_batch_norm_2d(self.trunk)
-        else:
-            # NOTE: partial freeze requires latest timm (master) branch and is subject to change
-            try:
-                # FIXME import here until API stable and in an official release
-                from timm.models.helpers import group_parameters, group_modules
-            except ImportError:
-                raise RuntimeError(
-                    "Please install latest timm `pip install git+https://github.com/rwightman/pytorch-image-models`"
-                )
-            matcher = self.trunk.group_matcher()
-            gparams = group_parameters(self.trunk, matcher)
-            max_layer_id = max(gparams.keys())
-            max_layer_id = max_layer_id - unlocked_groups
-            for group_idx in range(max_layer_id + 1):
-                group = gparams[group_idx]
-                for param in group:
-                    self.trunk.get_parameter(param).requires_grad = False
-            if freeze_bn_stats:
-                gmodules = group_modules(self.trunk, matcher, reverse=True)
-                gmodules = {k for k, v in gmodules.items() if v <= max_layer_id}
-                freeze_batch_norm_2d(self.trunk, gmodules)
-
-    def forward(self, x):
-        x = self.trunk(x)
-        x = self.head(x)
-        return x
+""" timm model adapter
+
+Wraps timm (https://github.com/rwightman/pytorch-image-models) models for use as a vision tower in CLIP model.
+"""
+from collections import OrderedDict
+
+import torch.nn as nn
+
+try:
+    import timm
+    from timm.models.layers import Mlp, to_2tuple
+    from timm.models.layers.attention_pool2d import RotAttentionPool2d
+    from timm.models.layers.attention_pool2d import (
+        AttentionPool2d as AbsAttentionPool2d,
+    )
+except ImportError:
+    timm = None
+
+from .utils import freeze_batch_norm_2d
+
+
+class TimmModel(nn.Module):
+    """timm model adapter
+    # FIXME this adapter is a work in progress, may change in ways that break weight compat
+    """
+
+    def __init__(
+        self,
+        model_name,
+        embed_dim,
+        image_size=224,
+        pool="avg",
+        proj="linear",
+        drop=0.0,
+        pretrained=False,
+    ):
+        super().__init__()
+        if timm is None:
+            raise RuntimeError("Please `pip install timm` to use timm models.")
+
+        self.image_size = to_2tuple(image_size)
+        self.trunk = timm.create_model(model_name, pretrained=pretrained)
+        feat_size = self.trunk.default_cfg.get("pool_size", None)
+        feature_ndim = 1 if not feat_size else 2
+        if pool in ("abs_attn", "rot_attn"):
+            assert feature_ndim == 2
+            # if attn pooling used, remove both classifier and default pool
+            self.trunk.reset_classifier(0, global_pool="")
+        else:
+            # reset global pool if pool config set, otherwise leave as network default
+            reset_kwargs = dict(global_pool=pool) if pool else {}
+            self.trunk.reset_classifier(0, **reset_kwargs)
+        prev_chs = self.trunk.num_features
+
+        head_layers = OrderedDict()
+        if pool == "abs_attn":
+            head_layers["pool"] = AbsAttentionPool2d(
+                prev_chs, feat_size=feat_size, out_features=embed_dim
+            )
+            prev_chs = embed_dim
+        elif pool == "rot_attn":
+            head_layers["pool"] = RotAttentionPool2d(prev_chs, out_features=embed_dim)
+            prev_chs = embed_dim
+        else:
+            assert proj, "projection layer needed if non-attention pooling is used."
+
+        # NOTE attention pool ends with a projection layer, so proj should usually be set to '' if such pooling is used
+        if proj == "linear":
+            head_layers["drop"] = nn.Dropout(drop)
+            head_layers["proj"] = nn.Linear(prev_chs, embed_dim)
+        elif proj == "mlp":
+            head_layers["mlp"] = Mlp(prev_chs, 2 * embed_dim, embed_dim, drop=drop)
+
+        self.head = nn.Sequential(head_layers)
+
+    def lock(self, unlocked_groups=0, freeze_bn_stats=False):
+        """lock modules
+        Args:
+            unlocked_groups (int): leave last n layer groups unlocked (default: 0)
+        """
+        if not unlocked_groups:
+            # lock full model
+            for param in self.trunk.parameters():
+                param.requires_grad = False
+            if freeze_bn_stats:
+                freeze_batch_norm_2d(self.trunk)
+        else:
+            # NOTE: partial freeze requires latest timm (master) branch and is subject to change
+            try:
+                # FIXME import here until API stable and in an official release
+                from timm.models.helpers import group_parameters, group_modules
+            except ImportError:
+                raise RuntimeError(
+                    "Please install latest timm `pip install git+https://github.com/rwightman/pytorch-image-models`"
+                )
+            matcher = self.trunk.group_matcher()
+            gparams = group_parameters(self.trunk, matcher)
+            max_layer_id = max(gparams.keys())
+            max_layer_id = max_layer_id - unlocked_groups
+            for group_idx in range(max_layer_id + 1):
+                group = gparams[group_idx]
+                for param in group:
+                    self.trunk.get_parameter(param).requires_grad = False
+            if freeze_bn_stats:
+                gmodules = group_modules(self.trunk, matcher, reverse=True)
+                gmodules = {k for k, v in gmodules.items() if v <= max_layer_id}
+                freeze_batch_norm_2d(self.trunk, gmodules)
+
+    def forward(self, x):
+        x = self.trunk(x)
+        x = self.head(x)
+        return x

audiosr/clap/open_clip/tokenizer.py

@@ -1,197 +1,197 @@
-""" CLIP tokenizer
-
-Copied from https://github.com/openai/CLIP. Originally MIT License, Copyright (c) 2021 OpenAI.
-"""
-import gzip
-import html
-import os
-from functools import lru_cache
-from typing import Union, List
-
-import ftfy
-import regex as re
-import torch
-
-
-@lru_cache()
-def default_bpe():
-    return os.path.join(
-        os.path.dirname(os.path.abspath(__file__)), "bpe_simple_vocab_16e6.txt.gz"
-    )
-
-
-@lru_cache()
-def bytes_to_unicode():
-    """
-    Returns list of utf-8 byte and a corresponding list of unicode strings.
-    The reversible bpe codes work on unicode strings.
-    This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
-    When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
-    This is a signficant percentage of your normal, say, 32K bpe vocab.
-    To avoid that, we want lookup tables between utf-8 bytes and unicode strings.
-    And avoids mapping to whitespace/control characters the bpe code barfs on.
-    """
-    bs = (
-        list(range(ord("!"), ord("~") + 1))
-        + list(range(ord("¡"), ord("¬") + 1))
-        + list(range(ord("®"), ord("ÿ") + 1))
-    )
-    cs = bs[:]
-    n = 0
-    for b in range(2**8):
-        if b not in bs:
-            bs.append(b)
-            cs.append(2**8 + n)
-            n += 1
-    cs = [chr(n) for n in cs]
-    return dict(zip(bs, cs))
-
-
-def get_pairs(word):
-    """Return set of symbol pairs in a word.
-    Word is represented as tuple of symbols (symbols being variable-length strings).
-    """
-    pairs = set()
-    prev_char = word[0]
-    for char in word[1:]:
-        pairs.add((prev_char, char))
-        prev_char = char
-    return pairs
-
-
-def basic_clean(text):
-    text = ftfy.fix_text(text)
-    text = html.unescape(html.unescape(text))
-    return text.strip()
-
-
-def whitespace_clean(text):
-    text = re.sub(r"\s+", " ", text)
-    text = text.strip()
-    return text
-
-
-class SimpleTokenizer(object):
-    def __init__(self, bpe_path: str = default_bpe(), special_tokens=None):
-        self.byte_encoder = bytes_to_unicode()
-        self.byte_decoder = {v: k for k, v in self.byte_encoder.items()}
-        merges = gzip.open(bpe_path).read().decode("utf-8").split("\n")
-        merges = merges[1 : 49152 - 256 - 2 + 1]
-        merges = [tuple(merge.split()) for merge in merges]
-        vocab = list(bytes_to_unicode().values())
-        vocab = vocab + [v + "</w>" for v in vocab]
-        for merge in merges:
-            vocab.append("".join(merge))
-        if not special_tokens:
-            special_tokens = ["<start_of_text>", "<end_of_text>"]
-        else:
-            special_tokens = ["<start_of_text>", "<end_of_text>"] + special_tokens
-        vocab.extend(special_tokens)
-        self.encoder = dict(zip(vocab, range(len(vocab))))
-        self.decoder = {v: k for k, v in self.encoder.items()}
-        self.bpe_ranks = dict(zip(merges, range(len(merges))))
-        self.cache = {t: t for t in special_tokens}
-        special = "|".join(special_tokens)
-        self.pat = re.compile(
-            special + r"""|'s|'t|'re|'ve|'m|'ll|'d|[\p{L}]+|[\p{N}]|[^\s\p{L}\p{N}]+""",
-            re.IGNORECASE,
-        )
-
-        self.vocab_size = len(self.encoder)
-        self.all_special_ids = [self.encoder[t] for t in special_tokens]
-
-    def bpe(self, token):
-        if token in self.cache:
-            return self.cache[token]
-        word = tuple(token[:-1]) + (token[-1] + "</w>",)
-        pairs = get_pairs(word)
-
-        if not pairs:
-            return token + "</w>"
-
-        while True:
-            bigram = min(pairs, key=lambda pair: self.bpe_ranks.get(pair, float("inf")))
-            if bigram not in self.bpe_ranks:
-                break
-            first, second = bigram
-            new_word = []
-            i = 0
-            while i < len(word):
-                try:
-                    j = word.index(first, i)
-                    new_word.extend(word[i:j])
-                    i = j
-                except:
-                    new_word.extend(word[i:])
-                    break
-
-                if word[i] == first and i < len(word) - 1 and word[i + 1] == second:
-                    new_word.append(first + second)
-                    i += 2
-                else:
-                    new_word.append(word[i])
-                    i += 1
-            new_word = tuple(new_word)
-            word = new_word
-            if len(word) == 1:
-                break
-            else:
-                pairs = get_pairs(word)
-        word = " ".join(word)
-        self.cache[token] = word
-        return word
-
-    def encode(self, text):
-        bpe_tokens = []
-        text = whitespace_clean(basic_clean(text)).lower()
-        for token in re.findall(self.pat, text):
-            token = "".join(self.byte_encoder[b] for b in token.encode("utf-8"))
-            bpe_tokens.extend(
-                self.encoder[bpe_token] for bpe_token in self.bpe(token).split(" ")
-            )
-        return bpe_tokens
-
-    def decode(self, tokens):
-        text = "".join([self.decoder[token] for token in tokens])
-        text = (
-            bytearray([self.byte_decoder[c] for c in text])
-            .decode("utf-8", errors="replace")
-            .replace("</w>", " ")
-        )
-        return text
-
-
-_tokenizer = SimpleTokenizer()
-
-
-def tokenize(
-    texts: Union[str, List[str]], context_length: int = 77
-) -> torch.LongTensor:
-    """
-    Returns the tokenized representation of given input string(s)
-
-    Parameters
-    ----------
-    texts : Union[str, List[str]]
-        An input string or a list of input strings to tokenize
-    context_length : int
-        The context length to use; all CLIP models use 77 as the context length
-
-    Returns
-    -------
-    A two-dimensional tensor containing the resulting tokens, shape = [number of input strings, context_length]
-    """
-    if isinstance(texts, str):
-        texts = [texts]
-
-    sot_token = _tokenizer.encoder["<start_of_text>"]
-    eot_token = _tokenizer.encoder["<end_of_text>"]
-    all_tokens = [[sot_token] + _tokenizer.encode(text) + [eot_token] for text in texts]
-    result = torch.zeros(len(all_tokens), context_length, dtype=torch.long)
-
-    for i, tokens in enumerate(all_tokens):
-        if len(tokens) > context_length:
-            tokens = tokens[:context_length]  # Truncate
-        result[i, : len(tokens)] = torch.tensor(tokens)
-
-    return result
+""" CLIP tokenizer
+
+Copied from https://github.com/openai/CLIP. Originally MIT License, Copyright (c) 2021 OpenAI.
+"""
+import gzip
+import html
+import os
+from functools import lru_cache
+from typing import Union, List
+
+import ftfy
+import regex as re
+import torch
+
+
+@lru_cache()
+def default_bpe():
+    return os.path.join(
+        os.path.dirname(os.path.abspath(__file__)), "bpe_simple_vocab_16e6.txt.gz"
+    )
+
+
+@lru_cache()
+def bytes_to_unicode():
+    """
+    Returns list of utf-8 byte and a corresponding list of unicode strings.
+    The reversible bpe codes work on unicode strings.
+    This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
+    When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
+    This is a signficant percentage of your normal, say, 32K bpe vocab.
+    To avoid that, we want lookup tables between utf-8 bytes and unicode strings.
+    And avoids mapping to whitespace/control characters the bpe code barfs on.
+    """
+    bs = (
+        list(range(ord("!"), ord("~") + 1))
+        + list(range(ord("¡"), ord("¬") + 1))
+        + list(range(ord("®"), ord("ÿ") + 1))
+    )
+    cs = bs[:]
+    n = 0
+    for b in range(2**8):
+        if b not in bs:
+            bs.append(b)
+            cs.append(2**8 + n)
+            n += 1
+    cs = [chr(n) for n in cs]
+    return dict(zip(bs, cs))
+
+
+def get_pairs(word):
+    """Return set of symbol pairs in a word.
+    Word is represented as tuple of symbols (symbols being variable-length strings).
+    """
+    pairs = set()
+    prev_char = word[0]
+    for char in word[1:]:
+        pairs.add((prev_char, char))
+        prev_char = char
+    return pairs
+
+
+def basic_clean(text):
+    text = ftfy.fix_text(text)
+    text = html.unescape(html.unescape(text))
+    return text.strip()
+
+
+def whitespace_clean(text):
+    text = re.sub(r"\s+", " ", text)
+    text = text.strip()
+    return text
+
+
+class SimpleTokenizer(object):
+    def __init__(self, bpe_path: str = default_bpe(), special_tokens=None):
+        self.byte_encoder = bytes_to_unicode()
+        self.byte_decoder = {v: k for k, v in self.byte_encoder.items()}
+        merges = gzip.open(bpe_path).read().decode("utf-8").split("\n")
+        merges = merges[1 : 49152 - 256 - 2 + 1]
+        merges = [tuple(merge.split()) for merge in merges]
+        vocab = list(bytes_to_unicode().values())
+        vocab = vocab + [v + "</w>" for v in vocab]
+        for merge in merges:
+            vocab.append("".join(merge))
+        if not special_tokens:
+            special_tokens = ["<start_of_text>", "<end_of_text>"]
+        else:
+            special_tokens = ["<start_of_text>", "<end_of_text>"] + special_tokens
+        vocab.extend(special_tokens)
+        self.encoder = dict(zip(vocab, range(len(vocab))))
+        self.decoder = {v: k for k, v in self.encoder.items()}
+        self.bpe_ranks = dict(zip(merges, range(len(merges))))
+        self.cache = {t: t for t in special_tokens}
+        special = "|".join(special_tokens)
+        self.pat = re.compile(
+            special + r"""|'s|'t|'re|'ve|'m|'ll|'d|[\p{L}]+|[\p{N}]|[^\s\p{L}\p{N}]+""",
+            re.IGNORECASE,
+        )
+
+        self.vocab_size = len(self.encoder)
+        self.all_special_ids = [self.encoder[t] for t in special_tokens]
+
+    def bpe(self, token):
+        if token in self.cache:
+            return self.cache[token]
+        word = tuple(token[:-1]) + (token[-1] + "</w>",)
+        pairs = get_pairs(word)
+
+        if not pairs:
+            return token + "</w>"
+
+        while True:
+            bigram = min(pairs, key=lambda pair: self.bpe_ranks.get(pair, float("inf")))
+            if bigram not in self.bpe_ranks:
+                break
+            first, second = bigram
+            new_word = []
+            i = 0
+            while i < len(word):
+                try:
+                    j = word.index(first, i)
+                    new_word.extend(word[i:j])
+                    i = j
+                except:
+                    new_word.extend(word[i:])
+                    break
+
+                if word[i] == first and i < len(word) - 1 and word[i + 1] == second:
+                    new_word.append(first + second)
+                    i += 2
+                else:
+                    new_word.append(word[i])
+                    i += 1
+            new_word = tuple(new_word)
+            word = new_word
+            if len(word) == 1:
+                break
+            else:
+                pairs = get_pairs(word)
+        word = " ".join(word)
+        self.cache[token] = word
+        return word
+
+    def encode(self, text):
+        bpe_tokens = []
+        text = whitespace_clean(basic_clean(text)).lower()
+        for token in re.findall(self.pat, text):
+            token = "".join(self.byte_encoder[b] for b in token.encode("utf-8"))
+            bpe_tokens.extend(
+                self.encoder[bpe_token] for bpe_token in self.bpe(token).split(" ")
+            )
+        return bpe_tokens
+
+    def decode(self, tokens):
+        text = "".join([self.decoder[token] for token in tokens])
+        text = (
+            bytearray([self.byte_decoder[c] for c in text])
+            .decode("utf-8", errors="replace")
+            .replace("</w>", " ")
+        )
+        return text
+
+
+_tokenizer = SimpleTokenizer()
+
+
+def tokenize(
+    texts: Union[str, List[str]], context_length: int = 77
+) -> torch.LongTensor:
+    """
+    Returns the tokenized representation of given input string(s)
+
+    Parameters
+    ----------
+    texts : Union[str, List[str]]
+        An input string or a list of input strings to tokenize
+    context_length : int
+        The context length to use; all CLIP models use 77 as the context length
+
+    Returns
+    -------
+    A two-dimensional tensor containing the resulting tokens, shape = [number of input strings, context_length]
+    """
+    if isinstance(texts, str):
+        texts = [texts]
+
+    sot_token = _tokenizer.encoder["<start_of_text>"]
+    eot_token = _tokenizer.encoder["<end_of_text>"]
+    all_tokens = [[sot_token] + _tokenizer.encode(text) + [eot_token] for text in texts]
+    result = torch.zeros(len(all_tokens), context_length, dtype=torch.long)
+
+    for i, tokens in enumerate(all_tokens):
+        if len(tokens) > context_length:
+            tokens = tokens[:context_length]  # Truncate
+        result[i, : len(tokens)] = torch.tensor(tokens)
+
+    return result

audiosr/clap/open_clip/transform.py

@@ -1,45 +1,45 @@
-from torchvision.transforms import (
-    Normalize,
-    Compose,
-    RandomResizedCrop,
-    InterpolationMode,
-    ToTensor,
-    Resize,
-    CenterCrop,
-)
-
-
-def _convert_to_rgb(image):
-    return image.convert("RGB")
-
-
-def image_transform(
-    image_size: int,
-    is_train: bool,
-    mean=(0.48145466, 0.4578275, 0.40821073),
-    std=(0.26862954, 0.26130258, 0.27577711),
-):
-    normalize = Normalize(mean=mean, std=std)
-    if is_train:
-        return Compose(
-            [
-                RandomResizedCrop(
-                    image_size,
-                    scale=(0.9, 1.0),
-                    interpolation=InterpolationMode.BICUBIC,
-                ),
-                _convert_to_rgb,
-                ToTensor(),
-                normalize,
-            ]
-        )
-    else:
-        return Compose(
-            [
-                Resize(image_size, interpolation=InterpolationMode.BICUBIC),
-                CenterCrop(image_size),
-                _convert_to_rgb,
-                ToTensor(),
-                normalize,
-            ]
-        )
+from torchvision.transforms import (
+    Normalize,
+    Compose,
+    RandomResizedCrop,
+    InterpolationMode,
+    ToTensor,
+    Resize,
+    CenterCrop,
+)
+
+
+def _convert_to_rgb(image):
+    return image.convert("RGB")
+
+
+def image_transform(
+    image_size: int,
+    is_train: bool,
+    mean=(0.48145466, 0.4578275, 0.40821073),
+    std=(0.26862954, 0.26130258, 0.27577711),
+):
+    normalize = Normalize(mean=mean, std=std)
+    if is_train:
+        return Compose(
+            [
+                RandomResizedCrop(
+                    image_size,
+                    scale=(0.9, 1.0),
+                    interpolation=InterpolationMode.BICUBIC,
+                ),
+                _convert_to_rgb,
+                ToTensor(),
+                normalize,
+            ]
+        )
+    else:
+        return Compose(
+            [
+                Resize(image_size, interpolation=InterpolationMode.BICUBIC),
+                CenterCrop(image_size),
+                _convert_to_rgb,
+                ToTensor(),
+                normalize,
+            ]
+        )

audiosr/clap/open_clip/utils.py

@@ -1,355 +1,355 @@
-import numpy as np
-import torch
-from torch import nn as nn
-from torchvision.ops.misc import FrozenBatchNorm2d
-import logging
-from tqdm import tqdm
-import random
-import json
-import os
-import pathlib
-
-# TODO: (yusong) this not a good place to store those information and does not scale. Need to be fixed later.
-dataset_split = {
-    "audiocaps": ["train", "valid", "test"],
-    "audioset": ["balanced_train", "unbalanced_train", "eval"],
-    "BBCSoundEffects": ["train", "test"],
-    "Clotho": ["train", "test", "valid"],
-    "free_to_use_sounds": ["train", "test"],
-    "paramount_motion": ["train", "test"],
-    "sonniss_game_effects": ["train", "test"],
-    "wesoundeffects": ["train", "test"],
-    "MACS": ["train", "test"],
-    "freesound": ["train", "test"],
-    "FSD50K": ["train", "test", "valid"],
-    "fsd50k_class_label": ["train", "test", "valid"],
-    "esc50": ["train", "test"],
-    "audiostock": ["train", "test"],
-    "freesound_no_overlap_noesc50": ["train", "test"],
-    "epidemic_sound_effects": ["train", "test"],
-    "VGGSound": ["train", "test"],
-    "urbansound8k_class_label": ["train", "test"],
-    "audioset_t5": ["balanced_train", "unbalanced_train", "eval"],
-    "epidemic_sound_effects_t5": ["train", "test"],
-    "WavText5K": ["train", "test"],
-    "esc50_no_overlap": ["train", "test"],
-    "usd8k_no_overlap": ["train", "test"],
-    "fsd50k_200_class_label": ["train", "test", "valid"],
-}
-
-
-def freeze_batch_norm_2d(module, module_match={}, name=""):
-    """
-    Converts all `BatchNorm2d` and `SyncBatchNorm` layers of provided module into `FrozenBatchNorm2d`. If `module` is
-    itself an instance of either `BatchNorm2d` or `SyncBatchNorm`, it is converted into `FrozenBatchNorm2d` and
-    returned. Otherwise, the module is walked recursively and submodules are converted in place.
-
-    Args:
-        module (torch.nn.Module): Any PyTorch module.
-        module_match (dict): Dictionary of full module names to freeze (all if empty)
-        name (str): Full module name (prefix)
-
-    Returns:
-        torch.nn.Module: Resulting module
-
-    Inspired by https://github.com/pytorch/pytorch/blob/a5895f85be0f10212791145bfedc0261d364f103/torch/nn/modules/batchnorm.py#L762
-    """
-    res = module
-    is_match = True
-    if module_match:
-        is_match = name in module_match
-    if is_match and isinstance(
-        module, (nn.modules.batchnorm.BatchNorm2d, nn.modules.batchnorm.SyncBatchNorm)
-    ):
-        res = FrozenBatchNorm2d(module.num_features)
-        res.num_features = module.num_features
-        res.affine = module.affine
-        if module.affine:
-            res.weight.data = module.weight.data.clone().detach()
-            res.bias.data = module.bias.data.clone().detach()
-        res.running_mean.data = module.running_mean.data
-        res.running_var.data = module.running_var.data
-        res.eps = module.eps
-    else:
-        for child_name, child in module.named_children():
-            full_child_name = ".".join([name, child_name]) if name else child_name
-            new_child = freeze_batch_norm_2d(child, module_match, full_child_name)
-            if new_child is not child:
-                res.add_module(child_name, new_child)
-    return res
-
-
-def exist(dataset_name, dataset_type):
-    """
-    Check if dataset exists
-    """
-    if dataset_type in dataset_split[dataset_name]:
-        return True
-    else:
-        return False
-
-
-def get_tar_path_from_dataset_name(
-    dataset_names, dataset_types, islocal, dataset_path, proportion=1, full_dataset=None
-):
-    """
-    Get tar path from dataset name and type
-    """
-    output = []
-    for n in dataset_names:
-        if full_dataset is not None and n in full_dataset:
-            current_dataset_types = dataset_split[n]
-        else:
-            current_dataset_types = dataset_types
-        for s in current_dataset_types:
-            tmp = []
-            if islocal:
-                sizefilepath_ = f"{dataset_path}/{n}/{s}/sizes.json"
-                if not os.path.exists(sizefilepath_):
-                    sizefilepath_ = f"./json_files/{n}/{s}/sizes.json"
-            else:
-                sizefilepath_ = f"./json_files/{n}/{s}/sizes.json"
-            if not os.path.exists(sizefilepath_):
-                continue
-            sizes = json.load(open(sizefilepath_, "r"))
-            for k in sizes.keys():
-                if islocal:
-                    tmp.append(f"{dataset_path}/{n}/{s}/{k}")
-                else:
-                    tmp.append(
-                        f"pipe:aws s3 --cli-connect-timeout 0 cp s3://s-laion-audio/webdataset_tar/{n}/{s}/{k} -"
-                    )
-            if proportion != 1:
-                tmp = random.sample(tmp, int(proportion * len(tmp)))
-            output.append(tmp)
-    return sum(output, [])
-
-
-def get_tar_path_from_txts(txt_path, islocal, proportion=1):
-    """
-    Get tar path from txt path
-    """
-    if isinstance(txt_path, (list, tuple)):
-        return sum(
-            [
-                get_tar_path_from_txts(
-                    txt_path[i], islocal=islocal, proportion=proportion
-                )
-                for i in range(len(txt_path))
-            ],
-            [],
-        )
-    if isinstance(txt_path, str):
-        with open(txt_path) as f:
-            lines = f.readlines()
-        if islocal:
-            lines = [
-                lines[i]
-                .split("\n")[0]
-                .replace("pipe:aws s3 cp s3://s-laion-audio/", "/mnt/audio_clip/")
-                for i in range(len(lines))
-            ]
-        else:
-            lines = [
-                lines[i].split("\n")[0].replace(".tar", ".tar -")
-                for i in range(len(lines))
-            ]
-        if proportion != 1:
-            print("Sampling tars with proportion of {}".format(proportion))
-            lines = random.sample(lines, int(proportion * len(lines)))
-        return lines
-
-
-def get_mix_lambda(mixup_alpha, batch_size):
-    mixup_lambdas = [
-        np.random.beta(mixup_alpha, mixup_alpha, 1)[0] for _ in range(batch_size)
-    ]
-    return np.array(mixup_lambdas).astype(np.float32)
-
-
-def do_mixup(x, mixup_lambda):
-    """
-    Args:
-      x: (batch_size , ...)
-      mixup_lambda: (batch_size,)
-    Returns:
-      out: (batch_size, ...)
-    """
-    out = (
-        x.transpose(0, -1) * mixup_lambda
-        + torch.flip(x, dims=[0]).transpose(0, -1) * (1 - mixup_lambda)
-    ).transpose(0, -1)
-    return out
-
-
-def interpolate(x, ratio):
-    """Interpolate data in time domain. This is used to compensate the
-    resolution reduction in downsampling of a CNN.
-
-    Args:
-      x: (batch_size, time_steps, classes_num)
-      ratio: int, ratio to interpolate
-    Returns:
-      upsampled: (batch_size, time_steps * ratio, classes_num)
-    """
-    (batch_size, time_steps, classes_num) = x.shape
-    upsampled = x[:, :, None, :].repeat(1, 1, ratio, 1)
-    upsampled = upsampled.reshape(batch_size, time_steps * ratio, classes_num)
-    return upsampled
-
-
-def pad_framewise_output(framewise_output, frames_num):
-    """Pad framewise_output to the same length as input frames. The pad value
-    is the same as the value of the last frame.
-    Args:
-      framewise_output: (batch_size, frames_num, classes_num)
-      frames_num: int, number of frames to pad
-    Outputs:
-      output: (batch_size, frames_num, classes_num)
-    """
-    pad = framewise_output[:, -1:, :].repeat(
-        1, frames_num - framewise_output.shape[1], 1
-    )
-    """tensor for padding"""
-
-    output = torch.cat((framewise_output, pad), dim=1)
-    """(batch_size, frames_num, classes_num)"""
-
-
-# def process_ipc(index_path, classes_num, filename):
-#     # load data
-#     logging.info("Load Data...............")
-#     ipc = [[] for _ in range(classes_num)]
-#     with h5py.File(index_path, "r") as f:
-#         for i in tqdm(range(len(f["target"]))):
-#             t_class = np.where(f["target"][i])[0]
-#             for t in t_class:
-#                 ipc[t].append(i)
-#     print(ipc)
-#     np.save(filename, ipc)
-#     logging.info("Load Data Succeed...............")
-
-
-def save_to_dict(s, o_={}):
-    sp = s.split(": ")
-    o_.update({sp[0]: float(sp[1])})
-    return o_
-
-
-def get_data_from_log(txt_path):
-    """
-    Output dictionary from out.txt log file
-    """
-    with open(txt_path) as f:
-        lines = f.readlines()
-    val_data = {}
-    train_data = {}
-    train_losses = []
-    train_losses_epoch = []
-    for i in range(len(lines)):
-        if "| INFO |" in lines[i]:
-            if "Eval Epoch" in lines[i]:
-                if "val_loss" in lines[i]:
-                    # float(regex.sub("", lines[310].split("	")[-1]).replace(" ", ""))
-                    line = lines[i].split("Eval Epoch: ")[-1]
-                    num_epoch = int(line.split("	")[0].split(" ")[0])
-                    d = {
-                        line.split("	")[0]
-                        .split(" ")[1]
-                        .replace(":", ""): float(line.split("	")[0].split(" ")[-1])
-                    }
-                    for i in range(1, len(line.split("	"))):
-                        d = save_to_dict(line.split("	")[i], d)
-                    val_data[num_epoch] = d
-            elif "Train Epoch" in lines[i]:
-                num_epoch = int(lines[i].split("Train Epoch: ")[1][0])
-                loss = float(lines[i].split("Loss: ")[-1].split(" (")[0])
-                train_losses.append(loss)
-                train_losses_epoch.append(num_epoch)
-    for i in range(len(train_losses)):
-        train_data[i] = {
-            "num_epoch": train_losses_epoch[i],
-            "train_loss": train_losses[i],
-        }
-    return train_data, val_data
-
-
-def save_p(obj, filename):
-    import pickle
-
-    try:
-        from deepdiff import DeepDiff
-    except:
-        os.system("pip install deepdiff")
-        from deepdiff import DeepDiff
-    with open(filename, "wb") as file:
-        pickle.dump(obj, file, protocol=pickle.HIGHEST_PROTOCOL)  # highest protocol
-    with open(filename, "rb") as file:
-        z = pickle.load(file)
-    assert (
-        DeepDiff(obj, z, ignore_string_case=True) == {}
-    ), "there is something wrong with the saving process"
-    return
-
-
-def load_p(filename):
-    import pickle
-
-    with open(filename, "rb") as file:
-        z = pickle.load(file)
-    return z
-
-
-def save_json(data, name="data.json"):
-    import json
-
-    with open(name, "w") as fp:
-        json.dump(data, fp)
-    return
-
-
-def load_json(name):
-    import json
-
-    with open(name, "r") as fp:
-        data = json.load(fp)
-    return data
-
-
-def load_class_label(path):
-    # https://stackoverflow.com/questions/48004243/how-to-share-large-read-only-dictionary-list-across-processes-in-multiprocessing
-    # https://stackoverflow.com/questions/45693949/storing-strings-in-a-multiprocessing-sharedctypes-array
-    out = None
-    if path is not None:
-        if pathlib.Path(path).suffix in [".pkl", ".pickle"]:
-            out = load_p(path)
-        elif pathlib.Path(path).suffix in [".json", ".txt"]:
-            out = load_json(path)
-        elif pathlib.Path(path).suffix in [".npy", ".npz"]:
-            out = np.load(path)
-        elif pathlib.Path(path).suffix in [".csv"]:
-            import pandas as pd
-
-            out = pd.read_csv(path)
-    return out
-    # if out is None:
-    #     return None
-    # else:
-    #     key = Array(c_wchar, '\n'.join(list(out.keys())), lock=False)
-    #     val = Array('i', out.values(), lock=False)
-    #     return (key, val)
-
-
-from torch import optim
-
-
-def get_optimizer(params, lr, betas, eps, momentum, optimizer_name):
-    if optimizer_name.lower() == "adamw":
-        optimizer = optim.AdamW(params, lr=lr, betas=betas, eps=eps)
-    elif optimizer_name.lower() == "sgd":
-        optimizer = optim.SGD(params, lr=lr, momentum=momentum)
-    elif optimizer_name.lower() == "adam":
-        optimizer = optim.Adam(params, lr=lr, betas=betas, eps=eps)
-    else:
-        raise ValueError("optimizer name is not correct")
-    return optimizer
+import numpy as np
+import torch
+from torch import nn as nn
+from torchvision.ops.misc import FrozenBatchNorm2d
+import logging
+from tqdm import tqdm
+import random
+import json
+import os
+import pathlib
+
+# TODO: (yusong) this not a good place to store those information and does not scale. Need to be fixed later.
+dataset_split = {
+    "audiocaps": ["train", "valid", "test"],
+    "audioset": ["balanced_train", "unbalanced_train", "eval"],
+    "BBCSoundEffects": ["train", "test"],
+    "Clotho": ["train", "test", "valid"],
+    "free_to_use_sounds": ["train", "test"],
+    "paramount_motion": ["train", "test"],
+    "sonniss_game_effects": ["train", "test"],
+    "wesoundeffects": ["train", "test"],
+    "MACS": ["train", "test"],
+    "freesound": ["train", "test"],
+    "FSD50K": ["train", "test", "valid"],
+    "fsd50k_class_label": ["train", "test", "valid"],
+    "esc50": ["train", "test"],
+    "audiostock": ["train", "test"],
+    "freesound_no_overlap_noesc50": ["train", "test"],
+    "epidemic_sound_effects": ["train", "test"],
+    "VGGSound": ["train", "test"],
+    "urbansound8k_class_label": ["train", "test"],
+    "audioset_t5": ["balanced_train", "unbalanced_train", "eval"],
+    "epidemic_sound_effects_t5": ["train", "test"],
+    "WavText5K": ["train", "test"],
+    "esc50_no_overlap": ["train", "test"],
+    "usd8k_no_overlap": ["train", "test"],
+    "fsd50k_200_class_label": ["train", "test", "valid"],
+}
+
+
+def freeze_batch_norm_2d(module, module_match={}, name=""):
+    """
+    Converts all `BatchNorm2d` and `SyncBatchNorm` layers of provided module into `FrozenBatchNorm2d`. If `module` is
+    itself an instance of either `BatchNorm2d` or `SyncBatchNorm`, it is converted into `FrozenBatchNorm2d` and
+    returned. Otherwise, the module is walked recursively and submodules are converted in place.
+
+    Args:
+        module (torch.nn.Module): Any PyTorch module.
+        module_match (dict): Dictionary of full module names to freeze (all if empty)
+        name (str): Full module name (prefix)
+
+    Returns:
+        torch.nn.Module: Resulting module
+
+    Inspired by https://github.com/pytorch/pytorch/blob/a5895f85be0f10212791145bfedc0261d364f103/torch/nn/modules/batchnorm.py#L762
+    """
+    res = module
+    is_match = True
+    if module_match:
+        is_match = name in module_match
+    if is_match and isinstance(
+        module, (nn.modules.batchnorm.BatchNorm2d, nn.modules.batchnorm.SyncBatchNorm)
+    ):
+        res = FrozenBatchNorm2d(module.num_features)
+        res.num_features = module.num_features
+        res.affine = module.affine
+        if module.affine:
+            res.weight.data = module.weight.data.clone().detach()
+            res.bias.data = module.bias.data.clone().detach()
+        res.running_mean.data = module.running_mean.data
+        res.running_var.data = module.running_var.data
+        res.eps = module.eps
+    else:
+        for child_name, child in module.named_children():
+            full_child_name = ".".join([name, child_name]) if name else child_name
+            new_child = freeze_batch_norm_2d(child, module_match, full_child_name)
+            if new_child is not child:
+                res.add_module(child_name, new_child)
+    return res
+
+
+def exist(dataset_name, dataset_type):
+    """
+    Check if dataset exists
+    """
+    if dataset_type in dataset_split[dataset_name]:
+        return True
+    else:
+        return False
+
+
+def get_tar_path_from_dataset_name(
+    dataset_names, dataset_types, islocal, dataset_path, proportion=1, full_dataset=None
+):
+    """
+    Get tar path from dataset name and type
+    """
+    output = []
+    for n in dataset_names:
+        if full_dataset is not None and n in full_dataset:
+            current_dataset_types = dataset_split[n]
+        else:
+            current_dataset_types = dataset_types
+        for s in current_dataset_types:
+            tmp = []
+            if islocal:
+                sizefilepath_ = f"{dataset_path}/{n}/{s}/sizes.json"
+                if not os.path.exists(sizefilepath_):
+                    sizefilepath_ = f"./json_files/{n}/{s}/sizes.json"
+            else:
+                sizefilepath_ = f"./json_files/{n}/{s}/sizes.json"
+            if not os.path.exists(sizefilepath_):
+                continue
+            sizes = json.load(open(sizefilepath_, "r"))
+            for k in sizes.keys():
+                if islocal:
+                    tmp.append(f"{dataset_path}/{n}/{s}/{k}")
+                else:
+                    tmp.append(
+                        f"pipe:aws s3 --cli-connect-timeout 0 cp s3://s-laion-audio/webdataset_tar/{n}/{s}/{k} -"
+                    )
+            if proportion != 1:
+                tmp = random.sample(tmp, int(proportion * len(tmp)))
+            output.append(tmp)
+    return sum(output, [])
+
+
+def get_tar_path_from_txts(txt_path, islocal, proportion=1):
+    """
+    Get tar path from txt path
+    """
+    if isinstance(txt_path, (list, tuple)):
+        return sum(
+            [
+                get_tar_path_from_txts(
+                    txt_path[i], islocal=islocal, proportion=proportion
+                )
+                for i in range(len(txt_path))
+            ],
+            [],
+        )
+    if isinstance(txt_path, str):
+        with open(txt_path) as f:
+            lines = f.readlines()
+        if islocal:
+            lines = [
+                lines[i]
+                .split("\n")[0]
+                .replace("pipe:aws s3 cp s3://s-laion-audio/", "/mnt/audio_clip/")
+                for i in range(len(lines))
+            ]
+        else:
+            lines = [
+                lines[i].split("\n")[0].replace(".tar", ".tar -")
+                for i in range(len(lines))
+            ]
+        if proportion != 1:
+            print("Sampling tars with proportion of {}".format(proportion))
+            lines = random.sample(lines, int(proportion * len(lines)))
+        return lines
+
+
+def get_mix_lambda(mixup_alpha, batch_size):
+    mixup_lambdas = [
+        np.random.beta(mixup_alpha, mixup_alpha, 1)[0] for _ in range(batch_size)
+    ]
+    return np.array(mixup_lambdas).astype(np.float32)
+
+
+def do_mixup(x, mixup_lambda):
+    """
+    Args:
+      x: (batch_size , ...)
+      mixup_lambda: (batch_size,)
+    Returns:
+      out: (batch_size, ...)
+    """
+    out = (
+        x.transpose(0, -1) * mixup_lambda
+        + torch.flip(x, dims=[0]).transpose(0, -1) * (1 - mixup_lambda)
+    ).transpose(0, -1)
+    return out
+
+
+def interpolate(x, ratio):
+    """Interpolate data in time domain. This is used to compensate the
+    resolution reduction in downsampling of a CNN.
+
+    Args:
+      x: (batch_size, time_steps, classes_num)
+      ratio: int, ratio to interpolate
+    Returns:
+      upsampled: (batch_size, time_steps * ratio, classes_num)
+    """
+    (batch_size, time_steps, classes_num) = x.shape
+    upsampled = x[:, :, None, :].repeat(1, 1, ratio, 1)
+    upsampled = upsampled.reshape(batch_size, time_steps * ratio, classes_num)
+    return upsampled
+
+
+def pad_framewise_output(framewise_output, frames_num):
+    """Pad framewise_output to the same length as input frames. The pad value
+    is the same as the value of the last frame.
+    Args:
+      framewise_output: (batch_size, frames_num, classes_num)
+      frames_num: int, number of frames to pad
+    Outputs:
+      output: (batch_size, frames_num, classes_num)
+    """
+    pad = framewise_output[:, -1:, :].repeat(
+        1, frames_num - framewise_output.shape[1], 1
+    )
+    """tensor for padding"""
+
+    output = torch.cat((framewise_output, pad), dim=1)
+    """(batch_size, frames_num, classes_num)"""
+
+
+# def process_ipc(index_path, classes_num, filename):
+#     # load data
+#     logging.info("Load Data...............")
+#     ipc = [[] for _ in range(classes_num)]
+#     with h5py.File(index_path, "r") as f:
+#         for i in tqdm(range(len(f["target"]))):
+#             t_class = np.where(f["target"][i])[0]
+#             for t in t_class:
+#                 ipc[t].append(i)
+#     print(ipc)
+#     np.save(filename, ipc)
+#     logging.info("Load Data Succeed...............")
+
+
+def save_to_dict(s, o_={}):
+    sp = s.split(": ")
+    o_.update({sp[0]: float(sp[1])})
+    return o_
+
+
+def get_data_from_log(txt_path):
+    """
+    Output dictionary from out.txt log file
+    """
+    with open(txt_path) as f:
+        lines = f.readlines()
+    val_data = {}
+    train_data = {}
+    train_losses = []
+    train_losses_epoch = []
+    for i in range(len(lines)):
+        if "| INFO |" in lines[i]:
+            if "Eval Epoch" in lines[i]:
+                if "val_loss" in lines[i]:
+                    # float(regex.sub("", lines[310].split("	")[-1]).replace(" ", ""))
+                    line = lines[i].split("Eval Epoch: ")[-1]
+                    num_epoch = int(line.split("	")[0].split(" ")[0])
+                    d = {
+                        line.split("	")[0]
+                        .split(" ")[1]
+                        .replace(":", ""): float(line.split("	")[0].split(" ")[-1])
+                    }
+                    for i in range(1, len(line.split("	"))):
+                        d = save_to_dict(line.split("	")[i], d)
+                    val_data[num_epoch] = d
+            elif "Train Epoch" in lines[i]:
+                num_epoch = int(lines[i].split("Train Epoch: ")[1][0])
+                loss = float(lines[i].split("Loss: ")[-1].split(" (")[0])
+                train_losses.append(loss)
+                train_losses_epoch.append(num_epoch)
+    for i in range(len(train_losses)):
+        train_data[i] = {
+            "num_epoch": train_losses_epoch[i],
+            "train_loss": train_losses[i],
+        }
+    return train_data, val_data
+
+
+def save_p(obj, filename):
+    import pickle
+
+    try:
+        from deepdiff import DeepDiff
+    except:
+        os.system("pip install deepdiff")
+        from deepdiff import DeepDiff
+    with open(filename, "wb") as file:
+        pickle.dump(obj, file, protocol=pickle.HIGHEST_PROTOCOL)  # highest protocol
+    with open(filename, "rb") as file:
+        z = pickle.load(file)
+    assert (
+        DeepDiff(obj, z, ignore_string_case=True) == {}
+    ), "there is something wrong with the saving process"
+    return
+
+
+def load_p(filename):
+    import pickle
+
+    with open(filename, "rb") as file:
+        z = pickle.load(file)
+    return z
+
+
+def save_json(data, name="data.json"):
+    import json
+
+    with open(name, "w") as fp:
+        json.dump(data, fp)
+    return
+
+
+def load_json(name):
+    import json
+
+    with open(name, "r") as fp:
+        data = json.load(fp)
+    return data
+
+
+def load_class_label(path):
+    # https://stackoverflow.com/questions/48004243/how-to-share-large-read-only-dictionary-list-across-processes-in-multiprocessing
+    # https://stackoverflow.com/questions/45693949/storing-strings-in-a-multiprocessing-sharedctypes-array
+    out = None
+    if path is not None:
+        if pathlib.Path(path).suffix in [".pkl", ".pickle"]:
+            out = load_p(path)
+        elif pathlib.Path(path).suffix in [".json", ".txt"]:
+            out = load_json(path)
+        elif pathlib.Path(path).suffix in [".npy", ".npz"]:
+            out = np.load(path)
+        elif pathlib.Path(path).suffix in [".csv"]:
+            import pandas as pd
+
+            out = pd.read_csv(path)
+    return out
+    # if out is None:
+    #     return None
+    # else:
+    #     key = Array(c_wchar, '\n'.join(list(out.keys())), lock=False)
+    #     val = Array('i', out.values(), lock=False)
+    #     return (key, val)
+
+
+from torch import optim
+
+
+def get_optimizer(params, lr, betas, eps, momentum, optimizer_name):
+    if optimizer_name.lower() == "adamw":
+        optimizer = optim.AdamW(params, lr=lr, betas=betas, eps=eps)
+    elif optimizer_name.lower() == "sgd":
+        optimizer = optim.SGD(params, lr=lr, momentum=momentum)
+    elif optimizer_name.lower() == "adam":
+        optimizer = optim.Adam(params, lr=lr, betas=betas, eps=eps)
+    else:
+        raise ValueError("optimizer name is not correct")
+    return optimizer

