Initial commit

.gitattributes

@@ -33,3 +33,5 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+*.jpg filter=lfs diff=lfs merge=lfs -text
+*.png filter=lfs diff=lfs merge=lfs -text

.gitignore

@@ -0,0 +1,146 @@
+datasets/*
+experiments/*
+results/*
+tb_logger/*
+wandb/*
+tmp/*
+weights/*
+inputs/*
+models/*
+comparisons/*
+flux_dev_fp8_quantized_model.pth
+array_outputs/*
+
+*.DS_Store
+
+# Byte-compiled / optimized / DLL files
+__pycache__/
+*.py[cod]
+*$py.class
+
+# C extensions
+*.so
+
+# Distribution / packaging
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+pip-wheel-metadata/
+share/python-wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+MANIFEST
+
+# PyInstaller
+#  Usually these files are written by a python script from a template
+#  before PyInstaller builds the exe, so as to inject date/other infos into it.
+*.manifest
+*.spec
+
+# Installer logs
+pip-log.txt
+pip-delete-this-directory.txt
+
+# Unit test / coverage reports
+htmlcov/
+.tox/
+.nox/
+.coverage
+.coverage.*
+.cache
+nosetests.xml
+coverage.xml
+*.cover
+*.py,cover
+.hypothesis/
+.pytest_cache/
+
+# Translations
+*.mo
+*.pot
+
+# Django stuff:
+*.log
+local_settings.py
+db.sqlite3
+db.sqlite3-journal
+
+# Flask stuff:
+instance/
+.webassets-cache
+
+# Scrapy stuff:
+.scrapy
+
+# Sphinx documentation
+docs/_build/
+
+# PyBuilder
+target/
+
+# Jupyter Notebook
+.ipynb_checkpoints
+
+# IPython
+profile_default/
+ipython_config.py
+
+# pyenv
+.python-version
+
+# pipenv
+#   According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
+#   However, in case of collaboration, if having platform-specific dependencies or dependencies
+#   having no cross-platform support, pipenv may install dependencies that don't work, or not
+#   install all needed dependencies.
+#Pipfile.lock
+
+# PEP 582; used by e.g. github.com/David-OConnor/pyflow
+__pypackages__/
+
+# Celery stuff
+celerybeat-schedule
+celerybeat.pid
+
+# SageMath parsed files
+*.sage.py
+
+# Environments
+.env
+.venv
+env/
+venv/
+ENV/
+env.bak/
+venv.bak/
+
+# Spyder project settings
+.spyderproject
+.spyproject
+
+# Rope project settings
+.ropeproject
+
+# mkdocs documentation
+/site
+
+# mypy
+.mypy_cache/
+.dmypy.json
+dmypy.json
+
+# Pyre type checker
+.pyre/
+
+.idea/

app.py

@@ -0,0 +1,461 @@
+import time
+
+import gradio as gr
+import spaces
+import numpy as np
+import torch
+from einops import rearrange, repeat
+from PIL import Image
+
+from flux.sampling import denoise, get_noise, get_schedule, prepare, rf_denoise, rf_inversion, unpack
+from flux.util import (
+    SamplingOptions,
+    load_ae,
+    load_clip,
+    load_flow_model,
+    load_flow_model_quintized,
+    load_t5,
+)
+from pulid.pipeline_flux import PuLIDPipeline
+from pulid.utils import resize_numpy_image_long, seed_everything
+
+
+def get_models(name: str, device: torch.device, offload: bool, fp8: bool):
+    t5 = load_t5(device, max_length=128)
+    clip = load_clip(device)
+    if fp8:
+        model = load_flow_model_quintized(name, device="cpu" if offload else device)
+    else:
+        model = load_flow_model(name, device="cpu" if offload else device)
+    model.eval()
+    ae = load_ae(name, device="cpu" if offload else device)
+    return model, ae, t5, clip
+
+
+class FluxGenerator:
+    def __init__(self, model_name: str, device: str, offload: bool, aggressive_offload: bool, args):
+        self.device = torch.device(device)
+        self.offload = offload
+        self.aggressive_offload = aggressive_offload
+        self.model_name = model_name
+        self.model, self.ae, self.t5, self.clip_model = get_models(
+            model_name,
+            device=self.device,
+            offload=self.offload,
+            fp8=args.fp8,
+        )
+        self.pulid_model = PuLIDPipeline(self.model, device="cpu" if offload else device, weight_dtype=torch.bfloat16,
+                                         onnx_provider=args.onnx_provider)
+        if offload:
+            self.pulid_model.face_helper.face_det.mean_tensor = self.pulid_model.face_helper.face_det.mean_tensor.to(torch.device("cuda"))
+            self.pulid_model.face_helper.face_det.device = torch.device("cuda")
+            self.pulid_model.face_helper.device = torch.device("cuda")
+            self.pulid_model.device = torch.device("cuda")
+        self.pulid_model.load_pretrain(args.pretrained_model, version=args.version)
+
+    # function to encode an image into latents
+    def encode_image_to_latents(self, img, opts):
+        """
+        Opposite of decode: Takes a PIL image and encodes it into latents (x).
+        """
+        t0 = time.perf_counter()
+
+        # Resize if necessary, or use opts.height / opts.width if you want a fixed size:
+        img = img.resize((opts.width, opts.height), resample=Image.LANCZOS)
+
+        # Convert image to torch.Tensor and scale to [-1, 1]
+        # Image is in [0, 255] → scale to [0,1] → then map to [-1,1].
+        x = np.array(img).astype(np.float32)
+        x = torch.from_numpy(x)  # shape: (H, W, C)
+        x = (x / 127.5) - 1.0    # now in [-1, 1]
+        x = rearrange(x, "h w c -> 1 c h w")  # shape: (1, C, H, W)
+
+        # Move encoder to device if you are offloading
+        if self.offload:
+            self.ae.encoder.to(self.device)
+
+        x = x.to(self.device, dtype=torch.bfloat16)
+
+        # 2) Encode with autocast
+        with torch.autocast(device_type=self.device.type, dtype=torch.bfloat16):
+            x = self.ae.encode(x)
+
+        x = x.to(torch.bfloat16)
+
+
+        # 3) Offload if needed
+        if self.offload:
+            self.ae.encoder.cpu()
+            torch.cuda.empty_cache()
+
+        t1 = time.perf_counter()
+        print(f"Encoded in {t1 - t0:.2f} seconds.")
+
+        return x
+
+    @spaces.GPU
+    @torch.inference_mode()
+    def generate_image(
+        self,
+        prompt: str,
+        id_image = None,
+        width: int = 512,
+        height: int = 512,
+        num_steps: int = 20,
+        start_step: int = 0,
+        guidance: float = 4.0,
+        seed: int = -1,
+        id_weight: float = 1.0,
+        neg_prompt: str = "",
+        true_cfg: float = 1.0,
+        timestep_to_start_cfg: int = 1,
+        max_sequence_length: int = 128,
+        gamma: float = 0.5,
+        eta: float = 0.7,
+        s: float = 0,
+        tau: float = 5,
+        perform_inversion: bool = True,
+        perform_reconstruction: bool = False,
+        perform_editing: bool = True,
+        inversion_true_cfg: float = 1.0,
+    ):
+        """
+        Core function that performs the image generation.
+        """
+        self.t5.max_length = max_sequence_length
+
+        # If seed == -1, random
+        seed = int(seed)
+        if seed == -1:
+            seed = None
+
+        opts = SamplingOptions(
+            prompt=prompt,
+            width=width,
+            height=height,
+            num_steps=num_steps,
+            guidance=guidance,
+            seed=seed,
+        )
+
+        if opts.seed is None:
+            opts.seed = torch.Generator(device="cpu").seed()
+
+        seed_everything(opts.seed)
+
+        print(f"Generating prompt: '{opts.prompt}' (seed={opts.seed})...")
+        t0 = time.perf_counter()
+
+        use_true_cfg = abs(true_cfg - 1.0) > 1e-6
+
+
+        # 1) Prepare input noise
+        noise = get_noise(
+            num_samples=1,
+            height=opts.height,
+            width=opts.width,
+            device=self.device,
+            dtype=torch.bfloat16,
+            seed=opts.seed,
+        )
+        bs, c, h, w = noise.shape
+        noise = rearrange(noise, "b c (h ph) (w pw) -> b (h w) (c ph pw)", ph=2, pw=2)
+        if noise.shape[0] == 1 and bs > 1:
+            noise = repeat(noise, "1 ... -> bs ...", bs=bs)
+        # encode
+        x = self.encode_image_to_latents(id_image, opts)
+        
+        timesteps = get_schedule(opts.num_steps, x.shape[-1] * x.shape[-2] // 4, shift=False)
+
+        # 2) Prepare text embeddings
+        if self.offload:
+            self.t5 = self.t5.to(self.device)
+            self.clip_model = self.clip_model.to(self.device)
+
+        inp = prepare(t5=self.t5, clip=self.clip_model, img=x, prompt=opts.prompt)
+        inp_inversion = prepare(t5=self.t5, clip=self.clip_model, img=x, prompt="")
+        inp_neg = None
+        if use_true_cfg:
+            inp_neg = prepare(t5=self.t5, clip=self.clip_model, img=x, prompt=neg_prompt)
+
+        # Offload text encoders, load ID detection to GPU
+        if self.offload:
+            self.t5 = self.t5.cpu()
+            self.clip_model = self.clip_model.cpu()
+            torch.cuda.empty_cache()
+            self.pulid_model.components_to_device(torch.device("cuda"))
+
+        # 3) ID Embeddings (optional)
+        id_embeddings = None
+        uncond_id_embeddings = None
+        if id_image is not None:
+            id_image = np.array(id_image)
+            id_image = resize_numpy_image_long(id_image, 1024)
+            id_embeddings, uncond_id_embeddings = self.pulid_model.get_id_embedding(id_image, cal_uncond=use_true_cfg)
+        else:
+            id_embeddings = None
+            uncond_id_embeddings = None
+
+        # Offload ID pipeline, load main FLUX model to GPU
+        if self.offload:
+            self.pulid_model.components_to_device(torch.device("cpu"))
+            torch.cuda.empty_cache()
+
+            if self.aggressive_offload:
+                self.model.components_to_gpu()
+            else:
+                self.model = self.model.to(self.device)
+
+        y_0 = inp["img"].clone().detach()
+
+        inverted = None
+        if perform_inversion:
+            inverted = rf_inversion(
+                self.model,
+                **inp_inversion,
+                timesteps=timesteps,
+                guidance=opts.guidance,
+                id=id_embeddings,
+                id_weight=id_weight,
+                start_step=start_step,
+                uncond_id=uncond_id_embeddings,
+                true_cfg=inversion_true_cfg,
+                timestep_to_start_cfg=timestep_to_start_cfg,
+                neg_txt=inp_neg["txt"] if use_true_cfg else None,
+                neg_txt_ids=inp_neg["txt_ids"] if use_true_cfg else None,
+                neg_vec=inp_neg["vec"] if use_true_cfg else None,
+                aggressive_offload=self.aggressive_offload,
+                y_1=noise,
+                gamma=gamma
+            )
+
+            img = inverted
+        else:
+            img = noise
+        inp["img"] = img
+        inp_inversion["img"] = img
+
+        recon = None
+        if perform_reconstruction:
+            recon = rf_denoise(
+                self.model,
+                **inp_inversion,
+                timesteps=timesteps,
+                guidance=opts.guidance,
+                id=id_embeddings,
+                id_weight=id_weight,
+                start_step=start_step,
+                uncond_id=uncond_id_embeddings,
+                true_cfg=inversion_true_cfg,
+                timestep_to_start_cfg=timestep_to_start_cfg,
+                neg_txt=inp_neg["txt"] if use_true_cfg else None,
+                neg_txt_ids=inp_neg["txt_ids"] if use_true_cfg else None,
+                neg_vec=inp_neg["vec"] if use_true_cfg else None,
+                aggressive_offload=self.aggressive_offload,
+                y_0=y_0,
+                eta=eta,
+                s=s,
+                tau=tau,
+            )
+
+        edited = None
+        if perform_editing:
+            edited = rf_denoise(
+                self.model,
+                **inp,
+                timesteps=timesteps,
+                guidance=opts.guidance,
+                id=id_embeddings,
+                id_weight=id_weight,
+                start_step=start_step,
+                uncond_id=uncond_id_embeddings,
+                true_cfg=true_cfg,
+                timestep_to_start_cfg=timestep_to_start_cfg,
+                neg_txt=inp_neg["txt"] if use_true_cfg else None,
+                neg_txt_ids=inp_neg["txt_ids"] if use_true_cfg else None,
+                neg_vec=inp_neg["vec"] if use_true_cfg else None,
+                aggressive_offload=self.aggressive_offload,
+                y_0=y_0,
+                eta=eta,
+                s=s,
+                tau=tau,
+            )
+
+        # Offload flux model, load auto-decoder
+        if self.offload:
+            self.model.cpu()
+            torch.cuda.empty_cache()
+            self.ae.decoder.to(x.device)
+
+        # 5) Decode latents
+        if edited is not None:
+            edited = unpack(edited.float(), opts.height, opts.width)
+            with torch.autocast(device_type=self.device.type, dtype=torch.bfloat16):
+                edited = self.ae.decode(edited)
+
+        if inverted is not None:
+            inverted = unpack(inverted.float(), opts.height, opts.width)
+            with torch.autocast(device_type=self.device.type, dtype=torch.bfloat16):
+                inverted = self.ae.decode(inverted)
+        
+        if recon is not None:
+            recon = unpack(recon.float(), opts.height, opts.width)
+            with torch.autocast(device_type=self.device.type, dtype=torch.bfloat16):
+                recon = self.ae.decode(recon)
+
+        if self.offload:
+            self.ae.decoder.cpu()
+            torch.cuda.empty_cache()
+
+        t1 = time.perf_counter()
+        print(f"Done in {t1 - t0:.2f} seconds.")
+
+        # Convert to PIL
+        if edited is not None:
+            edited = edited.clamp(-1, 1)
+            edited = rearrange(edited[0], "c h w -> h w c")
+            edited = Image.fromarray((127.5 * (edited + 1.0)).cpu().byte().numpy())
+
+        if inverted is not None:
+            inverted = inverted.clamp(-1, 1)
+            inverted = rearrange(inverted[0], "c h w -> h w c")
+            inverted = Image.fromarray((127.5 * (inverted + 1.0)).cpu().byte().numpy())
+        
+        if recon is not None:
+            recon = recon.clamp(-1, 1)
+            recon = rearrange(recon[0], "c h w -> h w c")
+            recon = Image.fromarray((127.5 * (recon + 1.0)).cpu().byte().numpy())
+
+        return edited, str(opts.seed), self.pulid_model.debug_img_list
+
+# <p style="font-size: 1rem; margin-bottom: 1.5rem;">Paper: <a href='https://arxiv.org/abs/2404.16022' target='_blank'>PuLID: Pure and Lightning ID Customization via Contrastive Alignment</a> | Codes: <a href='https://github.com/ToTheBeginning/PuLID' target='_blank'>GitHub</a></p>
+_HEADER_ = '''
+<div style="text-align: center; max-width: 650px; margin: 0 auto;">
+    <h1 style="font-size: 2.5rem; font-weight: 700; margin-bottom: 1rem; display: contents;">Tight Inversion for Portrait Editing with FLUX</h1>
+</div>
+
+❗️❗️❗️**Tips:**
+Provide a portrait image and an edit prompt. You can try the examples below or upload your own image.
+Adjust the id weight to control the faithfulness of the generated image to the input image.
+'''  # noqa E501
+_CITE_ = r"""
+"""  # noqa E501
+
+
+def create_demo(args, model_name: str, device: str = "cuda" if torch.cuda.is_available() else "cpu",
+                offload: bool = False, aggressive_offload: bool = False):
+    generator = FluxGenerator(model_name, device, offload, aggressive_offload, args)
+
+    with gr.Blocks() as demo:
+        gr.Markdown(_HEADER_)
+
+        with gr.Row():
+            with gr.Column():
+                prompt = gr.Textbox(label="Prompt", value="portrait, color, cinematic")
+                id_image = gr.Image(label="ID Image", type="pil")
+                id_weight = gr.Slider(0.0, 1.0, 0.4, step=0.05, label="id weight")
+
+                width = gr.Slider(256, 1536, 1024, step=16, label="Width", visible=args.dev)
+                height = gr.Slider(256, 1536, 1024, step=16, label="Height", visible=args.dev)
+                num_steps = gr.Slider(1, 28, 16, step=1, label="Number of steps")
+                guidance = gr.Slider(1.0, 10.0, 3.5, step=0.1, label="Guidance")
+
+                with gr.Accordion("Advanced Options (True CFG, true_cfg_scale=1 means use fake CFG, >1 means use true CFG", open=False):    # noqa E501
+                    neg_prompt = gr.Textbox(
+                        label="Negative Prompt",
+                        value="")
+                    true_cfg = gr.Slider(1.0, 10.0, 3.5, step=0.1, label="true CFG scale")
+                    timestep_to_start_cfg = gr.Slider(0, 20, 1, step=1, label="timestep to start cfg", visible=args.dev)
+                    start_step = gr.Slider(0, 10, 0, step=1, label="timestep to start inserting ID")
+                    seed = gr.Textbox(-1, label="Seed (-1 for random)")
+                    max_sequence_length = gr.Slider(128, 512, 128, step=128,
+                                                    label="max_sequence_length for prompt (T5), small will be faster")
+                    gr.Markdown("### RF Inversion Options")
+                    gamma = gr.Slider(0.0, 1.0, 0.5, step=0.1, label="gamma")
+                    eta = gr.Slider(0.0, 1.0, 0.7, step=0.1, label="eta")
+                    s = gr.Slider(0.0, 1.0, 0.0, step=0.1, label="s")
+                    tau = gr.Slider(0, 20, 2, step=1, label="tau")
+
+                generate_btn = gr.Button("Generate")
+
+            with gr.Column():
+                output_image = gr.Image(label="Generated Image")
+                seed_output = gr.Textbox(label="Used Seed")
+                intermediate_output = gr.Gallery(label='Output', elem_id="gallery", visible=args.dev)
+                gr.Markdown(_CITE_)
+
+        with gr.Row(), gr.Column():
+                gr.Markdown("## Examples")
+                example_inps = [
+                    # [
+                    #     'a portrait of a vampire',
+                    #     'example_inputs/unsplash/krisna-putra-pratama-lKF-MdtuIss-unsplash.jpg',
+                    #     0.4, 3.5, 42, 3.5
+                    # ],
+                    [
+                        'a portrait of a zombie',
+                        'example_inputs/unsplash/baruk-granda-cfLL_jHQ-Iw-unsplash.jpg',
+                        0.4, 3.5, 42, 5.0
+                    ],
+                    [
+                        'a portrait of an elf',
+                        'example_inputs/unsplash/rahmat-alizada-7PwFKOgyoKo-unsplash.jpg',
+                        0.5, 3.5, 42, 5.0
+                    ],
+                    [
+                        'a portrait of a clown',
+                        'example_inputs/unsplash/lhon-karwan-11tbHtK5STE-unsplash.jpg',
+                        0.5, 3.5, 42, 3.5
+                    ],
+                    [
+                        'a portrait of an elf',
+                        'example_inputs/unsplash/masoud-razeghi--qsrZhXPius-unsplash.jpg',
+                        0.5, 3.5, 42, 5.0
+                    ],
+                    # [
+                    #     'a portrait of a pirate',
+                    #     'example_inputs/unsplash/mina-rad-AEVUFpDGxZM-unsplash.jpg',
+                    #     0.3, 3.5, 42, 3.5
+                    # ],
+                    [
+                        'a portrait of a superhero',
+                        'example_inputs/unsplash/gus-tu-njana-Mf4MN7MZqcE-unsplash.jpg',
+                        0.2, 3.5, 42, 5.0
+                    ],
+                ]
+                gr.Examples(examples=example_inps, inputs=[prompt, id_image, id_weight, guidance, seed, true_cfg])
+
+        generate_btn.click(
+            fn=generator.generate_image,
+            inputs=[prompt, id_image, width, height, num_steps, start_step, guidance, seed, id_weight, neg_prompt,
+                    true_cfg, timestep_to_start_cfg, max_sequence_length, gamma, eta, s, tau],
+            outputs=[output_image, seed_output, intermediate_output],
+        )
+
+    return demo
+
+
+if __name__ == "__main__":
+    import argparse
+
+    parser = argparse.ArgumentParser(description="PuLID for FLUX.1-dev")
+    parser.add_argument('--version', type=str, default='v0.9.1', help='version of the model', choices=['v0.9.0', 'v0.9.1'])
+    parser.add_argument("--name", type=str, default="flux-dev", choices=list('flux-dev'),
+                        help="currently only support flux-dev")
+    parser.add_argument("--device", type=str, default="cuda", help="Device to use")
+    parser.add_argument("--offload", action="store_true", help="Offload model to CPU when not in use")
+    parser.add_argument("--aggressive_offload", action="store_true", help="Offload model more aggressively to CPU when not in use, for 24G GPUs")
+    parser.add_argument("--fp8", action="store_true", help="use flux-dev-fp8 model")
+    parser.add_argument("--onnx_provider", type=str, default="gpu", choices=["gpu", "cpu"],
+                        help="set onnx_provider to cpu (default gpu) can help reduce RAM usage, and when combined with"
+                             "fp8 option, the peak RAM is under 15GB")
+    parser.add_argument("--port", type=int, default=8080, help="Port to use")
+    parser.add_argument("--dev", action='store_true', help="Development mode")
+    parser.add_argument("--pretrained_model", type=str, help='for development')
+    args = parser.parse_args()
+
+    if args.aggressive_offload:
+        args.offload = True
+
+    demo = create_demo(args, args.name, args.device, args.offload, args.aggressive_offload)
+    demo.launch(server_name='0.0.0.0', server_port=args.port, ssr_mode=False)

eva_clip/__init__.py

@@ -0,0 +1,11 @@
+from .constants import OPENAI_DATASET_MEAN, OPENAI_DATASET_STD
+from .factory import create_model, create_model_and_transforms, create_model_from_pretrained, get_tokenizer, create_transforms
+from .factory import list_models, add_model_config, get_model_config, load_checkpoint
+from .loss import ClipLoss
+from .model import CLIP, CustomCLIP, CLIPTextCfg, CLIPVisionCfg,\
+    convert_weights_to_lp, convert_weights_to_fp16, trace_model, get_cast_dtype
+from .openai import load_openai_model, list_openai_models
+from .pretrained import list_pretrained, list_pretrained_models_by_tag, list_pretrained_tags_by_model,\
+    get_pretrained_url, download_pretrained_from_url, is_pretrained_cfg, get_pretrained_cfg, download_pretrained
+from .tokenizer import SimpleTokenizer, tokenize
+from .transform import image_transform

eva_clip/bpe_simple_vocab_16e6.txt.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:924691ac288e54409236115652ad4aa250f48203de50a9e4722a6ecd48d6804a
+size 1356917

eva_clip/constants.py

@@ -0,0 +1,2 @@
+OPENAI_DATASET_MEAN = (0.48145466, 0.4578275, 0.40821073)
+OPENAI_DATASET_STD = (0.26862954, 0.26130258, 0.27577711)