audiosr/clap/training/data.py

@@ -1,865 +1,865 @@
-import json
-import logging
-import os
-import random
-from dataclasses import dataclass
-import numpy as np
-import pandas as pd
-import torch
-import torchvision.datasets as datasets
-from PIL import Image
-from torch.utils.data import Dataset, DataLoader, SubsetRandomSampler
-from torch.utils.data.distributed import DistributedSampler
-import soundfile as sf
-import io
-from pathlib import Path
-
-# import wget
-
-from audiosr.clap.open_clip.utils import get_tar_path_from_dataset_name
-from audiosr.clap.open_clip.utils import load_class_label
-
-try:
-    import horovod.torch as hvd
-except ImportError:
-    hvd = None
-
-try:
-    import torchaudio
-except ImportError:
-    torchaudio = None
-
-from audiosr.clap.open_clip import tokenize
-
-
-def tokenizer(text):
-    return tokenize(text).squeeze(0)
-
-
-from transformers import RobertaTokenizer
-
-tokenize = RobertaTokenizer.from_pretrained("roberta-base")
-
-
-def tokenizer(text):
-    result = tokenize(
-        text,
-        padding="max_length",
-        truncation=True,
-        max_length=77,
-        return_tensors="pt",
-    )
-    return {k: v.squeeze(0) for k, v in result.items()}
-
-
-# initizlied the audioset map
-_AUDIOSET_MAP_PATH = os.path.join(Path(__file__).parent, "audioset_textmap.npy")
-_AUDIOSET_MAP = np.load(_AUDIOSET_MAP_PATH, allow_pickle=True)
-
-
-def int16_to_float32(x):
-    return (x / 32767.0).astype(np.float32)
-
-
-def float32_to_int16(x):
-    x = np.clip(x, a_min=-1.0, a_max=1.0)
-    return (x * 32767.0).astype(np.int16)
-
-
-# For Toy Dataset
-# class ToyDataset(Dataset):
-#     def __init__(self, index_path, ipc, config, eval_mode=False):
-#         """Toy Dataset for testing the audioset input with text labels
-#         Parameters
-#         ----------
-#             index_path: str
-#                 the link to the h5 file of each audio
-#             idc: str
-#                 the link to the npy file, the number of samples in each class
-#             config: dict
-#                 the audio cfg file
-#            eval_model (bool): to indicate if the dataset is a testing dataset
-#         """
-#         self.audio_cfg = config["audio_cfg"]
-#         self.text_cfg = config["text_cfg"]
-#         self.fp = h5py.File(index_path, "r")
-#         self.ipc = np.load(ipc, allow_pickle=True)
-#         self.total_size = len(self.fp["audio_name"])
-#         self.classes_num = self.audio_cfg["class_num"]
-#         self.eval_mode = eval_mode
-
-#         if not eval_mode:
-#             self.generate_queue()
-#         else:
-#             self.queue = []
-#             for i in range(self.total_size):
-#                 target = self.fp["target"][i]
-#                 if np.sum(target) > 0:
-#                     self.queue.append(i)
-#             self.total_size = len(self.queue)
-#         logging.info("total dataset size: %d" % (self.total_size))
-#         logging.info("class num: %d" % (self.classes_num))
-
-#     def time_shifting(self, x):
-#         frame_num = len(x)
-#         shift_len = random.randint(0, frame_num - 1)
-#         new_sample = np.concatenate([x[shift_len:], x[:shift_len]], axis=0)
-#         return new_sample
-
-#     def generate_queue(self):
-#         self.queue = []
-#         while len(self.queue) < self.total_size:
-#             class_set = [*range(self.classes_num)]
-#             random.shuffle(class_set)
-#             self.queue += [
-#                 self.ipc[d][random.randint(0, len(self.ipc[d]) - 1)] for d in class_set
-#             ]
-#         self.queue = self.queue[: self.total_size]
-
-#         logging.info("queue regenerated:%s" % (self.queue[-5:]))
-
-#     def crop_wav(self, x):
-#         crop_size = self.audio_cfg["crop_size"]
-#         crop_pos = random.randint(0, len(x) - crop_size - 1)
-#         return x[crop_pos : crop_pos + crop_size]
-
-#     def prompt_text(self, target):
-#         events = _AUDIOSET_MAP[np.where(target > 0)]
-#         event_text = "The sounds of " + ", ".join(events[:-1]) + " and " + events[-1]
-#         text = tokenize(event_text)[0]
-#         return text
-
-#     def __getitem__(self, index):
-#         """Load waveform, text, and target of an audio clip
-
-#         Parameters
-#         ----------
-#             index: int
-#                 the index number
-#         Return
-#         ------
-#             output: dict {
-#                 "hdf5_path": str,
-#                 "index_in_hdf5": int,
-#                 "audio_name": str,
-#                 "waveform": list (audio_length,),
-#                 "target": list (class_num, ),
-#                 "text": torch.tensor (context_length,)
-#             }
-#                 the output dictionary
-#         """
-#         s_index = self.queue[index]
-
-#         audio_name = self.fp["audio_name"][s_index].decode()
-#         # Hardcode here CHANGE
-#         hdf5_path = (
-#             self.fp["hdf5_path"][s_index]
-#             .decode()
-#             .replace(
-#                 "../workspace",
-#                 "/home/la/kechen/Research/ke_zsasp/workspace",
-#             )
-#         )
-#         r_idx = self.fp["index_in_hdf5"][s_index]
-#         target = self.fp["target"][s_index].astype(np.float32)
-#         text = self.prompt_text(target)
-#         with h5py.File(hdf5_path, "r") as f:
-#             waveform = int16_to_float32(f["waveform"][r_idx])[
-#                 : self.audio_cfg["clip_samples"]
-#             ]
-#         assert (
-#             len(waveform) == self.audio_cfg["clip_samples"]
-#         ), "The sample length is not match"
-#         # Time shift
-#         # if (self.config.enable_time_shift) and (not self.eval_mode):
-#         #     waveform = self.time_shifting(waveform)
-#         # # Label Enhance
-#         # if (self.config.crop_size is not None) and (not self.eval_mode):
-#         #     waveform = self.crop_wav(waveform)
-#         # # the label enhance rate is fixed 0.5
-#         # if (self.config.enable_label_enhance) and (not self.eval_mode) and random.random() < 0.5:
-#         #     kidx = np.where(target)[0]
-#         #     for k in kidx:
-#         #         for add_key in self.class_map[k][1]:
-#         #             target[add_key] = 1.0
-#         #         if len(self.class_map[k][2]) > 0:
-#         #             add_key = random.choice(self.class_map[k][2])
-#         #             target[add_key] = 1.0
-
-#         # missing the text input
-#         mel_spec = get_mel(torch.from_numpy(waveform), self.audio_cfg)[None, :, :]
-#         mel_spec = (
-#             torch.cat(
-#                 [mel_spec, mel_spec.clone(), mel_spec.clone(), mel_spec.clone()], dim=0
-#             )
-#             .cpu()
-#             .numpy()
-#         )
-#         longer = random.choice([True, False])
-#         if longer == False:
-#             mel_spec[1:, :, :] = 0.0
-#         data_dict = {
-#             "hdf5_path": hdf5_path,
-#             "index_in_hdf5": r_idx,
-#             "audio_name": audio_name,
-#             "waveform": waveform,
-#             "class_label": target,
-#             "text": text,
-#             "longer": longer,
-#             "mel_fusion": mel_spec,
-#         }
-#         return data_dict
-
-#     def __len__(self):
-#         return self.total_size
-
-
-class CsvDataset(Dataset):
-    def __init__(self, input_filename, transforms, img_key, caption_key, sep="\t"):
-        logging.debug(f"Loading csv data from {input_filename}.")
-        df = pd.read_csv(input_filename, sep=sep)
-
-        self.images = df[img_key].tolist()
-        self.captions = df[caption_key].tolist()
-        self.transforms = transforms
-        logging.debug("Done loading data.")
-
-    def __len__(self):
-        return len(self.captions)
-
-    def __getitem__(self, idx):
-        images = self.transforms(Image.open(str(self.images[idx])))
-        texts = tokenize([str(self.captions[idx])])[0]
-        return images, texts
-
-
-@dataclass
-class DataInfo:
-    dataloader: DataLoader
-    sampler: DistributedSampler
-
-
-def preprocess_txt(text):
-    return tokenize([str(text)])[0]
-
-
-# def get_dataset_size(shards, sizefilepath_=None, is_local=True):
-#     if isinstance(shards, list):
-#         size_list = []
-#         for s in shards:
-#             size_list.append(
-#                 get_dataset_size(s, sizefilepath_=sizefilepath_, is_local=is_local)[0]
-#             )
-#     else:
-#         if not is_local:
-#             for n in dataset_split.keys():
-#                 if n in shards.split("/"):
-#                     break
-#             for s in dataset_split[n]:
-#                 if s in shards.split("/"):
-#                     break
-#             sizefilepath_ = f"./json_files/{n}/{s}/sizes.json"
-#         shards_list = list(braceexpand.braceexpand(shards))
-#         dir_path = os.path.dirname(shards)
-#         if sizefilepath_ is not None:
-#             sizes = json.load(open(sizefilepath_, "r"))
-#             total_size = sum(
-#                 [
-#                     int(sizes[os.path.basename(shard.replace(".tar -", ".tar"))])
-#                     for shard in shards_list
-#                 ]
-#             )
-#         else:
-#             sizes_filename = os.path.join(dir_path, "sizes.json")
-#             len_filename = os.path.join(dir_path, "__len__")
-#             if os.path.exists(sizes_filename):
-#                 sizes = json.load(open(sizes_filename, "r"))
-#                 total_size = sum(
-#                     [int(sizes[os.path.basename(shard)]) for shard in shards_list]
-#                 )
-#             elif os.path.exists(len_filename):
-#                 # FIXME this used to be eval(open(...)) but that seemed rather unsafe
-#                 total_size = ast.literal_eval(open(len_filename, "r").read())
-#             else:
-#                 raise Exception(
-#                     "Cannot find sizes file for dataset. Please specify the path to the file."
-#                 )
-#                 # total_size = None  # num samples undefined
-#                 # some common dataset sizes (at time of authors last download)
-#                 # cc3m-train: 2905954
-#                 # cc12m: 10968539
-#                 # LAION-400m: 407332084
-#         num_shards = len(shards_list)
-#     if isinstance(shards, list):
-#         return sum(size_list), len(shards)
-#     else:
-#         return total_size, num_shards
-
-
-def get_imagenet(args, preprocess_fns, split):
-    assert split in ["train", "val", "v2"]
-    is_train = split == "train"
-    preprocess_train, preprocess_val = preprocess_fns
-
-    if split == "v2":
-        from imagenetv2_pytorch import ImageNetV2Dataset
-
-        dataset = ImageNetV2Dataset(location=args.imagenet_v2, transform=preprocess_val)
-    else:
-        if is_train:
-            data_path = args.imagenet_train
-            preprocess_fn = preprocess_train
-        else:
-            data_path = args.imagenet_val
-            preprocess_fn = preprocess_val
-        assert data_path
-
-        dataset = datasets.ImageFolder(data_path, transform=preprocess_fn)
-
-    if is_train:
-        idxs = np.zeros(len(dataset.targets))
-        target_array = np.array(dataset.targets)
-        k = 50
-        for c in range(1000):
-            m = target_array == c
-            n = len(idxs[m])
-            arr = np.zeros(n)
-            arr[:k] = 1
-            np.random.shuffle(arr)
-            idxs[m] = arr
-
-        idxs = idxs.astype("int")
-        sampler = SubsetRandomSampler(np.where(idxs)[0])
-    else:
-        sampler = None
-
-    dataloader = torch.utils.data.DataLoader(
-        dataset,
-        batch_size=args.batch_size,
-        num_workers=args.workers,
-        sampler=sampler,
-    )
-
-    return DataInfo(dataloader, sampler)
-
-
-def count_samples(dataloader):
-    os.environ["WDS_EPOCH"] = "0"
-    n_elements, n_batches = 0, 0
-    for images, texts in dataloader:
-        n_batches += 1
-        n_elements += len(images)
-        assert len(images) == len(texts)
-    return n_elements, n_batches
-
-
-def filter_no_caption(sample):
-    return "txt" in sample
-
-
-def log_and_continue(exn):
-    """Call in an exception handler to ignore any exception, isssue a warning, and continue."""
-    logging.warning(f"Handling webdataset error ({repr(exn)}). Ignoring.")
-    return True
-
-
-_SHARD_SHUFFLE_SIZE = 2000
-_SHARD_SHUFFLE_INITIAL = 500
-_SAMPLE_SHUFFLE_SIZE = 5000
-_SAMPLE_SHUFFLE_INITIAL = 1000
-
-
-# def sample_prop(sizefile, inputs, proportion, is_local=True):
-#     """
-#     Sample a proportion of the data.
-#     """
-#     file_path_dict = {
-#         os.path.split(inputs[i])[1]: os.path.split(inputs[i])[0]
-#         for i in range(len(inputs))
-#     }
-#     sampled_filepath_dict = {}
-#     sampled_size_dict = {}
-#     if not is_local:
-#         if os.path.exists("sizes.json"):
-#             os.remove("sizes.json")
-#         wget.download(sizefile, "sizes.json")
-#         sizefile = "sizes.json"
-#     with open(sizefile, "r", encoding="UTF-8") as f:
-#         load_dict = json.load(f)
-#     L = int(len(file_path_dict) * proportion)
-#     subkeys = random.sample(file_path_dict.keys(), L)
-#     for k in subkeys:
-#         sampled_size_dict[k] = load_dict[k]
-#         sampled_filepath_dict[k] = file_path_dict[k]
-#     return (
-#         sum(sampled_size_dict.values()),
-#         L,
-#         [os.path.join(v, k) for k, v in sampled_filepath_dict.items()],
-#         sampled_size_dict,
-#     )
-
-
-def get_mel(audio_data, audio_cfg):
-    # mel shape: (n_mels, T)
-    mel = torchaudio.transforms.MelSpectrogram(
-        sample_rate=audio_cfg["sample_rate"],
-        n_fft=audio_cfg["window_size"],
-        win_length=audio_cfg["window_size"],
-        hop_length=audio_cfg["hop_size"],
-        center=True,
-        pad_mode="reflect",
-        power=2.0,
-        norm=None,
-        onesided=True,
-        n_mels=64,
-        f_min=audio_cfg["fmin"],
-        f_max=audio_cfg["fmax"],
-    ).to(audio_data.device)
-    mel = mel(audio_data)
-    # we use log mel spectrogram as input
-    mel = torchaudio.transforms.AmplitudeToDB(top_db=None)(mel)
-    return mel.T  # (T, n_mels)
-
-
-def get_audio_features(
-    audio_data, mel, max_len, data_truncating, data_filling, audio_cfg
-):
-    """
-    Calculate and add audio features to sample.
-    Sample: a dict containing all the data of current sample.
-    audio_data: a tensor of shape (T) containing audio data.
-    max_len: the maximum length of audio data.
-    data_truncating: the method of truncating data.
-    data_filling: the method of filling data.
-    audio_cfg: a dict containing audio configuration. Comes from model_cfg['audio_cfg'].
-    """
-    sample = {}
-
-    # assert audio_data.size(-1) <= max_len, str(audio_data.size())
-
-    # split to three parts
-    chunk_frames = (
-        max_len // audio_cfg["hop_size"] + 1
-    )  # the +1 related to how the spectrogram is computed
-    mel = mel[:chunk_frames]
-
-    audio_data = audio_data[..., :max_len]
-    sample["mel_fusion"] = mel
-    longer = torch.tensor([True])
-
-    sample["longer"] = longer
-    sample["waveform"] = audio_data
-
-    return sample
-
-
-def preprocess(
-    sample,
-    audio_ext,
-    text_ext,
-    max_len,
-    audio_cfg,
-    class_index_dict=None,
-    data_filling="pad",
-    data_truncating="rand_trunc",
-    text_augment_selection=None,
-):
-    """
-    Preprocess a single sample for wdsdataloader.
-    """
-    audio_data, orig_sr = sf.read(io.BytesIO(sample[audio_ext]))
-    audio_data = int16_to_float32(float32_to_int16(audio_data))
-    audio_data = torch.tensor(audio_data).float()
-
-    # TODO: (yusong) to be include in the future
-    # # if torchaudio not installed, use soundfile to load audio
-    # if torchaudio is None:
-    #     audio_data, orig_sr = sf.read(io.BytesIO(sample[audio_ext]))
-    #     audio_data = torch.tensor(audio_data).float()
-    # else:
-    #     # https://github.com/webdataset/webdataset/blob/main/webdataset/autodecode.py
-    #     with tempfile.TemporaryDirectory() as dirname:
-    #         os.makedirs(dirname, exist_ok=True)
-    #         fname = os.path.join(dirname, f"file.flac")
-    #         with open(fname, "wb") as stream:
-    #             stream.write(sample[audio_ext])
-    #         audio_data, orig_sr = torchaudio.load(fname)
-    #         audio_data = audio_data[0, :].float()
-
-    sample = get_audio_features(
-        sample, audio_data, max_len, data_truncating, data_filling, audio_cfg
-    )
-    del sample[audio_ext]
-
-    try:
-        json_dict_raw = json.loads(sample[text_ext].decode("utf-8"))
-    except:
-        print("sample[__url__]:", sample["__url__"])
-
-    # For selecting augmented text from dataset
-    if text_augment_selection is None or text_augment_selection == "none":
-        texts = json_dict_raw["text"]
-    elif text_augment_selection == "all":
-        if "text_augment_all" in json_dict_raw.keys():
-            texts = json_dict_raw["text_augment_all"]
-        else:
-            texts = json_dict_raw["text"]
-    elif text_augment_selection == "augment_only":
-        if "text_augment_all" in json_dict_raw.keys():
-            if json_dict_raw["text_augment_t5"] is None:
-                texts = json_dict_raw["text"]
-            else:
-                texts = json_dict_raw["text_augment_t5"]
-        else:
-            texts = json_dict_raw["text"]
-    else:
-        raise NotImplementedError(
-            f"text_augment_selection {text_augment_selection} not implemented"
-        )
-    sample["full_text"] = texts
-
-    if isinstance(texts, list) and isinstance(texts[0], str) and len(texts) > 1:
-        texts = random.choice(texts)
-    sample["raw_text"] = texts
-    sample["text"] = tokenizer(texts)  # text shape: [num_token]
-    if class_index_dict is not None:
-        # https://stackoverflow.com/questions/48004243/how-to-share-large-read-only-dictionary-list-across-processes-in-multiprocessing
-        # https://stackoverflow.com/questions/45693949/storing-strings-in-a-multiprocessing-sharedctypes-array
-        # key, val = class_index_dict
-        # key = key[:].split('\n')
-        # _dict = {k: v for k, v in zip(key, val)}
-        sample["class_label"] = np.zeros(len(class_index_dict.keys()))
-        for x in json_dict_raw["tag"]:
-            sample["class_label"][class_index_dict[x]] = 1
-        sample["class_label"] = torch.tensor(sample["class_label"]).float()
-    del sample[text_ext]
-    sample["audio_name"] = sample["__key__"].split("/")[-1] + "." + audio_ext
-    sample["text_name"] = sample["__key__"].split("/")[-1] + "." + text_ext
-    sample["audio_orig_sr"] = orig_sr
-    return sample
-
-
-def collate_fn(batch):
-    """
-    Collate function for wdsdataloader.
-    batch: a list of dict, each dict is a sample
-    """
-    # concatenate values in each dictionary. if it is a tensor, concatenate. if it is a list, extend.
-    batch_dict = {}
-    for k in batch[0].keys():
-        if isinstance(batch[0][k], dict):  # dealwith bert tokenizer output
-            batch_dict[k] = {}
-            for kk in batch[0][k].keys():
-                tmp = []
-                for i in range(len(batch)):
-                    tmp.append(batch[i][k][kk])
-                batch_dict[k][kk] = torch.vstack(tmp)
-        elif isinstance(batch[0][k], torch.Tensor):
-            batch_dict[k] = torch.stack([sample[k] for sample in batch])
-        elif isinstance(batch[0][k], np.ndarray):
-            batch_dict[k] = torch.tensor(np.stack([sample[k] for sample in batch]))
-        else:
-            batch_dict[k] = [sample[k] for sample in batch]
-    return batch_dict
-
-
-# def get_wds_dataset(
-#     args,
-#     model_cfg,
-#     is_train,
-#     audio_ext="flac",
-#     text_ext="json",
-#     max_len=480000,
-#     proportion=1.0,
-#     sizefilepath_=None,
-#     is_local=None,
-# ):
-#     """
-#     Get a dataset for wdsdataloader.
-#     """
-#     if is_local is None and (not args.remotedata is None):
-#         is_local = not args.remotedata
-
-#     input_shards = args.train_data if is_train else args.val_data
-#     assert input_shards is not None
-
-#     if not sizefilepath_ is None:
-#         sizefilepath = sizefilepath_
-#     else:
-#         sizefilepath = os.path.join(os.path.dirname(input_shards[0]), "sizes.json")
-
-#     if proportion != 1.0:
-#         num_samples, num_shards, input_shards, _ = sample_prop(
-#             sizefilepath, input_shards, proportion, is_local=is_local
-#         )
-#     else:
-#         num_samples, num_shards = get_dataset_size(
-#             input_shards, sizefilepath_=sizefilepath_, is_local=is_local
-#         )
-
-#     if not num_samples:
-#         if is_train:
-#             num_samples = args.train_num_samples
-#             if not num_samples:
-#                 raise RuntimeError(
-#                     "Currently, number of dataset samples must be specified for training dataset. "
-#                     "Please specify via `--train-num-samples` if no dataset length info present."
-#                 )
-#         else:
-#             num_samples = (
-#                 args.val_num_samples or 0
-#             )  # eval will just exhaust the iterator if not specified
-
-#     pipeline = [wds.SimpleShardList(input_shards)]
-#     # at this point we have an iterator over all the shards
-#     # TODO: (yusong): add a if statement of distributed. If not, we don't need to split_by_node
-#     if is_train or args.parallel_eval:
-#         pipeline.extend(
-#             [
-#                 wds.detshuffle(
-#                     bufsize=_SHARD_SHUFFLE_SIZE,
-#                     initial=_SHARD_SHUFFLE_INITIAL,
-#                     seed=args.seed,
-#                 ),
-#                 wds.split_by_node,
-#                 wds.split_by_worker,
-#                 # at this point, we have an iterator over the shards assigned to each worker at each node
-#                 wds.tarfile_to_samples(handler=log_and_continue),
-#                 wds.shuffle(
-#                     bufsize=_SAMPLE_SHUFFLE_SIZE,
-#                     initial=_SAMPLE_SHUFFLE_INITIAL,
-#                     rng=random.Random(args.seed),
-#                 ),
-#                 # wds.repeatedly,  # FIXME determine if this is beneficial
-#             ]
-#         )
-#     else:
-#         pipeline.extend(
-#             [
-#                 wds.split_by_worker,
-#                 # at this point, we have an iterator over the shards assigned to each worker
-#                 wds.tarfile_to_samples(handler=log_and_continue),
-#             ]
-#         )
-#     pipeline.append(
-#         wds.map(
-#             partial(
-#                 preprocess,
-#                 audio_ext=audio_ext,
-#                 text_ext=text_ext,
-#                 max_len=max_len,
-#                 audio_cfg=model_cfg["audio_cfg"],
-#                 class_index_dict=copy.deepcopy(args.class_index_dict),
-#                 data_filling=args.data_filling,
-#                 data_truncating=args.data_truncating,
-#                 text_augment_selection=args.text_augment_selection,
-#             )
-#         ),
-#     )
-
-#     pipeline.append(
-#         wds.batched(
-#             args.batch_size,
-#             partial=not (is_train or args.parallel_eval),
-#             collation_fn=collate_fn,
-#         )
-#     )
-
-#     dataset = wds.DataPipeline(*pipeline)
-#     if is_train or args.parallel_eval:
-#         # (yusong): Currently parallel evaluation will be not precise as we are repeat the last few samples.
-#         # (yusong): See comments below.
-#         # roll over and repeat a few samples to get same number of full batches on each node
-#         global_batch_size = args.batch_size * args.world_size
-#         num_batches = math.ceil(num_samples / global_batch_size)
-#         num_workers = max(1, args.workers)
-#         num_worker_batches = math.ceil(
-#             num_batches / num_workers
-#         )  # per dataloader worker
-#         num_batches = num_worker_batches * num_workers
-#         num_samples = num_batches * global_batch_size
-#         dataset = dataset.with_epoch(
-#             num_worker_batches
-#         )  # each worker is iterating over this
-#     else:
-#         # last batches are partial, eval is done on single (master) node
-#         num_batches = math.ceil(num_samples / args.batch_size)
-
-#     kwargs = {}
-#     if args.horovod:  # multi-node training on summit
-#         kwargs["multiprocessing_context"] = "forkserver"
-
-#     dataloader = wds.WebLoader(
-#         dataset, batch_size=None, shuffle=False, num_workers=args.workers, **kwargs
-#     )
-
-#     # FIXME not clear which approach is better, with_epoch before vs after dataloader?
-#     # hoping to resolve via https://github.com/webdataset/webdataset/issues/169
-#     # if is_train:
-#     #     # roll over and repeat a few samples to get same number of full batches on each node
-#     #     global_batch_size = args.batch_size * args.world_size
-#     #     num_batches = math.ceil(num_samples / global_batch_size)
-#     #     num_workers = max(1, args.workers)
-#     #     num_batches = math.ceil(num_batches / num_workers) * num_workers
-#     #     num_samples = num_batches * global_batch_size
-#     #     dataloader = dataloader.with_epoch(num_batches)
-#     # else:
-#     #     # last batches are partial, eval is done on single (master) node
-#     #     num_batches = math.ceil(num_samples / args.batch_size)
-
-#     # add meta-data to dataloader instance for convenience
-#     dataloader.num_batches = num_batches
-#     dataloader.num_samples = num_samples
-
-#     return DataInfo(dataloader, None)
-
-
-def wds_batch_list2dict(
-    batch,
-    keys=[
-        "__url__",
-        "__key__",
-        "waveform",
-        "text",
-        "raw_text",
-        "audio_name",
-        "text_name",
-        "audio_orig_sr",
-    ],
-):
-    """
-    Return a dictionary of the batch, with keys as the names of the fields.
-    """
-    assert len(keys) == len(
-        batch
-    ), "batch must have same number of keys as keys argument"
-    return {keys[i]: batch[i] for i in range(len(batch))}
-
-
-def get_csv_dataset(args, preprocess_fn, is_train):
-    input_filename = args.train_data if is_train else args.val_data
-    assert input_filename
-    dataset = CsvDataset(
-        input_filename,
-        preprocess_fn,
-        img_key=args.csv_img_key,
-        caption_key=args.csv_caption_key,
-        sep=args.csv_separator,
-    )
-    num_samples = len(dataset)
-    sampler = DistributedSampler(dataset) if args.distributed and is_train else None
-    shuffle = is_train and sampler is None
-
-    dataloader = DataLoader(
-        dataset,
-        batch_size=args.batch_size,
-        shuffle=shuffle,
-        num_workers=args.workers,
-        pin_memory=True,
-        sampler=sampler,
-        drop_last=is_train,
-    )
-    dataloader.num_samples = num_samples
-    dataloader.num_batches = len(dataloader)
-
-    return DataInfo(dataloader, sampler)
-
-
-def get_toy_dataset(args, model_cfg, is_train):
-    index_path = args.train_data if is_train else args.val_data
-    ipc_path = args.train_ipc if is_train else args.val_ipc
-    assert index_path and ipc_path
-    eval_mode = not is_train
-    dataset = ToyDataset(index_path, ipc_path, model_cfg, eval_mode=eval_mode)
-
-    num_samples = len(dataset)
-    sampler = (
-        DistributedSampler(dataset, shuffle=False)
-        if args.distributed and is_train
-        else None
-    )
-
-    dataloader = DataLoader(
-        dataset,
-        batch_size=args.batch_size,
-        shuffle=False,
-        num_workers=args.workers,
-        sampler=sampler,
-        drop_last=is_train,
-    )
-    dataloader.num_samples = num_samples
-    dataloader.num_batches = len(dataloader)
-
-    return DataInfo(dataloader, sampler)
-
-
-def get_dataset_fn(data_path, dataset_type):
-    if dataset_type == "webdataset":
-        return get_wds_dataset
-    elif dataset_type == "csv":
-        return get_csv_dataset
-    elif dataset_type == "auto":
-        ext = data_path.split(".")[-1]
-        if ext in ["csv", "tsv"]:
-            return get_csv_dataset
-        elif ext in ["tar"]:
-            return get_wds_dataset
-        else:
-            raise ValueError(
-                f"Tried to figure out dataset type, but failed for extension {ext}."
-            )
-    elif dataset_type == "toy":
-        return get_toy_dataset
-    else:
-        raise ValueError(f"Unsupported dataset type: {dataset_type}")
-
-
-def get_data(args, model_cfg):
-    data = {}
-
-    args.class_index_dict = load_class_label(args.class_label_path)
-
-    if args.datasetinfos is None:
-        args.datasetinfos = ["train", "unbalanced_train", "balanced_train"]
-    if args.dataset_type == "webdataset":
-        args.train_data = get_tar_path_from_dataset_name(
-            args.datasetnames,
-            args.datasetinfos,
-            islocal=not args.remotedata,
-            proportion=args.dataset_proportion,
-            dataset_path=args.datasetpath,
-            full_dataset=args.full_train_dataset,
-        )
-
-        if args.full_train_dataset is None:
-            args.full_train_dataset = []
-        if args.exclude_eval_dataset is None:
-            args.exclude_eval_dataset = []
-        excluded_eval_datasets = args.full_train_dataset + args.exclude_eval_dataset
-
-        val_dataset_names = (
-            [n for n in args.datasetnames if n not in excluded_eval_datasets]
-            if excluded_eval_datasets
-            else args.datasetnames
-        )
-        args.val_dataset_names = val_dataset_names
-        args.val_data = get_tar_path_from_dataset_name(
-            val_dataset_names,
-            ["valid", "test", "eval"],
-            islocal=not args.remotedata,
-            proportion=1,
-            dataset_path=args.datasetpath,
-            full_dataset=None,
-        )
-
-    if args.train_data:
-        data["train"] = get_dataset_fn(args.train_data, args.dataset_type)(
-            args, model_cfg, is_train=True
-        )
-
-    if args.val_data:
-        data["val"] = get_dataset_fn(args.val_data, args.dataset_type)(
-            args, model_cfg, is_train=False
-        )
-
-    return data
+import json
+import logging
+import os
+import random
+from dataclasses import dataclass
+import numpy as np
+import pandas as pd
+import torch
+import torchvision.datasets as datasets
+from PIL import Image
+from torch.utils.data import Dataset, DataLoader, SubsetRandomSampler
+from torch.utils.data.distributed import DistributedSampler
+import soundfile as sf
+import io
+from pathlib import Path
+
+# import wget
+
+from audiosr.clap.open_clip.utils import get_tar_path_from_dataset_name
+from audiosr.clap.open_clip.utils import load_class_label
+
+try:
+    import horovod.torch as hvd
+except ImportError:
+    hvd = None
+
+try:
+    import torchaudio
+except ImportError:
+    torchaudio = None
+
+from audiosr.clap.open_clip import tokenize
+
+
+def tokenizer(text):
+    return tokenize(text).squeeze(0)
+
+
+from transformers import RobertaTokenizer
+
+tokenize = RobertaTokenizer.from_pretrained("roberta-base")
+
+
+def tokenizer(text):
+    result = tokenize(
+        text,
+        padding="max_length",
+        truncation=True,
+        max_length=77,
+        return_tensors="pt",
+    )
+    return {k: v.squeeze(0) for k, v in result.items()}
+
+
+# initizlied the audioset map
+_AUDIOSET_MAP_PATH = os.path.join(Path(__file__).parent, "audioset_textmap.npy")
+_AUDIOSET_MAP = np.load(_AUDIOSET_MAP_PATH, allow_pickle=True)
+
+
+def int16_to_float32(x):
+    return (x / 32767.0).astype(np.float32)
+
+
+def float32_to_int16(x):
+    x = np.clip(x, a_min=-1.0, a_max=1.0)
+    return (x * 32767.0).astype(np.int16)
+
+
+# For Toy Dataset
+# class ToyDataset(Dataset):
+#     def __init__(self, index_path, ipc, config, eval_mode=False):
+#         """Toy Dataset for testing the audioset input with text labels
+#         Parameters
+#         ----------
+#             index_path: str
+#                 the link to the h5 file of each audio
+#             idc: str
+#                 the link to the npy file, the number of samples in each class
+#             config: dict
+#                 the audio cfg file
+#            eval_model (bool): to indicate if the dataset is a testing dataset
+#         """
+#         self.audio_cfg = config["audio_cfg"]
+#         self.text_cfg = config["text_cfg"]
+#         self.fp = h5py.File(index_path, "r")
+#         self.ipc = np.load(ipc, allow_pickle=True)
+#         self.total_size = len(self.fp["audio_name"])
+#         self.classes_num = self.audio_cfg["class_num"]
+#         self.eval_mode = eval_mode
+
+#         if not eval_mode:
+#             self.generate_queue()
+#         else:
+#             self.queue = []
+#             for i in range(self.total_size):
+#                 target = self.fp["target"][i]
+#                 if np.sum(target) > 0:
+#                     self.queue.append(i)
+#             self.total_size = len(self.queue)
+#         logging.info("total dataset size: %d" % (self.total_size))
+#         logging.info("class num: %d" % (self.classes_num))
+
+#     def time_shifting(self, x):
+#         frame_num = len(x)
+#         shift_len = random.randint(0, frame_num - 1)
+#         new_sample = np.concatenate([x[shift_len:], x[:shift_len]], axis=0)
+#         return new_sample
+
+#     def generate_queue(self):
+#         self.queue = []
+#         while len(self.queue) < self.total_size:
+#             class_set = [*range(self.classes_num)]
+#             random.shuffle(class_set)
+#             self.queue += [
+#                 self.ipc[d][random.randint(0, len(self.ipc[d]) - 1)] for d in class_set
+#             ]
+#         self.queue = self.queue[: self.total_size]
+
+#         logging.info("queue regenerated:%s" % (self.queue[-5:]))
+
+#     def crop_wav(self, x):
+#         crop_size = self.audio_cfg["crop_size"]
+#         crop_pos = random.randint(0, len(x) - crop_size - 1)
+#         return x[crop_pos : crop_pos + crop_size]
+
+#     def prompt_text(self, target):
+#         events = _AUDIOSET_MAP[np.where(target > 0)]
+#         event_text = "The sounds of " + ", ".join(events[:-1]) + " and " + events[-1]
+#         text = tokenize(event_text)[0]
+#         return text
+
+#     def __getitem__(self, index):
+#         """Load waveform, text, and target of an audio clip
+
+#         Parameters
+#         ----------
+#             index: int
+#                 the index number
+#         Return
+#         ------
+#             output: dict {
+#                 "hdf5_path": str,
+#                 "index_in_hdf5": int,
+#                 "audio_name": str,
+#                 "waveform": list (audio_length,),
+#                 "target": list (class_num, ),
+#                 "text": torch.tensor (context_length,)
+#             }
+#                 the output dictionary
+#         """
+#         s_index = self.queue[index]
+
+#         audio_name = self.fp["audio_name"][s_index].decode()
+#         # Hardcode here CHANGE
+#         hdf5_path = (
+#             self.fp["hdf5_path"][s_index]
+#             .decode()
+#             .replace(
+#                 "../workspace",
+#                 "/home/la/kechen/Research/ke_zsasp/workspace",
+#             )
+#         )
+#         r_idx = self.fp["index_in_hdf5"][s_index]
+#         target = self.fp["target"][s_index].astype(np.float32)
+#         text = self.prompt_text(target)
+#         with h5py.File(hdf5_path, "r") as f:
+#             waveform = int16_to_float32(f["waveform"][r_idx])[
+#                 : self.audio_cfg["clip_samples"]
+#             ]
+#         assert (
+#             len(waveform) == self.audio_cfg["clip_samples"]
+#         ), "The sample length is not match"
+#         # Time shift
+#         # if (self.config.enable_time_shift) and (not self.eval_mode):
+#         #     waveform = self.time_shifting(waveform)
+#         # # Label Enhance
+#         # if (self.config.crop_size is not None) and (not self.eval_mode):
+#         #     waveform = self.crop_wav(waveform)
+#         # # the label enhance rate is fixed 0.5
+#         # if (self.config.enable_label_enhance) and (not self.eval_mode) and random.random() < 0.5:
+#         #     kidx = np.where(target)[0]
+#         #     for k in kidx:
+#         #         for add_key in self.class_map[k][1]:
+#         #             target[add_key] = 1.0
+#         #         if len(self.class_map[k][2]) > 0:
+#         #             add_key = random.choice(self.class_map[k][2])
+#         #             target[add_key] = 1.0
+
+#         # missing the text input
+#         mel_spec = get_mel(torch.from_numpy(waveform), self.audio_cfg)[None, :, :]
+#         mel_spec = (
+#             torch.cat(
+#                 [mel_spec, mel_spec.clone(), mel_spec.clone(), mel_spec.clone()], dim=0
+#             )
+#             .cpu()
+#             .numpy()
+#         )
+#         longer = random.choice([True, False])
+#         if longer == False:
+#             mel_spec[1:, :, :] = 0.0
+#         data_dict = {
+#             "hdf5_path": hdf5_path,
+#             "index_in_hdf5": r_idx,
+#             "audio_name": audio_name,
+#             "waveform": waveform,
+#             "class_label": target,
+#             "text": text,
+#             "longer": longer,
+#             "mel_fusion": mel_spec,
+#         }
+#         return data_dict
+
+#     def __len__(self):
+#         return self.total_size
+
+
+class CsvDataset(Dataset):
+    def __init__(self, input_filename, transforms, img_key, caption_key, sep="\t"):
+        logging.debug(f"Loading csv data from {input_filename}.")
+        df = pd.read_csv(input_filename, sep=sep)
+
+        self.images = df[img_key].tolist()
+        self.captions = df[caption_key].tolist()
+        self.transforms = transforms
+        logging.debug("Done loading data.")
+
+    def __len__(self):
+        return len(self.captions)
+
+    def __getitem__(self, idx):
+        images = self.transforms(Image.open(str(self.images[idx])))
+        texts = tokenize([str(self.captions[idx])])[0]
+        return images, texts
+
+
+@dataclass
+class DataInfo:
+    dataloader: DataLoader
+    sampler: DistributedSampler
+
+
+def preprocess_txt(text):
+    return tokenize([str(text)])[0]
+
+
+# def get_dataset_size(shards, sizefilepath_=None, is_local=True):
+#     if isinstance(shards, list):
+#         size_list = []
+#         for s in shards:
+#             size_list.append(
+#                 get_dataset_size(s, sizefilepath_=sizefilepath_, is_local=is_local)[0]
+#             )
+#     else:
+#         if not is_local:
+#             for n in dataset_split.keys():
+#                 if n in shards.split("/"):
+#                     break
+#             for s in dataset_split[n]:
+#                 if s in shards.split("/"):
+#                     break
+#             sizefilepath_ = f"./json_files/{n}/{s}/sizes.json"
+#         shards_list = list(braceexpand.braceexpand(shards))
+#         dir_path = os.path.dirname(shards)
+#         if sizefilepath_ is not None:
+#             sizes = json.load(open(sizefilepath_, "r"))
+#             total_size = sum(
+#                 [
+#                     int(sizes[os.path.basename(shard.replace(".tar -", ".tar"))])
+#                     for shard in shards_list
+#                 ]
+#             )
+#         else:
+#             sizes_filename = os.path.join(dir_path, "sizes.json")
+#             len_filename = os.path.join(dir_path, "__len__")
+#             if os.path.exists(sizes_filename):
+#                 sizes = json.load(open(sizes_filename, "r"))
+#                 total_size = sum(
+#                     [int(sizes[os.path.basename(shard)]) for shard in shards_list]
+#                 )
+#             elif os.path.exists(len_filename):
+#                 # FIXME this used to be eval(open(...)) but that seemed rather unsafe
+#                 total_size = ast.literal_eval(open(len_filename, "r").read())
+#             else:
+#                 raise Exception(
+#                     "Cannot find sizes file for dataset. Please specify the path to the file."
+#                 )
+#                 # total_size = None  # num samples undefined
+#                 # some common dataset sizes (at time of authors last download)
+#                 # cc3m-train: 2905954
+#                 # cc12m: 10968539
+#                 # LAION-400m: 407332084
+#         num_shards = len(shards_list)
+#     if isinstance(shards, list):
+#         return sum(size_list), len(shards)
+#     else:
+#         return total_size, num_shards
+
+
+def get_imagenet(args, preprocess_fns, split):
+    assert split in ["train", "val", "v2"]
+    is_train = split == "train"
+    preprocess_train, preprocess_val = preprocess_fns
+
+    if split == "v2":
+        from imagenetv2_pytorch import ImageNetV2Dataset
+
+        dataset = ImageNetV2Dataset(location=args.imagenet_v2, transform=preprocess_val)
+    else:
+        if is_train:
+            data_path = args.imagenet_train
+            preprocess_fn = preprocess_train
+        else:
+            data_path = args.imagenet_val
+            preprocess_fn = preprocess_val
+        assert data_path
+
+        dataset = datasets.ImageFolder(data_path, transform=preprocess_fn)
+
+    if is_train:
+        idxs = np.zeros(len(dataset.targets))
+        target_array = np.array(dataset.targets)
+        k = 50
+        for c in range(1000):
+            m = target_array == c
+            n = len(idxs[m])
+            arr = np.zeros(n)
+            arr[:k] = 1
+            np.random.shuffle(arr)
+            idxs[m] = arr
+
+        idxs = idxs.astype("int")
+        sampler = SubsetRandomSampler(np.where(idxs)[0])
+    else:
+        sampler = None
+
+    dataloader = torch.utils.data.DataLoader(
+        dataset,
+        batch_size=args.batch_size,
+        num_workers=args.workers,
+        sampler=sampler,
+    )
+
+    return DataInfo(dataloader, sampler)
+
+
+def count_samples(dataloader):
+    os.environ["WDS_EPOCH"] = "0"
+    n_elements, n_batches = 0, 0
+    for images, texts in dataloader:
+        n_batches += 1
+        n_elements += len(images)
+        assert len(images) == len(texts)
+    return n_elements, n_batches
+
+
+def filter_no_caption(sample):
+    return "txt" in sample
+
+
+def log_and_continue(exn):
+    """Call in an exception handler to ignore any exception, isssue a warning, and continue."""
+    logging.warning(f"Handling webdataset error ({repr(exn)}). Ignoring.")
+    return True
+
+
+_SHARD_SHUFFLE_SIZE = 2000
+_SHARD_SHUFFLE_INITIAL = 500
+_SAMPLE_SHUFFLE_SIZE = 5000
+_SAMPLE_SHUFFLE_INITIAL = 1000
+
+
+# def sample_prop(sizefile, inputs, proportion, is_local=True):
+#     """
+#     Sample a proportion of the data.
+#     """
+#     file_path_dict = {
+#         os.path.split(inputs[i])[1]: os.path.split(inputs[i])[0]
+#         for i in range(len(inputs))
+#     }
+#     sampled_filepath_dict = {}
+#     sampled_size_dict = {}
+#     if not is_local:
+#         if os.path.exists("sizes.json"):
+#             os.remove("sizes.json")
+#         wget.download(sizefile, "sizes.json")
+#         sizefile = "sizes.json"
+#     with open(sizefile, "r", encoding="UTF-8") as f:
+#         load_dict = json.load(f)
+#     L = int(len(file_path_dict) * proportion)
+#     subkeys = random.sample(file_path_dict.keys(), L)
+#     for k in subkeys:
+#         sampled_size_dict[k] = load_dict[k]
+#         sampled_filepath_dict[k] = file_path_dict[k]
+#     return (
+#         sum(sampled_size_dict.values()),
+#         L,
+#         [os.path.join(v, k) for k, v in sampled_filepath_dict.items()],
+#         sampled_size_dict,
+#     )
+
+
+def get_mel(audio_data, audio_cfg):
+    # mel shape: (n_mels, T)
+    mel = torchaudio.transforms.MelSpectrogram(
+        sample_rate=audio_cfg["sample_rate"],
+        n_fft=audio_cfg["window_size"],
+        win_length=audio_cfg["window_size"],
+        hop_length=audio_cfg["hop_size"],
+        center=True,
+        pad_mode="reflect",
+        power=2.0,
+        norm=None,
+        onesided=True,
+        n_mels=64,
+        f_min=audio_cfg["fmin"],
+        f_max=audio_cfg["fmax"],
+    ).to(audio_data.device)
+    mel = mel(audio_data)
+    # we use log mel spectrogram as input
+    mel = torchaudio.transforms.AmplitudeToDB(top_db=None)(mel)
+    return mel.T  # (T, n_mels)
+
+
+def get_audio_features(
+    audio_data, mel, max_len, data_truncating, data_filling, audio_cfg
+):
+    """
+    Calculate and add audio features to sample.
+    Sample: a dict containing all the data of current sample.
+    audio_data: a tensor of shape (T) containing audio data.
+    max_len: the maximum length of audio data.
+    data_truncating: the method of truncating data.
+    data_filling: the method of filling data.
+    audio_cfg: a dict containing audio configuration. Comes from model_cfg['audio_cfg'].
+    """
+    sample = {}
+
+    # assert audio_data.size(-1) <= max_len, str(audio_data.size())
+
+    # split to three parts
+    chunk_frames = (
+        max_len // audio_cfg["hop_size"] + 1
+    )  # the +1 related to how the spectrogram is computed
+    mel = mel[:chunk_frames]
+
+    audio_data = audio_data[..., :max_len]
+    sample["mel_fusion"] = mel
+    longer = torch.tensor([True])
+
+    sample["longer"] = longer
+    sample["waveform"] = audio_data
+
+    return sample
+
+
+def preprocess(
+    sample,
+    audio_ext,
+    text_ext,
+    max_len,
+    audio_cfg,
+    class_index_dict=None,
+    data_filling="pad",
+    data_truncating="rand_trunc",
+    text_augment_selection=None,
+):
+    """
+    Preprocess a single sample for wdsdataloader.
+    """
+    audio_data, orig_sr = sf.read(io.BytesIO(sample[audio_ext]))
+    audio_data = int16_to_float32(float32_to_int16(audio_data))
+    audio_data = torch.tensor(audio_data).float()
+
+    # TODO: (yusong) to be include in the future
+    # # if torchaudio not installed, use soundfile to load audio
+    # if torchaudio is None:
+    #     audio_data, orig_sr = sf.read(io.BytesIO(sample[audio_ext]))
+    #     audio_data = torch.tensor(audio_data).float()
+    # else:
+    #     # https://github.com/webdataset/webdataset/blob/main/webdataset/autodecode.py
+    #     with tempfile.TemporaryDirectory() as dirname:
+    #         os.makedirs(dirname, exist_ok=True)
+    #         fname = os.path.join(dirname, f"file.flac")
+    #         with open(fname, "wb") as stream:
+    #             stream.write(sample[audio_ext])
+    #         audio_data, orig_sr = torchaudio.load(fname)
+    #         audio_data = audio_data[0, :].float()
+
+    sample = get_audio_features(
+        sample, audio_data, max_len, data_truncating, data_filling, audio_cfg
+    )
+    del sample[audio_ext]
+
+    try:
+        json_dict_raw = json.loads(sample[text_ext].decode("utf-8"))
+    except:
+        print("sample[__url__]:", sample["__url__"])
+
+    # For selecting augmented text from dataset
+    if text_augment_selection is None or text_augment_selection == "none":
+        texts = json_dict_raw["text"]
+    elif text_augment_selection == "all":
+        if "text_augment_all" in json_dict_raw.keys():
+            texts = json_dict_raw["text_augment_all"]
+        else:
+            texts = json_dict_raw["text"]
+    elif text_augment_selection == "augment_only":
+        if "text_augment_all" in json_dict_raw.keys():
+            if json_dict_raw["text_augment_t5"] is None:
+                texts = json_dict_raw["text"]
+            else:
+                texts = json_dict_raw["text_augment_t5"]
+        else:
+            texts = json_dict_raw["text"]
+    else:
+        raise NotImplementedError(
+            f"text_augment_selection {text_augment_selection} not implemented"
+        )
+    sample["full_text"] = texts
+
+    if isinstance(texts, list) and isinstance(texts[0], str) and len(texts) > 1:
+        texts = random.choice(texts)
+    sample["raw_text"] = texts
+    sample["text"] = tokenizer(texts)  # text shape: [num_token]
+    if class_index_dict is not None:
+        # https://stackoverflow.com/questions/48004243/how-to-share-large-read-only-dictionary-list-across-processes-in-multiprocessing
+        # https://stackoverflow.com/questions/45693949/storing-strings-in-a-multiprocessing-sharedctypes-array
+        # key, val = class_index_dict
+        # key = key[:].split('\n')
+        # _dict = {k: v for k, v in zip(key, val)}
+        sample["class_label"] = np.zeros(len(class_index_dict.keys()))
+        for x in json_dict_raw["tag"]:
+            sample["class_label"][class_index_dict[x]] = 1
+        sample["class_label"] = torch.tensor(sample["class_label"]).float()
+    del sample[text_ext]
+    sample["audio_name"] = sample["__key__"].split("/")[-1] + "." + audio_ext
+    sample["text_name"] = sample["__key__"].split("/")[-1] + "." + text_ext
+    sample["audio_orig_sr"] = orig_sr
+    return sample
+
+
+def collate_fn(batch):
+    """
+    Collate function for wdsdataloader.
+    batch: a list of dict, each dict is a sample
+    """
+    # concatenate values in each dictionary. if it is a tensor, concatenate. if it is a list, extend.
+    batch_dict = {}
+    for k in batch[0].keys():
+        if isinstance(batch[0][k], dict):  # dealwith bert tokenizer output
+            batch_dict[k] = {}
+            for kk in batch[0][k].keys():
+                tmp = []
+                for i in range(len(batch)):
+                    tmp.append(batch[i][k][kk])
+                batch_dict[k][kk] = torch.vstack(tmp)
+        elif isinstance(batch[0][k], torch.Tensor):
+            batch_dict[k] = torch.stack([sample[k] for sample in batch])
+        elif isinstance(batch[0][k], np.ndarray):
+            batch_dict[k] = torch.tensor(np.stack([sample[k] for sample in batch]))
+        else:
+            batch_dict[k] = [sample[k] for sample in batch]
+    return batch_dict
+
+
+# def get_wds_dataset(
+#     args,
+#     model_cfg,
+#     is_train,
+#     audio_ext="flac",
+#     text_ext="json",
+#     max_len=480000,
+#     proportion=1.0,
+#     sizefilepath_=None,
+#     is_local=None,
+# ):
+#     """
+#     Get a dataset for wdsdataloader.
+#     """
+#     if is_local is None and (not args.remotedata is None):
+#         is_local = not args.remotedata
+
+#     input_shards = args.train_data if is_train else args.val_data
+#     assert input_shards is not None
+
+#     if not sizefilepath_ is None:
+#         sizefilepath = sizefilepath_
+#     else:
+#         sizefilepath = os.path.join(os.path.dirname(input_shards[0]), "sizes.json")
+
+#     if proportion != 1.0:
+#         num_samples, num_shards, input_shards, _ = sample_prop(
+#             sizefilepath, input_shards, proportion, is_local=is_local
+#         )
+#     else:
+#         num_samples, num_shards = get_dataset_size(
+#             input_shards, sizefilepath_=sizefilepath_, is_local=is_local
+#         )
+
+#     if not num_samples:
+#         if is_train:
+#             num_samples = args.train_num_samples
+#             if not num_samples:
+#                 raise RuntimeError(
+#                     "Currently, number of dataset samples must be specified for training dataset. "
+#                     "Please specify via `--train-num-samples` if no dataset length info present."
+#                 )
+#         else:
+#             num_samples = (
+#                 args.val_num_samples or 0
+#             )  # eval will just exhaust the iterator if not specified
+
+#     pipeline = [wds.SimpleShardList(input_shards)]
+#     # at this point we have an iterator over all the shards
+#     # TODO: (yusong): add a if statement of distributed. If not, we don't need to split_by_node
+#     if is_train or args.parallel_eval:
+#         pipeline.extend(
+#             [
+#                 wds.detshuffle(
+#                     bufsize=_SHARD_SHUFFLE_SIZE,
+#                     initial=_SHARD_SHUFFLE_INITIAL,
+#                     seed=args.seed,
+#                 ),
+#                 wds.split_by_node,
+#                 wds.split_by_worker,
+#                 # at this point, we have an iterator over the shards assigned to each worker at each node
+#                 wds.tarfile_to_samples(handler=log_and_continue),
+#                 wds.shuffle(
+#                     bufsize=_SAMPLE_SHUFFLE_SIZE,
+#                     initial=_SAMPLE_SHUFFLE_INITIAL,
+#                     rng=random.Random(args.seed),
+#                 ),
+#                 # wds.repeatedly,  # FIXME determine if this is beneficial
+#             ]
+#         )
+#     else:
+#         pipeline.extend(
+#             [
+#                 wds.split_by_worker,
+#                 # at this point, we have an iterator over the shards assigned to each worker
+#                 wds.tarfile_to_samples(handler=log_and_continue),
+#             ]
+#         )
+#     pipeline.append(
+#         wds.map(
+#             partial(
+#                 preprocess,
+#                 audio_ext=audio_ext,
+#                 text_ext=text_ext,
+#                 max_len=max_len,
+#                 audio_cfg=model_cfg["audio_cfg"],
+#                 class_index_dict=copy.deepcopy(args.class_index_dict),
+#                 data_filling=args.data_filling,
+#                 data_truncating=args.data_truncating,
+#                 text_augment_selection=args.text_augment_selection,
+#             )
+#         ),
+#     )
+
+#     pipeline.append(
+#         wds.batched(
+#             args.batch_size,
+#             partial=not (is_train or args.parallel_eval),
+#             collation_fn=collate_fn,
+#         )
+#     )
+
+#     dataset = wds.DataPipeline(*pipeline)
+#     if is_train or args.parallel_eval:
+#         # (yusong): Currently parallel evaluation will be not precise as we are repeat the last few samples.
+#         # (yusong): See comments below.
+#         # roll over and repeat a few samples to get same number of full batches on each node
+#         global_batch_size = args.batch_size * args.world_size
+#         num_batches = math.ceil(num_samples / global_batch_size)
+#         num_workers = max(1, args.workers)
+#         num_worker_batches = math.ceil(
+#             num_batches / num_workers
+#         )  # per dataloader worker
+#         num_batches = num_worker_batches * num_workers
+#         num_samples = num_batches * global_batch_size
+#         dataset = dataset.with_epoch(
+#             num_worker_batches
+#         )  # each worker is iterating over this
+#     else:
+#         # last batches are partial, eval is done on single (master) node
+#         num_batches = math.ceil(num_samples / args.batch_size)
+
+#     kwargs = {}
+#     if args.horovod:  # multi-node training on summit
+#         kwargs["multiprocessing_context"] = "forkserver"
+
+#     dataloader = wds.WebLoader(
+#         dataset, batch_size=None, shuffle=False, num_workers=args.workers, **kwargs
+#     )
+
+#     # FIXME not clear which approach is better, with_epoch before vs after dataloader?
+#     # hoping to resolve via https://github.com/webdataset/webdataset/issues/169
+#     # if is_train:
+#     #     # roll over and repeat a few samples to get same number of full batches on each node
+#     #     global_batch_size = args.batch_size * args.world_size
+#     #     num_batches = math.ceil(num_samples / global_batch_size)
+#     #     num_workers = max(1, args.workers)
+#     #     num_batches = math.ceil(num_batches / num_workers) * num_workers
+#     #     num_samples = num_batches * global_batch_size
+#     #     dataloader = dataloader.with_epoch(num_batches)
+#     # else:
+#     #     # last batches are partial, eval is done on single (master) node
+#     #     num_batches = math.ceil(num_samples / args.batch_size)
+
+#     # add meta-data to dataloader instance for convenience
+#     dataloader.num_batches = num_batches
+#     dataloader.num_samples = num_samples
+
+#     return DataInfo(dataloader, None)
+
+
+def wds_batch_list2dict(
+    batch,
+    keys=[
+        "__url__",
+        "__key__",
+        "waveform",
+        "text",
+        "raw_text",
+        "audio_name",
+        "text_name",
+        "audio_orig_sr",
+    ],
+):
+    """
+    Return a dictionary of the batch, with keys as the names of the fields.
+    """
+    assert len(keys) == len(
+        batch
+    ), "batch must have same number of keys as keys argument"
+    return {keys[i]: batch[i] for i in range(len(batch))}
+
+
+def get_csv_dataset(args, preprocess_fn, is_train):
+    input_filename = args.train_data if is_train else args.val_data
+    assert input_filename
+    dataset = CsvDataset(
+        input_filename,
+        preprocess_fn,
+        img_key=args.csv_img_key,
+        caption_key=args.csv_caption_key,
+        sep=args.csv_separator,
+    )
+    num_samples = len(dataset)
+    sampler = DistributedSampler(dataset) if args.distributed and is_train else None
+    shuffle = is_train and sampler is None
+
+    dataloader = DataLoader(
+        dataset,
+        batch_size=args.batch_size,
+        shuffle=shuffle,
+        num_workers=args.workers,
+        pin_memory=True,
+        sampler=sampler,
+        drop_last=is_train,
+    )
+    dataloader.num_samples = num_samples
+    dataloader.num_batches = len(dataloader)
+
+    return DataInfo(dataloader, sampler)
+
+
+def get_toy_dataset(args, model_cfg, is_train):
+    index_path = args.train_data if is_train else args.val_data
+    ipc_path = args.train_ipc if is_train else args.val_ipc
+    assert index_path and ipc_path
+    eval_mode = not is_train
+    dataset = ToyDataset(index_path, ipc_path, model_cfg, eval_mode=eval_mode)
+
+    num_samples = len(dataset)
+    sampler = (
+        DistributedSampler(dataset, shuffle=False)
+        if args.distributed and is_train
+        else None
+    )
+
+    dataloader = DataLoader(
+        dataset,
+        batch_size=args.batch_size,
+        shuffle=False,
+        num_workers=args.workers,
+        sampler=sampler,
+        drop_last=is_train,
+    )
+    dataloader.num_samples = num_samples
+    dataloader.num_batches = len(dataloader)
+
+    return DataInfo(dataloader, sampler)
+
+
+def get_dataset_fn(data_path, dataset_type):
+    if dataset_type == "webdataset":
+        return get_wds_dataset
+    elif dataset_type == "csv":
+        return get_csv_dataset
+    elif dataset_type == "auto":
+        ext = data_path.split(".")[-1]
+        if ext in ["csv", "tsv"]:
+            return get_csv_dataset
+        elif ext in ["tar"]:
+            return get_wds_dataset
+        else:
+            raise ValueError(
+                f"Tried to figure out dataset type, but failed for extension {ext}."
+            )
+    elif dataset_type == "toy":
+        return get_toy_dataset
+    else:
+        raise ValueError(f"Unsupported dataset type: {dataset_type}")
+
+
+def get_data(args, model_cfg):
+    data = {}
+
+    args.class_index_dict = load_class_label(args.class_label_path)
+
+    if args.datasetinfos is None:
+        args.datasetinfos = ["train", "unbalanced_train", "balanced_train"]
+    if args.dataset_type == "webdataset":
+        args.train_data = get_tar_path_from_dataset_name(
+            args.datasetnames,
+            args.datasetinfos,
+            islocal=not args.remotedata,
+            proportion=args.dataset_proportion,
+            dataset_path=args.datasetpath,
+            full_dataset=args.full_train_dataset,
+        )
+
+        if args.full_train_dataset is None:
+            args.full_train_dataset = []
+        if args.exclude_eval_dataset is None:
+            args.exclude_eval_dataset = []
+        excluded_eval_datasets = args.full_train_dataset + args.exclude_eval_dataset
+
+        val_dataset_names = (
+            [n for n in args.datasetnames if n not in excluded_eval_datasets]
+            if excluded_eval_datasets
+            else args.datasetnames
+        )
+        args.val_dataset_names = val_dataset_names
+        args.val_data = get_tar_path_from_dataset_name(
+            val_dataset_names,
+            ["valid", "test", "eval"],
+            islocal=not args.remotedata,
+            proportion=1,
+            dataset_path=args.datasetpath,
+            full_dataset=None,
+        )
+
+    if args.train_data:
+        data["train"] = get_dataset_fn(args.train_data, args.dataset_type)(
+            args, model_cfg, is_train=True
+        )
+
+    if args.val_data:
+        data["val"] = get_dataset_fn(args.val_data, args.dataset_type)(
+            args, model_cfg, is_train=False
+        )
+
+    return data