eva_clip/eva_vit_model.py

@@ -0,0 +1,548 @@
+# --------------------------------------------------------
+# Adapted from  https://github.com/microsoft/unilm/tree/master/beit
+# --------------------------------------------------------
+import math
+import os
+from functools import partial
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+try:
+    from timm.models.layers import drop_path, to_2tuple, trunc_normal_
+except:
+    from timm.layers import drop_path, to_2tuple, trunc_normal_
+    
+from .transformer import PatchDropout
+from .rope import VisionRotaryEmbedding, VisionRotaryEmbeddingFast
+
+if os.getenv('ENV_TYPE') == 'deepspeed':
+    try:
+        from deepspeed.runtime.activation_checkpointing.checkpointing import checkpoint
+    except:
+        from torch.utils.checkpoint import checkpoint
+else:
+    from torch.utils.checkpoint import checkpoint
+
+try:
+    import xformers
+    import xformers.ops as xops
+    XFORMERS_IS_AVAILBLE = True
+except:
+    XFORMERS_IS_AVAILBLE = False
+
+class DropPath(nn.Module):
+    """Drop paths (Stochastic Depth) per sample  (when applied in main path of residual blocks).
+    """
+    def __init__(self, drop_prob=None):
+        super(DropPath, self).__init__()
+        self.drop_prob = drop_prob
+
+    def forward(self, x):
+        return drop_path(x, self.drop_prob, self.training)
+    
+    def extra_repr(self) -> str:
+        return 'p={}'.format(self.drop_prob)
+
+
+class Mlp(nn.Module):
+    def __init__(
+        self, 
+        in_features, 
+        hidden_features=None, 
+        out_features=None, 
+        act_layer=nn.GELU, 
+        norm_layer=nn.LayerNorm, 
+        drop=0.,
+        subln=False,
+
+        ):
+        super().__init__()
+        out_features = out_features or in_features
+        hidden_features = hidden_features or in_features
+        self.fc1 = nn.Linear(in_features, hidden_features)
+        self.act = act_layer()
+
+        self.ffn_ln = norm_layer(hidden_features) if subln else nn.Identity()
+
+        self.fc2 = nn.Linear(hidden_features, out_features)
+        self.drop = nn.Dropout(drop)
+
+    def forward(self, x):
+        x = self.fc1(x)
+        x = self.act(x)
+        # x = self.drop(x)
+        # commit this for the orignal BERT implement 
+        x = self.ffn_ln(x)
+
+        x = self.fc2(x)
+        x = self.drop(x)
+        return x
+
+class SwiGLU(nn.Module):
+    def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.SiLU, drop=0., 
+                norm_layer=nn.LayerNorm, subln=False):
+        super().__init__()
+        out_features = out_features or in_features
+        hidden_features = hidden_features or in_features
+
+        self.w1 = nn.Linear(in_features, hidden_features)
+        self.w2 = nn.Linear(in_features, hidden_features)
+
+        self.act = act_layer()
+        self.ffn_ln = norm_layer(hidden_features) if subln else nn.Identity()
+        self.w3 = nn.Linear(hidden_features, out_features)
+        
+        self.drop = nn.Dropout(drop)
+
+    def forward(self, x):
+        x1 = self.w1(x)
+        x2 = self.w2(x)
+        hidden = self.act(x1) * x2
+        x = self.ffn_ln(hidden)
+        x = self.w3(x)
+        x = self.drop(x)
+        return x
+
+class Attention(nn.Module):
+    def __init__(
+            self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0.,
+            proj_drop=0., window_size=None, attn_head_dim=None, xattn=False, rope=None, subln=False, norm_layer=nn.LayerNorm):
+        super().__init__()
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+        if attn_head_dim is not None:
+            head_dim = attn_head_dim
+        all_head_dim = head_dim * self.num_heads
+        self.scale = qk_scale or head_dim ** -0.5
+
+        self.subln = subln
+        if self.subln:
+            self.q_proj = nn.Linear(dim, all_head_dim, bias=False)
+            self.k_proj = nn.Linear(dim, all_head_dim, bias=False)
+            self.v_proj = nn.Linear(dim, all_head_dim, bias=False)
+        else:
+            self.qkv = nn.Linear(dim, all_head_dim * 3, bias=False)
+
+        if qkv_bias:
+            self.q_bias = nn.Parameter(torch.zeros(all_head_dim))
+            self.v_bias = nn.Parameter(torch.zeros(all_head_dim))
+        else:
+            self.q_bias = None
+            self.v_bias = None
+
+        if window_size:
+            self.window_size = window_size
+            self.num_relative_distance = (2 * window_size[0] - 1) * (2 * window_size[1] - 1) + 3
+            self.relative_position_bias_table = nn.Parameter(
+                torch.zeros(self.num_relative_distance, num_heads))  # 2*Wh-1 * 2*Ww-1, nH
+            # cls to token & token 2 cls & cls to cls
+
+            # get pair-wise relative position index for each token inside the window
+            coords_h = torch.arange(window_size[0])
+            coords_w = torch.arange(window_size[1])
+            coords = torch.stack(torch.meshgrid([coords_h, coords_w]))  # 2, Wh, Ww
+            coords_flatten = torch.flatten(coords, 1)  # 2, Wh*Ww
+            relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :]  # 2, Wh*Ww, Wh*Ww
+            relative_coords = relative_coords.permute(1, 2, 0).contiguous()  # Wh*Ww, Wh*Ww, 2
+            relative_coords[:, :, 0] += window_size[0] - 1  # shift to start from 0
+            relative_coords[:, :, 1] += window_size[1] - 1
+            relative_coords[:, :, 0] *= 2 * window_size[1] - 1
+            relative_position_index = \
+                torch.zeros(size=(window_size[0] * window_size[1] + 1, ) * 2, dtype=relative_coords.dtype)
+            relative_position_index[1:, 1:] = relative_coords.sum(-1)  # Wh*Ww, Wh*Ww
+            relative_position_index[0, 0:] = self.num_relative_distance - 3
+            relative_position_index[0:, 0] = self.num_relative_distance - 2
+            relative_position_index[0, 0] = self.num_relative_distance - 1
+
+            self.register_buffer("relative_position_index", relative_position_index)
+        else:
+            self.window_size = None
+            self.relative_position_bias_table = None
+            self.relative_position_index = None
+
+        self.attn_drop = nn.Dropout(attn_drop)
+        self.inner_attn_ln = norm_layer(all_head_dim) if subln else nn.Identity()
+        # self.proj = nn.Linear(all_head_dim, all_head_dim)
+        self.proj = nn.Linear(all_head_dim, dim)
+        self.proj_drop = nn.Dropout(proj_drop)
+        self.xattn = xattn
+        self.xattn_drop = attn_drop
+
+        self.rope = rope
+
+    def forward(self, x, rel_pos_bias=None, attn_mask=None):
+        B, N, C = x.shape
+        if self.subln: 
+            q = F.linear(input=x, weight=self.q_proj.weight, bias=self.q_bias)
+            k = F.linear(input=x, weight=self.k_proj.weight, bias=None)
+            v = F.linear(input=x, weight=self.v_proj.weight, bias=self.v_bias)
+
+            q = q.reshape(B, N, self.num_heads, -1).permute(0, 2, 1, 3)     # B, num_heads, N, C
+            k = k.reshape(B, N, self.num_heads, -1).permute(0, 2, 1, 3)  
+            v = v.reshape(B, N, self.num_heads, -1).permute(0, 2, 1, 3) 
+        else: 
+
+            qkv_bias = None
+            if self.q_bias is not None:
+                qkv_bias = torch.cat((self.q_bias, torch.zeros_like(self.v_bias, requires_grad=False), self.v_bias))
+            
+            qkv = F.linear(input=x, weight=self.qkv.weight, bias=qkv_bias)
+            qkv = qkv.reshape(B, N, 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)   # 3, B, num_heads, N, C
+            q, k, v = qkv[0], qkv[1], qkv[2]
+
+        if self.rope:
+            # slightly fast impl
+            q_t = q[:, :, 1:, :]
+            ro_q_t = self.rope(q_t)
+            q = torch.cat((q[:, :, :1, :], ro_q_t), -2).type_as(v)
+
+            k_t = k[:, :, 1:, :]
+            ro_k_t = self.rope(k_t)
+            k = torch.cat((k[:, :, :1, :], ro_k_t), -2).type_as(v)
+
+        if self.xattn:
+            q = q.permute(0, 2, 1, 3)   # B, num_heads, N, C -> B, N, num_heads, C
+            k = k.permute(0, 2, 1, 3)
+            v = v.permute(0, 2, 1, 3)
+
+            x = xops.memory_efficient_attention(
+                q, k, v,
+                p=self.xattn_drop,
+                scale=self.scale,
+                )
+            x = x.reshape(B, N, -1)
+            x = self.inner_attn_ln(x)
+            x = self.proj(x)
+            x = self.proj_drop(x)
+        else:
+            q = q * self.scale
+            attn = (q @ k.transpose(-2, -1))
+
+            if self.relative_position_bias_table is not None:
+                relative_position_bias = \
+                    self.relative_position_bias_table[self.relative_position_index.view(-1)].view(
+                        self.window_size[0] * self.window_size[1] + 1,
+                        self.window_size[0] * self.window_size[1] + 1, -1)  # Wh*Ww,Wh*Ww,nH
+                relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous()  # nH, Wh*Ww, Wh*Ww
+                attn = attn + relative_position_bias.unsqueeze(0).type_as(attn)
+
+            if rel_pos_bias is not None:
+                attn = attn + rel_pos_bias.type_as(attn)
+
+            if attn_mask is not None:
+                attn_mask = attn_mask.bool()
+                attn = attn.masked_fill(~attn_mask[:, None, None, :], float("-inf"))
+            
+            attn = attn.softmax(dim=-1)
+            attn = self.attn_drop(attn)
+
+            x = (attn @ v).transpose(1, 2).reshape(B, N, -1)
+            x = self.inner_attn_ln(x)
+            x = self.proj(x)
+            x = self.proj_drop(x)
+        return x
+
+
+class Block(nn.Module):
+
+    def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0.,
+                 drop_path=0., init_values=None, act_layer=nn.GELU, norm_layer=nn.LayerNorm,
+                 window_size=None, attn_head_dim=None, xattn=False, rope=None, postnorm=False,
+                 subln=False, naiveswiglu=False):
+        super().__init__()
+        self.norm1 = norm_layer(dim)
+        self.attn = Attention(
+            dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale,
+            attn_drop=attn_drop, proj_drop=drop, window_size=window_size, attn_head_dim=attn_head_dim,
+            xattn=xattn, rope=rope, subln=subln, norm_layer=norm_layer)
+        # NOTE: drop path for stochastic depth, we shall see if this is better than dropout here
+        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
+        self.norm2 = norm_layer(dim)
+        mlp_hidden_dim = int(dim * mlp_ratio)
+
+        if naiveswiglu:
+            self.mlp = SwiGLU(
+                in_features=dim, 
+                hidden_features=mlp_hidden_dim, 
+                subln=subln,
+                norm_layer=norm_layer,
+            )
+        else:
+            self.mlp = Mlp(
+                in_features=dim, 
+                hidden_features=mlp_hidden_dim, 
+                act_layer=act_layer,
+                subln=subln,
+                drop=drop
+            )
+
+        if init_values is not None and init_values > 0:
+            self.gamma_1 = nn.Parameter(init_values * torch.ones((dim)),requires_grad=True)
+            self.gamma_2 = nn.Parameter(init_values * torch.ones((dim)),requires_grad=True)
+        else:
+            self.gamma_1, self.gamma_2 = None, None
+
+        self.postnorm = postnorm
+
+    def forward(self, x, rel_pos_bias=None, attn_mask=None):
+        if self.gamma_1 is None:
+            if self.postnorm:
+                x = x + self.drop_path(self.norm1(self.attn(x, rel_pos_bias=rel_pos_bias, attn_mask=attn_mask)))
+                x = x + self.drop_path(self.norm2(self.mlp(x)))
+            else:
+                x = x + self.drop_path(self.attn(self.norm1(x), rel_pos_bias=rel_pos_bias, attn_mask=attn_mask))
+                x = x + self.drop_path(self.mlp(self.norm2(x)))
+        else:
+            if self.postnorm:
+                x = x + self.drop_path(self.gamma_1 * self.norm1(self.attn(x, rel_pos_bias=rel_pos_bias, attn_mask=attn_mask)))
+                x = x + self.drop_path(self.gamma_2 * self.norm2(self.mlp(x)))
+            else:
+                x = x + self.drop_path(self.gamma_1 * self.attn(self.norm1(x), rel_pos_bias=rel_pos_bias, attn_mask=attn_mask))
+                x = x + self.drop_path(self.gamma_2 * self.mlp(self.norm2(x)))
+        return x
+
+
+class PatchEmbed(nn.Module):
+    """ Image to Patch Embedding
+    """
+    def __init__(self, img_size=224, patch_size=16, in_chans=3, embed_dim=768):
+        super().__init__()
+        img_size = to_2tuple(img_size)
+        patch_size = to_2tuple(patch_size)
+        num_patches = (img_size[1] // patch_size[1]) * (img_size[0] // patch_size[0])
+        self.patch_shape = (img_size[0] // patch_size[0], img_size[1] // patch_size[1])
+        self.img_size = img_size
+        self.patch_size = patch_size
+        self.num_patches = num_patches
+
+        self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size)
+
+    def forward(self, x, **kwargs):
+        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).flatten(2).transpose(1, 2)
+        return x
+
+
+class RelativePositionBias(nn.Module):
+
+    def __init__(self, window_size, num_heads):
+        super().__init__()
+        self.window_size = window_size
+        self.num_relative_distance = (2 * window_size[0] - 1) * (2 * window_size[1] - 1) + 3
+        self.relative_position_bias_table = nn.Parameter(
+            torch.zeros(self.num_relative_distance, num_heads))  # 2*Wh-1 * 2*Ww-1, nH
+        # cls to token & token 2 cls & cls to cls
+
+        # get pair-wise relative position index for each token inside the window
+        coords_h = torch.arange(window_size[0])
+        coords_w = torch.arange(window_size[1])
+        coords = torch.stack(torch.meshgrid([coords_h, coords_w]))  # 2, Wh, Ww
+        coords_flatten = torch.flatten(coords, 1)  # 2, Wh*Ww
+        relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :]  # 2, Wh*Ww, Wh*Ww
+        relative_coords = relative_coords.permute(1, 2, 0).contiguous()  # Wh*Ww, Wh*Ww, 2
+        relative_coords[:, :, 0] += window_size[0] - 1  # shift to start from 0
+        relative_coords[:, :, 1] += window_size[1] - 1
+        relative_coords[:, :, 0] *= 2 * window_size[1] - 1
+        relative_position_index = \
+            torch.zeros(size=(window_size[0] * window_size[1] + 1,) * 2, dtype=relative_coords.dtype)
+        relative_position_index[1:, 1:] = relative_coords.sum(-1)  # Wh*Ww, Wh*Ww
+        relative_position_index[0, 0:] = self.num_relative_distance - 3
+        relative_position_index[0:, 0] = self.num_relative_distance - 2
+        relative_position_index[0, 0] = self.num_relative_distance - 1
+
+        self.register_buffer("relative_position_index", relative_position_index)
+
+    def forward(self):
+        relative_position_bias = \
+            self.relative_position_bias_table[self.relative_position_index.view(-1)].view(
+                self.window_size[0] * self.window_size[1] + 1,
+                self.window_size[0] * self.window_size[1] + 1, -1)  # Wh*Ww,Wh*Ww,nH
+        return relative_position_bias.permute(2, 0, 1).contiguous()  # nH, Wh*Ww, Wh*Ww
+
+
+class EVAVisionTransformer(nn.Module):
+    """ Vision Transformer with support for patch or hybrid CNN input stage
+    """
+    def __init__(self, img_size=224, patch_size=16, in_chans=3, num_classes=1000, embed_dim=768, depth=12,
+                 num_heads=12, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop_rate=0., attn_drop_rate=0.,
+                 drop_path_rate=0., norm_layer=nn.LayerNorm, init_values=None, patch_dropout=0.,
+                 use_abs_pos_emb=True, use_rel_pos_bias=False, use_shared_rel_pos_bias=False, rope=False,
+                 use_mean_pooling=True, init_scale=0.001, grad_checkpointing=False, xattn=False, postnorm=False,
+                 pt_hw_seq_len=16, intp_freq=False, naiveswiglu=False, subln=False):
+        super().__init__()
+
+        if not XFORMERS_IS_AVAILBLE:
+            xattn = False
+
+        self.image_size = img_size
+        self.num_classes = num_classes
+        self.num_features = self.embed_dim = embed_dim  # num_features for consistency with other models
+
+        self.patch_embed = PatchEmbed(
+            img_size=img_size, patch_size=patch_size, in_chans=in_chans, embed_dim=embed_dim)
+        num_patches = self.patch_embed.num_patches
+
+        self.cls_token = nn.Parameter(torch.zeros(1, 1, embed_dim))
+        # self.mask_token = nn.Parameter(torch.zeros(1, 1, embed_dim))
+        if use_abs_pos_emb:
+            self.pos_embed = nn.Parameter(torch.zeros(1, num_patches + 1, embed_dim))
+        else:
+            self.pos_embed = None
+        self.pos_drop = nn.Dropout(p=drop_rate)
+
+        if use_shared_rel_pos_bias:
+            self.rel_pos_bias = RelativePositionBias(window_size=self.patch_embed.patch_shape, num_heads=num_heads)
+        else:
+            self.rel_pos_bias = None
+
+        if rope:
+            half_head_dim = embed_dim // num_heads // 2
+            hw_seq_len = img_size // patch_size
+            self.rope = VisionRotaryEmbeddingFast(
+                dim=half_head_dim,
+                pt_seq_len=pt_hw_seq_len,
+                ft_seq_len=hw_seq_len if intp_freq else None,
+                # patch_dropout=patch_dropout
+            )
+        else: 
+            self.rope = None
+
+        self.naiveswiglu = naiveswiglu
+
+        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, depth)]  # stochastic depth decay rule
+        self.use_rel_pos_bias = use_rel_pos_bias
+        self.blocks = nn.ModuleList([
+            Block(
+                dim=embed_dim, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, qk_scale=qk_scale,
+                drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[i], norm_layer=norm_layer,
+                init_values=init_values, window_size=self.patch_embed.patch_shape if use_rel_pos_bias else None,
+                xattn=xattn, rope=self.rope, postnorm=postnorm, subln=subln, naiveswiglu=naiveswiglu)
+            for i in range(depth)])
+        self.norm = nn.Identity() if use_mean_pooling else norm_layer(embed_dim)
+        self.fc_norm = norm_layer(embed_dim) if use_mean_pooling else None
+        self.head = nn.Linear(embed_dim, num_classes) if num_classes > 0 else nn.Identity()
+
+        if self.pos_embed is not None:
+            trunc_normal_(self.pos_embed, std=.02)
+
+        trunc_normal_(self.cls_token, std=.02)
+        # trunc_normal_(self.mask_token, std=.02)
+
+        self.apply(self._init_weights)
+        self.fix_init_weight()
+
+        if isinstance(self.head, nn.Linear):
+            trunc_normal_(self.head.weight, std=.02)
+            self.head.weight.data.mul_(init_scale)
+            self.head.bias.data.mul_(init_scale)
+
+        # setting a patch_dropout of 0. would mean it is disabled and this function would be the identity fn
+        self.patch_dropout = PatchDropout(patch_dropout) if patch_dropout > 0. else nn.Identity()
+
+        self.grad_checkpointing = grad_checkpointing
+
+    def fix_init_weight(self):
+        def rescale(param, layer_id):
+            param.div_(math.sqrt(2.0 * layer_id))
+
+        for layer_id, layer in enumerate(self.blocks):
+            rescale(layer.attn.proj.weight.data, layer_id + 1)
+            if self.naiveswiglu:
+                rescale(layer.mlp.w3.weight.data, layer_id + 1)
+            else:
+                rescale(layer.mlp.fc2.weight.data, layer_id + 1)
+
+    def get_cast_dtype(self) -> torch.dtype:
+        return self.blocks[0].mlp.fc2.weight.dtype
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=.02)
+            if 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 get_num_layers(self):
+        return len(self.blocks)
+    
+    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
+
+    @torch.jit.ignore
+    def set_grad_checkpointing(self, enable=True):
+        self.grad_checkpointing = enable
+
+    @torch.jit.ignore
+    def no_weight_decay(self):
+        return {'pos_embed', 'cls_token'}
+
+    def get_classifier(self):
+        return self.head
+
+    def reset_classifier(self, num_classes, global_pool=''):
+        self.num_classes = num_classes
+        self.head = nn.Linear(self.embed_dim, num_classes) if num_classes > 0 else nn.Identity()
+
+    def forward_features(self, x, return_all_features=False, return_hidden=False, shuffle=False):
+        
+        x = self.patch_embed(x)
+        batch_size, seq_len, _ = x.size()
+
+        if shuffle:
+            idx = torch.randperm(x.shape[1]) + 1
+            zero = torch.LongTensor([0, ])
+            idx = torch.cat([zero, idx])
+            pos_embed = self.pos_embed[:, idx]
+
+        cls_tokens = self.cls_token.expand(batch_size, -1, -1)  # stole cls_tokens impl from Phil Wang, thanks
+        x = torch.cat((cls_tokens, x), dim=1)
+        if shuffle:
+            x = x + pos_embed
+        elif self.pos_embed is not None:
+            x = x + self.pos_embed
+        x = self.pos_drop(x)
+
+        # a patch_dropout of 0. would mean it is disabled and this function would do nothing but return what was passed in
+        if os.getenv('RoPE') == '1':
+            if self.training and not isinstance(self.patch_dropout, nn.Identity):
+                x, patch_indices_keep = self.patch_dropout(x)
+                self.rope.forward = partial(self.rope.forward, patch_indices_keep=patch_indices_keep)
+            else:
+                self.rope.forward = partial(self.rope.forward, patch_indices_keep=None)
+                x = self.patch_dropout(x)
+        else:
+            x = self.patch_dropout(x)
+
+        rel_pos_bias = self.rel_pos_bias() if self.rel_pos_bias is not None else None
+        hidden_states = []
+        for idx, blk in enumerate(self.blocks):
+            if (0 < idx <= 20) and (idx % 4 == 0) and return_hidden:
+                hidden_states.append(x)
+            if self.grad_checkpointing:
+                x = checkpoint(blk, x, (rel_pos_bias,))
+            else:
+                x = blk(x, rel_pos_bias=rel_pos_bias)
+
+        if not return_all_features:
+            x = self.norm(x)
+            if self.fc_norm is not None:
+                return self.fc_norm(x.mean(1)), hidden_states
+            else:
+                return x[:, 0], hidden_states
+        return x
+
+    def forward(self, x, return_all_features=False, return_hidden=False, shuffle=False):
+        if return_all_features:
+            return self.forward_features(x, return_all_features, return_hidden, shuffle)
+        x, hidden_states = self.forward_features(x, return_all_features, return_hidden, shuffle)
+        x = self.head(x)
+        if return_hidden:
+            return x, hidden_states
+        return x