audiosr/clap/training/params.py

@@ -1,563 +1,563 @@
-import argparse
-
-
-def get_default_params(model_name):
-    # Params from paper (https://arxiv.org/pdf/2103.00020.pdf)
-    model_name = model_name.lower()
-    if "vit" in model_name:
-        return {"lr": 5.0e-4, "beta1": 0.9, "beta2": 0.98, "eps": 1.0e-6}
-    else:
-        return {"lr": 5.0e-4, "beta1": 0.9, "beta2": 0.999, "eps": 1.0e-8}
-
-
-def parse_args():
-    parser = argparse.ArgumentParser()
-    parser.add_argument(
-        "--train-data",
-        type=str,
-        default=None,
-        help="Path to h5 filewith training data",
-    )
-    parser.add_argument(
-        "--val-data",
-        type=str,
-        default=None,
-        help="Path to h5 file with validation data",
-    )
-    parser.add_argument(
-        "--freeze-text",
-        default=False,
-        action="store_true",
-        help="if you need to freeze the text encoder, make this True",
-    )
-    parser.add_argument(
-        "--freeze-text-after",
-        type=int,
-        default=-1,
-        help="if you need to freeze the text encoder after (include) epoch x, set this param to x. Set -1 to disable it",
-    )
-    parser.add_argument(
-        "--train-ipc",
-        type=str,
-        default=None,
-        help="Path to npy file of the number of instance per class in training data",
-    )
-    parser.add_argument(
-        "--val-ipc",
-        type=str,
-        default=None,
-        help="Path to npy file of the number of instance per class in validation data",
-    )
-    parser.add_argument(
-        "--train-num-samples",
-        type=int,
-        default=None,
-        help="Number of samples in dataset. Required for webdataset if not available in info file.",
-    )
-    parser.add_argument(
-        "--val-num-samples",
-        type=int,
-        default=None,
-        help="Number of samples in dataset. Useful for webdataset if not available in info file.",
-    )
-    parser.add_argument(
-        "--dataset-type",
-        choices=["webdataset", "csv", "auto", "toy"],
-        default="auto",
-        help="Which type of dataset to process.",
-    )
-    parser.add_argument(
-        "--csv-separator",
-        type=str,
-        default="\t",
-        help="For csv-like datasets, which separator to use.",
-    )
-    parser.add_argument(
-        "--csv-img-key",
-        type=str,
-        default="filepath",
-        help="For csv-like datasets, the name of the key for the image paths.",
-    )
-    parser.add_argument(
-        "--csv-caption-key",
-        type=str,
-        default="title",
-        help="For csv-like datasets, the name of the key for the captions.",
-    )
-    parser.add_argument(
-        "--imagenet-val",
-        type=str,
-        default=None,
-        help="Path to imagenet val set for conducting zero shot evaluation.",
-    )
-    parser.add_argument(
-        "--imagenet-v2",
-        type=str,
-        default=None,
-        help="Path to imagenet v2 for conducting zero shot evaluation.",
-    )
-    parser.add_argument(
-        "--datasetnames",
-        nargs="+",
-        default=None,
-        help="If loading webdataset, spedify the dataset names to load. Can be some of these: Clotho, audioset, audiocaps, BBCSoundEffects",
-    )
-    parser.add_argument(
-        "--full-train-dataset",
-        nargs="+",
-        default=None,
-        help="Which dataset will be trained with all the subsets. (train+test)",
-    )
-    parser.add_argument(
-        "--exclude-eval-dataset",
-        nargs="+",
-        default=None,
-        help="Which dataset will be excluded with evaluation",
-    )
-    parser.add_argument(
-        "--datasetinfos",
-        nargs="+",
-        default=None,
-        help="If loading webdataset, spedify the dataset types to load. Can be some of these: train, test, valid, unbalanced_train, balanced_train, eval",
-    )
-    parser.add_argument(
-        "--dataset-proportion",
-        type=float,
-        default=1.0,
-        help="How much proportion of dataset we want to train.",
-    )
-    parser.add_argument(
-        "--remotedata",
-        default=False,
-        action="store_true",
-        help="if the dataset is remote, set this flag",
-    )
-    parser.add_argument(
-        "--class-label-path",
-        type=str,
-        default=None,
-        help="The path of the class label pickle or csv.",
-    )
-    parser.add_argument(
-        "--datasetpath",
-        type=str,
-        default="/mnt/audio_clip/webdataset_tar",
-        help="The path to the dataset",
-    )
-    parser.add_argument(
-        "--logs",
-        type=str,
-        default="./logs/",
-        help="Where to store tensorboard logs. Use None to avoid storing logs.",
-    )
-    parser.add_argument(
-        "--log-local",
-        action="store_true",
-        default=False,
-        help="log files on local master, otherwise global master only.",
-    )
-    parser.add_argument(
-        "--name",
-        type=str,
-        default=None,
-        help="Optional identifier for the experiment when storing logs. Otherwise use current time.",
-    )
-    parser.add_argument(
-        "--workers", type=int, default=1, help="Number of workers per GPU."
-    )
-    parser.add_argument(
-        "--batch-size", type=int, default=64, help="Batch size per GPU."
-    )
-    parser.add_argument(
-        "--epochs", type=int, default=32, help="Number of epochs to train for."
-    )
-    parser.add_argument("--lr", type=float, default=None, help="Learning rate.")
-    parser.add_argument("--beta1", type=float, default=None, help="Adam beta 1.")
-    parser.add_argument("--beta2", type=float, default=None, help="Adam beta 2.")
-    parser.add_argument("--eps", type=float, default=None, help="Adam epsilon.")
-    parser.add_argument("--momentum", type=float, default=None, help="SGD epsilon.")
-    parser.add_argument("--wd", type=float, default=0.2, help="Weight decay.")
-
-    parser.add_argument(
-        "--split-opt",
-        action="store_true",
-        default=False,
-        help="Use this flag to skip the learning rate decay.",
-    )
-    parser.add_argument(
-        "--lr-pretrained", type=float, default=None, help="Learning rate for text."
-    )
-    parser.add_argument(
-        "--beta1-pretrained", type=float, default=None, help="Adam beta 1 for text."
-    )
-    parser.add_argument(
-        "--beta2-pretrained", type=float, default=None, help="Adam beta 2 for text."
-    )
-    parser.add_argument(
-        "--eps-pretrained", type=float, default=None, help="Adam epsilon for text."
-    )
-    parser.add_argument(
-        "--wd-pretrained", type=float, default=0.2, help="Weight decay for text."
-    )
-    parser.add_argument(
-        "--momentum-pretrained", type=float, default=0.9, help="Momentum for text."
-    )
-    parser.add_argument(
-        "--lr-new", type=float, default=None, help="Learning rate for audio."
-    )
-    parser.add_argument(
-        "--beta1-new", type=float, default=None, help="Adam beta 1 for audio."
-    )
-    parser.add_argument(
-        "--beta2-new", type=float, default=None, help="Adam beta 2 for audio."
-    )
-    parser.add_argument(
-        "--eps-new", type=float, default=None, help="Adam epsilon for audio."
-    )
-    parser.add_argument(
-        "--wd-new", type=float, default=0.2, help="Weight decay for audio."
-    )
-    parser.add_argument(
-        "--momentum-new", type=float, default=0.9, help="Momentum for audio."
-    )
-    parser.add_argument(
-        "--warmup", type=int, default=10000, help="Number of steps to warmup for."
-    )
-    parser.add_argument(
-        "--use-bn-sync",
-        default=False,
-        action="store_true",
-        help="Whether to use batch norm sync.",
-    )
-    parser.add_argument(
-        "--skip-scheduler",
-        action="store_true",
-        default=False,
-        help="Use this flag to skip the learning rate decay.",
-    )
-    parser.add_argument(
-        "--save-frequency", type=int, default=1, help="How often to save checkpoints."
-    )
-    parser.add_argument(
-        "--save-top-performance",
-        type=int,
-        default=0,
-        help="Save the top x performance weights if the value >0",
-    )
-    parser.add_argument(
-        "--save-most-recent",
-        action="store_true",
-        default=False,
-        help="Always save the most recent model trained to epoch_latest.pt.",
-    )
-    parser.add_argument(
-        "--zeroshot-frequency", type=int, default=2, help="How often to run zero shot."
-    )
-    parser.add_argument(
-        "--val-frequency",
-        type=int,
-        default=1,
-        help="How often to run evaluation with val data.",
-    )
-    parser.add_argument(
-        "--resume",
-        default=None,
-        type=str,
-        help="path to latest checkpoint (default: none)",
-    )
-    parser.add_argument(
-        "--precision",
-        choices=["amp", "fp16", "fp32"],
-        default="amp",
-        help="Floating point precision.",
-    )
-    parser.add_argument(
-        "--amodel",
-        type=str,
-        default="RN50",
-        help="Name of the audio backbone to use.",
-    )
-    parser.add_argument(
-        "--tmodel",
-        type=str,
-        default="transformer",
-        help="Name of the text backbone to use. Can be [transformer, bert, roberta, bart]",
-    )
-    parser.add_argument(
-        "--pretrained-audio",
-        default="",
-        type=str,
-        help="Use a pretrained audio model weights for the audio encoder of CLAP",
-    )
-    parser.add_argument(
-        "--pretrained-text",
-        default="",
-        type=str,
-        help="Use a pretrained text model weights for the text encoder of CLAP",
-    )
-    parser.add_argument(
-        "--pretrained",
-        default="",
-        type=str,
-        help="Use a pretrained CLIP model weights with the specified tag or file path.",
-    )
-    parser.add_argument(
-        "--pretrained-image",
-        default=False,
-        action="store_true",
-        help="Load imagenet pretrained weights for image tower backbone if available.",
-    )
-    parser.add_argument(
-        "--lock-image",
-        default=False,
-        action="store_true",
-        help="Lock full image tower by disabling gradients.",
-    )
-    parser.add_argument(
-        "--lock-image-unlocked-groups",
-        type=int,
-        default=0,
-        help="Leave last n image tower layer groups unlocked.",
-    )
-    parser.add_argument(
-        "--lock-image-freeze-bn-stats",
-        default=False,
-        action="store_true",
-        help="Freeze BatchNorm running stats in image tower for any locked layers.",
-    )
-    parser.add_argument(
-        "--local-loss",
-        default=False,
-        action="store_true",
-        help="calculate loss w/ local features @ global (instead of realizing full global @ global matrix)",
-    )
-    parser.add_argument(
-        "--gather-with-grad",
-        default=False,
-        action="store_true",
-        help="enable full distributed gradient for feature gather",
-    )
-    parser.add_argument(
-        "--force-quick-gelu",
-        default=False,
-        action="store_true",
-        help="Force use of QuickGELU activation for non-OpenAI transformer models.",
-    )
-    parser.add_argument(
-        "--torchscript",
-        default=False,
-        action="store_true",
-        help="torch.jit.script the model, also uses jit version of OpenAI models if pretrained=='openai'",
-    )
-    parser.add_argument(
-        "--trace",
-        default=False,
-        action="store_true",
-        help="torch.jit.trace the model for inference / eval only",
-    )
-    # arguments for distributed training
-    parser.add_argument(
-        "--dist-url",
-        default="env://",
-        type=str,
-        help="url used to set up distributed training",
-    )
-    parser.add_argument(
-        "--dist-backend", default="nccl", type=str, help="distributed backend"
-    )
-    parser.add_argument(
-        "--report-to",
-        default="",
-        type=str,
-        help="Options are ['wandb', 'tensorboard', 'wandb,tensorboard']",
-    )
-    parser.add_argument(
-        "--wandb-notes", default="", type=str, help="Notes if logging with wandb"
-    )
-    parser.add_argument(
-        "--C", type=float, default=3.16, help="inverse regularizer for logistic reg."
-    )
-    parser.add_argument(
-        "--debug",
-        default=False,
-        action="store_true",
-        help="If true, more information is logged.",
-    )
-    parser.add_argument(
-        "--copy-codebase",
-        default=False,
-        action="store_true",
-        help="If true, we copy the entire base on the log diretory, and execute from there.",
-    )
-    parser.add_argument(
-        "--horovod",
-        default=False,
-        action="store_true",
-        help="Use horovod for distributed training.",
-    )
-    parser.add_argument(
-        "--ddp-static-graph",
-        default=False,
-        action="store_true",
-        help="Enable static graph optimization for DDP in PyTorch >= 1.11.",
-    )
-    parser.add_argument(
-        "--no-set-device-rank",
-        default=False,
-        action="store_true",
-        help="Don't set device index from local rank (when CUDA_VISIBLE_DEVICES restricted to one per proc).",
-    )
-    parser.add_argument("--seed", type=int, default=4242, help="Default random seed.")
-
-    parser.add_argument(
-        "--top-k-checkpoint-select-dataset",
-        type=str,
-        default="all",
-        help="The dataset of selecting top-k checkpoint.",
-    )
-
-    # @R10, @R@5, @R1, mAP@10
-    parser.add_argument(
-        "--top-k-checkpoint-select-metric",
-        type=str,
-        default="_R@10",
-        help="The metric for selecting top-k checkpoint.",
-    )
-    parser.add_argument(
-        "--openai-model-cache-dir",
-        type=str,
-        default="~/.cache/clip",
-        help="Directory to download OpenAI models.",
-    )
-    parser.add_argument(
-        "--optimizer",
-        type=str,
-        default="adamw",
-        help="can be AdamW or SGD",
-    )
-    parser.add_argument(
-        "--parallel-eval",
-        default=False,
-        action="store_true",
-        help="Eval in parallel (multi-GPU, multi-node).",
-    )
-
-    parser.add_argument(
-        "--no-eval",
-        default=False,
-        action="store_true",
-        help="Training without evaluation.",
-    )
-
-    parser.add_argument(
-        "--lp-mlp",
-        default=False,
-        action="store_true",
-        help="Linear Probe using MLP layer or not.",
-    )
-
-    parser.add_argument(
-        "--lp-freeze",
-        default=False,
-        action="store_true",
-        help="Linear Probe using Freeze CLAP or not",
-    )
-
-    parser.add_argument(
-        "--lp-act",
-        default="None",
-        type=str,
-        help="Options are ['relu','elu','prelu','softmax','sigmoid']",
-    )
-
-    parser.add_argument(
-        "--lp-loss", type=str, default="bce", help="Loss func of Linear Probe."
-    )
-
-    parser.add_argument(
-        "--lp-metrics",
-        type=str,
-        default="map,mauc,acc",
-        help="Metrics of Linear Probe.",
-    )
-
-    parser.add_argument(
-        "--lp-lr", type=float, default=1e-4, help="learning rate of linear probe"
-    )
-    parser.add_argument(
-        "--kappa",
-        type=float,
-        default=0,
-        help="the kappa in the weighted contrastive loss, default is to turn off the weighted contrastive loss",
-    )
-
-    parser.add_argument(
-        "--data-filling",
-        type=str,
-        default="pad",
-        help="type of data filling when the audio length is shorter than the max length."
-        "Can be one of the following: repeat, repeatpad, pad",
-    )
-    parser.add_argument(
-        "--data-truncating",
-        type=str,
-        default="rand_trunc",
-        help="type of data truncation when the audio length is longer than the max length."
-        "Can be one of the following: rand_trunc, fusion",
-    )
-
-    parser.add_argument(
-        "--clap-mlploss",
-        default=False,
-        action="store_true",
-        help="Using MLP loss for CLAP model or not",
-    )
-
-    parser.add_argument(
-        "--wandb-id",
-        type=str,
-        default=None,
-        help="the id of wandb experiment to restore.",
-    )
-
-    parser.add_argument(
-        "--sleep", type=float, default=0, help="sleep n seconds before start training"
-    )
-
-    # variable length processing
-    parser.add_argument(
-        "--enable-fusion",
-        default=False,
-        action="store_true",
-        help="Enable feature funsion for variable-length data",
-    )
-
-    parser.add_argument(
-        "--fusion-type",
-        type=str,
-        default="None",
-        help="Type is among ['channel_map', 'daf_1d','aff_1d','iaff_1d','daf_2d','aff_2d','iaff_2d']",
-    )
-
-    parser.add_argument(
-        "--mixup",
-        default=False,
-        action="store_true",
-        help="Enable mixup in finetuning training.",
-    )
-    parser.add_argument(
-        "--text-augment-selection",
-        type=str,
-        default=None,
-        help="For selecting levels of augmented text. Type is among ['all', 'augment_only', 'none']",
-    )
-
-    args = parser.parse_args()
-
-    # If some params are not passed, we use the default values based on model name.
-    default_params = get_default_params(args.amodel)
-    for name, val in default_params.items():
-        if getattr(args, name) is None:
-            setattr(args, name, val)
-
-    return args
+import argparse
+
+
+def get_default_params(model_name):
+    # Params from paper (https://arxiv.org/pdf/2103.00020.pdf)
+    model_name = model_name.lower()
+    if "vit" in model_name:
+        return {"lr": 5.0e-4, "beta1": 0.9, "beta2": 0.98, "eps": 1.0e-6}
+    else:
+        return {"lr": 5.0e-4, "beta1": 0.9, "beta2": 0.999, "eps": 1.0e-8}
+
+
+def parse_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--train-data",
+        type=str,
+        default=None,
+        help="Path to h5 filewith training data",
+    )
+    parser.add_argument(
+        "--val-data",
+        type=str,
+        default=None,
+        help="Path to h5 file with validation data",
+    )
+    parser.add_argument(
+        "--freeze-text",
+        default=False,
+        action="store_true",
+        help="if you need to freeze the text encoder, make this True",
+    )
+    parser.add_argument(
+        "--freeze-text-after",
+        type=int,
+        default=-1,
+        help="if you need to freeze the text encoder after (include) epoch x, set this param to x. Set -1 to disable it",
+    )
+    parser.add_argument(
+        "--train-ipc",
+        type=str,
+        default=None,
+        help="Path to npy file of the number of instance per class in training data",
+    )
+    parser.add_argument(
+        "--val-ipc",
+        type=str,
+        default=None,
+        help="Path to npy file of the number of instance per class in validation data",
+    )
+    parser.add_argument(
+        "--train-num-samples",
+        type=int,
+        default=None,
+        help="Number of samples in dataset. Required for webdataset if not available in info file.",
+    )
+    parser.add_argument(
+        "--val-num-samples",
+        type=int,
+        default=None,
+        help="Number of samples in dataset. Useful for webdataset if not available in info file.",
+    )
+    parser.add_argument(
+        "--dataset-type",
+        choices=["webdataset", "csv", "auto", "toy"],
+        default="auto",
+        help="Which type of dataset to process.",
+    )
+    parser.add_argument(
+        "--csv-separator",
+        type=str,
+        default="\t",
+        help="For csv-like datasets, which separator to use.",
+    )
+    parser.add_argument(
+        "--csv-img-key",
+        type=str,
+        default="filepath",
+        help="For csv-like datasets, the name of the key for the image paths.",
+    )
+    parser.add_argument(
+        "--csv-caption-key",
+        type=str,
+        default="title",
+        help="For csv-like datasets, the name of the key for the captions.",
+    )
+    parser.add_argument(
+        "--imagenet-val",
+        type=str,
+        default=None,
+        help="Path to imagenet val set for conducting zero shot evaluation.",
+    )
+    parser.add_argument(
+        "--imagenet-v2",
+        type=str,
+        default=None,
+        help="Path to imagenet v2 for conducting zero shot evaluation.",
+    )
+    parser.add_argument(
+        "--datasetnames",
+        nargs="+",
+        default=None,
+        help="If loading webdataset, spedify the dataset names to load. Can be some of these: Clotho, audioset, audiocaps, BBCSoundEffects",
+    )
+    parser.add_argument(
+        "--full-train-dataset",
+        nargs="+",
+        default=None,
+        help="Which dataset will be trained with all the subsets. (train+test)",
+    )
+    parser.add_argument(
+        "--exclude-eval-dataset",
+        nargs="+",
+        default=None,
+        help="Which dataset will be excluded with evaluation",
+    )
+    parser.add_argument(
+        "--datasetinfos",
+        nargs="+",
+        default=None,
+        help="If loading webdataset, spedify the dataset types to load. Can be some of these: train, test, valid, unbalanced_train, balanced_train, eval",
+    )
+    parser.add_argument(
+        "--dataset-proportion",
+        type=float,
+        default=1.0,
+        help="How much proportion of dataset we want to train.",
+    )
+    parser.add_argument(
+        "--remotedata",
+        default=False,
+        action="store_true",
+        help="if the dataset is remote, set this flag",
+    )
+    parser.add_argument(
+        "--class-label-path",
+        type=str,
+        default=None,
+        help="The path of the class label pickle or csv.",
+    )
+    parser.add_argument(
+        "--datasetpath",
+        type=str,
+        default="/mnt/audio_clip/webdataset_tar",
+        help="The path to the dataset",
+    )
+    parser.add_argument(
+        "--logs",
+        type=str,
+        default="./logs/",
+        help="Where to store tensorboard logs. Use None to avoid storing logs.",
+    )
+    parser.add_argument(
+        "--log-local",
+        action="store_true",
+        default=False,
+        help="log files on local master, otherwise global master only.",
+    )
+    parser.add_argument(
+        "--name",
+        type=str,
+        default=None,
+        help="Optional identifier for the experiment when storing logs. Otherwise use current time.",
+    )
+    parser.add_argument(
+        "--workers", type=int, default=1, help="Number of workers per GPU."
+    )
+    parser.add_argument(
+        "--batch-size", type=int, default=64, help="Batch size per GPU."
+    )
+    parser.add_argument(
+        "--epochs", type=int, default=32, help="Number of epochs to train for."
+    )
+    parser.add_argument("--lr", type=float, default=None, help="Learning rate.")
+    parser.add_argument("--beta1", type=float, default=None, help="Adam beta 1.")
+    parser.add_argument("--beta2", type=float, default=None, help="Adam beta 2.")
+    parser.add_argument("--eps", type=float, default=None, help="Adam epsilon.")
+    parser.add_argument("--momentum", type=float, default=None, help="SGD epsilon.")
+    parser.add_argument("--wd", type=float, default=0.2, help="Weight decay.")
+
+    parser.add_argument(
+        "--split-opt",
+        action="store_true",
+        default=False,
+        help="Use this flag to skip the learning rate decay.",
+    )
+    parser.add_argument(
+        "--lr-pretrained", type=float, default=None, help="Learning rate for text."
+    )
+    parser.add_argument(
+        "--beta1-pretrained", type=float, default=None, help="Adam beta 1 for text."
+    )
+    parser.add_argument(
+        "--beta2-pretrained", type=float, default=None, help="Adam beta 2 for text."
+    )
+    parser.add_argument(
+        "--eps-pretrained", type=float, default=None, help="Adam epsilon for text."
+    )
+    parser.add_argument(
+        "--wd-pretrained", type=float, default=0.2, help="Weight decay for text."
+    )
+    parser.add_argument(
+        "--momentum-pretrained", type=float, default=0.9, help="Momentum for text."
+    )
+    parser.add_argument(
+        "--lr-new", type=float, default=None, help="Learning rate for audio."
+    )
+    parser.add_argument(
+        "--beta1-new", type=float, default=None, help="Adam beta 1 for audio."
+    )
+    parser.add_argument(
+        "--beta2-new", type=float, default=None, help="Adam beta 2 for audio."
+    )
+    parser.add_argument(
+        "--eps-new", type=float, default=None, help="Adam epsilon for audio."
+    )
+    parser.add_argument(
+        "--wd-new", type=float, default=0.2, help="Weight decay for audio."
+    )
+    parser.add_argument(
+        "--momentum-new", type=float, default=0.9, help="Momentum for audio."
+    )
+    parser.add_argument(
+        "--warmup", type=int, default=10000, help="Number of steps to warmup for."
+    )
+    parser.add_argument(
+        "--use-bn-sync",
+        default=False,
+        action="store_true",
+        help="Whether to use batch norm sync.",
+    )
+    parser.add_argument(
+        "--skip-scheduler",
+        action="store_true",
+        default=False,
+        help="Use this flag to skip the learning rate decay.",
+    )
+    parser.add_argument(
+        "--save-frequency", type=int, default=1, help="How often to save checkpoints."
+    )
+    parser.add_argument(
+        "--save-top-performance",
+        type=int,
+        default=0,
+        help="Save the top x performance weights if the value >0",
+    )
+    parser.add_argument(
+        "--save-most-recent",
+        action="store_true",
+        default=False,
+        help="Always save the most recent model trained to epoch_latest.pt.",
+    )
+    parser.add_argument(
+        "--zeroshot-frequency", type=int, default=2, help="How often to run zero shot."
+    )
+    parser.add_argument(
+        "--val-frequency",
+        type=int,
+        default=1,
+        help="How often to run evaluation with val data.",
+    )
+    parser.add_argument(
+        "--resume",
+        default=None,
+        type=str,
+        help="path to latest checkpoint (default: none)",
+    )
+    parser.add_argument(
+        "--precision",
+        choices=["amp", "fp16", "fp32"],
+        default="amp",
+        help="Floating point precision.",
+    )
+    parser.add_argument(
+        "--amodel",
+        type=str,
+        default="RN50",
+        help="Name of the audio backbone to use.",
+    )
+    parser.add_argument(
+        "--tmodel",
+        type=str,
+        default="transformer",
+        help="Name of the text backbone to use. Can be [transformer, bert, roberta, bart]",
+    )
+    parser.add_argument(
+        "--pretrained-audio",
+        default="",
+        type=str,
+        help="Use a pretrained audio model weights for the audio encoder of CLAP",
+    )
+    parser.add_argument(
+        "--pretrained-text",
+        default="",
+        type=str,
+        help="Use a pretrained text model weights for the text encoder of CLAP",
+    )
+    parser.add_argument(
+        "--pretrained",
+        default="",
+        type=str,
+        help="Use a pretrained CLIP model weights with the specified tag or file path.",
+    )
+    parser.add_argument(
+        "--pretrained-image",
+        default=False,
+        action="store_true",
+        help="Load imagenet pretrained weights for image tower backbone if available.",
+    )
+    parser.add_argument(
+        "--lock-image",
+        default=False,
+        action="store_true",
+        help="Lock full image tower by disabling gradients.",
+    )
+    parser.add_argument(
+        "--lock-image-unlocked-groups",
+        type=int,
+        default=0,
+        help="Leave last n image tower layer groups unlocked.",
+    )
+    parser.add_argument(
+        "--lock-image-freeze-bn-stats",
+        default=False,
+        action="store_true",
+        help="Freeze BatchNorm running stats in image tower for any locked layers.",
+    )
+    parser.add_argument(
+        "--local-loss",
+        default=False,
+        action="store_true",
+        help="calculate loss w/ local features @ global (instead of realizing full global @ global matrix)",
+    )
+    parser.add_argument(
+        "--gather-with-grad",
+        default=False,
+        action="store_true",
+        help="enable full distributed gradient for feature gather",
+    )
+    parser.add_argument(
+        "--force-quick-gelu",
+        default=False,
+        action="store_true",
+        help="Force use of QuickGELU activation for non-OpenAI transformer models.",
+    )
+    parser.add_argument(
+        "--torchscript",
+        default=False,
+        action="store_true",
+        help="torch.jit.script the model, also uses jit version of OpenAI models if pretrained=='openai'",
+    )
+    parser.add_argument(
+        "--trace",
+        default=False,
+        action="store_true",
+        help="torch.jit.trace the model for inference / eval only",
+    )
+    # arguments for distributed training
+    parser.add_argument(
+        "--dist-url",
+        default="env://",
+        type=str,
+        help="url used to set up distributed training",
+    )
+    parser.add_argument(
+        "--dist-backend", default="nccl", type=str, help="distributed backend"
+    )
+    parser.add_argument(
+        "--report-to",
+        default="",
+        type=str,
+        help="Options are ['wandb', 'tensorboard', 'wandb,tensorboard']",
+    )
+    parser.add_argument(
+        "--wandb-notes", default="", type=str, help="Notes if logging with wandb"
+    )
+    parser.add_argument(
+        "--C", type=float, default=3.16, help="inverse regularizer for logistic reg."
+    )
+    parser.add_argument(
+        "--debug",
+        default=False,
+        action="store_true",
+        help="If true, more information is logged.",
+    )
+    parser.add_argument(
+        "--copy-codebase",
+        default=False,
+        action="store_true",
+        help="If true, we copy the entire base on the log diretory, and execute from there.",
+    )
+    parser.add_argument(
+        "--horovod",
+        default=False,
+        action="store_true",
+        help="Use horovod for distributed training.",
+    )
+    parser.add_argument(
+        "--ddp-static-graph",
+        default=False,
+        action="store_true",
+        help="Enable static graph optimization for DDP in PyTorch >= 1.11.",
+    )
+    parser.add_argument(
+        "--no-set-device-rank",
+        default=False,
+        action="store_true",
+        help="Don't set device index from local rank (when CUDA_VISIBLE_DEVICES restricted to one per proc).",
+    )
+    parser.add_argument("--seed", type=int, default=4242, help="Default random seed.")
+
+    parser.add_argument(
+        "--top-k-checkpoint-select-dataset",
+        type=str,
+        default="all",
+        help="The dataset of selecting top-k checkpoint.",
+    )
+
+    # @R10, @R@5, @R1, mAP@10
+    parser.add_argument(
+        "--top-k-checkpoint-select-metric",
+        type=str,
+        default="_R@10",
+        help="The metric for selecting top-k checkpoint.",
+    )
+    parser.add_argument(
+        "--openai-model-cache-dir",
+        type=str,
+        default="~/.cache/clip",
+        help="Directory to download OpenAI models.",
+    )
+    parser.add_argument(
+        "--optimizer",
+        type=str,
+        default="adamw",
+        help="can be AdamW or SGD",
+    )
+    parser.add_argument(
+        "--parallel-eval",
+        default=False,
+        action="store_true",
+        help="Eval in parallel (multi-GPU, multi-node).",
+    )
+
+    parser.add_argument(
+        "--no-eval",
+        default=False,
+        action="store_true",
+        help="Training without evaluation.",
+    )
+
+    parser.add_argument(
+        "--lp-mlp",
+        default=False,
+        action="store_true",
+        help="Linear Probe using MLP layer or not.",
+    )
+
+    parser.add_argument(
+        "--lp-freeze",
+        default=False,
+        action="store_true",
+        help="Linear Probe using Freeze CLAP or not",
+    )
+
+    parser.add_argument(
+        "--lp-act",
+        default="None",
+        type=str,
+        help="Options are ['relu','elu','prelu','softmax','sigmoid']",
+    )
+
+    parser.add_argument(
+        "--lp-loss", type=str, default="bce", help="Loss func of Linear Probe."
+    )
+
+    parser.add_argument(
+        "--lp-metrics",
+        type=str,
+        default="map,mauc,acc",
+        help="Metrics of Linear Probe.",
+    )
+
+    parser.add_argument(
+        "--lp-lr", type=float, default=1e-4, help="learning rate of linear probe"
+    )
+    parser.add_argument(
+        "--kappa",
+        type=float,
+        default=0,
+        help="the kappa in the weighted contrastive loss, default is to turn off the weighted contrastive loss",
+    )
+
+    parser.add_argument(
+        "--data-filling",
+        type=str,
+        default="pad",
+        help="type of data filling when the audio length is shorter than the max length."
+        "Can be one of the following: repeat, repeatpad, pad",
+    )
+    parser.add_argument(
+        "--data-truncating",
+        type=str,
+        default="rand_trunc",
+        help="type of data truncation when the audio length is longer than the max length."
+        "Can be one of the following: rand_trunc, fusion",
+    )
+
+    parser.add_argument(
+        "--clap-mlploss",
+        default=False,
+        action="store_true",
+        help="Using MLP loss for CLAP model or not",
+    )
+
+    parser.add_argument(
+        "--wandb-id",
+        type=str,
+        default=None,
+        help="the id of wandb experiment to restore.",
+    )
+
+    parser.add_argument(
+        "--sleep", type=float, default=0, help="sleep n seconds before start training"
+    )
+
+    # variable length processing
+    parser.add_argument(
+        "--enable-fusion",
+        default=False,
+        action="store_true",
+        help="Enable feature funsion for variable-length data",
+    )
+
+    parser.add_argument(
+        "--fusion-type",
+        type=str,
+        default="None",
+        help="Type is among ['channel_map', 'daf_1d','aff_1d','iaff_1d','daf_2d','aff_2d','iaff_2d']",
+    )
+
+    parser.add_argument(
+        "--mixup",
+        default=False,
+        action="store_true",
+        help="Enable mixup in finetuning training.",
+    )
+    parser.add_argument(
+        "--text-augment-selection",
+        type=str,
+        default=None,
+        help="For selecting levels of augmented text. Type is among ['all', 'augment_only', 'none']",
+    )
+
+    args = parser.parse_args()
+
+    # If some params are not passed, we use the default values based on model name.
+    default_params = get_default_params(args.amodel)
+    for name, val in default_params.items():
+        if getattr(args, name) is None:
+            setattr(args, name, val)
+
+    return args