eva_clip/factory.py

@@ -0,0 +1,517 @@
+import json
+import logging
+import os
+import pathlib
+import re
+from copy import deepcopy
+from pathlib import Path
+from typing import Optional, Tuple, Union, Dict, Any
+import torch
+
+from .constants import OPENAI_DATASET_MEAN, OPENAI_DATASET_STD
+from .model import CLIP, CustomCLIP, convert_weights_to_lp, convert_to_custom_text_state_dict,\
+    get_cast_dtype
+from .openai import load_openai_model
+from .pretrained import is_pretrained_cfg, get_pretrained_cfg, download_pretrained, list_pretrained_tags_by_model
+from .transform import image_transform
+from .tokenizer import HFTokenizer, tokenize
+from .utils import resize_clip_pos_embed, resize_evaclip_pos_embed, resize_visual_pos_embed, resize_eva_pos_embed
+
+
+_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:
+        with open(cf, "r", encoding="utf8") as f:
+            model_cfg = json.load(f)
+            if all(a in model_cfg for a in ('embed_dim', 'vision_cfg', 'text_cfg')):
+                _MODEL_CONFIGS[cf.stem] = model_cfg
+
+    _MODEL_CONFIGS = dict(sorted(_MODEL_CONFIGS.items(), key=lambda x: _natural_key(x[0])))
+
+
+_rescan_model_configs()  # initial populate of model config registry
+
+
+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()
+
+
+def get_model_config(model_name):
+    if model_name in _MODEL_CONFIGS:
+        return deepcopy(_MODEL_CONFIGS[model_name])
+    else:
+        return None
+
+
+def get_tokenizer(model_name):
+    config = get_model_config(model_name)
+    tokenizer = HFTokenizer(config['text_cfg']['hf_tokenizer_name']) if 'hf_tokenizer_name' in config['text_cfg'] else tokenize
+    return tokenizer
+
+
+# loading openai CLIP weights when is_openai=True for training
+def load_state_dict(checkpoint_path: str, map_location: str='cpu', model_key: str='model|module|state_dict', is_openai: bool=False, skip_list: list=[]):
+    if is_openai:
+        model = torch.jit.load(checkpoint_path, map_location="cpu").eval()
+        state_dict = model.state_dict()
+        for key in ["input_resolution", "context_length", "vocab_size"]:
+            state_dict.pop(key, None)
+    else:
+        checkpoint = torch.load(checkpoint_path, map_location=map_location)
+        for mk in model_key.split('|'):
+            if isinstance(checkpoint, dict) and mk in checkpoint:
+                state_dict = checkpoint[mk]
+                break
+            else:
+                state_dict = checkpoint
+        if next(iter(state_dict.items()))[0].startswith('module'):
+            state_dict = {k[7:]: v for k, v in state_dict.items()}
+    
+    for k in skip_list:
+        if k in list(state_dict.keys()):
+            logging.info(f"Removing key {k} from pretrained checkpoint")
+            del state_dict[k]
+
+    if os.getenv('RoPE') == '1':
+        for k in list(state_dict.keys()):
+            if 'freqs_cos' in k or 'freqs_sin' in k:
+                del state_dict[k]
+    return state_dict
+
+
+
+def load_checkpoint(model, checkpoint_path, model_key="model|module|state_dict", strict=True):
+    state_dict = load_state_dict(checkpoint_path, model_key=model_key, is_openai=False)
+    # detect old format and make compatible with new format
+    if 'positional_embedding' in state_dict and not hasattr(model, 'positional_embedding'):
+        state_dict = convert_to_custom_text_state_dict(state_dict)
+    if 'text.logit_scale' in state_dict and hasattr(model, 'logit_scale'):
+        state_dict['logit_scale'] = state_dict['text.logit_scale']
+        del state_dict['text.logit_scale']
+
+    # resize_clip_pos_embed for CLIP and open CLIP
+    if 'visual.positional_embedding' in state_dict:
+        resize_clip_pos_embed(state_dict, model)
+    # specified to eva_vit_model
+    elif 'visual.pos_embed' in state_dict:
+        resize_evaclip_pos_embed(state_dict, model)
+
+    # resize_clip_pos_embed(state_dict, model)
+    incompatible_keys = model.load_state_dict(state_dict, strict=strict)
+    logging.info(f"incompatible_keys.missing_keys: {incompatible_keys.missing_keys}")
+    return incompatible_keys
+
+def load_clip_visual_state_dict(checkpoint_path: str, map_location: str='cpu', is_openai: bool=False, skip_list:list=[]):
+    state_dict = load_state_dict(checkpoint_path, map_location=map_location, is_openai=is_openai, skip_list=skip_list)
+
+    for k in list(state_dict.keys()):
+        if not k.startswith('visual.'):
+            del state_dict[k]
+    for k in list(state_dict.keys()):
+        if k.startswith('visual.'):
+            new_k = k[7:]
+            state_dict[new_k] = state_dict[k]
+            del state_dict[k]
+    return state_dict
+
+def load_clip_text_state_dict(checkpoint_path: str, map_location: str='cpu', is_openai: bool=False, skip_list:list=[]):
+    state_dict = load_state_dict(checkpoint_path, map_location=map_location, is_openai=is_openai, skip_list=skip_list)
+
+    for k in list(state_dict.keys()):
+        if k.startswith('visual.'):
+            del state_dict[k]
+    return state_dict
+
+def get_pretrained_tag(pretrained_model):
+    pretrained_model = pretrained_model.lower()
+    if "laion" in pretrained_model or "open_clip" in pretrained_model:
+        return "open_clip"
+    elif "openai" in pretrained_model:
+        return "clip"
+    elif "eva" in pretrained_model and "clip" in pretrained_model:
+        return "eva_clip"
+    else:
+        return "other"
+
+def load_pretrained_checkpoint(
+        model,
+        visual_checkpoint_path,
+        text_checkpoint_path,
+        strict=True,
+        visual_model=None,
+        text_model=None,
+        model_key="model|module|state_dict",
+        skip_list=[]):
+    visual_tag = get_pretrained_tag(visual_model)
+    text_tag = get_pretrained_tag(text_model)
+
+    logging.info(f"num of model state_dict keys: {len(model.state_dict().keys())}")
+    visual_incompatible_keys, text_incompatible_keys = None, None
+    if visual_checkpoint_path:
+        if visual_tag == "eva_clip" or visual_tag == "open_clip":
+            visual_state_dict = load_clip_visual_state_dict(visual_checkpoint_path, is_openai=False, skip_list=skip_list)
+        elif visual_tag == "clip":
+            visual_state_dict = load_clip_visual_state_dict(visual_checkpoint_path, is_openai=True, skip_list=skip_list)
+        else:
+            visual_state_dict = load_state_dict(visual_checkpoint_path, model_key=model_key, is_openai=False, skip_list=skip_list)
+    
+        # resize_clip_pos_embed for CLIP and open CLIP
+        if 'positional_embedding' in visual_state_dict:
+            resize_visual_pos_embed(visual_state_dict, model)
+        # specified to EVA model
+        elif 'pos_embed' in visual_state_dict:
+            resize_eva_pos_embed(visual_state_dict, model)
+
+        visual_incompatible_keys = model.visual.load_state_dict(visual_state_dict, strict=strict)
+        logging.info(f"num of loaded visual_state_dict keys: {len(visual_state_dict.keys())}")
+        logging.info(f"visual_incompatible_keys.missing_keys: {visual_incompatible_keys.missing_keys}")
+
+    if text_checkpoint_path:
+        if text_tag == "eva_clip" or text_tag == "open_clip":
+            text_state_dict = load_clip_text_state_dict(text_checkpoint_path, is_openai=False, skip_list=skip_list)
+        elif text_tag == "clip":
+            text_state_dict = load_clip_text_state_dict(text_checkpoint_path, is_openai=True, skip_list=skip_list)
+        else:
+            text_state_dict = load_state_dict(visual_checkpoint_path, model_key=model_key, is_openai=False, skip_list=skip_list)
+
+        text_incompatible_keys = model.text.load_state_dict(text_state_dict, strict=strict)
+        
+        logging.info(f"num of loaded text_state_dict keys: {len(text_state_dict.keys())}")
+        logging.info(f"text_incompatible_keys.missing_keys: {text_incompatible_keys.missing_keys}")
+
+    return visual_incompatible_keys, text_incompatible_keys
+
+def create_model(
+        model_name: str,
+        pretrained: Optional[str] = None,
+        precision: str = 'fp32',
+        device: Union[str, torch.device] = 'cpu',
+        jit: bool = False,
+        force_quick_gelu: bool = False,
+        force_custom_clip: bool = False,
+        force_patch_dropout: Optional[float] = None,
+        pretrained_image: str = '',
+        pretrained_text: str = '',
+        pretrained_hf: bool = True,
+        pretrained_visual_model: str = None,
+        pretrained_text_model: str = None,
+        cache_dir: Optional[str] = None,
+        skip_list: list  = [],
+):
+    model_name = model_name.replace('/', '-')  # for callers using old naming with / in ViT names
+    if isinstance(device, str):
+        device = torch.device(device)
+
+    if pretrained and pretrained.lower() == 'openai':
+        logging.info(f'Loading pretrained {model_name} from OpenAI.')
+        model = load_openai_model(
+            model_name,
+            precision=precision,
+            device=device,
+            jit=jit,
+            cache_dir=cache_dir,
+        )
+    else:
+        model_cfg = get_model_config(model_name)
+        if model_cfg is not None:
+            logging.info(f'Loaded {model_name} model config.')
+        else:
+            logging.error(f'Model config for {model_name} not found; available models {list_models()}.')
+            raise RuntimeError(f'Model config for {model_name} not found.')
+
+        if 'rope' in model_cfg.get('vision_cfg', {}):
+            if model_cfg['vision_cfg']['rope']:
+                os.environ['RoPE'] = "1"
+        else:
+            os.environ['RoPE'] = "0"
+
+        if force_quick_gelu:
+            # override for use of QuickGELU on non-OpenAI transformer models
+            model_cfg["quick_gelu"] = True
+        
+        if force_patch_dropout is not None:
+            # override the default patch dropout value
+            model_cfg['vision_cfg']["patch_dropout"] = force_patch_dropout
+
+        cast_dtype = get_cast_dtype(precision)
+        custom_clip = model_cfg.pop('custom_text', False) or force_custom_clip or ('hf_model_name' in model_cfg['text_cfg'])
+
+
+        if custom_clip:
+            if 'hf_model_name' in model_cfg.get('text_cfg', {}):
+                model_cfg['text_cfg']['hf_model_pretrained'] = pretrained_hf
+            model = CustomCLIP(**model_cfg, cast_dtype=cast_dtype)
+        else:
+            model = CLIP(**model_cfg, cast_dtype=cast_dtype)
+
+        pretrained_cfg = {}
+        if pretrained:
+            checkpoint_path = ''
+            pretrained_cfg = get_pretrained_cfg(model_name, pretrained)
+            if pretrained_cfg:
+                checkpoint_path = download_pretrained(pretrained_cfg, cache_dir=cache_dir)
+            elif os.path.exists(pretrained):
+                checkpoint_path = pretrained
+
+            if checkpoint_path:
+                logging.info(f'Loading pretrained {model_name} weights ({pretrained}).')
+                load_checkpoint(model,
+                               checkpoint_path,
+                               model_key="model|module|state_dict",
+                               strict=False
+                               ) 
+            else:
+                error_str = (
+                    f'Pretrained weights ({pretrained}) not found for model {model_name}.'
+                    f'Available pretrained tags ({list_pretrained_tags_by_model(model_name)}.')
+                logging.warning(error_str)
+                raise RuntimeError(error_str)
+        else:
+            visual_checkpoint_path = ''
+            text_checkpoint_path = ''
+            
+            if pretrained_image:
+                pretrained_visual_model = pretrained_visual_model.replace('/', '-')  # for callers using old naming with / in ViT names
+                pretrained_image_cfg = get_pretrained_cfg(pretrained_visual_model, pretrained_image)
+                if 'timm_model_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
+                elif pretrained_image_cfg:
+                    visual_checkpoint_path = download_pretrained(pretrained_image_cfg, cache_dir=cache_dir)
+                elif os.path.exists(pretrained_image):
+                    visual_checkpoint_path = pretrained_image
+                else:
+                    logging.warning(f'Pretrained weights ({visual_checkpoint_path}) not found for model {model_name}.visual.')
+                    raise RuntimeError(f'Pretrained weights ({visual_checkpoint_path}) not found for model {model_name}.visual.')
+
+            if pretrained_text:
+                pretrained_text_model = pretrained_text_model.replace('/', '-')  # for callers using old naming with / in ViT names
+                pretrained_text_cfg = get_pretrained_cfg(pretrained_text_model, pretrained_text)
+                if pretrained_image_cfg:
+                    text_checkpoint_path = download_pretrained(pretrained_text_cfg, cache_dir=cache_dir)
+                elif os.path.exists(pretrained_text):
+                    text_checkpoint_path = pretrained_text
+                else:
+                    logging.warning(f'Pretrained weights ({text_checkpoint_path}) not found for model {model_name}.text.')
+                    raise RuntimeError(f'Pretrained weights ({text_checkpoint_path}) not found for model {model_name}.text.')
+            
+            if visual_checkpoint_path:
+                logging.info(f'Loading pretrained {model_name}.visual weights ({visual_checkpoint_path}).')
+            if text_checkpoint_path:
+                logging.info(f'Loading pretrained {model_name}.text weights ({text_checkpoint_path}).')
+
+            if visual_checkpoint_path or text_checkpoint_path:
+                load_pretrained_checkpoint(
+                    model,
+                    visual_checkpoint_path,
+                    text_checkpoint_path,
+                    strict=False,
+                    visual_model=pretrained_visual_model,
+                    text_model=pretrained_text_model,
+                    model_key="model|module|state_dict",
+                    skip_list=skip_list
+                )
+        
+        if "fp16" in precision or "bf16" in precision:
+            logging.info(f'convert precision to {precision}')
+            model = model.to(torch.bfloat16) if 'bf16' in precision else model.to(torch.float16)
+
+        model.to(device=device)
+
+        # set image / mean metadata from pretrained_cfg if available, or use default
+        model.visual.image_mean = pretrained_cfg.get('mean', None) or OPENAI_DATASET_MEAN
+        model.visual.image_std = pretrained_cfg.get('std', None) or OPENAI_DATASET_STD
+
+        if jit:
+            model = torch.jit.script(model)
+
+    return model
+
+
+def create_model_and_transforms(
+        model_name: str,
+        pretrained: Optional[str] = None,
+        precision: str = 'fp32',
+        device: Union[str, torch.device] = 'cpu',
+        jit: bool = False,
+        force_quick_gelu: bool = False,
+        force_custom_clip: bool = False,
+        force_patch_dropout: Optional[float] = None,
+        pretrained_image: str = '',
+        pretrained_text: str = '',
+        pretrained_hf: bool = True,
+        pretrained_visual_model: str = None,
+        pretrained_text_model: str = None,
+        image_mean: Optional[Tuple[float, ...]] = None,
+        image_std: Optional[Tuple[float, ...]] = None,
+        cache_dir: Optional[str] = None,
+        skip_list: list = [],
+):
+    model = create_model(
+        model_name,
+        pretrained,
+        precision=precision,
+        device=device,
+        jit=jit,
+        force_quick_gelu=force_quick_gelu,
+        force_custom_clip=force_custom_clip,
+        force_patch_dropout=force_patch_dropout,
+        pretrained_image=pretrained_image,
+        pretrained_text=pretrained_text,
+        pretrained_hf=pretrained_hf,
+        pretrained_visual_model=pretrained_visual_model,
+        pretrained_text_model=pretrained_text_model,
+        cache_dir=cache_dir,
+        skip_list=skip_list,
+    )
+
+    image_mean = image_mean or getattr(model.visual, 'image_mean', None)
+    image_std = image_std or getattr(model.visual, 'image_std', None)
+    preprocess_train = image_transform(
+        model.visual.image_size,
+        is_train=True,
+        mean=image_mean,
+        std=image_std
+    )
+    preprocess_val = image_transform(
+        model.visual.image_size,
+        is_train=False,
+        mean=image_mean,
+        std=image_std
+    )
+
+    return model, preprocess_train, preprocess_val
+
+
+def create_transforms(
+        model_name: str,
+        pretrained: Optional[str] = None,
+        precision: str = 'fp32',
+        device: Union[str, torch.device] = 'cpu',
+        jit: bool = False,
+        force_quick_gelu: bool = False,
+        force_custom_clip: bool = False,
+        force_patch_dropout: Optional[float] = None,
+        pretrained_image: str = '',
+        pretrained_text: str = '',
+        pretrained_hf: bool = True,
+        pretrained_visual_model: str = None,
+        pretrained_text_model: str = None,
+        image_mean: Optional[Tuple[float, ...]] = None,
+        image_std: Optional[Tuple[float, ...]] = None,
+        cache_dir: Optional[str] = None,
+        skip_list: list = [],
+):
+    model = create_model(
+        model_name,
+        pretrained,
+        precision=precision,
+        device=device,
+        jit=jit,
+        force_quick_gelu=force_quick_gelu,
+        force_custom_clip=force_custom_clip,
+        force_patch_dropout=force_patch_dropout,
+        pretrained_image=pretrained_image,
+        pretrained_text=pretrained_text,
+        pretrained_hf=pretrained_hf,
+        pretrained_visual_model=pretrained_visual_model,
+        pretrained_text_model=pretrained_text_model,
+        cache_dir=cache_dir,
+        skip_list=skip_list,
+    )
+
+
+    image_mean = image_mean or getattr(model.visual, 'image_mean', None)
+    image_std = image_std or getattr(model.visual, 'image_std', None)
+    preprocess_train = image_transform(
+        model.visual.image_size,
+        is_train=True,
+        mean=image_mean,
+        std=image_std
+    )
+    preprocess_val = image_transform(
+        model.visual.image_size,
+        is_train=False,
+        mean=image_mean,
+        std=image_std
+    )
+    del model
+
+    return preprocess_train, preprocess_val
+
+def create_model_from_pretrained(
+        model_name: str,
+        pretrained: str,
+        precision: str = 'fp32',
+        device: Union[str, torch.device] = 'cpu',
+        jit: bool = False,
+        force_quick_gelu: bool = False,
+        force_custom_clip: bool = False,
+        force_patch_dropout: Optional[float] = None,
+        return_transform: bool = True,
+        image_mean: Optional[Tuple[float, ...]] = None,
+        image_std: Optional[Tuple[float, ...]] = None,
+        cache_dir: Optional[str] = None,
+        is_frozen: bool = False,
+):
+    if not is_pretrained_cfg(model_name, pretrained) and not os.path.exists(pretrained):
+        raise RuntimeError(
+            f'{pretrained} is not a valid pretrained cfg or checkpoint for {model_name}.'
+            f' Use open_clip.list_pretrained() to find one.')
+
+    model = create_model(
+        model_name,
+        pretrained,
+        precision=precision,
+        device=device,
+        jit=jit,
+        force_quick_gelu=force_quick_gelu,
+        force_custom_clip=force_custom_clip,
+        force_patch_dropout=force_patch_dropout,
+        cache_dir=cache_dir,
+    )
+
+    if is_frozen:
+        for param in model.parameters():
+            param.requires_grad = False
+
+    if not return_transform:
+        return model
+
+    image_mean = image_mean or getattr(model.visual, 'image_mean', None)
+    image_std = image_std or getattr(model.visual, 'image_std', None)
+    preprocess = image_transform(
+        model.visual.image_size,
+        is_train=False,
+        mean=image_mean,
+        std=image_std
+    )
+
+    return model, preprocess