audiosr/hifigan/LICENSE

@@ -1,21 +1,21 @@
-MIT License
-
-Copyright (c) 2020 Jungil Kong
-
-Permission is hereby granted, free of charge, to any person obtaining a copy
-of this software and associated documentation files (the "Software"), to deal
-in the Software without restriction, including without limitation the rights
-to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
-copies of the Software, and to permit persons to whom the Software is
-furnished to do so, subject to the following conditions:
-
-The above copyright notice and this permission notice shall be included in all
-copies or substantial portions of the Software.
-
-THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
-IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
-FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
-AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
-LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
-OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+MIT License
+
+Copyright (c) 2020 Jungil Kong
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
 SOFTWARE.

audiosr/hifigan/__init__.py

@@ -1,8 +1,8 @@
-from .models_v2 import Generator
-from .models import Generator as Generator_old
-
-
-class AttrDict(dict):
-    def __init__(self, *args, **kwargs):
-        super(AttrDict, self).__init__(*args, **kwargs)
-        self.__dict__ = self
+from .models_v2 import Generator
+from .models import Generator as Generator_old
+
+
+class AttrDict(dict):
+    def __init__(self, *args, **kwargs):
+        super(AttrDict, self).__init__(*args, **kwargs)
+        self.__dict__ = self

audiosr/hifigan/models.py

@@ -1,174 +1,174 @@
-import torch
-import torch.nn as nn
-import torch.nn.functional as F
-from torch.nn import Conv1d, ConvTranspose1d
-from torch.nn.utils import weight_norm, remove_weight_norm
-
-LRELU_SLOPE = 0.1
-
-
-def init_weights(m, mean=0.0, std=0.01):
-    classname = m.__class__.__name__
-    if classname.find("Conv") != -1:
-        m.weight.data.normal_(mean, std)
-
-
-def get_padding(kernel_size, dilation=1):
-    return int((kernel_size * dilation - dilation) / 2)
-
-
-class ResBlock(torch.nn.Module):
-    def __init__(self, h, channels, kernel_size=3, dilation=(1, 3, 5)):
-        super(ResBlock, self).__init__()
-        self.h = h
-        self.convs1 = nn.ModuleList(
-            [
-                weight_norm(
-                    Conv1d(
-                        channels,
-                        channels,
-                        kernel_size,
-                        1,
-                        dilation=dilation[0],
-                        padding=get_padding(kernel_size, dilation[0]),
-                    )
-                ),
-                weight_norm(
-                    Conv1d(
-                        channels,
-                        channels,
-                        kernel_size,
-                        1,
-                        dilation=dilation[1],
-                        padding=get_padding(kernel_size, dilation[1]),
-                    )
-                ),
-                weight_norm(
-                    Conv1d(
-                        channels,
-                        channels,
-                        kernel_size,
-                        1,
-                        dilation=dilation[2],
-                        padding=get_padding(kernel_size, dilation[2]),
-                    )
-                ),
-            ]
-        )
-        self.convs1.apply(init_weights)
-
-        self.convs2 = nn.ModuleList(
-            [
-                weight_norm(
-                    Conv1d(
-                        channels,
-                        channels,
-                        kernel_size,
-                        1,
-                        dilation=1,
-                        padding=get_padding(kernel_size, 1),
-                    )
-                ),
-                weight_norm(
-                    Conv1d(
-                        channels,
-                        channels,
-                        kernel_size,
-                        1,
-                        dilation=1,
-                        padding=get_padding(kernel_size, 1),
-                    )
-                ),
-                weight_norm(
-                    Conv1d(
-                        channels,
-                        channels,
-                        kernel_size,
-                        1,
-                        dilation=1,
-                        padding=get_padding(kernel_size, 1),
-                    )
-                ),
-            ]
-        )
-        self.convs2.apply(init_weights)
-
-    def forward(self, x):
-        for c1, c2 in zip(self.convs1, self.convs2):
-            xt = F.leaky_relu(x, LRELU_SLOPE)
-            xt = c1(xt)
-            xt = F.leaky_relu(xt, LRELU_SLOPE)
-            xt = c2(xt)
-            x = xt + x
-        return x
-
-    def remove_weight_norm(self):
-        for l in self.convs1:
-            remove_weight_norm(l)
-        for l in self.convs2:
-            remove_weight_norm(l)
-
-
-class Generator(torch.nn.Module):
-    def __init__(self, h):
-        super(Generator, self).__init__()
-        self.h = h
-        self.num_kernels = len(h.resblock_kernel_sizes)
-        self.num_upsamples = len(h.upsample_rates)
-        self.conv_pre = weight_norm(
-            Conv1d(h.num_mels, h.upsample_initial_channel, 7, 1, padding=3)
-        )
-        resblock = ResBlock
-
-        self.ups = nn.ModuleList()
-        for i, (u, k) in enumerate(zip(h.upsample_rates, h.upsample_kernel_sizes)):
-            self.ups.append(
-                weight_norm(
-                    ConvTranspose1d(
-                        h.upsample_initial_channel // (2**i),
-                        h.upsample_initial_channel // (2 ** (i + 1)),
-                        k,
-                        u,
-                        padding=(k - u) // 2,
-                    )
-                )
-            )
-
-        self.resblocks = nn.ModuleList()
-        for i in range(len(self.ups)):
-            ch = h.upsample_initial_channel // (2 ** (i + 1))
-            for j, (k, d) in enumerate(
-                zip(h.resblock_kernel_sizes, h.resblock_dilation_sizes)
-            ):
-                self.resblocks.append(resblock(h, ch, k, d))
-
-        self.conv_post = weight_norm(Conv1d(ch, 1, 7, 1, padding=3))
-        self.ups.apply(init_weights)
-        self.conv_post.apply(init_weights)
-
-    def forward(self, x):
-        x = self.conv_pre(x)
-        for i in range(self.num_upsamples):
-            x = F.leaky_relu(x, LRELU_SLOPE)
-            x = self.ups[i](x)
-            xs = None
-            for j in range(self.num_kernels):
-                if xs is None:
-                    xs = self.resblocks[i * self.num_kernels + j](x)
-                else:
-                    xs += self.resblocks[i * self.num_kernels + j](x)
-            x = xs / self.num_kernels
-        x = F.leaky_relu(x)
-        x = self.conv_post(x)
-        x = torch.tanh(x)
-
-        return x
-
-    def remove_weight_norm(self):
-        # print("Removing weight norm...")
-        for l in self.ups:
-            remove_weight_norm(l)
-        for l in self.resblocks:
-            l.remove_weight_norm()
-        remove_weight_norm(self.conv_pre)
-        remove_weight_norm(self.conv_post)
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torch.nn import Conv1d, ConvTranspose1d
+from torch.nn.utils import weight_norm, remove_weight_norm
+
+LRELU_SLOPE = 0.1
+
+
+def init_weights(m, mean=0.0, std=0.01):
+    classname = m.__class__.__name__
+    if classname.find("Conv") != -1:
+        m.weight.data.normal_(mean, std)
+
+
+def get_padding(kernel_size, dilation=1):
+    return int((kernel_size * dilation - dilation) / 2)
+
+
+class ResBlock(torch.nn.Module):
+    def __init__(self, h, channels, kernel_size=3, dilation=(1, 3, 5)):
+        super(ResBlock, self).__init__()
+        self.h = h
+        self.convs1 = nn.ModuleList(
+            [
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=dilation[0],
+                        padding=get_padding(kernel_size, dilation[0]),
+                    )
+                ),
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=dilation[1],
+                        padding=get_padding(kernel_size, dilation[1]),
+                    )
+                ),
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=dilation[2],
+                        padding=get_padding(kernel_size, dilation[2]),
+                    )
+                ),
+            ]
+        )
+        self.convs1.apply(init_weights)
+
+        self.convs2 = nn.ModuleList(
+            [
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=1,
+                        padding=get_padding(kernel_size, 1),
+                    )
+                ),
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=1,
+                        padding=get_padding(kernel_size, 1),
+                    )
+                ),
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=1,
+                        padding=get_padding(kernel_size, 1),
+                    )
+                ),
+            ]
+        )
+        self.convs2.apply(init_weights)
+
+    def forward(self, x):
+        for c1, c2 in zip(self.convs1, self.convs2):
+            xt = F.leaky_relu(x, LRELU_SLOPE)
+            xt = c1(xt)
+            xt = F.leaky_relu(xt, LRELU_SLOPE)
+            xt = c2(xt)
+            x = xt + x
+        return x
+
+    def remove_weight_norm(self):
+        for l in self.convs1:
+            remove_weight_norm(l)
+        for l in self.convs2:
+            remove_weight_norm(l)
+
+
+class Generator(torch.nn.Module):
+    def __init__(self, h):
+        super(Generator, self).__init__()
+        self.h = h
+        self.num_kernels = len(h.resblock_kernel_sizes)
+        self.num_upsamples = len(h.upsample_rates)
+        self.conv_pre = weight_norm(
+            Conv1d(h.num_mels, h.upsample_initial_channel, 7, 1, padding=3)
+        )
+        resblock = ResBlock
+
+        self.ups = nn.ModuleList()
+        for i, (u, k) in enumerate(zip(h.upsample_rates, h.upsample_kernel_sizes)):
+            self.ups.append(
+                weight_norm(
+                    ConvTranspose1d(
+                        h.upsample_initial_channel // (2**i),
+                        h.upsample_initial_channel // (2 ** (i + 1)),
+                        k,
+                        u,
+                        padding=(k - u) // 2,
+                    )
+                )
+            )
+
+        self.resblocks = nn.ModuleList()
+        for i in range(len(self.ups)):
+            ch = h.upsample_initial_channel // (2 ** (i + 1))
+            for j, (k, d) in enumerate(
+                zip(h.resblock_kernel_sizes, h.resblock_dilation_sizes)
+            ):
+                self.resblocks.append(resblock(h, ch, k, d))
+
+        self.conv_post = weight_norm(Conv1d(ch, 1, 7, 1, padding=3))
+        self.ups.apply(init_weights)
+        self.conv_post.apply(init_weights)
+
+    def forward(self, x):
+        x = self.conv_pre(x)
+        for i in range(self.num_upsamples):
+            x = F.leaky_relu(x, LRELU_SLOPE)
+            x = self.ups[i](x)
+            xs = None
+            for j in range(self.num_kernels):
+                if xs is None:
+                    xs = self.resblocks[i * self.num_kernels + j](x)
+                else:
+                    xs += self.resblocks[i * self.num_kernels + j](x)
+            x = xs / self.num_kernels
+        x = F.leaky_relu(x)
+        x = self.conv_post(x)
+        x = torch.tanh(x)
+
+        return x
+
+    def remove_weight_norm(self):
+        # print("Removing weight norm...")
+        for l in self.ups:
+            remove_weight_norm(l)
+        for l in self.resblocks:
+            l.remove_weight_norm()
+        remove_weight_norm(self.conv_pre)
+        remove_weight_norm(self.conv_post)

audiosr/hifigan/models_v2.py

@@ -1,395 +1,395 @@
-import torch
-import torch.nn.functional as F
-import torch.nn as nn
-from torch.nn import Conv1d, ConvTranspose1d
-from torch.nn.utils import weight_norm, remove_weight_norm
-
-LRELU_SLOPE = 0.1
-
-
-def init_weights(m, mean=0.0, std=0.01):
-    classname = m.__class__.__name__
-    if classname.find("Conv") != -1:
-        m.weight.data.normal_(mean, std)
-
-
-def get_padding(kernel_size, dilation=1):
-    return int((kernel_size * dilation - dilation) / 2)
-
-
-class ResBlock1(torch.nn.Module):
-    def __init__(self, h, channels, kernel_size=3, dilation=(1, 3, 5)):
-        super(ResBlock1, self).__init__()
-        self.h = h
-        self.convs1 = nn.ModuleList(
-            [
-                weight_norm(
-                    Conv1d(
-                        channels,
-                        channels,
-                        kernel_size,
-                        1,
-                        dilation=dilation[0],
-                        padding=get_padding(kernel_size, dilation[0]),
-                    )
-                ),
-                weight_norm(
-                    Conv1d(
-                        channels,
-                        channels,
-                        kernel_size,
-                        1,
-                        dilation=dilation[1],
-                        padding=get_padding(kernel_size, dilation[1]),
-                    )
-                ),
-                weight_norm(
-                    Conv1d(
-                        channels,
-                        channels,
-                        kernel_size,
-                        1,
-                        dilation=dilation[2],
-                        padding=get_padding(kernel_size, dilation[2]),
-                    )
-                ),
-            ]
-        )
-        self.convs1.apply(init_weights)
-
-        self.convs2 = nn.ModuleList(
-            [
-                weight_norm(
-                    Conv1d(
-                        channels,
-                        channels,
-                        kernel_size,
-                        1,
-                        dilation=1,
-                        padding=get_padding(kernel_size, 1),
-                    )
-                ),
-                weight_norm(
-                    Conv1d(
-                        channels,
-                        channels,
-                        kernel_size,
-                        1,
-                        dilation=1,
-                        padding=get_padding(kernel_size, 1),
-                    )
-                ),
-                weight_norm(
-                    Conv1d(
-                        channels,
-                        channels,
-                        kernel_size,
-                        1,
-                        dilation=1,
-                        padding=get_padding(kernel_size, 1),
-                    )
-                ),
-            ]
-        )
-        self.convs2.apply(init_weights)
-
-    def forward(self, x):
-        for c1, c2 in zip(self.convs1, self.convs2):
-            xt = F.leaky_relu(x, LRELU_SLOPE)
-            xt = c1(xt)
-            xt = F.leaky_relu(xt, LRELU_SLOPE)
-            xt = c2(xt)
-            x = xt + x
-        return x
-
-    def remove_weight_norm(self):
-        for l in self.convs1:
-            remove_weight_norm(l)
-        for l in self.convs2:
-            remove_weight_norm(l)
-
-
-class ResBlock2(torch.nn.Module):
-    def __init__(self, h, channels, kernel_size=3, dilation=(1, 3)):
-        super(ResBlock2, self).__init__()
-        self.h = h
-        self.convs = nn.ModuleList(
-            [
-                weight_norm(
-                    Conv1d(
-                        channels,
-                        channels,
-                        kernel_size,
-                        1,
-                        dilation=dilation[0],
-                        padding=get_padding(kernel_size, dilation[0]),
-                    )
-                ),
-                weight_norm(
-                    Conv1d(
-                        channels,
-                        channels,
-                        kernel_size,
-                        1,
-                        dilation=dilation[1],
-                        padding=get_padding(kernel_size, dilation[1]),
-                    )
-                ),
-            ]
-        )
-        self.convs.apply(init_weights)
-
-    def forward(self, x):
-        for c in self.convs:
-            xt = F.leaky_relu(x, LRELU_SLOPE)
-            xt = c(xt)
-            x = xt + x
-        return x
-
-    def remove_weight_norm(self):
-        for l in self.convs:
-            remove_weight_norm(l)
-
-
-class Generator(torch.nn.Module):
-    def __init__(self, h):
-        super(Generator, self).__init__()
-        self.h = h
-        self.num_kernels = len(h.resblock_kernel_sizes)
-        self.num_upsamples = len(h.upsample_rates)
-        self.conv_pre = weight_norm(
-            Conv1d(256, h.upsample_initial_channel, 7, 1, padding=3)
-        )
-        resblock = ResBlock1 if h.resblock == "1" else ResBlock2
-
-        self.ups = nn.ModuleList()
-        for i, (u, k) in enumerate(zip(h.upsample_rates, h.upsample_kernel_sizes)):
-            self.ups.append(
-                weight_norm(
-                    ConvTranspose1d(
-                        h.upsample_initial_channel // (2**i),
-                        h.upsample_initial_channel // (2 ** (i + 1)),
-                        u * 2,
-                        u,
-                        padding=u // 2 + u % 2,
-                        output_padding=u % 2,
-                    )
-                )
-            )
-
-        self.resblocks = nn.ModuleList()
-        for i in range(len(self.ups)):
-            ch = h.upsample_initial_channel // (2 ** (i + 1))
-            for j, (k, d) in enumerate(
-                zip(h.resblock_kernel_sizes, h.resblock_dilation_sizes)
-            ):
-                self.resblocks.append(resblock(h, ch, k, d))
-
-        self.conv_post = weight_norm(Conv1d(ch, 1, 7, 1, padding=3))
-        self.ups.apply(init_weights)
-        self.conv_post.apply(init_weights)
-
-    def forward(self, x):
-        # import ipdb; ipdb.set_trace()
-        x = self.conv_pre(x)
-        for i in range(self.num_upsamples):
-            x = F.leaky_relu(x, LRELU_SLOPE)
-            x = self.ups[i](x)
-            xs = None
-            for j in range(self.num_kernels):
-                if xs is None:
-                    xs = self.resblocks[i * self.num_kernels + j](x)
-                else:
-                    xs += self.resblocks[i * self.num_kernels + j](x)
-            x = xs / self.num_kernels
-        x = F.leaky_relu(x)
-        x = self.conv_post(x)
-        x = torch.tanh(x)
-
-        return x
-
-    def remove_weight_norm(self):
-        # print('Removing weight norm...')
-        for l in self.ups:
-            remove_weight_norm(l)
-        for l in self.resblocks:
-            l.remove_weight_norm()
-        remove_weight_norm(self.conv_pre)
-        remove_weight_norm(self.conv_post)
-
-
-##################################################################################################
-
-# import torch
-# import torch.nn as nn
-# import torch.nn.functional as F
-# from torch.nn import Conv1d, ConvTranspose1d
-# from torch.nn.utils import weight_norm, remove_weight_norm
-
-# LRELU_SLOPE = 0.1
-
-
-# def init_weights(m, mean=0.0, std=0.01):
-#     classname = m.__class__.__name__
-#     if classname.find("Conv") != -1:
-#         m.weight.data.normal_(mean, std)
-
-
-# def get_padding(kernel_size, dilation=1):
-#     return int((kernel_size * dilation - dilation) / 2)
-
-
-# class ResBlock(torch.nn.Module):
-#     def __init__(self, h, channels, kernel_size=3, dilation=(1, 3, 5)):
-#         super(ResBlock, self).__init__()
-#         self.h = h
-#         self.convs1 = nn.ModuleList(
-#             [
-#                 weight_norm(
-#                     Conv1d(
-#                         channels,
-#                         channels,
-#                         kernel_size,
-#                         1,
-#                         dilation=dilation[0],
-#                         padding=get_padding(kernel_size, dilation[0]),
-#                     )
-#                 ),
-#                 weight_norm(
-#                     Conv1d(
-#                         channels,
-#                         channels,
-#                         kernel_size,
-#                         1,
-#                         dilation=dilation[1],
-#                         padding=get_padding(kernel_size, dilation[1]),
-#                     )
-#                 ),
-#                 weight_norm(
-#                     Conv1d(
-#                         channels,
-#                         channels,
-#                         kernel_size,
-#                         1,
-#                         dilation=dilation[2],
-#                         padding=get_padding(kernel_size, dilation[2]),
-#                     )
-#                 ),
-#             ]
-#         )
-#         self.convs1.apply(init_weights)
-
-#         self.convs2 = nn.ModuleList(
-#             [
-#                 weight_norm(
-#                     Conv1d(
-#                         channels,
-#                         channels,
-#                         kernel_size,
-#                         1,
-#                         dilation=1,
-#                         padding=get_padding(kernel_size, 1),
-#                     )
-#                 ),
-#                 weight_norm(
-#                     Conv1d(
-#                         channels,
-#                         channels,
-#                         kernel_size,
-#                         1,
-#                         dilation=1,
-#                         padding=get_padding(kernel_size, 1),
-#                     )
-#                 ),
-#                 weight_norm(
-#                     Conv1d(
-#                         channels,
-#                         channels,
-#                         kernel_size,
-#                         1,
-#                         dilation=1,
-#                         padding=get_padding(kernel_size, 1),
-#                     )
-#                 ),
-#             ]
-#         )
-#         self.convs2.apply(init_weights)
-
-#     def forward(self, x):
-#         for c1, c2 in zip(self.convs1, self.convs2):
-#             xt = F.leaky_relu(x, LRELU_SLOPE)
-#             xt = c1(xt)
-#             xt = F.leaky_relu(xt, LRELU_SLOPE)
-#             xt = c2(xt)
-#             x = xt + x
-#         return x
-
-#     def remove_weight_norm(self):
-#         for l in self.convs1:
-#             remove_weight_norm(l)
-#         for l in self.convs2:
-#             remove_weight_norm(l)
-
-# class Generator(torch.nn.Module):
-#     def __init__(self, h):
-#         super(Generator, self).__init__()
-#         self.h = h
-#         self.num_kernels = len(h.resblock_kernel_sizes)
-#         self.num_upsamples = len(h.upsample_rates)
-#         self.conv_pre = weight_norm(
-#             Conv1d(h.num_mels, h.upsample_initial_channel, 7, 1, padding=3)
-#         )
-#         resblock = ResBlock
-
-#         self.ups = nn.ModuleList()
-#         for i, (u, k) in enumerate(zip(h.upsample_rates, h.upsample_kernel_sizes)):
-#             self.ups.append(
-#                 weight_norm(
-#                     ConvTranspose1d(
-#                         h.upsample_initial_channel // (2**i),
-#                         h.upsample_initial_channel // (2 ** (i + 1)),
-#                         k,
-#                         u,
-#                         padding=(k - u) // 2,
-#                     )
-#                 )
-#             )
-
-#         self.resblocks = nn.ModuleList()
-#         for i in range(len(self.ups)):
-#             ch = h.upsample_initial_channel // (2 ** (i + 1))
-#             for j, (k, d) in enumerate(
-#                 zip(h.resblock_kernel_sizes, h.resblock_dilation_sizes)
-#             ):
-#                 self.resblocks.append(resblock(h, ch, k, d))
-
-#         self.conv_post = weight_norm(Conv1d(ch, 1, 7, 1, padding=3))
-#         self.ups.apply(init_weights)
-#         self.conv_post.apply(init_weights)
-
-#     def forward(self, x):
-#         x = self.conv_pre(x)
-#         for i in range(self.num_upsamples):
-#             x = F.leaky_relu(x, LRELU_SLOPE)
-#             x = self.ups[i](x)
-#             xs = None
-#             for j in range(self.num_kernels):
-#                 if xs is None:
-#                     xs = self.resblocks[i * self.num_kernels + j](x)
-#                 else:
-#                     xs += self.resblocks[i * self.num_kernels + j](x)
-#             x = xs / self.num_kernels
-#         x = F.leaky_relu(x)
-#         x = self.conv_post(x)
-#         x = torch.tanh(x)
-
-#         return x
-
-#     def remove_weight_norm(self):
-#         print("Removing weight norm...")
-#         for l in self.ups:
-#             remove_weight_norm(l)
-#         for l in self.resblocks:
-#             l.remove_weight_norm()
-#         remove_weight_norm(self.conv_pre)
-#         remove_weight_norm(self.conv_post)
+import torch
+import torch.nn.functional as F
+import torch.nn as nn
+from torch.nn import Conv1d, ConvTranspose1d
+from torch.nn.utils import weight_norm, remove_weight_norm
+
+LRELU_SLOPE = 0.1
+
+
+def init_weights(m, mean=0.0, std=0.01):
+    classname = m.__class__.__name__
+    if classname.find("Conv") != -1:
+        m.weight.data.normal_(mean, std)
+
+
+def get_padding(kernel_size, dilation=1):
+    return int((kernel_size * dilation - dilation) / 2)
+
+
+class ResBlock1(torch.nn.Module):
+    def __init__(self, h, channels, kernel_size=3, dilation=(1, 3, 5)):
+        super(ResBlock1, self).__init__()
+        self.h = h
+        self.convs1 = nn.ModuleList(
+            [
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=dilation[0],
+                        padding=get_padding(kernel_size, dilation[0]),
+                    )
+                ),
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=dilation[1],
+                        padding=get_padding(kernel_size, dilation[1]),
+                    )
+                ),
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=dilation[2],
+                        padding=get_padding(kernel_size, dilation[2]),
+                    )
+                ),
+            ]
+        )
+        self.convs1.apply(init_weights)
+
+        self.convs2 = nn.ModuleList(
+            [
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=1,
+                        padding=get_padding(kernel_size, 1),
+                    )
+                ),
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=1,
+                        padding=get_padding(kernel_size, 1),
+                    )
+                ),
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=1,
+                        padding=get_padding(kernel_size, 1),
+                    )
+                ),
+            ]
+        )
+        self.convs2.apply(init_weights)
+
+    def forward(self, x):
+        for c1, c2 in zip(self.convs1, self.convs2):
+            xt = F.leaky_relu(x, LRELU_SLOPE)
+            xt = c1(xt)
+            xt = F.leaky_relu(xt, LRELU_SLOPE)
+            xt = c2(xt)
+            x = xt + x
+        return x
+
+    def remove_weight_norm(self):
+        for l in self.convs1:
+            remove_weight_norm(l)
+        for l in self.convs2:
+            remove_weight_norm(l)
+
+
+class ResBlock2(torch.nn.Module):
+    def __init__(self, h, channels, kernel_size=3, dilation=(1, 3)):
+        super(ResBlock2, self).__init__()
+        self.h = h
+        self.convs = nn.ModuleList(
+            [
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=dilation[0],
+                        padding=get_padding(kernel_size, dilation[0]),
+                    )
+                ),
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=dilation[1],
+                        padding=get_padding(kernel_size, dilation[1]),
+                    )
+                ),
+            ]
+        )
+        self.convs.apply(init_weights)
+
+    def forward(self, x):
+        for c in self.convs:
+            xt = F.leaky_relu(x, LRELU_SLOPE)
+            xt = c(xt)
+            x = xt + x
+        return x
+
+    def remove_weight_norm(self):
+        for l in self.convs:
+            remove_weight_norm(l)
+
+
+class Generator(torch.nn.Module):
+    def __init__(self, h):
+        super(Generator, self).__init__()
+        self.h = h
+        self.num_kernels = len(h.resblock_kernel_sizes)
+        self.num_upsamples = len(h.upsample_rates)
+        self.conv_pre = weight_norm(
+            Conv1d(256, h.upsample_initial_channel, 7, 1, padding=3)
+        )
+        resblock = ResBlock1 if h.resblock == "1" else ResBlock2
+
+        self.ups = nn.ModuleList()
+        for i, (u, k) in enumerate(zip(h.upsample_rates, h.upsample_kernel_sizes)):
+            self.ups.append(
+                weight_norm(
+                    ConvTranspose1d(
+                        h.upsample_initial_channel // (2**i),
+                        h.upsample_initial_channel // (2 ** (i + 1)),
+                        u * 2,
+                        u,
+                        padding=u // 2 + u % 2,
+                        output_padding=u % 2,
+                    )
+                )
+            )
+
+        self.resblocks = nn.ModuleList()
+        for i in range(len(self.ups)):
+            ch = h.upsample_initial_channel // (2 ** (i + 1))
+            for j, (k, d) in enumerate(
+                zip(h.resblock_kernel_sizes, h.resblock_dilation_sizes)
+            ):
+                self.resblocks.append(resblock(h, ch, k, d))
+
+        self.conv_post = weight_norm(Conv1d(ch, 1, 7, 1, padding=3))
+        self.ups.apply(init_weights)
+        self.conv_post.apply(init_weights)
+
+    def forward(self, x):
+        # import ipdb; ipdb.set_trace()
+        x = self.conv_pre(x)
+        for i in range(self.num_upsamples):
+            x = F.leaky_relu(x, LRELU_SLOPE)
+            x = self.ups[i](x)
+            xs = None
+            for j in range(self.num_kernels):
+                if xs is None:
+                    xs = self.resblocks[i * self.num_kernels + j](x)
+                else:
+                    xs += self.resblocks[i * self.num_kernels + j](x)
+            x = xs / self.num_kernels
+        x = F.leaky_relu(x)
+        x = self.conv_post(x)
+        x = torch.tanh(x)
+
+        return x
+
+    def remove_weight_norm(self):
+        # print('Removing weight norm...')
+        for l in self.ups:
+            remove_weight_norm(l)
+        for l in self.resblocks:
+            l.remove_weight_norm()
+        remove_weight_norm(self.conv_pre)
+        remove_weight_norm(self.conv_post)
+
+
+##################################################################################################
+
+# import torch
+# import torch.nn as nn
+# import torch.nn.functional as F
+# from torch.nn import Conv1d, ConvTranspose1d
+# from torch.nn.utils import weight_norm, remove_weight_norm
+
+# LRELU_SLOPE = 0.1
+
+
+# def init_weights(m, mean=0.0, std=0.01):
+#     classname = m.__class__.__name__
+#     if classname.find("Conv") != -1:
+#         m.weight.data.normal_(mean, std)
+
+
+# def get_padding(kernel_size, dilation=1):
+#     return int((kernel_size * dilation - dilation) / 2)
+
+
+# class ResBlock(torch.nn.Module):
+#     def __init__(self, h, channels, kernel_size=3, dilation=(1, 3, 5)):
+#         super(ResBlock, self).__init__()
+#         self.h = h
+#         self.convs1 = nn.ModuleList(
+#             [
+#                 weight_norm(
+#                     Conv1d(
+#                         channels,
+#                         channels,
+#                         kernel_size,
+#                         1,
+#                         dilation=dilation[0],
+#                         padding=get_padding(kernel_size, dilation[0]),
+#                     )
+#                 ),
+#                 weight_norm(
+#                     Conv1d(
+#                         channels,
+#                         channels,
+#                         kernel_size,
+#                         1,
+#                         dilation=dilation[1],
+#                         padding=get_padding(kernel_size, dilation[1]),
+#                     )
+#                 ),
+#                 weight_norm(
+#                     Conv1d(
+#                         channels,
+#                         channels,
+#                         kernel_size,
+#                         1,
+#                         dilation=dilation[2],
+#                         padding=get_padding(kernel_size, dilation[2]),
+#                     )
+#                 ),
+#             ]
+#         )
+#         self.convs1.apply(init_weights)
+
+#         self.convs2 = nn.ModuleList(
+#             [
+#                 weight_norm(
+#                     Conv1d(
+#                         channels,
+#                         channels,
+#                         kernel_size,
+#                         1,
+#                         dilation=1,
+#                         padding=get_padding(kernel_size, 1),
+#                     )
+#                 ),
+#                 weight_norm(
+#                     Conv1d(
+#                         channels,
+#                         channels,
+#                         kernel_size,
+#                         1,
+#                         dilation=1,
+#                         padding=get_padding(kernel_size, 1),
+#                     )
+#                 ),
+#                 weight_norm(
+#                     Conv1d(
+#                         channels,
+#                         channels,
+#                         kernel_size,
+#                         1,
+#                         dilation=1,
+#                         padding=get_padding(kernel_size, 1),
+#                     )
+#                 ),
+#             ]
+#         )
+#         self.convs2.apply(init_weights)
+
+#     def forward(self, x):
+#         for c1, c2 in zip(self.convs1, self.convs2):
+#             xt = F.leaky_relu(x, LRELU_SLOPE)
+#             xt = c1(xt)
+#             xt = F.leaky_relu(xt, LRELU_SLOPE)
+#             xt = c2(xt)
+#             x = xt + x
+#         return x
+
+#     def remove_weight_norm(self):
+#         for l in self.convs1:
+#             remove_weight_norm(l)
+#         for l in self.convs2:
+#             remove_weight_norm(l)
+
+# class Generator(torch.nn.Module):
+#     def __init__(self, h):
+#         super(Generator, self).__init__()
+#         self.h = h
+#         self.num_kernels = len(h.resblock_kernel_sizes)
+#         self.num_upsamples = len(h.upsample_rates)
+#         self.conv_pre = weight_norm(
+#             Conv1d(h.num_mels, h.upsample_initial_channel, 7, 1, padding=3)
+#         )
+#         resblock = ResBlock
+
+#         self.ups = nn.ModuleList()
+#         for i, (u, k) in enumerate(zip(h.upsample_rates, h.upsample_kernel_sizes)):
+#             self.ups.append(
+#                 weight_norm(
+#                     ConvTranspose1d(
+#                         h.upsample_initial_channel // (2**i),
+#                         h.upsample_initial_channel // (2 ** (i + 1)),
+#                         k,
+#                         u,
+#                         padding=(k - u) // 2,
+#                     )
+#                 )
+#             )
+
+#         self.resblocks = nn.ModuleList()
+#         for i in range(len(self.ups)):
+#             ch = h.upsample_initial_channel // (2 ** (i + 1))
+#             for j, (k, d) in enumerate(
+#                 zip(h.resblock_kernel_sizes, h.resblock_dilation_sizes)
+#             ):
+#                 self.resblocks.append(resblock(h, ch, k, d))
+
+#         self.conv_post = weight_norm(Conv1d(ch, 1, 7, 1, padding=3))
+#         self.ups.apply(init_weights)
+#         self.conv_post.apply(init_weights)
+
+#     def forward(self, x):
+#         x = self.conv_pre(x)
+#         for i in range(self.num_upsamples):
+#             x = F.leaky_relu(x, LRELU_SLOPE)
+#             x = self.ups[i](x)
+#             xs = None
+#             for j in range(self.num_kernels):
+#                 if xs is None:
+#                     xs = self.resblocks[i * self.num_kernels + j](x)
+#                 else:
+#                     xs += self.resblocks[i * self.num_kernels + j](x)
+#             x = xs / self.num_kernels
+#         x = F.leaky_relu(x)
+#         x = self.conv_post(x)
+#         x = torch.tanh(x)
+
+#         return x
+
+#     def remove_weight_norm(self):
+#         print("Removing weight norm...")
+#         for l in self.ups:
+#             remove_weight_norm(l)
+#         for l in self.resblocks:
+#             l.remove_weight_norm()
+#         remove_weight_norm(self.conv_pre)
+#         remove_weight_norm(self.conv_post)

audiosr/latent_diffusion/models/ddim.py

@@ -1,492 +1,492 @@
-"""SAMPLING ONLY."""
-
-import torch
-import numpy as np
-from tqdm import tqdm
-
-from audiosr.latent_diffusion.modules.diffusionmodules.util import (
-    make_ddim_sampling_parameters,
-    make_ddim_timesteps,
-    noise_like,
-    extract_into_tensor,
-)
-
-
-class DDIMSampler(object):
-    def __init__(self, model, schedule="linear", device=torch.device("cuda"), **kwargs):
-        super().__init__()
-        self.model = model
-        self.ddpm_num_timesteps = model.num_timesteps
-        self.schedule = schedule
-        self.device = device
-
-    def register_buffer(self, name, attr):
-        if type(attr) == torch.Tensor:
-            if attr.device != self.device:
-                is_mps = self.device == "mps" or self.device == torch.device("mps")
-                if is_mps and attr.dtype == torch.float64:
-                    attr = attr.to(self.device, dtype=torch.float32)
-                else:
-                    attr = attr.to(self.device)
-        setattr(self, name, attr)
-
-    def make_schedule(
-        self, ddim_num_steps, ddim_discretize="uniform", ddim_eta=0.0, verbose=True
-    ):
-        self.ddim_timesteps = make_ddim_timesteps(
-            ddim_discr_method=ddim_discretize,
-            num_ddim_timesteps=ddim_num_steps,
-            num_ddpm_timesteps=self.ddpm_num_timesteps,
-            verbose=verbose,
-        )
-        alphas_cumprod = self.model.alphas_cumprod
-        assert (
-            alphas_cumprod.shape[0] == self.ddpm_num_timesteps
-        ), "alphas have to be defined for each timestep"
-        to_torch = lambda x: x.clone().detach().to(torch.float32).to(self.model.device)
-
-        self.register_buffer("betas", to_torch(self.model.betas))
-        self.register_buffer("alphas_cumprod", to_torch(alphas_cumprod))
-        self.register_buffer(
-            "alphas_cumprod_prev", to_torch(self.model.alphas_cumprod_prev)
-        )
-
-        # calculations for diffusion q(x_t | x_{t-1}) and others
-        self.register_buffer(
-            "sqrt_alphas_cumprod", to_torch(np.sqrt(alphas_cumprod.cpu()))
-        )
-        self.register_buffer(
-            "sqrt_one_minus_alphas_cumprod",
-            to_torch(np.sqrt(1.0 - alphas_cumprod.cpu())),
-        )
-        self.register_buffer(
-            "log_one_minus_alphas_cumprod", to_torch(np.log(1.0 - alphas_cumprod.cpu()))
-        )
-        self.register_buffer(
-            "sqrt_recip_alphas_cumprod", to_torch(np.sqrt(1.0 / alphas_cumprod.cpu()))
-        )
-        self.register_buffer(
-            "sqrt_recipm1_alphas_cumprod",
-            to_torch(np.sqrt(1.0 / alphas_cumprod.cpu() - 1)),
-        )
-
-        # ddim sampling parameters
-        ddim_sigmas, ddim_alphas, ddim_alphas_prev = make_ddim_sampling_parameters(
-            alphacums=alphas_cumprod.cpu(),
-            ddim_timesteps=self.ddim_timesteps,
-            eta=ddim_eta,
-            verbose=verbose,
-        )
-        self.register_buffer("ddim_sigmas", ddim_sigmas)
-        self.register_buffer("ddim_alphas", ddim_alphas)
-        self.register_buffer("ddim_alphas_prev", ddim_alphas_prev)
-        self.register_buffer("ddim_sqrt_one_minus_alphas", np.sqrt(1.0 - ddim_alphas))
-        sigmas_for_original_sampling_steps = ddim_eta * torch.sqrt(
-            (1 - self.alphas_cumprod_prev)
-            / (1 - self.alphas_cumprod)
-            * (1 - self.alphas_cumprod / self.alphas_cumprod_prev)
-        )
-        self.register_buffer(
-            "ddim_sigmas_for_original_num_steps", sigmas_for_original_sampling_steps
-        )
-
-    @torch.no_grad()
-    def sample(
-        self,
-        S,
-        batch_size,
-        shape,
-        conditioning=None,
-        callback=None,
-        normals_sequence=None,
-        img_callback=None,
-        quantize_x0=False,
-        eta=0.0,
-        mask=None,
-        x0=None,
-        temperature=1.0,
-        noise_dropout=0.0,
-        score_corrector=None,
-        corrector_kwargs=None,
-        verbose=True,
-        x_T=None,
-        log_every_t=100,
-        unconditional_guidance_scale=1.0,
-        unconditional_conditioning=None,  # this has to come in the same format as the conditioning, # e.g. as encoded tokens, ...
-        dynamic_threshold=None,
-        ucg_schedule=None,
-        **kwargs,
-    ):
-        # if conditioning is not None:
-        #     if isinstance(conditioning, dict):
-        #         ctmp = conditioning[list(conditioning.keys())[0]]
-        #         while isinstance(ctmp, list): ctmp = ctmp[0]
-        #         cbs = ctmp.shape[0]
-        #         if cbs != batch_size:
-        #             print(f"Warning: Got {cbs} conditionings but batch-size is {batch_size}")
-
-        #     elif isinstance(conditioning, list):
-        #         for ctmp in conditioning:
-        #             if ctmp.shape[0] != batch_size:
-        #                 print(f"Warning: Got {cbs} conditionings but batch-size is {batch_size}")
-
-        #     else:
-        #         if conditioning.shape[0] != batch_size:
-        #             print(f"Warning: Got {conditioning.shape[0]} conditionings but batch-size is {batch_size}")
-
-        self.make_schedule(ddim_num_steps=S, ddim_eta=eta, verbose=verbose)
-        # sampling
-        C, H, W = shape
-        size = (batch_size, C, H, W)
-        # print(f'Data shape for DDIM sampling is {size}, eta {eta}')
-
-        samples, intermediates = self.ddim_sampling(
-            conditioning,
-            size,
-            callback=callback,
-            img_callback=img_callback,
-            quantize_denoised=quantize_x0,
-            mask=mask,
-            x0=x0,
-            ddim_use_original_steps=False,
-            noise_dropout=noise_dropout,
-            temperature=temperature,
-            score_corrector=score_corrector,
-            corrector_kwargs=corrector_kwargs,
-            x_T=x_T,
-            log_every_t=log_every_t,
-            unconditional_guidance_scale=unconditional_guidance_scale,
-            unconditional_conditioning=unconditional_conditioning,
-            dynamic_threshold=dynamic_threshold,
-            ucg_schedule=ucg_schedule,
-        )
-        return samples, intermediates
-
-    @torch.no_grad()
-    def ddim_sampling(
-        self,
-        cond,
-        shape,
-        x_T=None,
-        ddim_use_original_steps=False,
-        callback=None,
-        timesteps=None,
-        quantize_denoised=False,
-        mask=None,
-        x0=None,
-        img_callback=None,
-        log_every_t=100,
-        temperature=1.0,
-        noise_dropout=0.0,
-        score_corrector=None,
-        corrector_kwargs=None,
-        unconditional_guidance_scale=1.0,
-        unconditional_conditioning=None,
-        dynamic_threshold=None,
-        ucg_schedule=None,
-    ):
-        device = self.model.betas.device
-        b = shape[0]
-        if x_T is None:
-            img = torch.randn(shape, device=device)
-        else:
-            img = x_T
-
-        if timesteps is None:
-            timesteps = (
-                self.ddpm_num_timesteps
-                if ddim_use_original_steps
-                else self.ddim_timesteps
-            )
-        elif timesteps is not None and not ddim_use_original_steps:
-            subset_end = (
-                int(
-                    min(timesteps / self.ddim_timesteps.shape[0], 1)
-                    * self.ddim_timesteps.shape[0]
-                )
-                - 1
-            )
-            timesteps = self.ddim_timesteps[:subset_end]
-
-        intermediates = {"x_inter": [img], "pred_x0": [img]}
-        time_range = (
-            reversed(range(0, timesteps))
-            if ddim_use_original_steps
-            else np.flip(timesteps)
-        )
-        total_steps = timesteps if ddim_use_original_steps else timesteps.shape[0]
-        print(f"Running DDIM Sampling with {total_steps} timesteps")
-
-        iterator = tqdm(time_range, desc="DDIM Sampler", total=total_steps)
-
-        for i, step in enumerate(iterator):
-            index = total_steps - i - 1
-            ts = torch.full((b,), step, device=device, dtype=torch.long)
-
-            if mask is not None:
-                assert x0 is not None
-                img_orig = self.model.q_sample(
-                    x0, ts
-                )  # TODO: deterministic forward pass?
-                img = img_orig * mask + (1.0 - mask) * img
-
-            if ucg_schedule is not None:
-                assert len(ucg_schedule) == len(time_range)
-                unconditional_guidance_scale = ucg_schedule[i]
-
-            outs = self.p_sample_ddim(
-                img,
-                cond,
-                ts,
-                index=index,
-                use_original_steps=ddim_use_original_steps,
-                quantize_denoised=quantize_denoised,
-                temperature=temperature,
-                noise_dropout=noise_dropout,
-                score_corrector=score_corrector,
-                corrector_kwargs=corrector_kwargs,
-                unconditional_guidance_scale=unconditional_guidance_scale,
-                unconditional_conditioning=unconditional_conditioning,
-                dynamic_threshold=dynamic_threshold,
-            )
-            img, pred_x0 = outs
-            if callback:
-                callback(i)
-            if img_callback:
-                img_callback(pred_x0, i)
-
-            if index % log_every_t == 0 or index == total_steps - 1:
-                intermediates["x_inter"].append(img)
-                intermediates["pred_x0"].append(pred_x0)
-
-        return img, intermediates
-
-    @torch.no_grad()
-    def p_sample_ddim(
-        self,
-        x,
-        c,
-        t,
-        index,
-        repeat_noise=False,
-        use_original_steps=False,
-        quantize_denoised=False,
-        temperature=1.0,
-        noise_dropout=0.0,
-        score_corrector=None,
-        corrector_kwargs=None,
-        unconditional_guidance_scale=1.0,
-        unconditional_conditioning=None,
-        dynamic_threshold=None,
-    ):
-        b, *_, device = *x.shape, x.device
-
-        if unconditional_conditioning is None or unconditional_guidance_scale == 1.0:
-            model_output = self.model.apply_model(x, t, c)
-        else:
-            x_in = x
-            t_in = t
-
-            assert isinstance(c, dict)
-            assert isinstance(unconditional_conditioning, dict)
-
-            model_t = self.model.apply_model(x_in, t_in, c)
-
-            model_uncond = self.model.apply_model(
-                x_in, t_in, unconditional_conditioning
-            )
-
-            model_output = model_uncond + unconditional_guidance_scale * (
-                model_t - model_uncond
-            )
-
-        if self.model.parameterization == "v":
-            e_t = self.model.predict_eps_from_z_and_v(x, t, model_output)
-        else:
-            e_t = model_output
-
-        if score_corrector is not None:
-            assert self.model.parameterization == "eps", "not implemented"
-            e_t = score_corrector.modify_score(
-                self.model, e_t, x, t, c, **corrector_kwargs
-            )
-
-        alphas = self.model.alphas_cumprod if use_original_steps else self.ddim_alphas
-        alphas_prev = (
-            self.model.alphas_cumprod_prev
-            if use_original_steps
-            else self.ddim_alphas_prev
-        )
-        sqrt_one_minus_alphas = (
-            self.model.sqrt_one_minus_alphas_cumprod
-            if use_original_steps
-            else self.ddim_sqrt_one_minus_alphas
-        )
-        sigmas = (
-            self.model.ddim_sigmas_for_original_num_steps
-            if use_original_steps
-            else self.ddim_sigmas
-        )
-        # select parameters corresponding to the currently considered timestep
-        a_t = torch.full((b, 1, 1, 1), alphas[index], device=device)
-        a_prev = torch.full((b, 1, 1, 1), alphas_prev[index], device=device)
-        sigma_t = torch.full((b, 1, 1, 1), sigmas[index], device=device)
-        sqrt_one_minus_at = torch.full(
-            (b, 1, 1, 1), sqrt_one_minus_alphas[index], device=device
-        )
-
-        # current prediction for x_0
-        if self.model.parameterization != "v":
-            pred_x0 = (x - sqrt_one_minus_at * e_t) / a_t.sqrt()
-        else:
-            pred_x0 = self.model.predict_start_from_z_and_v(x, t, model_output)
-
-        if quantize_denoised:
-            pred_x0, _, *_ = self.model.first_stage_model.quantize(pred_x0)
-
-        if dynamic_threshold is not None:
-            raise NotImplementedError()
-
-        # direction pointing to x_t
-        dir_xt = (1.0 - a_prev - sigma_t**2).sqrt() * e_t
-        noise = sigma_t * noise_like(x.shape, device, repeat_noise) * temperature
-        if noise_dropout > 0.0:
-            noise = torch.nn.functional.dropout(noise, p=noise_dropout)
-        x_prev = a_prev.sqrt() * pred_x0 + dir_xt + noise
-        return x_prev, pred_x0
-
-    @torch.no_grad()
-    def encode(
-        self,
-        x0,
-        c,
-        t_enc,
-        use_original_steps=False,
-        return_intermediates=None,
-        unconditional_guidance_scale=1.0,
-        unconditional_conditioning=None,
-        callback=None,
-    ):
-        num_reference_steps = (
-            self.ddpm_num_timesteps
-            if use_original_steps
-            else self.ddim_timesteps.shape[0]
-        )
-
-        assert t_enc <= num_reference_steps
-        num_steps = t_enc
-
-        if use_original_steps:
-            alphas_next = self.alphas_cumprod[:num_steps]
-            alphas = self.alphas_cumprod_prev[:num_steps]
-        else:
-            alphas_next = self.ddim_alphas[:num_steps]
-            alphas = torch.tensor(self.ddim_alphas_prev[:num_steps])
-
-        x_next = x0
-        intermediates = []
-        inter_steps = []
-        for i in tqdm(range(num_steps), desc="Encoding Image"):
-            t = torch.full(
-                (x0.shape[0],), i, device=self.model.device, dtype=torch.long
-            )
-            if unconditional_guidance_scale == 1.0:
-                noise_pred = self.model.apply_model(x_next, t, c)
-            else:
-                assert unconditional_conditioning is not None
-                e_t_uncond, noise_pred = torch.chunk(
-                    self.model.apply_model(
-                        torch.cat((x_next, x_next)),
-                        torch.cat((t, t)),
-                        torch.cat((unconditional_conditioning, c)),
-                    ),
-                    2,
-                )
-                noise_pred = e_t_uncond + unconditional_guidance_scale * (
-                    noise_pred - e_t_uncond
-                )
-
-            xt_weighted = (alphas_next[i] / alphas[i]).sqrt() * x_next
-            weighted_noise_pred = (
-                alphas_next[i].sqrt()
-                * ((1 / alphas_next[i] - 1).sqrt() - (1 / alphas[i] - 1).sqrt())
-                * noise_pred
-            )
-            x_next = xt_weighted + weighted_noise_pred
-            if (
-                return_intermediates
-                and i % (num_steps // return_intermediates) == 0
-                and i < num_steps - 1
-            ):
-                intermediates.append(x_next)
-                inter_steps.append(i)
-            elif return_intermediates and i >= num_steps - 2:
-                intermediates.append(x_next)
-                inter_steps.append(i)
-            if callback:
-                callback(i)
-
-        out = {"x_encoded": x_next, "intermediate_steps": inter_steps}
-        if return_intermediates:
-            out.update({"intermediates": intermediates})
-        return x_next, out
-
-    @torch.no_grad()
-    def stochastic_encode(self, x0, t, use_original_steps=False, noise=None):
-        # fast, but does not allow for exact reconstruction
-        # t serves as an index to gather the correct alphas
-        if use_original_steps:
-            sqrt_alphas_cumprod = self.sqrt_alphas_cumprod
-            sqrt_one_minus_alphas_cumprod = self.sqrt_one_minus_alphas_cumprod
-        else:
-            sqrt_alphas_cumprod = torch.sqrt(self.ddim_alphas)
-            sqrt_one_minus_alphas_cumprod = self.ddim_sqrt_one_minus_alphas
-
-        if noise is None:
-            noise = torch.randn_like(x0)
-        return (
-            extract_into_tensor(sqrt_alphas_cumprod, t, x0.shape) * x0
-            + extract_into_tensor(sqrt_one_minus_alphas_cumprod, t, x0.shape) * noise
-        )
-
-    @torch.no_grad()
-    def decode(
-        self,
-        x_latent,
-        cond,
-        t_start,
-        unconditional_guidance_scale=1.0,
-        unconditional_conditioning=None,
-        use_original_steps=False,
-        callback=None,
-    ):
-        timesteps = (
-            np.arange(self.ddpm_num_timesteps)
-            if use_original_steps
-            else self.ddim_timesteps
-        )
-        timesteps = timesteps[:t_start]
-
-        time_range = np.flip(timesteps)
-        total_steps = timesteps.shape[0]
-        print(f"Running DDIM Sampling with {total_steps} timesteps")
-
-        iterator = tqdm(time_range, desc="Decoding image", total=total_steps)
-        x_dec = x_latent
-        for i, step in enumerate(iterator):
-            index = total_steps - i - 1
-            ts = torch.full(
-                (x_latent.shape[0],), step, device=x_latent.device, dtype=torch.long
-            )
-            x_dec, _ = self.p_sample_ddim(
-                x_dec,
-                cond,
-                ts,
-                index=index,
-                use_original_steps=use_original_steps,
-                unconditional_guidance_scale=unconditional_guidance_scale,
-                unconditional_conditioning=unconditional_conditioning,
-            )
-            if callback:
-                callback(i)
-        return x_dec
+"""SAMPLING ONLY."""
+
+import torch
+import numpy as np
+from tqdm import tqdm
+
+from audiosr.latent_diffusion.modules.diffusionmodules.util import (
+    make_ddim_sampling_parameters,
+    make_ddim_timesteps,
+    noise_like,
+    extract_into_tensor,
+)
+
+
+class DDIMSampler(object):
+    def __init__(self, model, schedule="linear", device=torch.device("cuda"), **kwargs):
+        super().__init__()
+        self.model = model
+        self.ddpm_num_timesteps = model.num_timesteps
+        self.schedule = schedule
+        self.device = device
+
+    def register_buffer(self, name, attr):
+        if type(attr) == torch.Tensor:
+            if attr.device != self.device:
+                is_mps = self.device == "mps" or self.device == torch.device("mps")
+                if is_mps and attr.dtype == torch.float64:
+                    attr = attr.to(self.device, dtype=torch.float32)
+                else:
+                    attr = attr.to(self.device)
+        setattr(self, name, attr)
+
+    def make_schedule(
+        self, ddim_num_steps, ddim_discretize="uniform", ddim_eta=0.0, verbose=True
+    ):
+        self.ddim_timesteps = make_ddim_timesteps(
+            ddim_discr_method=ddim_discretize,
+            num_ddim_timesteps=ddim_num_steps,
+            num_ddpm_timesteps=self.ddpm_num_timesteps,
+            verbose=verbose,
+        )
+        alphas_cumprod = self.model.alphas_cumprod
+        assert (
+            alphas_cumprod.shape[0] == self.ddpm_num_timesteps
+        ), "alphas have to be defined for each timestep"
+        to_torch = lambda x: x.clone().detach().to(torch.float32).to(self.model.device)
+
+        self.register_buffer("betas", to_torch(self.model.betas))
+        self.register_buffer("alphas_cumprod", to_torch(alphas_cumprod))
+        self.register_buffer(
+            "alphas_cumprod_prev", to_torch(self.model.alphas_cumprod_prev)
+        )
+
+        # calculations for diffusion q(x_t | x_{t-1}) and others
+        self.register_buffer(
+            "sqrt_alphas_cumprod", to_torch(np.sqrt(alphas_cumprod.cpu()))
+        )
+        self.register_buffer(
+            "sqrt_one_minus_alphas_cumprod",
+            to_torch(np.sqrt(1.0 - alphas_cumprod.cpu())),
+        )
+        self.register_buffer(
+            "log_one_minus_alphas_cumprod", to_torch(np.log(1.0 - alphas_cumprod.cpu()))
+        )
+        self.register_buffer(
+            "sqrt_recip_alphas_cumprod", to_torch(np.sqrt(1.0 / alphas_cumprod.cpu()))
+        )
+        self.register_buffer(
+            "sqrt_recipm1_alphas_cumprod",
+            to_torch(np.sqrt(1.0 / alphas_cumprod.cpu() - 1)),
+        )
+
+        # ddim sampling parameters
+        ddim_sigmas, ddim_alphas, ddim_alphas_prev = make_ddim_sampling_parameters(
+            alphacums=alphas_cumprod.cpu(),
+            ddim_timesteps=self.ddim_timesteps,
+            eta=ddim_eta,
+            verbose=verbose,
+        )
+        self.register_buffer("ddim_sigmas", ddim_sigmas)
+        self.register_buffer("ddim_alphas", ddim_alphas)
+        self.register_buffer("ddim_alphas_prev", ddim_alphas_prev)
+        self.register_buffer("ddim_sqrt_one_minus_alphas", np.sqrt(1.0 - ddim_alphas))
+        sigmas_for_original_sampling_steps = ddim_eta * torch.sqrt(
+            (1 - self.alphas_cumprod_prev)
+            / (1 - self.alphas_cumprod)
+            * (1 - self.alphas_cumprod / self.alphas_cumprod_prev)
+        )
+        self.register_buffer(
+            "ddim_sigmas_for_original_num_steps", sigmas_for_original_sampling_steps
+        )
+
+    @torch.no_grad()
+    def sample(
+        self,
+        S,
+        batch_size,
+        shape,
+        conditioning=None,
+        callback=None,
+        normals_sequence=None,
+        img_callback=None,
+        quantize_x0=False,
+        eta=0.0,
+        mask=None,
+        x0=None,
+        temperature=1.0,
+        noise_dropout=0.0,
+        score_corrector=None,
+        corrector_kwargs=None,
+        verbose=True,
+        x_T=None,
+        log_every_t=100,
+        unconditional_guidance_scale=1.0,
+        unconditional_conditioning=None,  # this has to come in the same format as the conditioning, # e.g. as encoded tokens, ...
+        dynamic_threshold=None,
+        ucg_schedule=None,
+        **kwargs,
+    ):
+        # if conditioning is not None:
+        #     if isinstance(conditioning, dict):
+        #         ctmp = conditioning[list(conditioning.keys())[0]]
+        #         while isinstance(ctmp, list): ctmp = ctmp[0]
+        #         cbs = ctmp.shape[0]
+        #         if cbs != batch_size:
+        #             print(f"Warning: Got {cbs} conditionings but batch-size is {batch_size}")
+
+        #     elif isinstance(conditioning, list):
+        #         for ctmp in conditioning:
+        #             if ctmp.shape[0] != batch_size:
+        #                 print(f"Warning: Got {cbs} conditionings but batch-size is {batch_size}")
+
+        #     else:
+        #         if conditioning.shape[0] != batch_size:
+        #             print(f"Warning: Got {conditioning.shape[0]} conditionings but batch-size is {batch_size}")
+
+        self.make_schedule(ddim_num_steps=S, ddim_eta=eta, verbose=verbose)
+        # sampling
+        C, H, W = shape
+        size = (batch_size, C, H, W)
+        # print(f'Data shape for DDIM sampling is {size}, eta {eta}')
+
+        samples, intermediates = self.ddim_sampling(
+            conditioning,
+            size,
+            callback=callback,
+            img_callback=img_callback,
+            quantize_denoised=quantize_x0,
+            mask=mask,
+            x0=x0,
+            ddim_use_original_steps=False,
+            noise_dropout=noise_dropout,
+            temperature=temperature,
+            score_corrector=score_corrector,
+            corrector_kwargs=corrector_kwargs,
+            x_T=x_T,
+            log_every_t=log_every_t,
+            unconditional_guidance_scale=unconditional_guidance_scale,
+            unconditional_conditioning=unconditional_conditioning,
+            dynamic_threshold=dynamic_threshold,
+            ucg_schedule=ucg_schedule,
+        )
+        return samples, intermediates
+
+    @torch.no_grad()
+    def ddim_sampling(
+        self,
+        cond,
+        shape,
+        x_T=None,
+        ddim_use_original_steps=False,
+        callback=None,
+        timesteps=None,
+        quantize_denoised=False,
+        mask=None,
+        x0=None,
+        img_callback=None,
+        log_every_t=100,
+        temperature=1.0,
+        noise_dropout=0.0,
+        score_corrector=None,
+        corrector_kwargs=None,
+        unconditional_guidance_scale=1.0,
+        unconditional_conditioning=None,
+        dynamic_threshold=None,
+        ucg_schedule=None,
+    ):
+        device = self.model.betas.device
+        b = shape[0]
+        if x_T is None:
+            img = torch.randn(shape, device=device)
+        else:
+            img = x_T
+
+        if timesteps is None:
+            timesteps = (
+                self.ddpm_num_timesteps
+                if ddim_use_original_steps
+                else self.ddim_timesteps
+            )
+        elif timesteps is not None and not ddim_use_original_steps:
+            subset_end = (
+                int(
+                    min(timesteps / self.ddim_timesteps.shape[0], 1)
+                    * self.ddim_timesteps.shape[0]
+                )
+                - 1
+            )
+            timesteps = self.ddim_timesteps[:subset_end]
+
+        intermediates = {"x_inter": [img], "pred_x0": [img]}
+        time_range = (
+            reversed(range(0, timesteps))
+            if ddim_use_original_steps
+            else np.flip(timesteps)
+        )
+        total_steps = timesteps if ddim_use_original_steps else timesteps.shape[0]
+        print(f"Running DDIM Sampling with {total_steps} timesteps")
+
+        iterator = tqdm(time_range, desc="DDIM Sampler", total=total_steps)
+
+        for i, step in enumerate(iterator):
+            index = total_steps - i - 1
+            ts = torch.full((b,), step, device=device, dtype=torch.long)
+
+            if mask is not None:
+                assert x0 is not None
+                img_orig = self.model.q_sample(
+                    x0, ts
+                )  # TODO: deterministic forward pass?
+                img = img_orig * mask + (1.0 - mask) * img
+
+            if ucg_schedule is not None:
+                assert len(ucg_schedule) == len(time_range)
+                unconditional_guidance_scale = ucg_schedule[i]
+
+            outs = self.p_sample_ddim(
+                img,
+                cond,
+                ts,
+                index=index,
+                use_original_steps=ddim_use_original_steps,
+                quantize_denoised=quantize_denoised,
+                temperature=temperature,
+                noise_dropout=noise_dropout,
+                score_corrector=score_corrector,
+                corrector_kwargs=corrector_kwargs,
+                unconditional_guidance_scale=unconditional_guidance_scale,
+                unconditional_conditioning=unconditional_conditioning,
+                dynamic_threshold=dynamic_threshold,
+            )
+            img, pred_x0 = outs
+            if callback:
+                callback(i)
+            if img_callback:
+                img_callback(pred_x0, i)
+
+            if index % log_every_t == 0 or index == total_steps - 1:
+                intermediates["x_inter"].append(img)
+                intermediates["pred_x0"].append(pred_x0)
+
+        return img, intermediates
+
+    @torch.no_grad()
+    def p_sample_ddim(
+        self,
+        x,
+        c,
+        t,
+        index,
+        repeat_noise=False,
+        use_original_steps=False,
+        quantize_denoised=False,
+        temperature=1.0,
+        noise_dropout=0.0,
+        score_corrector=None,
+        corrector_kwargs=None,
+        unconditional_guidance_scale=1.0,
+        unconditional_conditioning=None,
+        dynamic_threshold=None,
+    ):
+        b, *_, device = *x.shape, x.device
+
+        if unconditional_conditioning is None or unconditional_guidance_scale == 1.0:
+            model_output = self.model.apply_model(x, t, c)
+        else:
+            x_in = x
+            t_in = t
+
+            assert isinstance(c, dict)
+            assert isinstance(unconditional_conditioning, dict)
+
+            model_t = self.model.apply_model(x_in, t_in, c)
+
+            model_uncond = self.model.apply_model(
+                x_in, t_in, unconditional_conditioning
+            )
+
+            model_output = model_uncond + unconditional_guidance_scale * (
+                model_t - model_uncond
+            )
+
+        if self.model.parameterization == "v":
+            e_t = self.model.predict_eps_from_z_and_v(x, t, model_output)
+        else:
+            e_t = model_output
+
+        if score_corrector is not None:
+            assert self.model.parameterization == "eps", "not implemented"
+            e_t = score_corrector.modify_score(
+                self.model, e_t, x, t, c, **corrector_kwargs
+            )
+
+        alphas = self.model.alphas_cumprod if use_original_steps else self.ddim_alphas
+        alphas_prev = (
+            self.model.alphas_cumprod_prev
+            if use_original_steps
+            else self.ddim_alphas_prev
+        )
+        sqrt_one_minus_alphas = (
+            self.model.sqrt_one_minus_alphas_cumprod
+            if use_original_steps
+            else self.ddim_sqrt_one_minus_alphas
+        )
+        sigmas = (
+            self.model.ddim_sigmas_for_original_num_steps
+            if use_original_steps
+            else self.ddim_sigmas
+        )
+        # select parameters corresponding to the currently considered timestep
+        a_t = torch.full((b, 1, 1, 1), alphas[index], device=device)
+        a_prev = torch.full((b, 1, 1, 1), alphas_prev[index], device=device)
+        sigma_t = torch.full((b, 1, 1, 1), sigmas[index], device=device)
+        sqrt_one_minus_at = torch.full(
+            (b, 1, 1, 1), sqrt_one_minus_alphas[index], device=device
+        )
+
+        # current prediction for x_0
+        if self.model.parameterization != "v":
+            pred_x0 = (x - sqrt_one_minus_at * e_t) / a_t.sqrt()
+        else:
+            pred_x0 = self.model.predict_start_from_z_and_v(x, t, model_output)
+
+        if quantize_denoised:
+            pred_x0, _, *_ = self.model.first_stage_model.quantize(pred_x0)
+
+        if dynamic_threshold is not None:
+            raise NotImplementedError()
+
+        # direction pointing to x_t
+        dir_xt = (1.0 - a_prev - sigma_t**2).sqrt() * e_t
+        noise = sigma_t * noise_like(x.shape, device, repeat_noise) * temperature
+        if noise_dropout > 0.0:
+            noise = torch.nn.functional.dropout(noise, p=noise_dropout)
+        x_prev = a_prev.sqrt() * pred_x0 + dir_xt + noise
+        return x_prev, pred_x0
+
+    @torch.no_grad()
+    def encode(
+        self,
+        x0,
+        c,
+        t_enc,
+        use_original_steps=False,
+        return_intermediates=None,
+        unconditional_guidance_scale=1.0,
+        unconditional_conditioning=None,
+        callback=None,
+    ):
+        num_reference_steps = (
+            self.ddpm_num_timesteps
+            if use_original_steps
+            else self.ddim_timesteps.shape[0]
+        )
+
+        assert t_enc <= num_reference_steps
+        num_steps = t_enc
+
+        if use_original_steps:
+            alphas_next = self.alphas_cumprod[:num_steps]
+            alphas = self.alphas_cumprod_prev[:num_steps]
+        else:
+            alphas_next = self.ddim_alphas[:num_steps]
+            alphas = torch.tensor(self.ddim_alphas_prev[:num_steps])
+
+        x_next = x0
+        intermediates = []
+        inter_steps = []
+        for i in tqdm(range(num_steps), desc="Encoding Image"):
+            t = torch.full(
+                (x0.shape[0],), i, device=self.model.device, dtype=torch.long
+            )
+            if unconditional_guidance_scale == 1.0:
+                noise_pred = self.model.apply_model(x_next, t, c)
+            else:
+                assert unconditional_conditioning is not None
+                e_t_uncond, noise_pred = torch.chunk(
+                    self.model.apply_model(
+                        torch.cat((x_next, x_next)),
+                        torch.cat((t, t)),
+                        torch.cat((unconditional_conditioning, c)),
+                    ),
+                    2,
+                )
+                noise_pred = e_t_uncond + unconditional_guidance_scale * (
+                    noise_pred - e_t_uncond
+                )
+
+            xt_weighted = (alphas_next[i] / alphas[i]).sqrt() * x_next
+            weighted_noise_pred = (
+                alphas_next[i].sqrt()
+                * ((1 / alphas_next[i] - 1).sqrt() - (1 / alphas[i] - 1).sqrt())
+                * noise_pred
+            )
+            x_next = xt_weighted + weighted_noise_pred
+            if (
+                return_intermediates
+                and i % (num_steps // return_intermediates) == 0
+                and i < num_steps - 1
+            ):
+                intermediates.append(x_next)
+                inter_steps.append(i)
+            elif return_intermediates and i >= num_steps - 2:
+                intermediates.append(x_next)
+                inter_steps.append(i)
+            if callback:
+                callback(i)
+
+        out = {"x_encoded": x_next, "intermediate_steps": inter_steps}
+        if return_intermediates:
+            out.update({"intermediates": intermediates})
+        return x_next, out
+
+    @torch.no_grad()
+    def stochastic_encode(self, x0, t, use_original_steps=False, noise=None):
+        # fast, but does not allow for exact reconstruction
+        # t serves as an index to gather the correct alphas
+        if use_original_steps:
+            sqrt_alphas_cumprod = self.sqrt_alphas_cumprod
+            sqrt_one_minus_alphas_cumprod = self.sqrt_one_minus_alphas_cumprod
+        else:
+            sqrt_alphas_cumprod = torch.sqrt(self.ddim_alphas)
+            sqrt_one_minus_alphas_cumprod = self.ddim_sqrt_one_minus_alphas
+
+        if noise is None:
+            noise = torch.randn_like(x0)
+        return (
+            extract_into_tensor(sqrt_alphas_cumprod, t, x0.shape) * x0
+            + extract_into_tensor(sqrt_one_minus_alphas_cumprod, t, x0.shape) * noise
+        )
+
+    @torch.no_grad()
+    def decode(
+        self,
+        x_latent,
+        cond,
+        t_start,
+        unconditional_guidance_scale=1.0,
+        unconditional_conditioning=None,
+        use_original_steps=False,
+        callback=None,
+    ):
+        timesteps = (
+            np.arange(self.ddpm_num_timesteps)
+            if use_original_steps
+            else self.ddim_timesteps
+        )
+        timesteps = timesteps[:t_start]
+
+        time_range = np.flip(timesteps)
+        total_steps = timesteps.shape[0]
+        print(f"Running DDIM Sampling with {total_steps} timesteps")
+
+        iterator = tqdm(time_range, desc="Decoding image", total=total_steps)
+        x_dec = x_latent
+        for i, step in enumerate(iterator):
+            index = total_steps - i - 1
+            ts = torch.full(
+                (x_latent.shape[0],), step, device=x_latent.device, dtype=torch.long
+            )
+            x_dec, _ = self.p_sample_ddim(
+                x_dec,
+                cond,
+                ts,
+                index=index,
+                use_original_steps=use_original_steps,
+                unconditional_guidance_scale=unconditional_guidance_scale,
+                unconditional_conditioning=unconditional_conditioning,
+            )
+            if callback:
+                callback(i)
+        return x_dec