eva_clip/hf_configs.py

@@ -0,0 +1,57 @@
+# HF architecture dict:
+arch_dict = {
+  # https://huggingface.co/docs/transformers/model_doc/roberta#roberta
+  "roberta": {
+      "config_names": {
+          "context_length": "max_position_embeddings",
+          "vocab_size": "vocab_size",
+          "width": "hidden_size",
+          "heads": "num_attention_heads",
+          "layers": "num_hidden_layers",
+          "layer_attr": "layer",
+          "token_embeddings_attr": "embeddings"
+      },
+      "pooler": "mean_pooler",
+  },
+  # https://huggingface.co/docs/transformers/model_doc/xlm-roberta#transformers.XLMRobertaConfig
+  "xlm-roberta": {
+      "config_names": {
+          "context_length": "max_position_embeddings",
+          "vocab_size": "vocab_size",
+          "width": "hidden_size",
+          "heads": "num_attention_heads",
+          "layers": "num_hidden_layers",
+          "layer_attr": "layer",
+          "token_embeddings_attr": "embeddings"
+      },
+      "pooler": "mean_pooler",
+  },
+  # https://huggingface.co/docs/transformers/model_doc/mt5#mt5
+  "mt5": {
+      "config_names": {
+          # unlimited seqlen
+          # https://github.com/google-research/text-to-text-transfer-transformer/issues/273
+          # https://github.com/huggingface/transformers/blob/v4.24.0/src/transformers/models/t5/modeling_t5.py#L374
+          "context_length": "",
+          "vocab_size": "vocab_size",
+          "width": "d_model",
+          "heads": "num_heads",
+          "layers": "num_layers",
+          "layer_attr": "block",
+          "token_embeddings_attr": "embed_tokens"
+      },
+      "pooler": "mean_pooler",
+  },
+  "bert": {
+    "config_names": {
+      "context_length": "max_position_embeddings",
+      "vocab_size": "vocab_size",
+      "width": "hidden_size",
+      "heads": "num_attention_heads",
+      "layers": "num_hidden_layers",
+      "layer_attr": "layer",
+      "token_embeddings_attr": "embeddings"
+    },
+    "pooler": "mean_pooler",
+  }
+}

eva_clip/hf_model.py

@@ -0,0 +1,248 @@
+""" huggingface model adapter
+
+Wraps HuggingFace transformers (https://github.com/huggingface/transformers) models for use as a text tower in CLIP model.
+"""
+
+import re
+
+import torch
+import torch.nn as nn
+from torch.nn import functional as F
+from torch import TensorType
+try:
+    import transformers
+    from transformers import AutoModel, AutoModelForMaskedLM, AutoTokenizer, AutoConfig, PretrainedConfig
+    from transformers.modeling_outputs import BaseModelOutput, BaseModelOutputWithPooling, \
+        BaseModelOutputWithPoolingAndCrossAttentions
+except ImportError as e:
+    transformers = None
+
+
+    class BaseModelOutput:
+        pass
+
+
+    class PretrainedConfig:
+        pass
+
+from .hf_configs import arch_dict
+
+# utils
+def _camel2snake(s):
+    return re.sub(r'(?<!^)(?=[A-Z])', '_', s).lower()
+
+# TODO: ?last - for gpt-like models
+_POOLERS = {}
+
+def register_pooler(cls):
+    """Decorator registering pooler class"""
+    _POOLERS[_camel2snake(cls.__name__)] = cls
+    return cls
+
+
+@register_pooler
+class MeanPooler(nn.Module):
+    """Mean pooling"""
+    def forward(self, x:BaseModelOutput, attention_mask:TensorType):
+        masked_output = x.last_hidden_state * attention_mask.unsqueeze(-1)
+        return masked_output.sum(dim=1) / attention_mask.sum(-1, keepdim=True)
+
+@register_pooler
+class MaxPooler(nn.Module):
+    """Max pooling"""
+    def forward(self, x:BaseModelOutput, attention_mask:TensorType):
+        masked_output = x.last_hidden_state.masked_fill(attention_mask.unsqueeze(-1), -torch.inf)
+        return masked_output.max(1).values
+
+@register_pooler
+class ClsPooler(nn.Module):
+    """CLS token pooling"""
+    def __init__(self, use_pooler_output=True):
+        super().__init__()
+        self.cls_token_position = 0
+        self.use_pooler_output = use_pooler_output
+
+    def forward(self, x:BaseModelOutput, attention_mask:TensorType):
+        
+        if (self.use_pooler_output and 
+            isinstance(x, (BaseModelOutputWithPooling, BaseModelOutputWithPoolingAndCrossAttentions)) and
+            (x.pooler_output is not None)
+            ):
+            return x.pooler_output
+        
+        return x.last_hidden_state[:, self.cls_token_position, :]
+
+class HFTextEncoder(nn.Module):
+    """HuggingFace model adapter"""
+    def __init__(
+            self, 
+            model_name_or_path: str,
+            output_dim: int,
+            tokenizer_name: str = None,
+            config: PretrainedConfig = None,
+            pooler_type: str = None,
+            proj: str = None,
+            pretrained: bool = True,
+            masked_language_modeling: bool = False):
+        super().__init__()
+
+        self.output_dim = output_dim
+
+        # TODO: find better way to get this information
+        uses_transformer_pooler = (pooler_type == "cls_pooler")
+
+        if transformers is None:
+            raise RuntimeError("Please `pip install transformers` to use pre-trained HuggingFace models")
+        if config is None:
+            self.config = AutoConfig.from_pretrained(model_name_or_path)
+            if masked_language_modeling:
+                create_func, model_args = (AutoModelForMaskedLM.from_pretrained, model_name_or_path) if pretrained else (
+                    AutoModelForMaskedLM.from_config, self.config)
+            else:
+                create_func, model_args = (AutoModel.from_pretrained, model_name_or_path) if pretrained else (
+                    AutoModel.from_config, self.config)
+            # TODO: do all model configs have this attribute? PretrainedConfig does so yes??
+            if hasattr(self.config, "is_encoder_decoder") and self.config.is_encoder_decoder:
+                self.transformer = create_func(model_args)
+                self.transformer = self.transformer.encoder
+            else:
+                self.transformer = create_func(model_args, add_pooling_layer=uses_transformer_pooler)
+        else:
+            self.config = config
+            if masked_language_modeling:
+                self.transformer = AutoModelForMaskedLM.from_config(config)
+            else:
+                self.transformer = AutoModel.from_config(config)
+
+        if pooler_type is None: # get default arch pooler
+            self.pooler = _POOLERS[(arch_dict[self.config.model_type]["pooler"])]()
+        else:
+            self.pooler = _POOLERS[pooler_type]()
+
+        d_model = getattr(self.config, arch_dict[self.config.model_type]["config_names"]["width"])
+        if (d_model == output_dim) and (proj is None): # do we always need a proj?
+            self.proj = nn.Identity()
+        elif proj == 'linear':
+            self.proj = nn.Linear(d_model, output_dim, bias=False)
+        elif proj == 'mlp':
+            hidden_size = (d_model + output_dim) // 2
+            self.proj = nn.Sequential(
+                nn.Linear(d_model, hidden_size, bias=False),
+                nn.GELU(),
+                nn.Linear(hidden_size, output_dim, bias=False),
+            )
+
+        # self.itm_proj = nn.Linear(d_model, 2, bias=False)
+        # self.mlm_proj = nn.Linear(d_model, self.config.vocab_size), bias=False)
+        self.tokenizer = AutoTokenizer.from_pretrained(tokenizer_name)
+
+    # def forward_itm(self, x:TensorType, image_embeds:TensorType) -> TensorType:
+    #     image_atts = torch.ones(image_embeds.size()[:-1],dtype=torch.long).to(x.device)  
+    #     attn_mask = (x != self.config.pad_token_id).long()
+    #     out = self.transformer(
+    #         input_ids=x, 
+    #         attention_mask=attn_mask,
+    #         encoder_hidden_states = image_embeds,
+    #         encoder_attention_mask = image_atts,
+    #         )
+    #     pooled_out = self.pooler(out, attn_mask)
+
+    #     return self.itm_proj(pooled_out)
+
+    def mask(self, input_ids, vocab_size, device, targets=None, masked_indices=None, probability_matrix=None):
+        if masked_indices is None:                                       
+            masked_indices = torch.bernoulli(probability_matrix).bool()
+                                               
+        masked_indices[input_ids == self.tokenizer.pad_token_id] = False
+        masked_indices[input_ids == self.tokenizer.cls_token_id] = False
+        
+        if targets is not None:
+            targets[~masked_indices] = -100 # We only compute loss on masked tokens            
+
+        # 80% of the time, we replace masked input tokens with tokenizer.mask_token ([MASK])
+        indices_replaced = torch.bernoulli(torch.full(input_ids.shape, 0.8)).bool() & masked_indices
+        input_ids[indices_replaced] = self.tokenizer.mask_token_id
+
+        # 10% of the time, we replace masked input tokens with random word
+        indices_random = torch.bernoulli(torch.full(input_ids.shape, 0.5)).bool() & masked_indices & ~indices_replaced
+        random_words = torch.randint(vocab_size, input_ids.shape, dtype=torch.long).to(device)
+        input_ids[indices_random] = random_words[indices_random]                     
+        # The rest of the time (10% of the time) we keep the masked input tokens unchanged   
+        
+        if targets is not None:
+            return input_ids, targets
+        else:
+            return input_ids
+
+    def forward_mlm(self, input_ids, image_embeds, mlm_probability=0.25):
+        labels = input_ids.clone()
+        attn_mask = (input_ids != self.config.pad_token_id).long()
+        image_atts = torch.ones(image_embeds.size()[:-1],dtype=torch.long).to(input_ids.device) 
+        vocab_size = getattr(self.config, arch_dict[self.config.model_type]["config_names"]["vocab_size"])
+        probability_matrix = torch.full(labels.shape, mlm_probability)
+        input_ids, labels = self.mask(input_ids, vocab_size, input_ids.device, targets=labels,
+                                      probability_matrix = probability_matrix)
+        mlm_output = self.transformer(input_ids,
+                        attention_mask = attn_mask,
+                        encoder_hidden_states = image_embeds,
+                        encoder_attention_mask = image_atts,
+                        return_dict = True,
+                        labels = labels,
+                    )
+        return mlm_output.loss
+        # mlm_output = self.transformer(input_ids,
+        #                 attention_mask = attn_mask,
+        #                 encoder_hidden_states = image_embeds,
+        #                 encoder_attention_mask = image_atts,
+        #                 return_dict = True,
+        #             ).last_hidden_state
+        # logits = self.mlm_proj(mlm_output)
+
+        # # logits = logits[:, :-1, :].contiguous().view(-1, vocab_size)
+        # logits = logits[:, 1:, :].contiguous().view(-1, vocab_size)
+        # labels = labels[:, 1:].contiguous().view(-1)
+
+        # mlm_loss = F.cross_entropy(
+        #     logits,
+        #     labels,
+        #     # label_smoothing=0.1,
+        # )
+        # return mlm_loss
+
+
+    def forward(self, x:TensorType) -> TensorType:
+        attn_mask = (x != self.config.pad_token_id).long()
+        out = self.transformer(input_ids=x, attention_mask=attn_mask)
+        pooled_out = self.pooler(out, attn_mask)
+
+        return self.proj(pooled_out)
+
+    def lock(self, unlocked_layers:int=0, freeze_layer_norm:bool=True):
+        if not unlocked_layers: # full freezing
+             for n, p in self.transformer.named_parameters():
+                 p.requires_grad = (not freeze_layer_norm) if "LayerNorm" in n.split(".") else False
+             return
+
+        encoder = self.transformer.encoder if hasattr(self.transformer, 'encoder') else self.transformer
+        layer_list = getattr(encoder, arch_dict[self.config.model_type]["config_names"]["layer_attr"])
+        print(f"Unlocking {unlocked_layers}/{len(layer_list) + 1} layers of hf model")
+        embeddings = getattr(
+            self.transformer, arch_dict[self.config.model_type]["config_names"]["token_embeddings_attr"])
+        modules = [embeddings, *layer_list][:-unlocked_layers]
+        # freeze layers
+        for module in modules:
+            for n, p in module.named_parameters():
+                p.requires_grad = (not freeze_layer_norm) if "LayerNorm" in n.split(".") else False
+
+
+    @torch.jit.ignore
+    def set_grad_checkpointing(self, enable=True):
+        self.transformer.gradient_checkpointing_enable()
+
+    def get_num_layers(self):
+        encoder = self.transformer.encoder if hasattr(self.transformer, 'encoder') else self.transformer
+        layer_list = getattr(encoder, arch_dict[self.config.model_type]["config_names"]["layer_attr"])
+        return len(layer_list)
+
+    def init_parameters(self):
+        pass

eva_clip/loss.py

@@ -0,0 +1,138 @@
+import math
+import torch
+import torch.nn as nn
+from torch.nn import functional as F
+
+try:
+    import torch.distributed.nn
+    from torch import distributed as dist
+    has_distributed = True
+except ImportError:
+    has_distributed = False
+
+try:
+    import horovod.torch as hvd
+except ImportError:
+    hvd = None
+
+from timm.loss import LabelSmoothingCrossEntropy
+
+
+def gather_features(
+        image_features,
+        text_features,
+        local_loss=False,
+        gather_with_grad=False,
+        rank=0,
+        world_size=1,
+        use_horovod=False
+):
+    assert has_distributed, 'torch.distributed did not import correctly, please use a PyTorch version with support.'
+    if use_horovod:
+        assert hvd is not None, 'Please install horovod'
+        if gather_with_grad:
+            all_image_features = hvd.allgather(image_features)
+            all_text_features = hvd.allgather(text_features)
+        else:
+            with torch.no_grad():
+                all_image_features = hvd.allgather(image_features)
+                all_text_features = hvd.allgather(text_features)
+            if not local_loss:
+                # ensure grads for local rank when all_* features don't have a gradient
+                gathered_image_features = list(all_image_features.chunk(world_size, dim=0))
+                gathered_text_features = list(all_text_features.chunk(world_size, dim=0))
+                gathered_image_features[rank] = image_features
+                gathered_text_features[rank] = text_features
+                all_image_features = torch.cat(gathered_image_features, dim=0)
+                all_text_features = torch.cat(gathered_text_features, dim=0)
+    else:
+        # We gather tensors from all gpus
+        if gather_with_grad:
+            all_image_features = torch.cat(torch.distributed.nn.all_gather(image_features), dim=0)
+            all_text_features = torch.cat(torch.distributed.nn.all_gather(text_features), dim=0)
+            # all_image_features = torch.cat(torch.distributed.nn.all_gather(image_features, async_op=True), dim=0)
+            # all_text_features = torch.cat(torch.distributed.nn.all_gather(text_features, async_op=True), dim=0)
+        else:
+            gathered_image_features = [torch.zeros_like(image_features) for _ in range(world_size)]
+            gathered_text_features = [torch.zeros_like(text_features) for _ in range(world_size)]
+            dist.all_gather(gathered_image_features, image_features)
+            dist.all_gather(gathered_text_features, text_features)
+            if not local_loss:
+                # ensure grads for local rank when all_* features don't have a gradient
+                gathered_image_features[rank] = image_features
+                gathered_text_features[rank] = text_features
+            all_image_features = torch.cat(gathered_image_features, dim=0)
+            all_text_features = torch.cat(gathered_text_features, dim=0)
+
+    return all_image_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,
+            smoothing=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.label_smoothing_cross_entropy = LabelSmoothingCrossEntropy(smoothing=smoothing) if smoothing > 0 else None
+
+        # cache state
+        self.prev_num_logits = 0
+        self.labels = {}
+
+    def forward(self, image_features, text_features, logit_scale=1.):
+        device = image_features.device
+        if self.world_size > 1:
+            all_image_features, all_text_features = gather_features(
+                image_features, text_features,
+                self.local_loss, self.gather_with_grad, self.rank, self.world_size, self.use_horovod)
+
+            if self.local_loss:
+                logits_per_image = logit_scale * image_features @ all_text_features.T
+                logits_per_text = logit_scale * text_features @ all_image_features.T
+            else:
+                logits_per_image = logit_scale * all_image_features @ all_text_features.T
+                logits_per_text = logits_per_image.T
+        else:
+            logits_per_image = logit_scale * image_features @ text_features.T
+            logits_per_text = logit_scale * text_features @ image_features.T
+        # calculated ground-truth and cache if enabled
+        num_logits = logits_per_image.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 self.label_smoothing_cross_entropy:
+            total_loss = (
+                self.label_smoothing_cross_entropy(logits_per_image, labels) +
+                self.label_smoothing_cross_entropy(logits_per_text, labels)
+                ) / 2
+        else:
+            total_loss = (
+                F.cross_entropy(logits_per_image, labels) +
+                F.cross_entropy(logits_per_text, labels)
+                ) / 2
+            
+        acc = None
+        i2t_acc = (logits_per_image.argmax(-1) == labels).sum() / len(logits_per_image)
+        t2i_acc = (logits_per_text.argmax(-1) == labels).sum() / len(logits_per_text)
+        acc = {"i2t": i2t_acc, "t2i": t2i_acc}
+        return total_loss, acc