audiosr/latent_diffusion/models/plms.py

@@ -1,360 +1,360 @@
-"""SAMPLING ONLY."""
-
-import torch
-import numpy as np
-from tqdm import tqdm
-
-from audiosr.latent_diffusion.modules.diffusionmodules.util import (
-    make_ddim_sampling_parameters,
-    make_ddim_timesteps,
-    noise_like,
-)
-
-
-class PLMSSampler(object):
-    def __init__(self, model, schedule="linear", **kwargs):
-        super().__init__()
-        self.model = model
-        self.ddpm_num_timesteps = model.num_timesteps
-        self.schedule = schedule
-
-    def register_buffer(self, name, attr):
-        if type(attr) == torch.Tensor:
-            if attr.device != torch.device("cuda"):
-                attr = attr.to(torch.device("cuda"))
-        setattr(self, name, attr)
-
-    def make_schedule(
-        self, ddim_num_steps, ddim_discretize="uniform", ddim_eta=0.0, verbose=True
-    ):
-        if ddim_eta != 0:
-            ddim_eta = 0
-            # raise ValueError('ddim_eta must be 0 for PLMS')
-
-        self.ddim_timesteps = make_ddim_timesteps(
-            ddim_discr_method=ddim_discretize,
-            num_ddim_timesteps=ddim_num_steps,
-            num_ddpm_timesteps=self.ddpm_num_timesteps,
-            verbose=verbose,
-        )
-        alphas_cumprod = self.model.alphas_cumprod
-        assert (
-            alphas_cumprod.shape[0] == self.ddpm_num_timesteps
-        ), "alphas have to be defined for each timestep"
-        to_torch = lambda x: x.clone().detach().to(torch.float32).to(self.model.device)
-
-        self.register_buffer("betas", to_torch(self.model.betas))
-        self.register_buffer("alphas_cumprod", to_torch(alphas_cumprod))
-        self.register_buffer(
-            "alphas_cumprod_prev", to_torch(self.model.alphas_cumprod_prev)
-        )
-
-        # calculations for diffusion q(x_t | x_{t-1}) and others
-        self.register_buffer(
-            "sqrt_alphas_cumprod", to_torch(np.sqrt(alphas_cumprod.cpu()))
-        )
-        self.register_buffer(
-            "sqrt_one_minus_alphas_cumprod",
-            to_torch(np.sqrt(1.0 - alphas_cumprod.cpu())),
-        )
-        self.register_buffer(
-            "log_one_minus_alphas_cumprod", to_torch(np.log(1.0 - alphas_cumprod.cpu()))
-        )
-        self.register_buffer(
-            "sqrt_recip_alphas_cumprod", to_torch(np.sqrt(1.0 / alphas_cumprod.cpu()))
-        )
-        self.register_buffer(
-            "sqrt_recipm1_alphas_cumprod",
-            to_torch(np.sqrt(1.0 / alphas_cumprod.cpu() - 1)),
-        )
-
-        # ddim sampling parameters
-        ddim_sigmas, ddim_alphas, ddim_alphas_prev = make_ddim_sampling_parameters(
-            alphacums=alphas_cumprod.cpu(),
-            ddim_timesteps=self.ddim_timesteps,
-            eta=ddim_eta,
-            verbose=verbose,
-        )
-        self.register_buffer("ddim_sigmas", ddim_sigmas)
-        self.register_buffer("ddim_alphas", ddim_alphas)
-        self.register_buffer("ddim_alphas_prev", ddim_alphas_prev)
-        self.register_buffer("ddim_sqrt_one_minus_alphas", np.sqrt(1.0 - ddim_alphas))
-        sigmas_for_original_sampling_steps = ddim_eta * torch.sqrt(
-            (1 - self.alphas_cumprod_prev)
-            / (1 - self.alphas_cumprod)
-            * (1 - self.alphas_cumprod / self.alphas_cumprod_prev)
-        )
-        self.register_buffer(
-            "ddim_sigmas_for_original_num_steps", sigmas_for_original_sampling_steps
-        )
-
-    @torch.no_grad()
-    def sample(
-        self,
-        S,
-        batch_size,
-        shape,
-        conditioning=None,
-        callback=None,
-        normals_sequence=None,
-        img_callback=None,
-        quantize_x0=False,
-        eta=0.0,
-        mask=None,
-        x0=None,
-        temperature=1.0,
-        noise_dropout=0.0,
-        score_corrector=None,
-        corrector_kwargs=None,
-        verbose=True,
-        x_T=None,
-        log_every_t=100,
-        unconditional_guidance_scale=1.0,
-        unconditional_conditioning=None,
-        # this has to come in the same format as the conditioning, # e.g. as encoded tokens, ...
-        **kwargs,
-    ):
-        if conditioning is not None:
-            if isinstance(conditioning, dict):
-                cbs = conditioning[list(conditioning.keys())[0]].shape[0]
-                if cbs != batch_size:
-                    print(
-                        f"Warning: Got {cbs} conditionings but batch-size is {batch_size}"
-                    )
-            else:
-                if conditioning.shape[0] != batch_size:
-                    print(
-                        f"Warning: Got {conditioning.shape[0]} conditionings but batch-size is {batch_size}"
-                    )
-
-        self.make_schedule(ddim_num_steps=S, ddim_eta=eta, verbose=verbose)
-        # sampling
-        C, H, W = shape
-        size = (batch_size, C, H, W)
-        print(f"Data shape for PLMS sampling is {size}")
-
-        samples, intermediates = self.plms_sampling(
-            conditioning,
-            size,
-            callback=callback,
-            img_callback=img_callback,
-            quantize_denoised=quantize_x0,
-            mask=mask,
-            x0=x0,
-            ddim_use_original_steps=False,
-            noise_dropout=noise_dropout,
-            temperature=temperature,
-            score_corrector=score_corrector,
-            corrector_kwargs=corrector_kwargs,
-            x_T=x_T,
-            log_every_t=log_every_t,
-            unconditional_guidance_scale=unconditional_guidance_scale,
-            unconditional_conditioning=unconditional_conditioning,
-        )
-        return samples, intermediates
-
-    @torch.no_grad()
-    def plms_sampling(
-        self,
-        cond,
-        shape,
-        x_T=None,
-        ddim_use_original_steps=False,
-        callback=None,
-        timesteps=None,
-        quantize_denoised=False,
-        mask=None,
-        x0=None,
-        img_callback=None,
-        log_every_t=100,
-        temperature=1.0,
-        noise_dropout=0.0,
-        score_corrector=None,
-        corrector_kwargs=None,
-        unconditional_guidance_scale=1.0,
-        unconditional_conditioning=None,
-    ):
-        device = self.model.betas.device
-        b = shape[0]
-        if x_T is None:
-            img = torch.randn(shape, device=device)
-        else:
-            img = x_T
-
-        if timesteps is None:
-            timesteps = (
-                self.ddpm_num_timesteps
-                if ddim_use_original_steps
-                else self.ddim_timesteps
-            )
-        elif timesteps is not None and not ddim_use_original_steps:
-            subset_end = (
-                int(
-                    min(timesteps / self.ddim_timesteps.shape[0], 1)
-                    * self.ddim_timesteps.shape[0]
-                )
-                - 1
-            )
-            timesteps = self.ddim_timesteps[:subset_end]
-
-        intermediates = {"x_inter": [img], "pred_x0": [img]}
-        time_range = (
-            list(reversed(range(0, timesteps)))
-            if ddim_use_original_steps
-            else np.flip(timesteps)
-        )
-        total_steps = timesteps if ddim_use_original_steps else timesteps.shape[0]
-        print(f"Running PLMS Sampling with {total_steps} timesteps")
-
-        iterator = tqdm(time_range, desc="PLMS Sampler", total=total_steps)
-        old_eps = []
-
-        for i, step in enumerate(iterator):
-            index = total_steps - i - 1
-            ts = torch.full((b,), step, device=device, dtype=torch.long)
-            ts_next = torch.full(
-                (b,),
-                time_range[min(i + 1, len(time_range) - 1)],
-                device=device,
-                dtype=torch.long,
-            )
-
-            if mask is not None:
-                assert x0 is not None
-                img_orig = self.model.q_sample(
-                    x0, ts
-                )  # TODO: deterministic forward pass?
-                img = img_orig * mask + (1.0 - mask) * img
-
-            outs = self.p_sample_plms(
-                img,
-                cond,
-                ts,
-                index=index,
-                use_original_steps=ddim_use_original_steps,
-                quantize_denoised=quantize_denoised,
-                temperature=temperature,
-                noise_dropout=noise_dropout,
-                score_corrector=score_corrector,
-                corrector_kwargs=corrector_kwargs,
-                unconditional_guidance_scale=unconditional_guidance_scale,
-                unconditional_conditioning=unconditional_conditioning,
-                old_eps=old_eps,
-                t_next=ts_next,
-            )
-            img, pred_x0, e_t = outs
-            old_eps.append(e_t)
-            if len(old_eps) >= 4:
-                old_eps.pop(0)
-            if callback:
-                callback(i)
-            if img_callback:
-                img_callback(pred_x0, i)
-
-            if index % log_every_t == 0 or index == total_steps - 1:
-                intermediates["x_inter"].append(img)
-                intermediates["pred_x0"].append(pred_x0)
-
-        return img, intermediates
-
-    @torch.no_grad()
-    def p_sample_plms(
-        self,
-        x,
-        c,
-        t,
-        index,
-        repeat_noise=False,
-        use_original_steps=False,
-        quantize_denoised=False,
-        temperature=1.0,
-        noise_dropout=0.0,
-        score_corrector=None,
-        corrector_kwargs=None,
-        unconditional_guidance_scale=1.0,
-        unconditional_conditioning=None,
-        old_eps=None,
-        t_next=None,
-    ):
-        b, *_, device = *x.shape, x.device
-
-        def get_model_output(x, t):
-            if (
-                unconditional_conditioning is None
-                or unconditional_guidance_scale == 1.0
-            ):
-                e_t = self.model.apply_model(x, t, c)
-            else:
-                x_in = torch.cat([x] * 2)
-                t_in = torch.cat([t] * 2)
-                c_in = torch.cat([unconditional_conditioning, c])
-                e_t_uncond, e_t = self.model.apply_model(x_in, t_in, c_in).chunk(2)
-                e_t = e_t_uncond + unconditional_guidance_scale * (e_t - e_t_uncond)
-
-            if score_corrector is not None:
-                assert self.model.parameterization == "eps"
-                e_t = score_corrector.modify_score(
-                    self.model, e_t, x, t, c, **corrector_kwargs
-                )
-
-            return e_t
-
-        alphas = self.model.alphas_cumprod if use_original_steps else self.ddim_alphas
-        alphas_prev = (
-            self.model.alphas_cumprod_prev
-            if use_original_steps
-            else self.ddim_alphas_prev
-        )
-        sqrt_one_minus_alphas = (
-            self.model.sqrt_one_minus_alphas_cumprod
-            if use_original_steps
-            else self.ddim_sqrt_one_minus_alphas
-        )
-        sigmas = (
-            self.model.ddim_sigmas_for_original_num_steps
-            if use_original_steps
-            else self.ddim_sigmas
-        )
-
-        def get_x_prev_and_pred_x0(e_t, index):
-            # select parameters corresponding to the currently considered timestep
-            a_t = torch.full((b, 1, 1, 1), alphas[index], device=device)
-            a_prev = torch.full((b, 1, 1, 1), alphas_prev[index], device=device)
-            sigma_t = torch.full((b, 1, 1, 1), sigmas[index], device=device)
-            sqrt_one_minus_at = torch.full(
-                (b, 1, 1, 1), sqrt_one_minus_alphas[index], device=device
-            )
-
-            # current prediction for x_0
-            pred_x0 = (x - sqrt_one_minus_at * e_t) / a_t.sqrt()
-            if quantize_denoised:
-                pred_x0, _, *_ = self.model.first_stage_model.quantize(pred_x0)
-            # direction pointing to x_t
-            dir_xt = (1.0 - a_prev - sigma_t**2).sqrt() * e_t
-            noise = sigma_t * noise_like(x.shape, device, repeat_noise) * temperature
-            if noise_dropout > 0.0:
-                noise = torch.nn.functional.dropout(noise, p=noise_dropout)
-            x_prev = a_prev.sqrt() * pred_x0 + dir_xt + noise
-            return x_prev, pred_x0
-
-        e_t = get_model_output(x, t)
-        if len(old_eps) == 0:
-            # Pseudo Improved Euler (2nd order)
-            x_prev, pred_x0 = get_x_prev_and_pred_x0(e_t, index)
-            e_t_next = get_model_output(x_prev, t_next)
-            e_t_prime = (e_t + e_t_next) / 2
-        elif len(old_eps) == 1:
-            # 2nd order Pseudo Linear Multistep (Adams-Bashforth)
-            e_t_prime = (3 * e_t - old_eps[-1]) / 2
-        elif len(old_eps) == 2:
-            # 3nd order Pseudo Linear Multistep (Adams-Bashforth)
-            e_t_prime = (23 * e_t - 16 * old_eps[-1] + 5 * old_eps[-2]) / 12
-        elif len(old_eps) >= 3:
-            # 4nd order Pseudo Linear Multistep (Adams-Bashforth)
-            e_t_prime = (
-                55 * e_t - 59 * old_eps[-1] + 37 * old_eps[-2] - 9 * old_eps[-3]
-            ) / 24
-
-        x_prev, pred_x0 = get_x_prev_and_pred_x0(e_t_prime, index)
-
-        return x_prev, pred_x0, e_t
+"""SAMPLING ONLY."""
+
+import torch
+import numpy as np
+from tqdm import tqdm
+
+from audiosr.latent_diffusion.modules.diffusionmodules.util import (
+    make_ddim_sampling_parameters,
+    make_ddim_timesteps,
+    noise_like,
+)
+
+
+class PLMSSampler(object):
+    def __init__(self, model, schedule="linear", **kwargs):
+        super().__init__()
+        self.model = model
+        self.ddpm_num_timesteps = model.num_timesteps
+        self.schedule = schedule
+
+    def register_buffer(self, name, attr):
+        if type(attr) == torch.Tensor:
+            if attr.device != torch.device("cuda"):
+                attr = attr.to(torch.device("cuda"))
+        setattr(self, name, attr)
+
+    def make_schedule(
+        self, ddim_num_steps, ddim_discretize="uniform", ddim_eta=0.0, verbose=True
+    ):
+        if ddim_eta != 0:
+            ddim_eta = 0
+            # raise ValueError('ddim_eta must be 0 for PLMS')
+
+        self.ddim_timesteps = make_ddim_timesteps(
+            ddim_discr_method=ddim_discretize,
+            num_ddim_timesteps=ddim_num_steps,
+            num_ddpm_timesteps=self.ddpm_num_timesteps,
+            verbose=verbose,
+        )
+        alphas_cumprod = self.model.alphas_cumprod
+        assert (
+            alphas_cumprod.shape[0] == self.ddpm_num_timesteps
+        ), "alphas have to be defined for each timestep"
+        to_torch = lambda x: x.clone().detach().to(torch.float32).to(self.model.device)
+
+        self.register_buffer("betas", to_torch(self.model.betas))
+        self.register_buffer("alphas_cumprod", to_torch(alphas_cumprod))
+        self.register_buffer(
+            "alphas_cumprod_prev", to_torch(self.model.alphas_cumprod_prev)
+        )
+
+        # calculations for diffusion q(x_t | x_{t-1}) and others
+        self.register_buffer(
+            "sqrt_alphas_cumprod", to_torch(np.sqrt(alphas_cumprod.cpu()))
+        )
+        self.register_buffer(
+            "sqrt_one_minus_alphas_cumprod",
+            to_torch(np.sqrt(1.0 - alphas_cumprod.cpu())),
+        )
+        self.register_buffer(
+            "log_one_minus_alphas_cumprod", to_torch(np.log(1.0 - alphas_cumprod.cpu()))
+        )
+        self.register_buffer(
+            "sqrt_recip_alphas_cumprod", to_torch(np.sqrt(1.0 / alphas_cumprod.cpu()))
+        )
+        self.register_buffer(
+            "sqrt_recipm1_alphas_cumprod",
+            to_torch(np.sqrt(1.0 / alphas_cumprod.cpu() - 1)),
+        )
+
+        # ddim sampling parameters
+        ddim_sigmas, ddim_alphas, ddim_alphas_prev = make_ddim_sampling_parameters(
+            alphacums=alphas_cumprod.cpu(),
+            ddim_timesteps=self.ddim_timesteps,
+            eta=ddim_eta,
+            verbose=verbose,
+        )
+        self.register_buffer("ddim_sigmas", ddim_sigmas)
+        self.register_buffer("ddim_alphas", ddim_alphas)
+        self.register_buffer("ddim_alphas_prev", ddim_alphas_prev)
+        self.register_buffer("ddim_sqrt_one_minus_alphas", np.sqrt(1.0 - ddim_alphas))
+        sigmas_for_original_sampling_steps = ddim_eta * torch.sqrt(
+            (1 - self.alphas_cumprod_prev)
+            / (1 - self.alphas_cumprod)
+            * (1 - self.alphas_cumprod / self.alphas_cumprod_prev)
+        )
+        self.register_buffer(
+            "ddim_sigmas_for_original_num_steps", sigmas_for_original_sampling_steps
+        )
+
+    @torch.no_grad()
+    def sample(
+        self,
+        S,
+        batch_size,
+        shape,
+        conditioning=None,
+        callback=None,
+        normals_sequence=None,
+        img_callback=None,
+        quantize_x0=False,
+        eta=0.0,
+        mask=None,
+        x0=None,
+        temperature=1.0,
+        noise_dropout=0.0,
+        score_corrector=None,
+        corrector_kwargs=None,
+        verbose=True,
+        x_T=None,
+        log_every_t=100,
+        unconditional_guidance_scale=1.0,
+        unconditional_conditioning=None,
+        # this has to come in the same format as the conditioning, # e.g. as encoded tokens, ...
+        **kwargs,
+    ):
+        if conditioning is not None:
+            if isinstance(conditioning, dict):
+                cbs = conditioning[list(conditioning.keys())[0]].shape[0]
+                if cbs != batch_size:
+                    print(
+                        f"Warning: Got {cbs} conditionings but batch-size is {batch_size}"
+                    )
+            else:
+                if conditioning.shape[0] != batch_size:
+                    print(
+                        f"Warning: Got {conditioning.shape[0]} conditionings but batch-size is {batch_size}"
+                    )
+
+        self.make_schedule(ddim_num_steps=S, ddim_eta=eta, verbose=verbose)
+        # sampling
+        C, H, W = shape
+        size = (batch_size, C, H, W)
+        print(f"Data shape for PLMS sampling is {size}")
+
+        samples, intermediates = self.plms_sampling(
+            conditioning,
+            size,
+            callback=callback,
+            img_callback=img_callback,
+            quantize_denoised=quantize_x0,
+            mask=mask,
+            x0=x0,
+            ddim_use_original_steps=False,
+            noise_dropout=noise_dropout,
+            temperature=temperature,
+            score_corrector=score_corrector,
+            corrector_kwargs=corrector_kwargs,
+            x_T=x_T,
+            log_every_t=log_every_t,
+            unconditional_guidance_scale=unconditional_guidance_scale,
+            unconditional_conditioning=unconditional_conditioning,
+        )
+        return samples, intermediates
+
+    @torch.no_grad()
+    def plms_sampling(
+        self,
+        cond,
+        shape,
+        x_T=None,
+        ddim_use_original_steps=False,
+        callback=None,
+        timesteps=None,
+        quantize_denoised=False,
+        mask=None,
+        x0=None,
+        img_callback=None,
+        log_every_t=100,
+        temperature=1.0,
+        noise_dropout=0.0,
+        score_corrector=None,
+        corrector_kwargs=None,
+        unconditional_guidance_scale=1.0,
+        unconditional_conditioning=None,
+    ):
+        device = self.model.betas.device
+        b = shape[0]
+        if x_T is None:
+            img = torch.randn(shape, device=device)
+        else:
+            img = x_T
+
+        if timesteps is None:
+            timesteps = (
+                self.ddpm_num_timesteps
+                if ddim_use_original_steps
+                else self.ddim_timesteps
+            )
+        elif timesteps is not None and not ddim_use_original_steps:
+            subset_end = (
+                int(
+                    min(timesteps / self.ddim_timesteps.shape[0], 1)
+                    * self.ddim_timesteps.shape[0]
+                )
+                - 1
+            )
+            timesteps = self.ddim_timesteps[:subset_end]
+
+        intermediates = {"x_inter": [img], "pred_x0": [img]}
+        time_range = (
+            list(reversed(range(0, timesteps)))
+            if ddim_use_original_steps
+            else np.flip(timesteps)
+        )
+        total_steps = timesteps if ddim_use_original_steps else timesteps.shape[0]
+        print(f"Running PLMS Sampling with {total_steps} timesteps")
+
+        iterator = tqdm(time_range, desc="PLMS Sampler", total=total_steps)
+        old_eps = []
+
+        for i, step in enumerate(iterator):
+            index = total_steps - i - 1
+            ts = torch.full((b,), step, device=device, dtype=torch.long)
+            ts_next = torch.full(
+                (b,),
+                time_range[min(i + 1, len(time_range) - 1)],
+                device=device,
+                dtype=torch.long,
+            )
+
+            if mask is not None:
+                assert x0 is not None
+                img_orig = self.model.q_sample(
+                    x0, ts
+                )  # TODO: deterministic forward pass?
+                img = img_orig * mask + (1.0 - mask) * img
+
+            outs = self.p_sample_plms(
+                img,
+                cond,
+                ts,
+                index=index,
+                use_original_steps=ddim_use_original_steps,
+                quantize_denoised=quantize_denoised,
+                temperature=temperature,
+                noise_dropout=noise_dropout,
+                score_corrector=score_corrector,
+                corrector_kwargs=corrector_kwargs,
+                unconditional_guidance_scale=unconditional_guidance_scale,
+                unconditional_conditioning=unconditional_conditioning,
+                old_eps=old_eps,
+                t_next=ts_next,
+            )
+            img, pred_x0, e_t = outs
+            old_eps.append(e_t)
+            if len(old_eps) >= 4:
+                old_eps.pop(0)
+            if callback:
+                callback(i)
+            if img_callback:
+                img_callback(pred_x0, i)
+
+            if index % log_every_t == 0 or index == total_steps - 1:
+                intermediates["x_inter"].append(img)
+                intermediates["pred_x0"].append(pred_x0)
+
+        return img, intermediates
+
+    @torch.no_grad()
+    def p_sample_plms(
+        self,
+        x,
+        c,
+        t,
+        index,
+        repeat_noise=False,
+        use_original_steps=False,
+        quantize_denoised=False,
+        temperature=1.0,
+        noise_dropout=0.0,
+        score_corrector=None,
+        corrector_kwargs=None,
+        unconditional_guidance_scale=1.0,
+        unconditional_conditioning=None,
+        old_eps=None,
+        t_next=None,
+    ):
+        b, *_, device = *x.shape, x.device
+
+        def get_model_output(x, t):
+            if (
+                unconditional_conditioning is None
+                or unconditional_guidance_scale == 1.0
+            ):
+                e_t = self.model.apply_model(x, t, c)
+            else:
+                x_in = torch.cat([x] * 2)
+                t_in = torch.cat([t] * 2)
+                c_in = torch.cat([unconditional_conditioning, c])
+                e_t_uncond, e_t = self.model.apply_model(x_in, t_in, c_in).chunk(2)
+                e_t = e_t_uncond + unconditional_guidance_scale * (e_t - e_t_uncond)
+
+            if score_corrector is not None:
+                assert self.model.parameterization == "eps"
+                e_t = score_corrector.modify_score(
+                    self.model, e_t, x, t, c, **corrector_kwargs
+                )
+
+            return e_t
+
+        alphas = self.model.alphas_cumprod if use_original_steps else self.ddim_alphas
+        alphas_prev = (
+            self.model.alphas_cumprod_prev
+            if use_original_steps
+            else self.ddim_alphas_prev
+        )
+        sqrt_one_minus_alphas = (
+            self.model.sqrt_one_minus_alphas_cumprod
+            if use_original_steps
+            else self.ddim_sqrt_one_minus_alphas
+        )
+        sigmas = (
+            self.model.ddim_sigmas_for_original_num_steps
+            if use_original_steps
+            else self.ddim_sigmas
+        )
+
+        def get_x_prev_and_pred_x0(e_t, index):
+            # select parameters corresponding to the currently considered timestep
+            a_t = torch.full((b, 1, 1, 1), alphas[index], device=device)
+            a_prev = torch.full((b, 1, 1, 1), alphas_prev[index], device=device)
+            sigma_t = torch.full((b, 1, 1, 1), sigmas[index], device=device)
+            sqrt_one_minus_at = torch.full(
+                (b, 1, 1, 1), sqrt_one_minus_alphas[index], device=device
+            )
+
+            # current prediction for x_0
+            pred_x0 = (x - sqrt_one_minus_at * e_t) / a_t.sqrt()
+            if quantize_denoised:
+                pred_x0, _, *_ = self.model.first_stage_model.quantize(pred_x0)
+            # direction pointing to x_t
+            dir_xt = (1.0 - a_prev - sigma_t**2).sqrt() * e_t
+            noise = sigma_t * noise_like(x.shape, device, repeat_noise) * temperature
+            if noise_dropout > 0.0:
+                noise = torch.nn.functional.dropout(noise, p=noise_dropout)
+            x_prev = a_prev.sqrt() * pred_x0 + dir_xt + noise
+            return x_prev, pred_x0
+
+        e_t = get_model_output(x, t)
+        if len(old_eps) == 0:
+            # Pseudo Improved Euler (2nd order)
+            x_prev, pred_x0 = get_x_prev_and_pred_x0(e_t, index)
+            e_t_next = get_model_output(x_prev, t_next)
+            e_t_prime = (e_t + e_t_next) / 2
+        elif len(old_eps) == 1:
+            # 2nd order Pseudo Linear Multistep (Adams-Bashforth)
+            e_t_prime = (3 * e_t - old_eps[-1]) / 2
+        elif len(old_eps) == 2:
+            # 3nd order Pseudo Linear Multistep (Adams-Bashforth)
+            e_t_prime = (23 * e_t - 16 * old_eps[-1] + 5 * old_eps[-2]) / 12
+        elif len(old_eps) >= 3:
+            # 4nd order Pseudo Linear Multistep (Adams-Bashforth)
+            e_t_prime = (
+                55 * e_t - 59 * old_eps[-1] + 37 * old_eps[-2] - 9 * old_eps[-3]
+            ) / 24
+
+        x_prev, pred_x0 = get_x_prev_and_pred_x0(e_t_prime, index)
+
+        return x_prev, pred_x0, e_t

audiosr/latent_diffusion/modules/attention.py

@@ -1,467 +1,467 @@
-from inspect import isfunction
-import math
-import torch
-import torch.nn.functional as F
-from torch import nn, einsum
-from einops import rearrange, repeat
-
-from audiosr.latent_diffusion.modules.diffusionmodules.util import checkpoint
-
-
-def exists(val):
-    return val is not None
-
-
-def uniq(arr):
-    return {el: True for el in arr}.keys()
-
-
-def default(val, d):
-    if exists(val):
-        return val
-    return d() if isfunction(d) else d
-
-
-def max_neg_value(t):
-    return -torch.finfo(t.dtype).max
-
-
-def init_(tensor):
-    dim = tensor.shape[-1]
-    std = 1 / math.sqrt(dim)
-    tensor.uniform_(-std, std)
-    return tensor
-
-
-# feedforward
-class GEGLU(nn.Module):
-    def __init__(self, dim_in, dim_out):
-        super().__init__()
-        self.proj = nn.Linear(dim_in, dim_out * 2)
-
-    def forward(self, x):
-        x, gate = self.proj(x).chunk(2, dim=-1)
-        return x * F.gelu(gate)
-
-
-class FeedForward(nn.Module):
-    def __init__(self, dim, dim_out=None, mult=4, glu=False, dropout=0.0):
-        super().__init__()
-        inner_dim = int(dim * mult)
-        dim_out = default(dim_out, dim)
-        project_in = (
-            nn.Sequential(nn.Linear(dim, inner_dim), nn.GELU())
-            if not glu
-            else GEGLU(dim, inner_dim)
-        )
-
-        self.net = nn.Sequential(
-            project_in, nn.Dropout(dropout), nn.Linear(inner_dim, dim_out)
-        )
-
-    def forward(self, x):
-        return self.net(x)
-
-
-def zero_module(module):
-    """
-    Zero out the parameters of a module and return it.
-    """
-    for p in module.parameters():
-        p.detach().zero_()
-    return module
-
-
-def Normalize(in_channels):
-    return torch.nn.GroupNorm(
-        num_groups=32, num_channels=in_channels, eps=1e-6, affine=True
-    )
-
-
-class LinearAttention(nn.Module):
-    def __init__(self, dim, heads=4, dim_head=32):
-        super().__init__()
-        self.heads = heads
-        hidden_dim = dim_head * heads
-        self.to_qkv = nn.Conv2d(dim, hidden_dim * 3, 1, bias=False)
-        self.to_out = nn.Conv2d(hidden_dim, dim, 1)
-
-    def forward(self, x):
-        b, c, h, w = x.shape
-        qkv = self.to_qkv(x)
-        q, k, v = rearrange(
-            qkv, "b (qkv heads c) h w -> qkv b heads c (h w)", heads=self.heads, qkv=3
-        )
-        k = k.softmax(dim=-1)
-        context = torch.einsum("bhdn,bhen->bhde", k, v)
-        out = torch.einsum("bhde,bhdn->bhen", context, q)
-        out = rearrange(
-            out, "b heads c (h w) -> b (heads c) h w", heads=self.heads, h=h, w=w
-        )
-        return self.to_out(out)
-
-
-class SpatialSelfAttention(nn.Module):
-    def __init__(self, in_channels):
-        super().__init__()
-        self.in_channels = in_channels
-
-        self.norm = Normalize(in_channels)
-        self.q = torch.nn.Conv2d(
-            in_channels, in_channels, kernel_size=1, stride=1, padding=0
-        )
-        self.k = torch.nn.Conv2d(
-            in_channels, in_channels, kernel_size=1, stride=1, padding=0
-        )
-        self.v = torch.nn.Conv2d(
-            in_channels, in_channels, kernel_size=1, stride=1, padding=0
-        )
-        self.proj_out = torch.nn.Conv2d(
-            in_channels, in_channels, kernel_size=1, stride=1, padding=0
-        )
-
-    def forward(self, x):
-        h_ = x
-        h_ = self.norm(h_)
-        q = self.q(h_)
-        k = self.k(h_)
-        v = self.v(h_)
-
-        # compute attention
-        b, c, h, w = q.shape
-        q = rearrange(q, "b c h w -> b (h w) c")
-        k = rearrange(k, "b c h w -> b c (h w)")
-        w_ = torch.einsum("bij,bjk->bik", q, k)
-
-        w_ = w_ * (int(c) ** (-0.5))
-        w_ = torch.nn.functional.softmax(w_, dim=2)
-
-        # attend to values
-        v = rearrange(v, "b c h w -> b c (h w)")
-        w_ = rearrange(w_, "b i j -> b j i")
-        h_ = torch.einsum("bij,bjk->bik", v, w_)
-        h_ = rearrange(h_, "b c (h w) -> b c h w", h=h)
-        h_ = self.proj_out(h_)
-
-        return x + h_
-
-
-# class CrossAttention(nn.Module):
-#     """
-#     ### Cross Attention Layer
-#     This falls-back to self-attention when conditional embeddings are not specified.
-#     """
-
-#     use_flash_attention: bool = True
-
-#     # use_flash_attention: bool = False
-#     def __init__(
-#         self,
-#         query_dim,
-#         context_dim=None,
-#         heads=8,
-#         dim_head=64,
-#         dropout=0.0,
-#         is_inplace: bool = True,
-#     ):
-#         # def __init__(self, d_model: int, d_cond: int, n_heads: int, d_head: int, is_inplace: bool = True):
-#         """
-#         :param d_model: is the input embedding size
-#         :param n_heads: is the number of attention heads
-#         :param d_head: is the size of a attention head
-#         :param d_cond: is the size of the conditional embeddings
-#         :param is_inplace: specifies whether to perform the attention softmax computation inplace to
-#             save memory
-#         """
-#         super().__init__()
-
-#         self.is_inplace = is_inplace
-#         self.n_heads = heads
-#         self.d_head = dim_head
-
-#         # Attention scaling factor
-#         self.scale = dim_head**-0.5
-
-#         # The normal self-attention layer
-#         if context_dim is None:
-#             context_dim = query_dim
-
-#         # Query, key and value mappings
-#         d_attn = dim_head * heads
-#         self.to_q = nn.Linear(query_dim, d_attn, bias=False)
-#         self.to_k = nn.Linear(context_dim, d_attn, bias=False)
-#         self.to_v = nn.Linear(context_dim, d_attn, bias=False)
-
-#         # Final linear layer
-#         self.to_out = nn.Sequential(nn.Linear(d_attn, query_dim), nn.Dropout(dropout))
-
-#         # Setup [flash attention](https://github.com/HazyResearch/flash-attention).
-#         # Flash attention is only used if it's installed
-#         # and `CrossAttention.use_flash_attention` is set to `True`.
-#         try:
-#             # You can install flash attention by cloning their Github repo,
-#             # [https://github.com/HazyResearch/flash-attention](https://github.com/HazyResearch/flash-attention)
-#             # and then running `python setup.py install`
-#             from flash_attn.flash_attention import FlashAttention
-
-#             self.flash = FlashAttention()
-#             # Set the scale for scaled dot-product attention.
-#             self.flash.softmax_scale = self.scale
-#         # Set to `None` if it's not installed
-#         except ImportError:
-#             self.flash = None
-
-#     def forward(self, x, context=None, mask=None):
-#         """
-#         :param x: are the input embeddings of shape `[batch_size, height * width, d_model]`
-#         :param cond: is the conditional embeddings of shape `[batch_size, n_cond, d_cond]`
-#         """
-
-#         # If `cond` is `None` we perform self attention
-#         has_cond = context is not None
-#         if not has_cond:
-#             context = x
-
-#         # Get query, key and value vectors
-#         q = self.to_q(x)
-#         k = self.to_k(context)
-#         v = self.to_v(context)
-
-#         # Use flash attention if it's available and the head size is less than or equal to `128`
-#         if (
-#             CrossAttention.use_flash_attention
-#             and self.flash is not None
-#             and not has_cond
-#             and self.d_head <= 128
-#         ):
-#             return self.flash_attention(q, k, v)
-#         # Otherwise, fallback to normal attention
-#         else:
-#             return self.normal_attention(q, k, v)
-
-#     def flash_attention(self, q: torch.Tensor, k: torch.Tensor, v: torch.Tensor):
-#         """
-#         #### Flash Attention
-#         :param q: are the query vectors before splitting heads, of shape `[batch_size, seq, d_attn]`
-#         :param k: are the query vectors before splitting heads, of shape `[batch_size, seq, d_attn]`
-#         :param v: are the query vectors before splitting heads, of shape `[batch_size, seq, d_attn]`
-#         """
-
-#         # Get batch size and number of elements along sequence axis (`width * height`)
-#         batch_size, seq_len, _ = q.shape
-
-#         # Stack `q`, `k`, `v` vectors for flash attention, to get a single tensor of
-#         # shape `[batch_size, seq_len, 3, n_heads * d_head]`
-#         qkv = torch.stack((q, k, v), dim=2)
-#         # Split the heads
-#         qkv = qkv.view(batch_size, seq_len, 3, self.n_heads, self.d_head)
-
-#         # Flash attention works for head sizes `32`, `64` and `128`, so we have to pad the heads to
-#         # fit this size.
-#         if self.d_head <= 32:
-#             pad = 32 - self.d_head
-#         elif self.d_head <= 64:
-#             pad = 64 - self.d_head
-#         elif self.d_head <= 128:
-#             pad = 128 - self.d_head
-#         else:
-#             raise ValueError(f"Head size ${self.d_head} too large for Flash Attention")
-
-#         # Pad the heads
-#         if pad:
-#             qkv = torch.cat(
-#                 (qkv, qkv.new_zeros(batch_size, seq_len, 3, self.n_heads, pad)), dim=-1
-#             )
-
-#         # Compute attention
-#         # $$\underset{seq}{softmax}\Bigg(\frac{Q K^\top}{\sqrt{d_{key}}}\Bigg)V$$
-#         # This gives a tensor of shape `[batch_size, seq_len, n_heads, d_padded]`
-#         # TODO here I add the dtype changing
-#         out, _ = self.flash(qkv.type(torch.float16))
-#         # Truncate the extra head size
-#         out = out[:, :, :, : self.d_head].float()
-#         # Reshape to `[batch_size, seq_len, n_heads * d_head]`
-#         out = out.reshape(batch_size, seq_len, self.n_heads * self.d_head)
-
-#         # Map to `[batch_size, height * width, d_model]` with a linear layer
-#         return self.to_out(out)
-
-#     def normal_attention(self, q: torch.Tensor, k: torch.Tensor, v: torch.Tensor):
-#         """
-#         #### Normal Attention
-
-#         :param q: are the query vectors before splitting heads, of shape `[batch_size, seq, d_attn]`
-#         :param k: are the query vectors before splitting heads, of shape `[batch_size, seq, d_attn]`
-#         :param v: are the query vectors before splitting heads, of shape `[batch_size, seq, d_attn]`
-#         """
-
-#         # Split them to heads of shape `[batch_size, seq_len, n_heads, d_head]`
-#         q = q.view(*q.shape[:2], self.n_heads, -1)  # [bs, 64, 20, 32]
-#         k = k.view(*k.shape[:2], self.n_heads, -1)  # [bs, 1, 20, 32]
-#         v = v.view(*v.shape[:2], self.n_heads, -1)
-
-#         # Calculate attention $\frac{Q K^\top}{\sqrt{d_{key}}}$
-#         attn = torch.einsum("bihd,bjhd->bhij", q, k) * self.scale
-
-#         # Compute softmax
-#         # $$\underset{seq}{softmax}\Bigg(\frac{Q K^\top}{\sqrt{d_{key}}}\Bigg)$$
-#         if self.is_inplace:
-#             half = attn.shape[0] // 2
-#             attn[half:] = attn[half:].softmax(dim=-1)
-#             attn[:half] = attn[:half].softmax(dim=-1)
-#         else:
-#             attn = attn.softmax(dim=-1)
-
-#         # Compute attention output
-#         # $$\underset{seq}{softmax}\Bigg(\frac{Q K^\top}{\sqrt{d_{key}}}\Bigg)V$$
-#         # attn: [bs, 20, 64, 1]
-#         # v: [bs, 1, 20, 32]
-#         out = torch.einsum("bhij,bjhd->bihd", attn, v)
-#         # Reshape to `[batch_size, height * width, n_heads * d_head]`
-#         out = out.reshape(*out.shape[:2], -1)
-#         # Map to `[batch_size, height * width, d_model]` with a linear layer
-#         return self.to_out(out)
-
-
-class CrossAttention(nn.Module):
-    def __init__(self, query_dim, context_dim=None, heads=8, dim_head=64, dropout=0.0):
-        super().__init__()
-        inner_dim = dim_head * heads
-        context_dim = default(context_dim, query_dim)
-
-        self.scale = dim_head**-0.5
-        self.heads = heads
-
-        self.to_q = nn.Linear(query_dim, inner_dim, bias=False)
-        self.to_k = nn.Linear(context_dim, inner_dim, bias=False)
-        self.to_v = nn.Linear(context_dim, inner_dim, bias=False)
-
-        self.to_out = nn.Sequential(
-            nn.Linear(inner_dim, query_dim), nn.Dropout(dropout)
-        )
-
-    def forward(self, x, context=None, mask=None):
-        h = self.heads
-
-        q = self.to_q(x)
-        context = default(context, x)
-
-        k = self.to_k(context)
-        v = self.to_v(context)
-
-        q, k, v = map(lambda t: rearrange(t, "b n (h d) -> (b h) n d", h=h), (q, k, v))
-
-        sim = einsum("b i d, b j d -> b i j", q, k) * self.scale
-
-        if exists(mask):
-            mask = rearrange(mask, "b ... -> b (...)")
-            max_neg_value = -torch.finfo(sim.dtype).max
-            mask = repeat(mask, "b j -> (b h) () j", h=h)
-            sim.masked_fill_(~(mask == 1), max_neg_value)
-
-        # attention, what we cannot get enough of
-        attn = sim.softmax(dim=-1)
-
-        out = einsum("b i j, b j d -> b i d", attn, v)
-        out = rearrange(out, "(b h) n d -> b n (h d)", h=h)
-        return self.to_out(out)
-
-
-class BasicTransformerBlock(nn.Module):
-    def __init__(
-        self,
-        dim,
-        n_heads,
-        d_head,
-        dropout=0.0,
-        context_dim=None,
-        gated_ff=True,
-        checkpoint=True,
-    ):
-        super().__init__()
-        self.attn1 = CrossAttention(
-            query_dim=dim, heads=n_heads, dim_head=d_head, dropout=dropout
-        )  # is a self-attention
-        self.ff = FeedForward(dim, dropout=dropout, glu=gated_ff)
-        self.attn2 = CrossAttention(
-            query_dim=dim,
-            context_dim=context_dim,
-            heads=n_heads,
-            dim_head=d_head,
-            dropout=dropout,
-        )  # is self-attn if context is none
-        self.norm1 = nn.LayerNorm(dim)
-        self.norm2 = nn.LayerNorm(dim)
-        self.norm3 = nn.LayerNorm(dim)
-        self.checkpoint = checkpoint
-
-    def forward(self, x, context=None, mask=None):
-        if context is None:
-            return checkpoint(self._forward, (x,), self.parameters(), self.checkpoint)
-        else:
-            return checkpoint(
-                self._forward, (x, context, mask), self.parameters(), self.checkpoint
-            )
-
-    def _forward(self, x, context=None, mask=None):
-        x = self.attn1(self.norm1(x)) + x
-        x = self.attn2(self.norm2(x), context=context, mask=mask) + x
-        x = self.ff(self.norm3(x)) + x
-        return x
-
-
-class SpatialTransformer(nn.Module):
-    """
-    Transformer block for image-like data.
-    First, project the input (aka embedding)
-    and reshape to b, t, d.
-    Then apply standard transformer action.
-    Finally, reshape to image
-    """
-
-    def __init__(
-        self,
-        in_channels,
-        n_heads,
-        d_head,
-        depth=1,
-        dropout=0.0,
-        context_dim=None,
-    ):
-        super().__init__()
-
-        context_dim = context_dim
-
-        self.in_channels = in_channels
-        inner_dim = n_heads * d_head
-        self.norm = Normalize(in_channels)
-
-        self.proj_in = nn.Conv2d(
-            in_channels, inner_dim, kernel_size=1, stride=1, padding=0
-        )
-
-        self.transformer_blocks = nn.ModuleList(
-            [
-                BasicTransformerBlock(
-                    inner_dim, n_heads, d_head, dropout=dropout, context_dim=context_dim
-                )
-                for d in range(depth)
-            ]
-        )
-
-        self.proj_out = zero_module(
-            nn.Conv2d(inner_dim, in_channels, kernel_size=1, stride=1, padding=0)
-        )
-
-    def forward(self, x, context=None, mask=None):
-        # note: if no context is given, cross-attention defaults to self-attention
-        b, c, h, w = x.shape
-        x_in = x
-        x = self.norm(x)
-        x = self.proj_in(x)
-        x = rearrange(x, "b c h w -> b (h w) c")
-        for block in self.transformer_blocks:
-            x = block(x, context=context, mask=mask)
-        x = rearrange(x, "b (h w) c -> b c h w", h=h, w=w)
-        x = self.proj_out(x)
-        return x + x_in
+from inspect import isfunction
+import math
+import torch
+import torch.nn.functional as F
+from torch import nn, einsum
+from einops import rearrange, repeat
+
+from audiosr.latent_diffusion.modules.diffusionmodules.util import checkpoint
+
+
+def exists(val):
+    return val is not None
+
+
+def uniq(arr):
+    return {el: True for el in arr}.keys()
+
+
+def default(val, d):
+    if exists(val):
+        return val
+    return d() if isfunction(d) else d
+
+
+def max_neg_value(t):
+    return -torch.finfo(t.dtype).max
+
+
+def init_(tensor):
+    dim = tensor.shape[-1]
+    std = 1 / math.sqrt(dim)
+    tensor.uniform_(-std, std)
+    return tensor
+
+
+# feedforward
+class GEGLU(nn.Module):
+    def __init__(self, dim_in, dim_out):
+        super().__init__()
+        self.proj = nn.Linear(dim_in, dim_out * 2)
+
+    def forward(self, x):
+        x, gate = self.proj(x).chunk(2, dim=-1)
+        return x * F.gelu(gate)
+
+
+class FeedForward(nn.Module):
+    def __init__(self, dim, dim_out=None, mult=4, glu=False, dropout=0.0):
+        super().__init__()
+        inner_dim = int(dim * mult)
+        dim_out = default(dim_out, dim)
+        project_in = (
+            nn.Sequential(nn.Linear(dim, inner_dim), nn.GELU())
+            if not glu
+            else GEGLU(dim, inner_dim)
+        )
+
+        self.net = nn.Sequential(
+            project_in, nn.Dropout(dropout), nn.Linear(inner_dim, dim_out)
+        )
+
+    def forward(self, x):
+        return self.net(x)
+
+
+def zero_module(module):
+    """
+    Zero out the parameters of a module and return it.
+    """
+    for p in module.parameters():
+        p.detach().zero_()
+    return module
+
+
+def Normalize(in_channels):
+    return torch.nn.GroupNorm(
+        num_groups=32, num_channels=in_channels, eps=1e-6, affine=True
+    )
+
+
+class LinearAttention(nn.Module):
+    def __init__(self, dim, heads=4, dim_head=32):
+        super().__init__()
+        self.heads = heads
+        hidden_dim = dim_head * heads
+        self.to_qkv = nn.Conv2d(dim, hidden_dim * 3, 1, bias=False)
+        self.to_out = nn.Conv2d(hidden_dim, dim, 1)
+
+    def forward(self, x):
+        b, c, h, w = x.shape
+        qkv = self.to_qkv(x)
+        q, k, v = rearrange(
+            qkv, "b (qkv heads c) h w -> qkv b heads c (h w)", heads=self.heads, qkv=3
+        )
+        k = k.softmax(dim=-1)
+        context = torch.einsum("bhdn,bhen->bhde", k, v)
+        out = torch.einsum("bhde,bhdn->bhen", context, q)
+        out = rearrange(
+            out, "b heads c (h w) -> b (heads c) h w", heads=self.heads, h=h, w=w
+        )
+        return self.to_out(out)
+
+
+class SpatialSelfAttention(nn.Module):
+    def __init__(self, in_channels):
+        super().__init__()
+        self.in_channels = in_channels
+
+        self.norm = Normalize(in_channels)
+        self.q = torch.nn.Conv2d(
+            in_channels, in_channels, kernel_size=1, stride=1, padding=0
+        )
+        self.k = torch.nn.Conv2d(
+            in_channels, in_channels, kernel_size=1, stride=1, padding=0
+        )
+        self.v = torch.nn.Conv2d(
+            in_channels, in_channels, kernel_size=1, stride=1, padding=0
+        )
+        self.proj_out = torch.nn.Conv2d(
+            in_channels, in_channels, kernel_size=1, stride=1, padding=0
+        )
+
+    def forward(self, x):
+        h_ = x
+        h_ = self.norm(h_)
+        q = self.q(h_)
+        k = self.k(h_)
+        v = self.v(h_)
+
+        # compute attention
+        b, c, h, w = q.shape
+        q = rearrange(q, "b c h w -> b (h w) c")
+        k = rearrange(k, "b c h w -> b c (h w)")
+        w_ = torch.einsum("bij,bjk->bik", q, k)
+
+        w_ = w_ * (int(c) ** (-0.5))
+        w_ = torch.nn.functional.softmax(w_, dim=2)
+
+        # attend to values
+        v = rearrange(v, "b c h w -> b c (h w)")
+        w_ = rearrange(w_, "b i j -> b j i")
+        h_ = torch.einsum("bij,bjk->bik", v, w_)
+        h_ = rearrange(h_, "b c (h w) -> b c h w", h=h)
+        h_ = self.proj_out(h_)
+
+        return x + h_
+
+
+# class CrossAttention(nn.Module):
+#     """
+#     ### Cross Attention Layer
+#     This falls-back to self-attention when conditional embeddings are not specified.
+#     """
+
+#     use_flash_attention: bool = True
+
+#     # use_flash_attention: bool = False
+#     def __init__(
+#         self,
+#         query_dim,
+#         context_dim=None,
+#         heads=8,
+#         dim_head=64,
+#         dropout=0.0,
+#         is_inplace: bool = True,
+#     ):
+#         # def __init__(self, d_model: int, d_cond: int, n_heads: int, d_head: int, is_inplace: bool = True):
+#         """
+#         :param d_model: is the input embedding size
+#         :param n_heads: is the number of attention heads
+#         :param d_head: is the size of a attention head
+#         :param d_cond: is the size of the conditional embeddings
+#         :param is_inplace: specifies whether to perform the attention softmax computation inplace to
+#             save memory
+#         """
+#         super().__init__()
+
+#         self.is_inplace = is_inplace
+#         self.n_heads = heads
+#         self.d_head = dim_head
+
+#         # Attention scaling factor
+#         self.scale = dim_head**-0.5
+
+#         # The normal self-attention layer
+#         if context_dim is None:
+#             context_dim = query_dim
+
+#         # Query, key and value mappings
+#         d_attn = dim_head * heads
+#         self.to_q = nn.Linear(query_dim, d_attn, bias=False)
+#         self.to_k = nn.Linear(context_dim, d_attn, bias=False)
+#         self.to_v = nn.Linear(context_dim, d_attn, bias=False)
+
+#         # Final linear layer
+#         self.to_out = nn.Sequential(nn.Linear(d_attn, query_dim), nn.Dropout(dropout))
+
+#         # Setup [flash attention](https://github.com/HazyResearch/flash-attention).
+#         # Flash attention is only used if it's installed
+#         # and `CrossAttention.use_flash_attention` is set to `True`.
+#         try:
+#             # You can install flash attention by cloning their Github repo,
+#             # [https://github.com/HazyResearch/flash-attention](https://github.com/HazyResearch/flash-attention)
+#             # and then running `python setup.py install`
+#             from flash_attn.flash_attention import FlashAttention
+
+#             self.flash = FlashAttention()
+#             # Set the scale for scaled dot-product attention.
+#             self.flash.softmax_scale = self.scale
+#         # Set to `None` if it's not installed
+#         except ImportError:
+#             self.flash = None
+
+#     def forward(self, x, context=None, mask=None):
+#         """
+#         :param x: are the input embeddings of shape `[batch_size, height * width, d_model]`
+#         :param cond: is the conditional embeddings of shape `[batch_size, n_cond, d_cond]`
+#         """
+
+#         # If `cond` is `None` we perform self attention
+#         has_cond = context is not None
+#         if not has_cond:
+#             context = x
+
+#         # Get query, key and value vectors
+#         q = self.to_q(x)
+#         k = self.to_k(context)
+#         v = self.to_v(context)
+
+#         # Use flash attention if it's available and the head size is less than or equal to `128`
+#         if (
+#             CrossAttention.use_flash_attention
+#             and self.flash is not None
+#             and not has_cond
+#             and self.d_head <= 128
+#         ):
+#             return self.flash_attention(q, k, v)
+#         # Otherwise, fallback to normal attention
+#         else:
+#             return self.normal_attention(q, k, v)
+
+#     def flash_attention(self, q: torch.Tensor, k: torch.Tensor, v: torch.Tensor):
+#         """
+#         #### Flash Attention
+#         :param q: are the query vectors before splitting heads, of shape `[batch_size, seq, d_attn]`
+#         :param k: are the query vectors before splitting heads, of shape `[batch_size, seq, d_attn]`
+#         :param v: are the query vectors before splitting heads, of shape `[batch_size, seq, d_attn]`
+#         """
+
+#         # Get batch size and number of elements along sequence axis (`width * height`)
+#         batch_size, seq_len, _ = q.shape
+
+#         # Stack `q`, `k`, `v` vectors for flash attention, to get a single tensor of
+#         # shape `[batch_size, seq_len, 3, n_heads * d_head]`
+#         qkv = torch.stack((q, k, v), dim=2)
+#         # Split the heads
+#         qkv = qkv.view(batch_size, seq_len, 3, self.n_heads, self.d_head)
+
+#         # Flash attention works for head sizes `32`, `64` and `128`, so we have to pad the heads to
+#         # fit this size.
+#         if self.d_head <= 32:
+#             pad = 32 - self.d_head
+#         elif self.d_head <= 64:
+#             pad = 64 - self.d_head
+#         elif self.d_head <= 128:
+#             pad = 128 - self.d_head
+#         else:
+#             raise ValueError(f"Head size ${self.d_head} too large for Flash Attention")
+
+#         # Pad the heads
+#         if pad:
+#             qkv = torch.cat(
+#                 (qkv, qkv.new_zeros(batch_size, seq_len, 3, self.n_heads, pad)), dim=-1
+#             )
+
+#         # Compute attention
+#         # $$\underset{seq}{softmax}\Bigg(\frac{Q K^\top}{\sqrt{d_{key}}}\Bigg)V$$
+#         # This gives a tensor of shape `[batch_size, seq_len, n_heads, d_padded]`
+#         # TODO here I add the dtype changing
+#         out, _ = self.flash(qkv.type(torch.float16))
+#         # Truncate the extra head size
+#         out = out[:, :, :, : self.d_head].float()
+#         # Reshape to `[batch_size, seq_len, n_heads * d_head]`
+#         out = out.reshape(batch_size, seq_len, self.n_heads * self.d_head)
+
+#         # Map to `[batch_size, height * width, d_model]` with a linear layer
+#         return self.to_out(out)
+
+#     def normal_attention(self, q: torch.Tensor, k: torch.Tensor, v: torch.Tensor):
+#         """
+#         #### Normal Attention
+
+#         :param q: are the query vectors before splitting heads, of shape `[batch_size, seq, d_attn]`
+#         :param k: are the query vectors before splitting heads, of shape `[batch_size, seq, d_attn]`
+#         :param v: are the query vectors before splitting heads, of shape `[batch_size, seq, d_attn]`
+#         """
+
+#         # Split them to heads of shape `[batch_size, seq_len, n_heads, d_head]`
+#         q = q.view(*q.shape[:2], self.n_heads, -1)  # [bs, 64, 20, 32]
+#         k = k.view(*k.shape[:2], self.n_heads, -1)  # [bs, 1, 20, 32]
+#         v = v.view(*v.shape[:2], self.n_heads, -1)
+
+#         # Calculate attention $\frac{Q K^\top}{\sqrt{d_{key}}}$
+#         attn = torch.einsum("bihd,bjhd->bhij", q, k) * self.scale
+
+#         # Compute softmax
+#         # $$\underset{seq}{softmax}\Bigg(\frac{Q K^\top}{\sqrt{d_{key}}}\Bigg)$$
+#         if self.is_inplace:
+#             half = attn.shape[0] // 2
+#             attn[half:] = attn[half:].softmax(dim=-1)
+#             attn[:half] = attn[:half].softmax(dim=-1)
+#         else:
+#             attn = attn.softmax(dim=-1)
+
+#         # Compute attention output
+#         # $$\underset{seq}{softmax}\Bigg(\frac{Q K^\top}{\sqrt{d_{key}}}\Bigg)V$$
+#         # attn: [bs, 20, 64, 1]
+#         # v: [bs, 1, 20, 32]
+#         out = torch.einsum("bhij,bjhd->bihd", attn, v)
+#         # Reshape to `[batch_size, height * width, n_heads * d_head]`
+#         out = out.reshape(*out.shape[:2], -1)
+#         # Map to `[batch_size, height * width, d_model]` with a linear layer
+#         return self.to_out(out)
+
+
+class CrossAttention(nn.Module):
+    def __init__(self, query_dim, context_dim=None, heads=8, dim_head=64, dropout=0.0):
+        super().__init__()
+        inner_dim = dim_head * heads
+        context_dim = default(context_dim, query_dim)
+
+        self.scale = dim_head**-0.5
+        self.heads = heads
+
+        self.to_q = nn.Linear(query_dim, inner_dim, bias=False)
+        self.to_k = nn.Linear(context_dim, inner_dim, bias=False)
+        self.to_v = nn.Linear(context_dim, inner_dim, bias=False)
+
+        self.to_out = nn.Sequential(
+            nn.Linear(inner_dim, query_dim), nn.Dropout(dropout)
+        )
+
+    def forward(self, x, context=None, mask=None):
+        h = self.heads
+
+        q = self.to_q(x)
+        context = default(context, x)
+
+        k = self.to_k(context)
+        v = self.to_v(context)
+
+        q, k, v = map(lambda t: rearrange(t, "b n (h d) -> (b h) n d", h=h), (q, k, v))
+
+        sim = einsum("b i d, b j d -> b i j", q, k) * self.scale
+
+        if exists(mask):
+            mask = rearrange(mask, "b ... -> b (...)")
+            max_neg_value = -torch.finfo(sim.dtype).max
+            mask = repeat(mask, "b j -> (b h) () j", h=h)
+            sim.masked_fill_(~(mask == 1), max_neg_value)
+
+        # attention, what we cannot get enough of
+        attn = sim.softmax(dim=-1)
+
+        out = einsum("b i j, b j d -> b i d", attn, v)
+        out = rearrange(out, "(b h) n d -> b n (h d)", h=h)
+        return self.to_out(out)
+
+
+class BasicTransformerBlock(nn.Module):
+    def __init__(
+        self,
+        dim,
+        n_heads,
+        d_head,
+        dropout=0.0,
+        context_dim=None,
+        gated_ff=True,
+        checkpoint=True,
+    ):
+        super().__init__()
+        self.attn1 = CrossAttention(
+            query_dim=dim, heads=n_heads, dim_head=d_head, dropout=dropout
+        )  # is a self-attention
+        self.ff = FeedForward(dim, dropout=dropout, glu=gated_ff)
+        self.attn2 = CrossAttention(
+            query_dim=dim,
+            context_dim=context_dim,
+            heads=n_heads,
+            dim_head=d_head,
+            dropout=dropout,
+        )  # is self-attn if context is none
+        self.norm1 = nn.LayerNorm(dim)
+        self.norm2 = nn.LayerNorm(dim)
+        self.norm3 = nn.LayerNorm(dim)
+        self.checkpoint = checkpoint
+
+    def forward(self, x, context=None, mask=None):
+        if context is None:
+            return checkpoint(self._forward, (x,), self.parameters(), self.checkpoint)
+        else:
+            return checkpoint(
+                self._forward, (x, context, mask), self.parameters(), self.checkpoint
+            )
+
+    def _forward(self, x, context=None, mask=None):
+        x = self.attn1(self.norm1(x)) + x
+        x = self.attn2(self.norm2(x), context=context, mask=mask) + x
+        x = self.ff(self.norm3(x)) + x
+        return x
+
+
+class SpatialTransformer(nn.Module):
+    """
+    Transformer block for image-like data.
+    First, project the input (aka embedding)
+    and reshape to b, t, d.
+    Then apply standard transformer action.
+    Finally, reshape to image
+    """
+
+    def __init__(
+        self,
+        in_channels,
+        n_heads,
+        d_head,
+        depth=1,
+        dropout=0.0,
+        context_dim=None,
+    ):
+        super().__init__()
+
+        context_dim = context_dim
+
+        self.in_channels = in_channels
+        inner_dim = n_heads * d_head
+        self.norm = Normalize(in_channels)
+
+        self.proj_in = nn.Conv2d(
+            in_channels, inner_dim, kernel_size=1, stride=1, padding=0
+        )
+
+        self.transformer_blocks = nn.ModuleList(
+            [
+                BasicTransformerBlock(
+                    inner_dim, n_heads, d_head, dropout=dropout, context_dim=context_dim
+                )
+                for d in range(depth)
+            ]
+        )
+
+        self.proj_out = zero_module(
+            nn.Conv2d(inner_dim, in_channels, kernel_size=1, stride=1, padding=0)
+        )
+
+    def forward(self, x, context=None, mask=None):
+        # note: if no context is given, cross-attention defaults to self-attention
+        b, c, h, w = x.shape
+        x_in = x
+        x = self.norm(x)
+        x = self.proj_in(x)
+        x = rearrange(x, "b c h w -> b (h w) c")
+        for block in self.transformer_blocks:
+            x = block(x, context=context, mask=mask)
+        x = rearrange(x, "b (h w) c -> b c h w", h=h, w=w)
+        x = self.proj_out(x)
+        return x + x_in

audiosr/latent_diffusion/modules/audiomae/AudioMAE.py

@@ -1,149 +1,149 @@
-"""
-Reference Repo: https://github.com/facebookresearch/AudioMAE
-"""
-
-import torch
-import torch.nn as nn
-from timm.models.layers import to_2tuple
-import audiosr.latent_diffusion.modules.audiomae.models_vit as models_vit
-import audiosr.latent_diffusion.modules.audiomae.models_mae as models_mae
-
-# model = mae_vit_base_patch16(in_chans=1, audio_exp=True, img_size=(1024, 128))
-
-
-class PatchEmbed_new(nn.Module):
-    """Flexible Image to Patch Embedding"""
-
-    def __init__(
-        self, img_size=224, patch_size=16, in_chans=3, embed_dim=768, stride=10
-    ):
-        super().__init__()
-        img_size = to_2tuple(img_size)
-        patch_size = to_2tuple(patch_size)
-        stride = to_2tuple(stride)
-
-        self.img_size = img_size
-        self.patch_size = patch_size
-
-        self.proj = nn.Conv2d(
-            in_chans, embed_dim, kernel_size=patch_size, stride=stride
-        )  # with overlapped patches
-        # self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size)
-
-        # self.patch_hw = (img_size[1] // patch_size[1], img_size[0] // patch_size[0])
-        # self.num_patches = (img_size[1] // patch_size[1]) * (img_size[0] // patch_size[0])
-        _, _, h, w = self.get_output_shape(img_size)  # n, emb_dim, h, w
-        self.patch_hw = (h, w)
-        self.num_patches = h * w
-
-    def get_output_shape(self, img_size):
-        # todo: don't be lazy..
-        return self.proj(torch.randn(1, 1, img_size[0], img_size[1])).shape
-
-    def forward(self, x):
-        B, C, H, W = x.shape
-        # FIXME look at relaxing size constraints
-        # assert H == self.img_size[0] and W == self.img_size[1], \
-        #    f"Input image size ({H}*{W}) doesn't match model ({self.img_size[0]}*{self.img_size[1]})."
-        x = self.proj(x)
-        x = x.flatten(2).transpose(1, 2)
-        return x
-
-
-class AudioMAE(nn.Module):
-    """Audio Masked Autoencoder (MAE) pre-trained and finetuned on AudioSet (for SoundCLIP)"""
-
-    def __init__(
-        self,
-    ):
-        super().__init__()
-        model = models_vit.__dict__["vit_base_patch16"](
-            num_classes=527,
-            drop_path_rate=0.1,
-            global_pool=True,
-            mask_2d=True,
-            use_custom_patch=False,
-        )
-
-        img_size = (1024, 128)
-        emb_dim = 768
-
-        model.patch_embed = PatchEmbed_new(
-            img_size=img_size,
-            patch_size=(16, 16),
-            in_chans=1,
-            embed_dim=emb_dim,
-            stride=16,
-        )
-        num_patches = model.patch_embed.num_patches
-        # num_patches = 512 # assume audioset, 1024//16=64, 128//16=8, 512=64x8
-        model.pos_embed = nn.Parameter(
-            torch.zeros(1, num_patches + 1, emb_dim), requires_grad=False
-        )  # fixed sin-cos embedding
-
-        # checkpoint_path = '/mnt/bn/data-xubo/project/Masked_AudioEncoder/checkpoint/finetuned.pth'
-        # checkpoint = torch.load(checkpoint_path, map_location='cpu')
-        # msg = model.load_state_dict(checkpoint['model'], strict=False)
-        # print(f'Load AudioMAE from {checkpoint_path} / message: {msg}')
-
-        self.model = model
-
-    def forward(self, x, mask_t_prob=0.0, mask_f_prob=0.0):
-        """
-        x: mel fbank [Batch, 1, T, F]
-        mask_t_prob: 'T masking ratio (percentage of removed patches).'
-        mask_f_prob: 'F masking ratio (percentage of removed patches).'
-        """
-        return self.model(x=x, mask_t_prob=mask_t_prob, mask_f_prob=mask_f_prob)
-
-
-class Vanilla_AudioMAE(nn.Module):
-    """Audio Masked Autoencoder (MAE) pre-trained on AudioSet (for AudioLDM2)"""
-
-    def __init__(
-        self,
-    ):
-        super().__init__()
-        model = models_mae.__dict__["mae_vit_base_patch16"](
-            in_chans=1, audio_exp=True, img_size=(1024, 128)
-        )
-
-        # checkpoint_path = '/mnt/bn/lqhaoheliu/exps/checkpoints/audiomae/pretrained.pth'
-        # checkpoint = torch.load(checkpoint_path, map_location='cpu')
-        # msg = model.load_state_dict(checkpoint['model'], strict=False)
-
-        # Skip the missing keys of decoder modules (not required)
-        # print(f'Load AudioMAE from {checkpoint_path} / message: {msg}')
-
-        self.model = model.eval()
-
-    def forward(self, x, mask_ratio=0.0, no_mask=False, no_average=False):
-        """
-        x: mel fbank [Batch, 1, 1024 (T), 128 (F)]
-        mask_ratio: 'masking ratio (percentage of removed patches).'
-        """
-        with torch.no_grad():
-            # embed: [B, 513, 768] for mask_ratio=0.0
-            if no_mask:
-                if no_average:
-                    raise RuntimeError("This function is deprecated")
-                    embed = self.model.forward_encoder_no_random_mask_no_average(
-                        x
-                    )  # mask_ratio
-                else:
-                    embed = self.model.forward_encoder_no_mask(x)  # mask_ratio
-            else:
-                raise RuntimeError("This function is deprecated")
-                embed, _, _, _ = self.model.forward_encoder(x, mask_ratio=mask_ratio)
-        return embed
-
-
-if __name__ == "__main__":
-    model = Vanilla_AudioMAE().cuda()
-    input = torch.randn(4, 1, 1024, 128).cuda()
-    print("The first run")
-    embed = model(input, mask_ratio=0.0, no_mask=True)
-    print(embed)
-    print("The second run")
-    embed = model(input, mask_ratio=0.0)
-    print(embed)
+"""
+Reference Repo: https://github.com/facebookresearch/AudioMAE
+"""
+
+import torch
+import torch.nn as nn
+from timm.models.layers import to_2tuple
+import audiosr.latent_diffusion.modules.audiomae.models_vit as models_vit
+import audiosr.latent_diffusion.modules.audiomae.models_mae as models_mae
+
+# model = mae_vit_base_patch16(in_chans=1, audio_exp=True, img_size=(1024, 128))
+
+
+class PatchEmbed_new(nn.Module):
+    """Flexible Image to Patch Embedding"""
+
+    def __init__(
+        self, img_size=224, patch_size=16, in_chans=3, embed_dim=768, stride=10
+    ):
+        super().__init__()
+        img_size = to_2tuple(img_size)
+        patch_size = to_2tuple(patch_size)
+        stride = to_2tuple(stride)
+
+        self.img_size = img_size
+        self.patch_size = patch_size
+
+        self.proj = nn.Conv2d(
+            in_chans, embed_dim, kernel_size=patch_size, stride=stride
+        )  # with overlapped patches
+        # self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size)
+
+        # self.patch_hw = (img_size[1] // patch_size[1], img_size[0] // patch_size[0])
+        # self.num_patches = (img_size[1] // patch_size[1]) * (img_size[0] // patch_size[0])
+        _, _, h, w = self.get_output_shape(img_size)  # n, emb_dim, h, w
+        self.patch_hw = (h, w)
+        self.num_patches = h * w
+
+    def get_output_shape(self, img_size):
+        # todo: don't be lazy..
+        return self.proj(torch.randn(1, 1, img_size[0], img_size[1])).shape
+
+    def forward(self, x):
+        B, C, H, W = x.shape
+        # FIXME look at relaxing size constraints
+        # assert H == self.img_size[0] and W == self.img_size[1], \
+        #    f"Input image size ({H}*{W}) doesn't match model ({self.img_size[0]}*{self.img_size[1]})."
+        x = self.proj(x)
+        x = x.flatten(2).transpose(1, 2)
+        return x
+
+
+class AudioMAE(nn.Module):
+    """Audio Masked Autoencoder (MAE) pre-trained and finetuned on AudioSet (for SoundCLIP)"""
+
+    def __init__(
+        self,
+    ):
+        super().__init__()
+        model = models_vit.__dict__["vit_base_patch16"](
+            num_classes=527,
+            drop_path_rate=0.1,
+            global_pool=True,
+            mask_2d=True,
+            use_custom_patch=False,
+        )
+
+        img_size = (1024, 128)
+        emb_dim = 768
+
+        model.patch_embed = PatchEmbed_new(
+            img_size=img_size,
+            patch_size=(16, 16),
+            in_chans=1,
+            embed_dim=emb_dim,
+            stride=16,
+        )
+        num_patches = model.patch_embed.num_patches
+        # num_patches = 512 # assume audioset, 1024//16=64, 128//16=8, 512=64x8
+        model.pos_embed = nn.Parameter(
+            torch.zeros(1, num_patches + 1, emb_dim), requires_grad=False
+        )  # fixed sin-cos embedding
+
+        # checkpoint_path = '/mnt/bn/data-xubo/project/Masked_AudioEncoder/checkpoint/finetuned.pth'
+        # checkpoint = torch.load(checkpoint_path, map_location='cpu')
+        # msg = model.load_state_dict(checkpoint['model'], strict=False)
+        # print(f'Load AudioMAE from {checkpoint_path} / message: {msg}')
+
+        self.model = model
+
+    def forward(self, x, mask_t_prob=0.0, mask_f_prob=0.0):
+        """
+        x: mel fbank [Batch, 1, T, F]
+        mask_t_prob: 'T masking ratio (percentage of removed patches).'
+        mask_f_prob: 'F masking ratio (percentage of removed patches).'
+        """
+        return self.model(x=x, mask_t_prob=mask_t_prob, mask_f_prob=mask_f_prob)
+
+
+class Vanilla_AudioMAE(nn.Module):
+    """Audio Masked Autoencoder (MAE) pre-trained on AudioSet (for AudioLDM2)"""
+
+    def __init__(
+        self,
+    ):
+        super().__init__()
+        model = models_mae.__dict__["mae_vit_base_patch16"](
+            in_chans=1, audio_exp=True, img_size=(1024, 128)
+        )
+
+        # checkpoint_path = '/mnt/bn/lqhaoheliu/exps/checkpoints/audiomae/pretrained.pth'
+        # checkpoint = torch.load(checkpoint_path, map_location='cpu')
+        # msg = model.load_state_dict(checkpoint['model'], strict=False)
+
+        # Skip the missing keys of decoder modules (not required)
+        # print(f'Load AudioMAE from {checkpoint_path} / message: {msg}')
+
+        self.model = model.eval()
+
+    def forward(self, x, mask_ratio=0.0, no_mask=False, no_average=False):
+        """
+        x: mel fbank [Batch, 1, 1024 (T), 128 (F)]
+        mask_ratio: 'masking ratio (percentage of removed patches).'
+        """
+        with torch.no_grad():
+            # embed: [B, 513, 768] for mask_ratio=0.0
+            if no_mask:
+                if no_average:
+                    raise RuntimeError("This function is deprecated")
+                    embed = self.model.forward_encoder_no_random_mask_no_average(
+                        x
+                    )  # mask_ratio
+                else:
+                    embed = self.model.forward_encoder_no_mask(x)  # mask_ratio
+            else:
+                raise RuntimeError("This function is deprecated")
+                embed, _, _, _ = self.model.forward_encoder(x, mask_ratio=mask_ratio)
+        return embed
+
+
+if __name__ == "__main__":
+    model = Vanilla_AudioMAE().cuda()
+    input = torch.randn(4, 1, 1024, 128).cuda()
+    print("The first run")
+    embed = model(input, mask_ratio=0.0, no_mask=True)
+    print(embed)
+    print("The second run")
+    embed = model(input, mask_ratio=0.0)
+    print(embed)

audiosr/latent_diffusion/modules/audiomae/models_mae.py

@@ -1,613 +1,613 @@
-# Copyright (c) Meta Platforms, Inc. and affiliates.
-# All rights reserved.
-
-# This source code is licensed under the license found in the
-# LICENSE file in the root directory of this source tree.
-# --------------------------------------------------------
-# References:
-# timm: https://github.com/rwightman/pytorch-image-models/tree/master/timm
-# DeiT: https://github.com/facebookresearch/deit
-# --------------------------------------------------------
-
-from functools import partial
-
-import torch
-import torch.nn as nn
-
-from timm.models.vision_transformer import Block
-from audiosr.latent_diffusion.modules.audiomae.util.pos_embed import (
-    get_2d_sincos_pos_embed,
-    get_2d_sincos_pos_embed_flexible,
-)
-from audiosr.latent_diffusion.modules.audiomae.util.patch_embed import (
-    PatchEmbed_new,
-    PatchEmbed_org,
-)
-
-
-class MaskedAutoencoderViT(nn.Module):
-    """Masked Autoencoder with VisionTransformer backbone"""
-
-    def __init__(
-        self,
-        img_size=224,
-        patch_size=16,
-        stride=10,
-        in_chans=3,
-        embed_dim=1024,
-        depth=24,
-        num_heads=16,
-        decoder_embed_dim=512,
-        decoder_depth=8,
-        decoder_num_heads=16,
-        mlp_ratio=4.0,
-        norm_layer=nn.LayerNorm,
-        norm_pix_loss=False,
-        audio_exp=False,
-        alpha=0.0,
-        temperature=0.2,
-        mode=0,
-        contextual_depth=8,
-        use_custom_patch=False,
-        split_pos=False,
-        pos_trainable=False,
-        use_nce=False,
-        beta=4.0,
-        decoder_mode=0,
-        mask_t_prob=0.6,
-        mask_f_prob=0.5,
-        mask_2d=False,
-        epoch=0,
-        no_shift=False,
-    ):
-        super().__init__()
-
-        self.audio_exp = audio_exp
-        self.embed_dim = embed_dim
-        self.decoder_embed_dim = decoder_embed_dim
-        # --------------------------------------------------------------------------
-        # MAE encoder specifics
-        if use_custom_patch:
-            print(
-                f"Use custom patch_emb with patch size: {patch_size}, stride: {stride}"
-            )
-            self.patch_embed = PatchEmbed_new(
-                img_size=img_size,
-                patch_size=patch_size,
-                in_chans=in_chans,
-                embed_dim=embed_dim,
-                stride=stride,
-            )
-        else:
-            self.patch_embed = PatchEmbed_org(img_size, patch_size, in_chans, embed_dim)
-        self.use_custom_patch = use_custom_patch
-        num_patches = self.patch_embed.num_patches
-
-        self.cls_token = nn.Parameter(torch.zeros(1, 1, embed_dim))
-
-        # self.split_pos = split_pos # not useful
-        self.pos_embed = nn.Parameter(
-            torch.zeros(1, num_patches + 1, embed_dim), requires_grad=pos_trainable
-        )  # fixed sin-cos embedding
-
-        self.encoder_depth = depth
-        self.contextual_depth = contextual_depth
-        self.blocks = nn.ModuleList(
-            [
-                Block(
-                    embed_dim,
-                    num_heads,
-                    mlp_ratio,
-                    qkv_bias=True,
-                    norm_layer=norm_layer,
-                )  # qk_scale=None
-                for i in range(depth)
-            ]
-        )
-        self.norm = norm_layer(embed_dim)
-
-        # --------------------------------------------------------------------------
-        # MAE decoder specifics
-        self.decoder_embed = nn.Linear(embed_dim, decoder_embed_dim, bias=True)
-
-        self.mask_token = nn.Parameter(torch.zeros(1, 1, decoder_embed_dim))
-        self.decoder_pos_embed = nn.Parameter(
-            torch.zeros(1, num_patches + 1, decoder_embed_dim),
-            requires_grad=pos_trainable,
-        )  # fixed sin-cos embedding
-
-        self.no_shift = no_shift
-
-        self.decoder_mode = decoder_mode
-        if (
-            self.use_custom_patch
-        ):  # overlapped patches as in AST. Similar performance yet compute heavy
-            window_size = (6, 6)
-            feat_size = (102, 12)
-        else:
-            window_size = (4, 4)
-            feat_size = (64, 8)
-        if self.decoder_mode == 1:
-            decoder_modules = []
-            for index in range(16):
-                if self.no_shift:
-                    shift_size = (0, 0)
-                else:
-                    if (index % 2) == 0:
-                        shift_size = (0, 0)
-                    else:
-                        shift_size = (2, 0)
-                    # shift_size = tuple([0 if ((index % 2) == 0) else w // 2 for w in window_size])
-                decoder_modules.append(
-                    SwinTransformerBlock(
-                        dim=decoder_embed_dim,
-                        num_heads=16,
-                        feat_size=feat_size,
-                        window_size=window_size,
-                        shift_size=shift_size,
-                        mlp_ratio=mlp_ratio,
-                        drop=0.0,
-                        drop_attn=0.0,
-                        drop_path=0.0,
-                        extra_norm=False,
-                        sequential_attn=False,
-                        norm_layer=norm_layer,  # nn.LayerNorm,
-                    )
-                )
-            self.decoder_blocks = nn.ModuleList(decoder_modules)
-        else:
-            # Transfomer
-            self.decoder_blocks = nn.ModuleList(
-                [
-                    Block(
-                        decoder_embed_dim,
-                        decoder_num_heads,
-                        mlp_ratio,
-                        qkv_bias=True,
-                        norm_layer=norm_layer,
-                    )  # qk_scale=None,
-                    for i in range(decoder_depth)
-                ]
-            )
-
-        self.decoder_norm = norm_layer(decoder_embed_dim)
-        self.decoder_pred = nn.Linear(
-            decoder_embed_dim, patch_size**2 * in_chans, bias=True
-        )  # decoder to patch
-
-        # --------------------------------------------------------------------------
-
-        self.norm_pix_loss = norm_pix_loss
-
-        self.patch_size = patch_size
-        self.stride = stride
-
-        # audio exps
-        self.alpha = alpha
-        self.T = temperature
-        self.mode = mode
-        self.use_nce = use_nce
-        self.beta = beta
-
-        self.log_softmax = nn.LogSoftmax(dim=-1)
-
-        self.mask_t_prob = mask_t_prob
-        self.mask_f_prob = mask_f_prob
-        self.mask_2d = mask_2d
-
-        self.epoch = epoch
-
-        self.initialize_weights()
-
-    def initialize_weights(self):
-        # initialization
-        # initialize (and freeze) pos_embed by sin-cos embedding
-        if self.audio_exp:
-            pos_embed = get_2d_sincos_pos_embed_flexible(
-                self.pos_embed.shape[-1], self.patch_embed.patch_hw, cls_token=True
-            )
-        else:
-            pos_embed = get_2d_sincos_pos_embed(
-                self.pos_embed.shape[-1],
-                int(self.patch_embed.num_patches**0.5),
-                cls_token=True,
-            )
-        self.pos_embed.data.copy_(torch.from_numpy(pos_embed).float().unsqueeze(0))
-
-        if self.audio_exp:
-            decoder_pos_embed = get_2d_sincos_pos_embed_flexible(
-                self.decoder_pos_embed.shape[-1],
-                self.patch_embed.patch_hw,
-                cls_token=True,
-            )
-        else:
-            decoder_pos_embed = get_2d_sincos_pos_embed(
-                self.decoder_pos_embed.shape[-1],
-                int(self.patch_embed.num_patches**0.5),
-                cls_token=True,
-            )
-        self.decoder_pos_embed.data.copy_(
-            torch.from_numpy(decoder_pos_embed).float().unsqueeze(0)
-        )
-
-        # initialize patch_embed like nn.Linear (instead of nn.Conv2d)
-        w = self.patch_embed.proj.weight.data
-        torch.nn.init.xavier_uniform_(w.view([w.shape[0], -1]))
-
-        # timm's trunc_normal_(std=.02) is effectively normal_(std=0.02) as cutoff is too big (2.)
-        torch.nn.init.normal_(self.cls_token, std=0.02)
-        torch.nn.init.normal_(self.mask_token, std=0.02)
-
-        # initialize nn.Linear and nn.LayerNorm
-        self.apply(self._init_weights)
-
-    def _init_weights(self, m):
-        if isinstance(m, nn.Linear):
-            # we use xavier_uniform following official JAX ViT:
-            torch.nn.init.xavier_uniform_(m.weight)
-            if isinstance(m, nn.Linear) and m.bias is not None:
-                nn.init.constant_(m.bias, 0)
-        elif isinstance(m, nn.LayerNorm):
-            nn.init.constant_(m.bias, 0)
-            nn.init.constant_(m.weight, 1.0)
-
-    def patchify(self, imgs):
-        """
-        imgs: (N, 3, H, W)
-        x: (N, L, patch_size**2 *3)
-        L = (H/p)*(W/p)
-        """
-        p = self.patch_embed.patch_size[0]
-        # assert imgs.shape[2] == imgs.shape[3] and imgs.shape[2] % p == 0
-
-        if self.audio_exp:
-            if self.use_custom_patch:  # overlapped patch
-                h, w = self.patch_embed.patch_hw
-                # todo: fixed h/w patch size and stride size. Make hw custom in the future
-                x = imgs.unfold(2, self.patch_size, self.stride).unfold(
-                    3, self.patch_size, self.stride
-                )  # n,1,H,W -> n,1,h,w,p,p
-                x = x.reshape(shape=(imgs.shape[0], h * w, p**2 * 1))
-                # x = imgs.reshape(shape=(imgs.shape[0], 1, h, p, w, p))
-                # x = torch.einsum('nchpwq->nhwpqc', x)
-                # x = x.reshape(shape=(imgs.shape[0], h * w, p**2 * 1))
-            else:
-                h = imgs.shape[2] // p
-                w = imgs.shape[3] // p
-                # h,w = self.patch_embed.patch_hw
-                x = imgs.reshape(shape=(imgs.shape[0], 1, h, p, w, p))
-                x = torch.einsum("nchpwq->nhwpqc", x)
-                x = x.reshape(shape=(imgs.shape[0], h * w, p**2 * 1))
-        else:
-            h = w = imgs.shape[2] // p
-            x = imgs.reshape(shape=(imgs.shape[0], 3, h, p, w, p))
-            x = torch.einsum("nchpwq->nhwpqc", x)
-            x = x.reshape(shape=(imgs.shape[0], h * w, p**2 * 3))
-
-        return x
-
-    def unpatchify(self, x):
-        """
-        x: (N, L, patch_size**2 *3)
-        specs: (N, 1, H, W)
-        """
-        p = self.patch_embed.patch_size[0]
-        h = 1024 // p
-        w = 128 // p
-        x = x.reshape(shape=(x.shape[0], h, w, p, p, 1))
-        x = torch.einsum("nhwpqc->nchpwq", x)
-        specs = x.reshape(shape=(x.shape[0], 1, h * p, w * p))
-        return specs
-
-    def random_masking(self, x, mask_ratio):
-        """
-        Perform per-sample random masking by per-sample shuffling.
-        Per-sample shuffling is done by argsort random noise.
-        x: [N, L, D], sequence
-        """
-        N, L, D = x.shape  # batch, length, dim
-        len_keep = int(L * (1 - mask_ratio))
-
-        noise = torch.rand(N, L, device=x.device)  # noise in [0, 1]
-
-        # sort noise for each sample
-        ids_shuffle = torch.argsort(
-            noise, dim=1
-        )  # ascend: small is keep, large is remove
-        ids_restore = torch.argsort(ids_shuffle, dim=1)
-
-        # keep the first subset
-        ids_keep = ids_shuffle[:, :len_keep]
-        x_masked = torch.gather(x, dim=1, index=ids_keep.unsqueeze(-1).repeat(1, 1, D))
-
-        # generate the binary mask: 0 is keep, 1 is remove
-        mask = torch.ones([N, L], device=x.device)
-        mask[:, :len_keep] = 0
-        # unshuffle to get the binary mask
-        mask = torch.gather(mask, dim=1, index=ids_restore)
-
-        return x_masked, mask, ids_restore
-
-    def random_masking_2d(self, x, mask_t_prob, mask_f_prob):
-        """
-        2D: Spectrogram (msking t and f under mask_t_prob and mask_f_prob)
-        Perform per-sample random masking by per-sample shuffling.
-        Per-sample shuffling is done by argsort random noise.
-        x: [N, L, D], sequence
-        """
-        N, L, D = x.shape  # batch, length, dim
-        if self.use_custom_patch:  # overlapped patch
-            T = 101
-            F = 12
-        else:
-            T = 64
-            F = 8
-        # x = x.reshape(N, T, F, D)
-        len_keep_t = int(T * (1 - mask_t_prob))
-        len_keep_f = int(F * (1 - mask_f_prob))
-
-        # noise for mask in time
-        noise_t = torch.rand(N, T, device=x.device)  # noise in [0, 1]
-        # sort noise for each sample aling time
-        ids_shuffle_t = torch.argsort(
-            noise_t, dim=1
-        )  # ascend: small is keep, large is remove
-        ids_restore_t = torch.argsort(ids_shuffle_t, dim=1)
-        ids_keep_t = ids_shuffle_t[:, :len_keep_t]
-        # noise mask in freq
-        noise_f = torch.rand(N, F, device=x.device)  # noise in [0, 1]
-        ids_shuffle_f = torch.argsort(
-            noise_f, dim=1
-        )  # ascend: small is keep, large is remove
-        ids_restore_f = torch.argsort(ids_shuffle_f, dim=1)
-        ids_keep_f = ids_shuffle_f[:, :len_keep_f]  #
-
-        # generate the binary mask: 0 is keep, 1 is remove
-        # mask in freq
-        mask_f = torch.ones(N, F, device=x.device)
-        mask_f[:, :len_keep_f] = 0
-        mask_f = (
-            torch.gather(mask_f, dim=1, index=ids_restore_f)
-            .unsqueeze(1)
-            .repeat(1, T, 1)
-        )  # N,T,F
-        # mask in time
-        mask_t = torch.ones(N, T, device=x.device)
-        mask_t[:, :len_keep_t] = 0
-        mask_t = (
-            torch.gather(mask_t, dim=1, index=ids_restore_t)
-            .unsqueeze(1)
-            .repeat(1, F, 1)
-            .permute(0, 2, 1)
-        )  # N,T,F
-        mask = 1 - (1 - mask_t) * (1 - mask_f)  # N, T, F
-
-        # get masked x
-        id2res = torch.Tensor(list(range(N * T * F))).reshape(N, T, F).to(x.device)
-        id2res = id2res + 999 * mask  # add a large value for masked elements
-        id2res2 = torch.argsort(id2res.flatten(start_dim=1))
-        ids_keep = id2res2.flatten(start_dim=1)[:, : len_keep_f * len_keep_t]
-        x_masked = torch.gather(x, dim=1, index=ids_keep.unsqueeze(-1).repeat(1, 1, D))
-
-        ids_restore = torch.argsort(id2res2.flatten(start_dim=1))
-        mask = mask.flatten(start_dim=1)
-
-        return x_masked, mask, ids_restore
-
-    def forward_encoder(self, x, mask_ratio, mask_2d=False):
-        # embed patches
-        x = self.patch_embed(x)
-        # add pos embed w/o cls token
-        x = x + self.pos_embed[:, 1:, :]
-
-        # masking: length -> length * mask_ratio
-        if mask_2d:
-            x, mask, ids_restore = self.random_masking_2d(
-                x, mask_t_prob=self.mask_t_prob, mask_f_prob=self.mask_f_prob
-            )
-        else:
-            x, mask, ids_restore = self.random_masking(x, mask_ratio)
-
-        # append cls token
-        cls_token = self.cls_token + self.pos_embed[:, :1, :]
-        cls_tokens = cls_token.expand(x.shape[0], -1, -1)
-        x = torch.cat((cls_tokens, x), dim=1)
-
-        # apply Transformer blocks
-        for blk in self.blocks:
-            x = blk(x)
-        x = self.norm(x)
-
-        return x, mask, ids_restore, None
-
-    def forward_encoder_no_random_mask_no_average(self, x):
-        # embed patches
-        x = self.patch_embed(x)
-        # add pos embed w/o cls token
-        x = x + self.pos_embed[:, 1:, :]
-
-        # masking: length -> length * mask_ratio
-        # if mask_2d:
-        #     x, mask, ids_restore = self.random_masking_2d(x, mask_t_prob=self.mask_t_prob, mask_f_prob=self.mask_f_prob)
-        # else:
-        #     x, mask, ids_restore = self.random_masking(x, mask_ratio)
-
-        # append cls token
-        cls_token = self.cls_token + self.pos_embed[:, :1, :]
-        cls_tokens = cls_token.expand(x.shape[0], -1, -1)
-        x = torch.cat((cls_tokens, x), dim=1)
-
-        # apply Transformer blocks
-        for blk in self.blocks:
-            x = blk(x)
-        x = self.norm(x)
-
-        return x
-
-    def forward_encoder_no_mask(self, x):
-        # embed patches
-        x = self.patch_embed(x)
-
-        # add pos embed w/o cls token
-        x = x + self.pos_embed[:, 1:, :]
-
-        # masking: length -> length * mask_ratio
-        # x, mask, ids_restore = self.random_masking(x, mask_ratio)
-        # append cls token
-        cls_token = self.cls_token + self.pos_embed[:, :1, :]
-        cls_tokens = cls_token.expand(x.shape[0], -1, -1)
-        x = torch.cat((cls_tokens, x), dim=1)
-
-        # apply Transformer blocks
-        contextual_embs = []
-        for n, blk in enumerate(self.blocks):
-            x = blk(x)
-            if n > self.contextual_depth:
-                contextual_embs.append(self.norm(x))
-        # x = self.norm(x)
-        contextual_emb = torch.stack(contextual_embs, dim=0).mean(dim=0)
-
-        return contextual_emb
-
-    def forward_decoder(self, x, ids_restore):
-        # embed tokens
-        x = self.decoder_embed(x)
-
-        # append mask tokens to sequence
-        mask_tokens = self.mask_token.repeat(
-            x.shape[0], ids_restore.shape[1] + 1 - x.shape[1], 1
-        )
-        x_ = torch.cat([x[:, 1:, :], mask_tokens], dim=1)  # no cls token
-        x_ = torch.gather(
-            x_, dim=1, index=ids_restore.unsqueeze(-1).repeat(1, 1, x.shape[2])
-        )  # unshuffle
-        x = torch.cat([x[:, :1, :], x_], dim=1)  # append cls token
-
-        # add pos embed
-        x = x + self.decoder_pos_embed
-
-        if self.decoder_mode != 0:
-            B, L, D = x.shape
-            x = x[:, 1:, :]
-            if self.use_custom_patch:
-                x = x.reshape(B, 101, 12, D)
-                x = torch.cat([x, x[:, -1, :].unsqueeze(1)], dim=1)  # hack
-                x = x.reshape(B, 1224, D)
-        if self.decoder_mode > 3:  # mvit
-            x = self.decoder_blocks(x)
-        else:
-            # apply Transformer blocks
-            for blk in self.decoder_blocks:
-                x = blk(x)
-        x = self.decoder_norm(x)
-
-        # predictor projection
-        pred = self.decoder_pred(x)
-
-        # remove cls token
-        if self.decoder_mode != 0:
-            if self.use_custom_patch:
-                pred = pred.reshape(B, 102, 12, 256)
-                pred = pred[:, :101, :, :]
-                pred = pred.reshape(B, 1212, 256)
-            else:
-                pred = pred
-        else:
-            pred = pred[:, 1:, :]
-        return pred, None, None  # emb, emb_pixel
-
-    def forward_loss(self, imgs, pred, mask, norm_pix_loss=False):
-        """
-        imgs: [N, 3, H, W]
-        pred: [N, L, p*p*3]
-        mask: [N, L], 0 is keep, 1 is remove,
-        """
-        target = self.patchify(imgs)
-        if norm_pix_loss:
-            mean = target.mean(dim=-1, keepdim=True)
-            var = target.var(dim=-1, keepdim=True)
-            target = (target - mean) / (var + 1.0e-6) ** 0.5
-
-        loss = (pred - target) ** 2
-        loss = loss.mean(dim=-1)  # [N, L], mean loss per patch
-
-        loss = (loss * mask).sum() / mask.sum()  # mean loss on removed patches
-        return loss
-
-    def forward(self, imgs, mask_ratio=0.8):
-        emb_enc, mask, ids_restore, _ = self.forward_encoder(
-            imgs, mask_ratio, mask_2d=self.mask_2d
-        )
-        pred, _, _ = self.forward_decoder(emb_enc, ids_restore)  # [N, L, p*p*3]
-        loss_recon = self.forward_loss(
-            imgs, pred, mask, norm_pix_loss=self.norm_pix_loss
-        )
-        loss_contrastive = torch.FloatTensor([0.0]).cuda()
-        return loss_recon, pred, mask, loss_contrastive
-
-
-def mae_vit_small_patch16_dec512d8b(**kwargs):
-    model = MaskedAutoencoderViT(
-        patch_size=16,
-        embed_dim=384,
-        depth=12,
-        num_heads=6,
-        decoder_embed_dim=512,
-        decoder_num_heads=16,
-        mlp_ratio=4,
-        norm_layer=partial(nn.LayerNorm, eps=1e-6),
-        **kwargs,
-    )
-    return model
-
-
-def mae_vit_base_patch16_dec512d8b(**kwargs):
-    model = MaskedAutoencoderViT(
-        patch_size=16,
-        embed_dim=768,
-        depth=12,
-        num_heads=12,
-        decoder_embed_dim=512,
-        decoder_num_heads=16,
-        mlp_ratio=4,
-        norm_layer=partial(nn.LayerNorm, eps=1e-6),
-        **kwargs,
-    )
-    return model
-
-
-def mae_vit_large_patch16_dec512d8b(**kwargs):
-    model = MaskedAutoencoderViT(
-        patch_size=16,
-        embed_dim=1024,
-        depth=24,
-        num_heads=16,
-        decoder_embed_dim=512,
-        decoder_num_heads=16,
-        mlp_ratio=4,
-        norm_layer=partial(nn.LayerNorm, eps=1e-6),
-        **kwargs,
-    )
-    return model
-
-
-def mae_vit_huge_patch14_dec512d8b(**kwargs):
-    model = MaskedAutoencoderViT(
-        patch_size=14,
-        embed_dim=1280,
-        depth=32,
-        num_heads=16,
-        decoder_embed_dim=512,
-        decoder_num_heads=16,
-        mlp_ratio=4,
-        norm_layer=partial(nn.LayerNorm, eps=1e-6),
-        **kwargs,
-    )
-    return model
-
-
-# set recommended archs
-mae_vit_base_patch16 = mae_vit_base_patch16_dec512d8b  # decoder: 512 dim, 8 blocks
-mae_vit_large_patch16 = mae_vit_large_patch16_dec512d8b  # decoder: 512 dim, 8 blocks
-mae_vit_huge_patch14 = mae_vit_huge_patch14_dec512d8b  # decoder: 512 dim, 8 blocks
-mae_vit_small_patch16 = mae_vit_small_patch16_dec512d8b  # decoder: 512 dim, 8 blocks
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+# --------------------------------------------------------
+# References:
+# timm: https://github.com/rwightman/pytorch-image-models/tree/master/timm
+# DeiT: https://github.com/facebookresearch/deit
+# --------------------------------------------------------
+
+from functools import partial
+
+import torch
+import torch.nn as nn
+
+from timm.models.vision_transformer import Block
+from audiosr.latent_diffusion.modules.audiomae.util.pos_embed import (
+    get_2d_sincos_pos_embed,
+    get_2d_sincos_pos_embed_flexible,
+)
+from audiosr.latent_diffusion.modules.audiomae.util.patch_embed import (
+    PatchEmbed_new,
+    PatchEmbed_org,
+)
+
+
+class MaskedAutoencoderViT(nn.Module):
+    """Masked Autoencoder with VisionTransformer backbone"""
+
+    def __init__(
+        self,
+        img_size=224,
+        patch_size=16,
+        stride=10,
+        in_chans=3,
+        embed_dim=1024,
+        depth=24,
+        num_heads=16,
+        decoder_embed_dim=512,
+        decoder_depth=8,
+        decoder_num_heads=16,
+        mlp_ratio=4.0,
+        norm_layer=nn.LayerNorm,
+        norm_pix_loss=False,
+        audio_exp=False,
+        alpha=0.0,
+        temperature=0.2,
+        mode=0,
+        contextual_depth=8,
+        use_custom_patch=False,
+        split_pos=False,
+        pos_trainable=False,
+        use_nce=False,
+        beta=4.0,
+        decoder_mode=0,
+        mask_t_prob=0.6,
+        mask_f_prob=0.5,
+        mask_2d=False,
+        epoch=0,
+        no_shift=False,
+    ):
+        super().__init__()
+
+        self.audio_exp = audio_exp
+        self.embed_dim = embed_dim
+        self.decoder_embed_dim = decoder_embed_dim
+        # --------------------------------------------------------------------------
+        # MAE encoder specifics
+        if use_custom_patch:
+            print(
+                f"Use custom patch_emb with patch size: {patch_size}, stride: {stride}"
+            )
+            self.patch_embed = PatchEmbed_new(
+                img_size=img_size,
+                patch_size=patch_size,
+                in_chans=in_chans,
+                embed_dim=embed_dim,
+                stride=stride,
+            )
+        else:
+            self.patch_embed = PatchEmbed_org(img_size, patch_size, in_chans, embed_dim)
+        self.use_custom_patch = use_custom_patch
+        num_patches = self.patch_embed.num_patches
+
+        self.cls_token = nn.Parameter(torch.zeros(1, 1, embed_dim))
+
+        # self.split_pos = split_pos # not useful
+        self.pos_embed = nn.Parameter(
+            torch.zeros(1, num_patches + 1, embed_dim), requires_grad=pos_trainable
+        )  # fixed sin-cos embedding
+
+        self.encoder_depth = depth
+        self.contextual_depth = contextual_depth
+        self.blocks = nn.ModuleList(
+            [
+                Block(
+                    embed_dim,
+                    num_heads,
+                    mlp_ratio,
+                    qkv_bias=True,
+                    norm_layer=norm_layer,
+                )  # qk_scale=None
+                for i in range(depth)
+            ]
+        )
+        self.norm = norm_layer(embed_dim)
+
+        # --------------------------------------------------------------------------
+        # MAE decoder specifics
+        self.decoder_embed = nn.Linear(embed_dim, decoder_embed_dim, bias=True)
+
+        self.mask_token = nn.Parameter(torch.zeros(1, 1, decoder_embed_dim))
+        self.decoder_pos_embed = nn.Parameter(
+            torch.zeros(1, num_patches + 1, decoder_embed_dim),
+            requires_grad=pos_trainable,
+        )  # fixed sin-cos embedding
+
+        self.no_shift = no_shift
+
+        self.decoder_mode = decoder_mode
+        if (
+            self.use_custom_patch
+        ):  # overlapped patches as in AST. Similar performance yet compute heavy
+            window_size = (6, 6)
+            feat_size = (102, 12)
+        else:
+            window_size = (4, 4)
+            feat_size = (64, 8)
+        if self.decoder_mode == 1:
+            decoder_modules = []
+            for index in range(16):
+                if self.no_shift:
+                    shift_size = (0, 0)
+                else:
+                    if (index % 2) == 0:
+                        shift_size = (0, 0)
+                    else:
+                        shift_size = (2, 0)
+                    # shift_size = tuple([0 if ((index % 2) == 0) else w // 2 for w in window_size])
+                decoder_modules.append(
+                    SwinTransformerBlock(
+                        dim=decoder_embed_dim,
+                        num_heads=16,
+                        feat_size=feat_size,
+                        window_size=window_size,
+                        shift_size=shift_size,
+                        mlp_ratio=mlp_ratio,
+                        drop=0.0,
+                        drop_attn=0.0,
+                        drop_path=0.0,
+                        extra_norm=False,
+                        sequential_attn=False,
+                        norm_layer=norm_layer,  # nn.LayerNorm,
+                    )
+                )
+            self.decoder_blocks = nn.ModuleList(decoder_modules)
+        else:
+            # Transfomer
+            self.decoder_blocks = nn.ModuleList(
+                [
+                    Block(
+                        decoder_embed_dim,
+                        decoder_num_heads,
+                        mlp_ratio,
+                        qkv_bias=True,
+                        norm_layer=norm_layer,
+                    )  # qk_scale=None,
+                    for i in range(decoder_depth)
+                ]
+            )
+
+        self.decoder_norm = norm_layer(decoder_embed_dim)
+        self.decoder_pred = nn.Linear(
+            decoder_embed_dim, patch_size**2 * in_chans, bias=True
+        )  # decoder to patch
+
+        # --------------------------------------------------------------------------
+
+        self.norm_pix_loss = norm_pix_loss
+
+        self.patch_size = patch_size
+        self.stride = stride
+
+        # audio exps
+        self.alpha = alpha
+        self.T = temperature
+        self.mode = mode
+        self.use_nce = use_nce
+        self.beta = beta
+
+        self.log_softmax = nn.LogSoftmax(dim=-1)
+
+        self.mask_t_prob = mask_t_prob
+        self.mask_f_prob = mask_f_prob
+        self.mask_2d = mask_2d
+
+        self.epoch = epoch
+
+        self.initialize_weights()
+
+    def initialize_weights(self):
+        # initialization
+        # initialize (and freeze) pos_embed by sin-cos embedding
+        if self.audio_exp:
+            pos_embed = get_2d_sincos_pos_embed_flexible(
+                self.pos_embed.shape[-1], self.patch_embed.patch_hw, cls_token=True
+            )
+        else:
+            pos_embed = get_2d_sincos_pos_embed(
+                self.pos_embed.shape[-1],
+                int(self.patch_embed.num_patches**0.5),
+                cls_token=True,
+            )
+        self.pos_embed.data.copy_(torch.from_numpy(pos_embed).float().unsqueeze(0))
+
+        if self.audio_exp:
+            decoder_pos_embed = get_2d_sincos_pos_embed_flexible(
+                self.decoder_pos_embed.shape[-1],
+                self.patch_embed.patch_hw,
+                cls_token=True,
+            )
+        else:
+            decoder_pos_embed = get_2d_sincos_pos_embed(
+                self.decoder_pos_embed.shape[-1],
+                int(self.patch_embed.num_patches**0.5),
+                cls_token=True,
+            )
+        self.decoder_pos_embed.data.copy_(
+            torch.from_numpy(decoder_pos_embed).float().unsqueeze(0)
+        )
+
+        # initialize patch_embed like nn.Linear (instead of nn.Conv2d)
+        w = self.patch_embed.proj.weight.data
+        torch.nn.init.xavier_uniform_(w.view([w.shape[0], -1]))
+
+        # timm's trunc_normal_(std=.02) is effectively normal_(std=0.02) as cutoff is too big (2.)
+        torch.nn.init.normal_(self.cls_token, std=0.02)
+        torch.nn.init.normal_(self.mask_token, std=0.02)
+
+        # initialize nn.Linear and nn.LayerNorm
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            # we use xavier_uniform following official JAX ViT:
+            torch.nn.init.xavier_uniform_(m.weight)
+            if isinstance(m, nn.Linear) and m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+        elif isinstance(m, nn.LayerNorm):
+            nn.init.constant_(m.bias, 0)
+            nn.init.constant_(m.weight, 1.0)
+
+    def patchify(self, imgs):
+        """
+        imgs: (N, 3, H, W)
+        x: (N, L, patch_size**2 *3)
+        L = (H/p)*(W/p)
+        """
+        p = self.patch_embed.patch_size[0]
+        # assert imgs.shape[2] == imgs.shape[3] and imgs.shape[2] % p == 0
+
+        if self.audio_exp:
+            if self.use_custom_patch:  # overlapped patch
+                h, w = self.patch_embed.patch_hw
+                # todo: fixed h/w patch size and stride size. Make hw custom in the future
+                x = imgs.unfold(2, self.patch_size, self.stride).unfold(
+                    3, self.patch_size, self.stride
+                )  # n,1,H,W -> n,1,h,w,p,p
+                x = x.reshape(shape=(imgs.shape[0], h * w, p**2 * 1))
+                # x = imgs.reshape(shape=(imgs.shape[0], 1, h, p, w, p))
+                # x = torch.einsum('nchpwq->nhwpqc', x)
+                # x = x.reshape(shape=(imgs.shape[0], h * w, p**2 * 1))
+            else:
+                h = imgs.shape[2] // p
+                w = imgs.shape[3] // p
+                # h,w = self.patch_embed.patch_hw
+                x = imgs.reshape(shape=(imgs.shape[0], 1, h, p, w, p))
+                x = torch.einsum("nchpwq->nhwpqc", x)
+                x = x.reshape(shape=(imgs.shape[0], h * w, p**2 * 1))
+        else:
+            h = w = imgs.shape[2] // p
+            x = imgs.reshape(shape=(imgs.shape[0], 3, h, p, w, p))
+            x = torch.einsum("nchpwq->nhwpqc", x)
+            x = x.reshape(shape=(imgs.shape[0], h * w, p**2 * 3))
+
+        return x
+
+    def unpatchify(self, x):
+        """
+        x: (N, L, patch_size**2 *3)
+        specs: (N, 1, H, W)
+        """
+        p = self.patch_embed.patch_size[0]
+        h = 1024 // p
+        w = 128 // p
+        x = x.reshape(shape=(x.shape[0], h, w, p, p, 1))
+        x = torch.einsum("nhwpqc->nchpwq", x)
+        specs = x.reshape(shape=(x.shape[0], 1, h * p, w * p))
+        return specs
+
+    def random_masking(self, x, mask_ratio):
+        """
+        Perform per-sample random masking by per-sample shuffling.
+        Per-sample shuffling is done by argsort random noise.
+        x: [N, L, D], sequence
+        """
+        N, L, D = x.shape  # batch, length, dim
+        len_keep = int(L * (1 - mask_ratio))
+
+        noise = torch.rand(N, L, device=x.device)  # noise in [0, 1]
+
+        # sort noise for each sample
+        ids_shuffle = torch.argsort(
+            noise, dim=1
+        )  # ascend: small is keep, large is remove
+        ids_restore = torch.argsort(ids_shuffle, dim=1)
+
+        # keep the first subset
+        ids_keep = ids_shuffle[:, :len_keep]
+        x_masked = torch.gather(x, dim=1, index=ids_keep.unsqueeze(-1).repeat(1, 1, D))
+
+        # generate the binary mask: 0 is keep, 1 is remove
+        mask = torch.ones([N, L], device=x.device)
+        mask[:, :len_keep] = 0
+        # unshuffle to get the binary mask
+        mask = torch.gather(mask, dim=1, index=ids_restore)
+
+        return x_masked, mask, ids_restore
+
+    def random_masking_2d(self, x, mask_t_prob, mask_f_prob):
+        """
+        2D: Spectrogram (msking t and f under mask_t_prob and mask_f_prob)
+        Perform per-sample random masking by per-sample shuffling.
+        Per-sample shuffling is done by argsort random noise.
+        x: [N, L, D], sequence
+        """
+        N, L, D = x.shape  # batch, length, dim
+        if self.use_custom_patch:  # overlapped patch
+            T = 101
+            F = 12
+        else:
+            T = 64
+            F = 8
+        # x = x.reshape(N, T, F, D)
+        len_keep_t = int(T * (1 - mask_t_prob))
+        len_keep_f = int(F * (1 - mask_f_prob))
+
+        # noise for mask in time
+        noise_t = torch.rand(N, T, device=x.device)  # noise in [0, 1]
+        # sort noise for each sample aling time
+        ids_shuffle_t = torch.argsort(
+            noise_t, dim=1
+        )  # ascend: small is keep, large is remove
+        ids_restore_t = torch.argsort(ids_shuffle_t, dim=1)
+        ids_keep_t = ids_shuffle_t[:, :len_keep_t]
+        # noise mask in freq
+        noise_f = torch.rand(N, F, device=x.device)  # noise in [0, 1]
+        ids_shuffle_f = torch.argsort(
+            noise_f, dim=1
+        )  # ascend: small is keep, large is remove
+        ids_restore_f = torch.argsort(ids_shuffle_f, dim=1)
+        ids_keep_f = ids_shuffle_f[:, :len_keep_f]  #
+
+        # generate the binary mask: 0 is keep, 1 is remove
+        # mask in freq
+        mask_f = torch.ones(N, F, device=x.device)
+        mask_f[:, :len_keep_f] = 0
+        mask_f = (
+            torch.gather(mask_f, dim=1, index=ids_restore_f)
+            .unsqueeze(1)
+            .repeat(1, T, 1)
+        )  # N,T,F
+        # mask in time
+        mask_t = torch.ones(N, T, device=x.device)
+        mask_t[:, :len_keep_t] = 0
+        mask_t = (
+            torch.gather(mask_t, dim=1, index=ids_restore_t)
+            .unsqueeze(1)
+            .repeat(1, F, 1)
+            .permute(0, 2, 1)
+        )  # N,T,F
+        mask = 1 - (1 - mask_t) * (1 - mask_f)  # N, T, F
+
+        # get masked x
+        id2res = torch.Tensor(list(range(N * T * F))).reshape(N, T, F).to(x.device)
+        id2res = id2res + 999 * mask  # add a large value for masked elements
+        id2res2 = torch.argsort(id2res.flatten(start_dim=1))
+        ids_keep = id2res2.flatten(start_dim=1)[:, : len_keep_f * len_keep_t]
+        x_masked = torch.gather(x, dim=1, index=ids_keep.unsqueeze(-1).repeat(1, 1, D))
+
+        ids_restore = torch.argsort(id2res2.flatten(start_dim=1))
+        mask = mask.flatten(start_dim=1)
+
+        return x_masked, mask, ids_restore
+
+    def forward_encoder(self, x, mask_ratio, mask_2d=False):
+        # embed patches
+        x = self.patch_embed(x)
+        # add pos embed w/o cls token
+        x = x + self.pos_embed[:, 1:, :]
+
+        # masking: length -> length * mask_ratio
+        if mask_2d:
+            x, mask, ids_restore = self.random_masking_2d(
+                x, mask_t_prob=self.mask_t_prob, mask_f_prob=self.mask_f_prob
+            )
+        else:
+            x, mask, ids_restore = self.random_masking(x, mask_ratio)
+
+        # append cls token
+        cls_token = self.cls_token + self.pos_embed[:, :1, :]
+        cls_tokens = cls_token.expand(x.shape[0], -1, -1)
+        x = torch.cat((cls_tokens, x), dim=1)
+
+        # apply Transformer blocks
+        for blk in self.blocks:
+            x = blk(x)
+        x = self.norm(x)
+
+        return x, mask, ids_restore, None
+
+    def forward_encoder_no_random_mask_no_average(self, x):
+        # embed patches
+        x = self.patch_embed(x)
+        # add pos embed w/o cls token
+        x = x + self.pos_embed[:, 1:, :]
+
+        # masking: length -> length * mask_ratio
+        # if mask_2d:
+        #     x, mask, ids_restore = self.random_masking_2d(x, mask_t_prob=self.mask_t_prob, mask_f_prob=self.mask_f_prob)
+        # else:
+        #     x, mask, ids_restore = self.random_masking(x, mask_ratio)
+
+        # append cls token
+        cls_token = self.cls_token + self.pos_embed[:, :1, :]
+        cls_tokens = cls_token.expand(x.shape[0], -1, -1)
+        x = torch.cat((cls_tokens, x), dim=1)
+
+        # apply Transformer blocks
+        for blk in self.blocks:
+            x = blk(x)
+        x = self.norm(x)
+
+        return x
+
+    def forward_encoder_no_mask(self, x):
+        # embed patches
+        x = self.patch_embed(x)
+
+        # add pos embed w/o cls token
+        x = x + self.pos_embed[:, 1:, :]
+
+        # masking: length -> length * mask_ratio
+        # x, mask, ids_restore = self.random_masking(x, mask_ratio)
+        # append cls token
+        cls_token = self.cls_token + self.pos_embed[:, :1, :]
+        cls_tokens = cls_token.expand(x.shape[0], -1, -1)
+        x = torch.cat((cls_tokens, x), dim=1)
+
+        # apply Transformer blocks
+        contextual_embs = []
+        for n, blk in enumerate(self.blocks):
+            x = blk(x)
+            if n > self.contextual_depth:
+                contextual_embs.append(self.norm(x))
+        # x = self.norm(x)
+        contextual_emb = torch.stack(contextual_embs, dim=0).mean(dim=0)
+
+        return contextual_emb
+
+    def forward_decoder(self, x, ids_restore):
+        # embed tokens
+        x = self.decoder_embed(x)
+
+        # append mask tokens to sequence
+        mask_tokens = self.mask_token.repeat(
+            x.shape[0], ids_restore.shape[1] + 1 - x.shape[1], 1
+        )
+        x_ = torch.cat([x[:, 1:, :], mask_tokens], dim=1)  # no cls token
+        x_ = torch.gather(
+            x_, dim=1, index=ids_restore.unsqueeze(-1).repeat(1, 1, x.shape[2])
+        )  # unshuffle
+        x = torch.cat([x[:, :1, :], x_], dim=1)  # append cls token
+
+        # add pos embed
+        x = x + self.decoder_pos_embed
+
+        if self.decoder_mode != 0:
+            B, L, D = x.shape
+            x = x[:, 1:, :]
+            if self.use_custom_patch:
+                x = x.reshape(B, 101, 12, D)
+                x = torch.cat([x, x[:, -1, :].unsqueeze(1)], dim=1)  # hack
+                x = x.reshape(B, 1224, D)
+        if self.decoder_mode > 3:  # mvit
+            x = self.decoder_blocks(x)
+        else:
+            # apply Transformer blocks
+            for blk in self.decoder_blocks:
+                x = blk(x)
+        x = self.decoder_norm(x)
+
+        # predictor projection
+        pred = self.decoder_pred(x)
+
+        # remove cls token
+        if self.decoder_mode != 0:
+            if self.use_custom_patch:
+                pred = pred.reshape(B, 102, 12, 256)
+                pred = pred[:, :101, :, :]
+                pred = pred.reshape(B, 1212, 256)
+            else:
+                pred = pred
+        else:
+            pred = pred[:, 1:, :]
+        return pred, None, None  # emb, emb_pixel
+
+    def forward_loss(self, imgs, pred, mask, norm_pix_loss=False):
+        """
+        imgs: [N, 3, H, W]
+        pred: [N, L, p*p*3]
+        mask: [N, L], 0 is keep, 1 is remove,
+        """
+        target = self.patchify(imgs)
+        if norm_pix_loss:
+            mean = target.mean(dim=-1, keepdim=True)
+            var = target.var(dim=-1, keepdim=True)
+            target = (target - mean) / (var + 1.0e-6) ** 0.5
+
+        loss = (pred - target) ** 2
+        loss = loss.mean(dim=-1)  # [N, L], mean loss per patch
+
+        loss = (loss * mask).sum() / mask.sum()  # mean loss on removed patches
+        return loss
+
+    def forward(self, imgs, mask_ratio=0.8):
+        emb_enc, mask, ids_restore, _ = self.forward_encoder(
+            imgs, mask_ratio, mask_2d=self.mask_2d
+        )
+        pred, _, _ = self.forward_decoder(emb_enc, ids_restore)  # [N, L, p*p*3]
+        loss_recon = self.forward_loss(
+            imgs, pred, mask, norm_pix_loss=self.norm_pix_loss
+        )
+        loss_contrastive = torch.FloatTensor([0.0]).cuda()
+        return loss_recon, pred, mask, loss_contrastive
+
+
+def mae_vit_small_patch16_dec512d8b(**kwargs):
+    model = MaskedAutoencoderViT(
+        patch_size=16,
+        embed_dim=384,
+        depth=12,
+        num_heads=6,
+        decoder_embed_dim=512,
+        decoder_num_heads=16,
+        mlp_ratio=4,
+        norm_layer=partial(nn.LayerNorm, eps=1e-6),
+        **kwargs,
+    )
+    return model
+
+
+def mae_vit_base_patch16_dec512d8b(**kwargs):
+    model = MaskedAutoencoderViT(
+        patch_size=16,
+        embed_dim=768,
+        depth=12,
+        num_heads=12,
+        decoder_embed_dim=512,
+        decoder_num_heads=16,
+        mlp_ratio=4,
+        norm_layer=partial(nn.LayerNorm, eps=1e-6),
+        **kwargs,
+    )
+    return model
+
+
+def mae_vit_large_patch16_dec512d8b(**kwargs):
+    model = MaskedAutoencoderViT(
+        patch_size=16,
+        embed_dim=1024,
+        depth=24,
+        num_heads=16,
+        decoder_embed_dim=512,
+        decoder_num_heads=16,
+        mlp_ratio=4,
+        norm_layer=partial(nn.LayerNorm, eps=1e-6),
+        **kwargs,
+    )
+    return model
+
+
+def mae_vit_huge_patch14_dec512d8b(**kwargs):
+    model = MaskedAutoencoderViT(
+        patch_size=14,
+        embed_dim=1280,
+        depth=32,
+        num_heads=16,
+        decoder_embed_dim=512,
+        decoder_num_heads=16,
+        mlp_ratio=4,
+        norm_layer=partial(nn.LayerNorm, eps=1e-6),
+        **kwargs,
+    )
+    return model
+
+
+# set recommended archs
+mae_vit_base_patch16 = mae_vit_base_patch16_dec512d8b  # decoder: 512 dim, 8 blocks
+mae_vit_large_patch16 = mae_vit_large_patch16_dec512d8b  # decoder: 512 dim, 8 blocks
+mae_vit_huge_patch14 = mae_vit_huge_patch14_dec512d8b  # decoder: 512 dim, 8 blocks
+mae_vit_small_patch16 = mae_vit_small_patch16_dec512d8b  # decoder: 512 dim, 8 blocks

audiosr/latent_diffusion/modules/audiomae/models_vit.py

@@ -1,243 +1,243 @@
-# Copyright (c) Meta Platforms, Inc. and affiliates.
-# All rights reserved.
-
-# This source code is licensed under the license found in the
-# LICENSE file in the root directory of this source tree.
-# --------------------------------------------------------
-# References:
-# timm: https://github.com/rwightman/pytorch-image-models/tree/master/timm
-# DeiT: https://github.com/facebookresearch/deit
-# --------------------------------------------------------
-
-from functools import partial
-
-import torch
-import torch.nn as nn
-import timm.models.vision_transformer
-
-
-class VisionTransformer(timm.models.vision_transformer.VisionTransformer):
-    """Vision Transformer with support for global average pooling"""
-
-    def __init__(
-        self, global_pool=False, mask_2d=True, use_custom_patch=False, **kwargs
-    ):
-        super(VisionTransformer, self).__init__(**kwargs)
-
-        self.global_pool = global_pool
-        if self.global_pool:
-            norm_layer = kwargs["norm_layer"]
-            embed_dim = kwargs["embed_dim"]
-            self.fc_norm = norm_layer(embed_dim)
-        del self.norm  # remove the original norm
-        self.mask_2d = mask_2d
-        self.use_custom_patch = use_custom_patch
-
-    def forward_features(self, x):
-        B = x.shape[0]
-        x = self.patch_embed(x)
-        x = x + self.pos_embed[:, 1:, :]
-        cls_token = self.cls_token + self.pos_embed[:, :1, :]
-        cls_tokens = cls_token.expand(
-            B, -1, -1
-        )  # stole cls_tokens impl from Phil Wang, thanks
-        x = torch.cat((cls_tokens, x), dim=1)
-        x = self.pos_drop(x)
-
-        for blk in self.blocks:
-            x = blk(x)
-
-        if self.global_pool:
-            x = x[:, 1:, :].mean(dim=1)  # global pool without cls token
-            outcome = self.fc_norm(x)
-        else:
-            x = self.norm(x)
-            outcome = x[:, 0]
-
-        return outcome
-
-    def random_masking(self, x, mask_ratio):
-        """
-        Perform per-sample random masking by per-sample shuffling.
-        Per-sample shuffling is done by argsort random noise.
-        x: [N, L, D], sequence
-        """
-        N, L, D = x.shape  # batch, length, dim
-        len_keep = int(L * (1 - mask_ratio))
-
-        noise = torch.rand(N, L, device=x.device)  # noise in [0, 1]
-
-        # sort noise for each sample
-        ids_shuffle = torch.argsort(
-            noise, dim=1
-        )  # ascend: small is keep, large is remove
-        ids_restore = torch.argsort(ids_shuffle, dim=1)
-
-        # keep the first subset
-        ids_keep = ids_shuffle[:, :len_keep]
-        x_masked = torch.gather(x, dim=1, index=ids_keep.unsqueeze(-1).repeat(1, 1, D))
-
-        # generate the binary mask: 0 is keep, 1 is remove
-        mask = torch.ones([N, L], device=x.device)
-        mask[:, :len_keep] = 0
-        # unshuffle to get the binary mask
-        mask = torch.gather(mask, dim=1, index=ids_restore)
-
-        return x_masked, mask, ids_restore
-
-    def random_masking_2d(self, x, mask_t_prob, mask_f_prob):
-        """
-        2D: Spectrogram (msking t and f under mask_t_prob and mask_f_prob)
-        Perform per-sample random masking by per-sample shuffling.
-        Per-sample shuffling is done by argsort random noise.
-        x: [N, L, D], sequence
-        """
-
-        N, L, D = x.shape  # batch, length, dim
-        if self.use_custom_patch:
-            # # for AS
-            T = 101  # 64,101
-            F = 12  # 8,12
-            # # for ESC
-            # T=50
-            # F=12
-            # for SPC
-            # T=12
-            # F=12
-        else:
-            # ## for AS
-            T = 64
-            F = 8
-            # ## for ESC
-            # T=32
-            # F=8
-            ## for SPC
-            # T=8
-            # F=8
-
-        # mask T
-        x = x.reshape(N, T, F, D)
-        len_keep_T = int(T * (1 - mask_t_prob))
-        noise = torch.rand(N, T, device=x.device)  # noise in [0, 1]
-        # sort noise for each sample
-        ids_shuffle = torch.argsort(
-            noise, dim=1
-        )  # ascend: small is keep, large is remove
-        ids_keep = ids_shuffle[:, :len_keep_T]
-        index = ids_keep.unsqueeze(-1).unsqueeze(-1).repeat(1, 1, F, D)
-        # x_masked = torch.gather(x, dim=1, index=index)
-        # x_masked = x_masked.reshape(N,len_keep_T*F,D)
-        x = torch.gather(x, dim=1, index=index)  # N, len_keep_T(T'), F, D
-
-        # mask F
-        # x = x.reshape(N, T, F, D)
-        x = x.permute(0, 2, 1, 3)  # N T' F D => N F T' D
-        len_keep_F = int(F * (1 - mask_f_prob))
-        noise = torch.rand(N, F, device=x.device)  # noise in [0, 1]
-        # sort noise for each sample
-        ids_shuffle = torch.argsort(
-            noise, dim=1
-        )  # ascend: small is keep, large is remove
-        ids_keep = ids_shuffle[:, :len_keep_F]
-        # index = ids_keep.unsqueeze(-1).unsqueeze(-1).repeat(1, 1, T, D)
-        index = ids_keep.unsqueeze(-1).unsqueeze(-1).repeat(1, 1, len_keep_T, D)
-        x_masked = torch.gather(x, dim=1, index=index)
-        x_masked = x_masked.permute(0, 2, 1, 3)  # N F' T' D => N T' F' D
-        # x_masked = x_masked.reshape(N,len_keep*T,D)
-        x_masked = x_masked.reshape(N, len_keep_F * len_keep_T, D)
-
-        return x_masked, None, None
-
-    def forward_features_mask(self, x, mask_t_prob, mask_f_prob):
-        B = x.shape[0]  # 4,1,1024,128
-        x = self.patch_embed(x)  # 4, 512, 768
-
-        x = x + self.pos_embed[:, 1:, :]
-        if self.random_masking_2d:
-            x, mask, ids_restore = self.random_masking_2d(x, mask_t_prob, mask_f_prob)
-        else:
-            x, mask, ids_restore = self.random_masking(x, mask_t_prob)
-        cls_token = self.cls_token + self.pos_embed[:, :1, :]
-        cls_tokens = cls_token.expand(B, -1, -1)
-        x = torch.cat((cls_tokens, x), dim=1)
-        x = self.pos_drop(x)
-
-        # apply Transformer blocks
-        for blk in self.blocks:
-            x = blk(x)
-
-        if self.global_pool:
-            x = x[:, 1:, :].mean(dim=1)  # global pool without cls token
-            outcome = self.fc_norm(x)
-        else:
-            x = self.norm(x)
-            outcome = x[:, 0]
-
-        return outcome
-
-    # overwrite original timm
-    def forward(self, x, v=None, mask_t_prob=0.0, mask_f_prob=0.0):
-        if mask_t_prob > 0.0 or mask_f_prob > 0.0:
-            x = self.forward_features_mask(
-                x, mask_t_prob=mask_t_prob, mask_f_prob=mask_f_prob
-            )
-        else:
-            x = self.forward_features(x)
-        x = self.head(x)
-        return x
-
-
-def vit_small_patch16(**kwargs):
-    model = VisionTransformer(
-        patch_size=16,
-        embed_dim=384,
-        depth=12,
-        num_heads=6,
-        mlp_ratio=4,
-        qkv_bias=True,
-        norm_layer=partial(nn.LayerNorm, eps=1e-6),
-        **kwargs
-    )
-    return model
-
-
-def vit_base_patch16(**kwargs):
-    model = VisionTransformer(
-        patch_size=16,
-        embed_dim=768,
-        depth=12,
-        num_heads=12,
-        mlp_ratio=4,
-        qkv_bias=True,
-        norm_layer=partial(nn.LayerNorm, eps=1e-6),
-        **kwargs
-    )
-    return model
-
-
-def vit_large_patch16(**kwargs):
-    model = VisionTransformer(
-        patch_size=16,
-        embed_dim=1024,
-        depth=24,
-        num_heads=16,
-        mlp_ratio=4,
-        qkv_bias=True,
-        norm_layer=partial(nn.LayerNorm, eps=1e-6),
-        **kwargs
-    )
-    return model
-
-
-def vit_huge_patch14(**kwargs):
-    model = VisionTransformer(
-        patch_size=14,
-        embed_dim=1280,
-        depth=32,
-        num_heads=16,
-        mlp_ratio=4,
-        qkv_bias=True,
-        norm_layer=partial(nn.LayerNorm, eps=1e-6),
-        **kwargs
-    )
-    return model
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+# --------------------------------------------------------
+# References:
+# timm: https://github.com/rwightman/pytorch-image-models/tree/master/timm
+# DeiT: https://github.com/facebookresearch/deit
+# --------------------------------------------------------
+
+from functools import partial
+
+import torch
+import torch.nn as nn
+import timm.models.vision_transformer
+
+
+class VisionTransformer(timm.models.vision_transformer.VisionTransformer):
+    """Vision Transformer with support for global average pooling"""
+
+    def __init__(
+        self, global_pool=False, mask_2d=True, use_custom_patch=False, **kwargs
+    ):
+        super(VisionTransformer, self).__init__(**kwargs)
+
+        self.global_pool = global_pool
+        if self.global_pool:
+            norm_layer = kwargs["norm_layer"]
+            embed_dim = kwargs["embed_dim"]
+            self.fc_norm = norm_layer(embed_dim)
+        del self.norm  # remove the original norm
+        self.mask_2d = mask_2d
+        self.use_custom_patch = use_custom_patch
+
+    def forward_features(self, x):
+        B = x.shape[0]
+        x = self.patch_embed(x)
+        x = x + self.pos_embed[:, 1:, :]
+        cls_token = self.cls_token + self.pos_embed[:, :1, :]
+        cls_tokens = cls_token.expand(
+            B, -1, -1
+        )  # stole cls_tokens impl from Phil Wang, thanks
+        x = torch.cat((cls_tokens, x), dim=1)
+        x = self.pos_drop(x)
+
+        for blk in self.blocks:
+            x = blk(x)
+
+        if self.global_pool:
+            x = x[:, 1:, :].mean(dim=1)  # global pool without cls token
+            outcome = self.fc_norm(x)
+        else:
+            x = self.norm(x)
+            outcome = x[:, 0]
+
+        return outcome
+
+    def random_masking(self, x, mask_ratio):
+        """
+        Perform per-sample random masking by per-sample shuffling.
+        Per-sample shuffling is done by argsort random noise.
+        x: [N, L, D], sequence
+        """
+        N, L, D = x.shape  # batch, length, dim
+        len_keep = int(L * (1 - mask_ratio))
+
+        noise = torch.rand(N, L, device=x.device)  # noise in [0, 1]
+
+        # sort noise for each sample
+        ids_shuffle = torch.argsort(
+            noise, dim=1
+        )  # ascend: small is keep, large is remove
+        ids_restore = torch.argsort(ids_shuffle, dim=1)
+
+        # keep the first subset
+        ids_keep = ids_shuffle[:, :len_keep]
+        x_masked = torch.gather(x, dim=1, index=ids_keep.unsqueeze(-1).repeat(1, 1, D))
+
+        # generate the binary mask: 0 is keep, 1 is remove
+        mask = torch.ones([N, L], device=x.device)
+        mask[:, :len_keep] = 0
+        # unshuffle to get the binary mask
+        mask = torch.gather(mask, dim=1, index=ids_restore)
+
+        return x_masked, mask, ids_restore
+
+    def random_masking_2d(self, x, mask_t_prob, mask_f_prob):
+        """
+        2D: Spectrogram (msking t and f under mask_t_prob and mask_f_prob)
+        Perform per-sample random masking by per-sample shuffling.
+        Per-sample shuffling is done by argsort random noise.
+        x: [N, L, D], sequence
+        """
+
+        N, L, D = x.shape  # batch, length, dim
+        if self.use_custom_patch:
+            # # for AS
+            T = 101  # 64,101
+            F = 12  # 8,12
+            # # for ESC
+            # T=50
+            # F=12
+            # for SPC
+            # T=12
+            # F=12
+        else:
+            # ## for AS
+            T = 64
+            F = 8
+            # ## for ESC
+            # T=32
+            # F=8
+            ## for SPC
+            # T=8
+            # F=8
+
+        # mask T
+        x = x.reshape(N, T, F, D)
+        len_keep_T = int(T * (1 - mask_t_prob))
+        noise = torch.rand(N, T, device=x.device)  # noise in [0, 1]
+        # sort noise for each sample
+        ids_shuffle = torch.argsort(
+            noise, dim=1
+        )  # ascend: small is keep, large is remove
+        ids_keep = ids_shuffle[:, :len_keep_T]
+        index = ids_keep.unsqueeze(-1).unsqueeze(-1).repeat(1, 1, F, D)
+        # x_masked = torch.gather(x, dim=1, index=index)
+        # x_masked = x_masked.reshape(N,len_keep_T*F,D)
+        x = torch.gather(x, dim=1, index=index)  # N, len_keep_T(T'), F, D
+
+        # mask F
+        # x = x.reshape(N, T, F, D)
+        x = x.permute(0, 2, 1, 3)  # N T' F D => N F T' D
+        len_keep_F = int(F * (1 - mask_f_prob))
+        noise = torch.rand(N, F, device=x.device)  # noise in [0, 1]
+        # sort noise for each sample
+        ids_shuffle = torch.argsort(
+            noise, dim=1
+        )  # ascend: small is keep, large is remove
+        ids_keep = ids_shuffle[:, :len_keep_F]
+        # index = ids_keep.unsqueeze(-1).unsqueeze(-1).repeat(1, 1, T, D)
+        index = ids_keep.unsqueeze(-1).unsqueeze(-1).repeat(1, 1, len_keep_T, D)
+        x_masked = torch.gather(x, dim=1, index=index)
+        x_masked = x_masked.permute(0, 2, 1, 3)  # N F' T' D => N T' F' D
+        # x_masked = x_masked.reshape(N,len_keep*T,D)
+        x_masked = x_masked.reshape(N, len_keep_F * len_keep_T, D)
+
+        return x_masked, None, None
+
+    def forward_features_mask(self, x, mask_t_prob, mask_f_prob):
+        B = x.shape[0]  # 4,1,1024,128
+        x = self.patch_embed(x)  # 4, 512, 768
+
+        x = x + self.pos_embed[:, 1:, :]
+        if self.random_masking_2d:
+            x, mask, ids_restore = self.random_masking_2d(x, mask_t_prob, mask_f_prob)
+        else:
+            x, mask, ids_restore = self.random_masking(x, mask_t_prob)
+        cls_token = self.cls_token + self.pos_embed[:, :1, :]
+        cls_tokens = cls_token.expand(B, -1, -1)
+        x = torch.cat((cls_tokens, x), dim=1)
+        x = self.pos_drop(x)
+
+        # apply Transformer blocks
+        for blk in self.blocks:
+            x = blk(x)
+
+        if self.global_pool:
+            x = x[:, 1:, :].mean(dim=1)  # global pool without cls token
+            outcome = self.fc_norm(x)
+        else:
+            x = self.norm(x)
+            outcome = x[:, 0]
+
+        return outcome
+
+    # overwrite original timm
+    def forward(self, x, v=None, mask_t_prob=0.0, mask_f_prob=0.0):
+        if mask_t_prob > 0.0 or mask_f_prob > 0.0:
+            x = self.forward_features_mask(
+                x, mask_t_prob=mask_t_prob, mask_f_prob=mask_f_prob
+            )
+        else:
+            x = self.forward_features(x)
+        x = self.head(x)
+        return x
+
+
+def vit_small_patch16(**kwargs):
+    model = VisionTransformer(
+        patch_size=16,
+        embed_dim=384,
+        depth=12,
+        num_heads=6,
+        mlp_ratio=4,
+        qkv_bias=True,
+        norm_layer=partial(nn.LayerNorm, eps=1e-6),
+        **kwargs
+    )
+    return model
+
+
+def vit_base_patch16(**kwargs):
+    model = VisionTransformer(
+        patch_size=16,
+        embed_dim=768,
+        depth=12,
+        num_heads=12,
+        mlp_ratio=4,
+        qkv_bias=True,
+        norm_layer=partial(nn.LayerNorm, eps=1e-6),
+        **kwargs
+    )
+    return model
+
+
+def vit_large_patch16(**kwargs):
+    model = VisionTransformer(
+        patch_size=16,
+        embed_dim=1024,
+        depth=24,
+        num_heads=16,
+        mlp_ratio=4,
+        qkv_bias=True,
+        norm_layer=partial(nn.LayerNorm, eps=1e-6),
+        **kwargs
+    )
+    return model
+
+
+def vit_huge_patch14(**kwargs):
+    model = VisionTransformer(
+        patch_size=14,
+        embed_dim=1280,
+        depth=32,
+        num_heads=16,
+        mlp_ratio=4,
+        qkv_bias=True,
+        norm_layer=partial(nn.LayerNorm, eps=1e-6),
+        **kwargs
+    )
+    return model

audiosr/latent_diffusion/modules/audiomae/util/crop.py

@@ -1,43 +1,43 @@
-# Copyright (c) Meta Platforms, Inc. and affiliates.
-# All rights reserved.
-
-# This source code is licensed under the license found in the
-# LICENSE file in the root directory of this source tree.
-
-import math
-
-import torch
-
-from torchvision import transforms
-from torchvision.transforms import functional as F
-
-
-class RandomResizedCrop(transforms.RandomResizedCrop):
-    """
-    RandomResizedCrop for matching TF/TPU implementation: no for-loop is used.
-    This may lead to results different with torchvision's version.
-    Following BYOL's TF code:
-    https://github.com/deepmind/deepmind-research/blob/master/byol/utils/dataset.py#L206
-    """
-
-    @staticmethod
-    def get_params(img, scale, ratio):
-        width, height = F._get_image_size(img)
-        area = height * width
-
-        target_area = area * torch.empty(1).uniform_(scale[0], scale[1]).item()
-        log_ratio = torch.log(torch.tensor(ratio))
-        aspect_ratio = torch.exp(
-            torch.empty(1).uniform_(log_ratio[0], log_ratio[1])
-        ).item()
-
-        w = int(round(math.sqrt(target_area * aspect_ratio)))
-        h = int(round(math.sqrt(target_area / aspect_ratio)))
-
-        w = min(w, width)
-        h = min(h, height)
-
-        i = torch.randint(0, height - h + 1, size=(1,)).item()
-        j = torch.randint(0, width - w + 1, size=(1,)).item()
-
-        return i, j, h, w
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import math
+
+import torch
+
+from torchvision import transforms
+from torchvision.transforms import functional as F
+
+
+class RandomResizedCrop(transforms.RandomResizedCrop):
+    """
+    RandomResizedCrop for matching TF/TPU implementation: no for-loop is used.
+    This may lead to results different with torchvision's version.
+    Following BYOL's TF code:
+    https://github.com/deepmind/deepmind-research/blob/master/byol/utils/dataset.py#L206
+    """
+
+    @staticmethod
+    def get_params(img, scale, ratio):
+        width, height = F._get_image_size(img)
+        area = height * width
+
+        target_area = area * torch.empty(1).uniform_(scale[0], scale[1]).item()
+        log_ratio = torch.log(torch.tensor(ratio))
+        aspect_ratio = torch.exp(
+            torch.empty(1).uniform_(log_ratio[0], log_ratio[1])
+        ).item()
+
+        w = int(round(math.sqrt(target_area * aspect_ratio)))
+        h = int(round(math.sqrt(target_area / aspect_ratio)))
+
+        w = min(w, width)
+        h = min(h, height)
+
+        i = torch.randint(0, height - h + 1, size=(1,)).item()
+        j = torch.randint(0, width - w + 1, size=(1,)).item()
+
+        return i, j, h, w