eva_clip/model.py

@@ -0,0 +1,439 @@
+""" CLIP Model
+
+Adapted from https://github.com/openai/CLIP. Originally MIT License, Copyright (c) 2021 OpenAI.
+"""
+import os
+from dataclasses import dataclass
+from typing import Optional, Tuple, Union
+from functools import partial
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+try:
+    from .hf_model import HFTextEncoder
+except:
+    HFTextEncoder = None
+from .modified_resnet import ModifiedResNet
+from .timm_model import TimmModel
+from .eva_vit_model import EVAVisionTransformer
+from .transformer import LayerNorm, QuickGELU, Attention, VisionTransformer, TextTransformer
+
+try:
+    from apex.normalization import FusedLayerNorm
+except:
+    FusedLayerNorm = LayerNorm
+    print("Please 'pip install apex'")
+
+try:
+    import xformers.ops as xops
+except ImportError:
+    xops = None
+    print("Please 'pip install xformers'")
+
+@dataclass
+class CLIPVisionCfg:
+    layers: Union[Tuple[int, int, int, int], int] = 12
+    width: int = 768
+    head_width: int = 64
+    mlp_ratio: float = 4.0
+    patch_size: int = 16
+    image_size: Union[Tuple[int, int], int] = 224
+    ls_init_value: Optional[float] = None  # layer scale initial value
+    patch_dropout: float = 0. # what fraction of patches to dropout during training (0 would mean disabled and no patches dropped) - 0.5 to 0.75 recommended in the paper for optimal results
+    global_average_pool: bool = False # whether to global average pool the last embedding layer, instead of using CLS token (https://arxiv.org/abs/2205.01580)
+    drop_path_rate: Optional[float] = None  # drop path rate
+    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', '')
+    timm_proj_bias: bool = False  # enable bias final projection
+    eva_model_name: str = None # a valid eva model name overrides layers, width, patch_size
+    qkv_bias: bool = True
+    fusedLN: bool = False
+    xattn: bool = False
+    postnorm: bool = False
+    rope: bool = False
+    pt_hw_seq_len: int = 16   # 224/14
+    intp_freq: bool = False
+    naiveswiglu: bool = False
+    subln: bool = False
+
+
+@dataclass
+class CLIPTextCfg:
+    context_length: int = 77
+    vocab_size: int = 49408
+    width: int = 512
+    heads: int = 8
+    layers: int = 12
+    ls_init_value: Optional[float] = None  # layer scale initial value
+    hf_model_name: str = None
+    hf_tokenizer_name: str = None
+    hf_model_pretrained: bool = True
+    proj: str = 'mlp'
+    pooler_type: str = 'mean_pooler'
+    masked_language_modeling: bool = False
+    fusedLN: bool = False
+    xattn: bool = False
+    attn_mask: bool = True
+
+def get_cast_dtype(precision: str):
+    cast_dtype = None
+    if precision == 'bf16':
+        cast_dtype = torch.bfloat16
+    elif precision == 'fp16':
+        cast_dtype = torch.float16
+    return cast_dtype
+
+
+def _build_vision_tower(
+        embed_dim: int,
+        vision_cfg: CLIPVisionCfg,
+        quick_gelu: bool = False,
+        cast_dtype: Optional[torch.dtype] = None
+):
+    if isinstance(vision_cfg, dict):
+        vision_cfg = CLIPVisionCfg(**vision_cfg)
+
+    # 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 vision_cfg.eva_model_name:
+        vision_heads = vision_cfg.width // vision_cfg.head_width
+        norm_layer = LayerNorm
+        
+        visual = EVAVisionTransformer(
+            img_size=vision_cfg.image_size,
+            patch_size=vision_cfg.patch_size,
+            num_classes=embed_dim,
+            use_mean_pooling=vision_cfg.global_average_pool, #False
+            init_values=vision_cfg.ls_init_value,
+            patch_dropout=vision_cfg.patch_dropout,
+            embed_dim=vision_cfg.width,
+            depth=vision_cfg.layers,
+            num_heads=vision_heads,
+            mlp_ratio=vision_cfg.mlp_ratio,
+            qkv_bias=vision_cfg.qkv_bias,
+            drop_path_rate=vision_cfg.drop_path_rate,
+            norm_layer= partial(FusedLayerNorm, eps=1e-6) if vision_cfg.fusedLN else partial(norm_layer, eps=1e-6),
+            xattn=vision_cfg.xattn,
+            rope=vision_cfg.rope,
+            postnorm=vision_cfg.postnorm,
+            pt_hw_seq_len= vision_cfg.pt_hw_seq_len,   # 224/14
+            intp_freq= vision_cfg.intp_freq,
+            naiveswiglu= vision_cfg.naiveswiglu,
+            subln= vision_cfg.subln
+        )
+    elif vision_cfg.timm_model_name:
+        visual = TimmModel(
+            vision_cfg.timm_model_name,
+            pretrained=vision_cfg.timm_model_pretrained,
+            pool=vision_cfg.timm_pool,
+            proj=vision_cfg.timm_proj,
+            proj_bias=vision_cfg.timm_proj_bias,
+            embed_dim=embed_dim,
+            image_size=vision_cfg.image_size
+        )
+        act_layer = nn.GELU  # so that text transformer doesn't use QuickGELU w/ timm models
+    elif isinstance(vision_cfg.layers, (tuple, list)):
+        vision_heads = vision_cfg.width * 32 // vision_cfg.head_width
+        visual = ModifiedResNet(
+            layers=vision_cfg.layers,
+            output_dim=embed_dim,
+            heads=vision_heads,
+            image_size=vision_cfg.image_size,
+            width=vision_cfg.width
+        )
+    else:
+        vision_heads = vision_cfg.width // vision_cfg.head_width
+        norm_layer = LayerNormFp32 if cast_dtype in (torch.float16, torch.bfloat16) else LayerNorm
+        visual = VisionTransformer(
+            image_size=vision_cfg.image_size,
+            patch_size=vision_cfg.patch_size,
+            width=vision_cfg.width,
+            layers=vision_cfg.layers,
+            heads=vision_heads,
+            mlp_ratio=vision_cfg.mlp_ratio,
+            ls_init_value=vision_cfg.ls_init_value,
+            patch_dropout=vision_cfg.patch_dropout,
+            global_average_pool=vision_cfg.global_average_pool,
+            output_dim=embed_dim,
+            act_layer=act_layer,
+            norm_layer=norm_layer,
+        )
+
+    return visual
+
+
+def _build_text_tower(
+        embed_dim: int,
+        text_cfg: CLIPTextCfg,
+        quick_gelu: bool = False,
+        cast_dtype: Optional[torch.dtype] = None,
+):
+    if isinstance(text_cfg, dict):
+        text_cfg = CLIPTextCfg(**text_cfg)
+
+    if text_cfg.hf_model_name:
+        text = HFTextEncoder(
+            text_cfg.hf_model_name,
+            output_dim=embed_dim,
+            tokenizer_name=text_cfg.hf_tokenizer_name,
+            proj=text_cfg.proj,
+            pooler_type=text_cfg.pooler_type,
+            masked_language_modeling=text_cfg.masked_language_modeling
+       )
+    else:
+        act_layer = QuickGELU if quick_gelu else nn.GELU
+        norm_layer = LayerNorm
+
+        text = TextTransformer(
+            context_length=text_cfg.context_length,
+            vocab_size=text_cfg.vocab_size,
+            width=text_cfg.width,
+            heads=text_cfg.heads,
+            layers=text_cfg.layers,
+            ls_init_value=text_cfg.ls_init_value,
+            output_dim=embed_dim,
+            act_layer=act_layer,
+            norm_layer= FusedLayerNorm if text_cfg.fusedLN else norm_layer,
+            xattn=text_cfg.xattn,
+            attn_mask=text_cfg.attn_mask,
+        )
+    return text
+
+class CLIP(nn.Module):
+    def __init__(
+            self,
+            embed_dim: int,
+            vision_cfg: CLIPVisionCfg,
+            text_cfg: CLIPTextCfg,
+            quick_gelu: bool = False,
+            cast_dtype: Optional[torch.dtype] = None,
+    ):
+        super().__init__()
+        self.visual = _build_vision_tower(embed_dim, vision_cfg, quick_gelu, cast_dtype)
+
+        text = _build_text_tower(embed_dim, text_cfg, quick_gelu, cast_dtype)
+        self.transformer = text.transformer
+        self.vocab_size = text.vocab_size
+        self.token_embedding = text.token_embedding
+        self.positional_embedding = text.positional_embedding
+        self.ln_final = text.ln_final
+        self.text_projection = text.text_projection
+        self.register_buffer('attn_mask', text.attn_mask, persistent=False)
+
+        self.logit_scale = nn.Parameter(torch.ones([]) * np.log(1 / 0.07))
+
+    def lock_image_tower(self, unlocked_groups=0, freeze_bn_stats=False):
+        # lock image tower as per LiT - https://arxiv.org/abs/2111.07991
+        self.visual.lock(unlocked_groups=unlocked_groups, freeze_bn_stats=freeze_bn_stats)
+
+    @torch.jit.ignore
+    def set_grad_checkpointing(self, enable=True):
+        self.visual.set_grad_checkpointing(enable)
+        self.transformer.grad_checkpointing = enable
+    
+    @torch.jit.ignore
+    def no_weight_decay(self):
+        return {'logit_scale'}
+
+    def encode_image(self, image, normalize: bool = False):
+        features = self.visual(image)
+        return F.normalize(features, dim=-1) if normalize else features
+
+    def encode_text(self, text, normalize: bool = False):
+        cast_dtype = self.transformer.get_cast_dtype()
+
+        x = self.token_embedding(text).to(cast_dtype)  # [batch_size, n_ctx, d_model]
+
+        x = x + self.positional_embedding.to(cast_dtype)
+        x = x.permute(1, 0, 2)  # NLD -> LND
+        x = self.transformer(x, attn_mask=self.attn_mask)
+        x = x.permute(1, 0, 2)  # LND -> NLD
+        x = self.ln_final(x)  # [batch_size, n_ctx, transformer.width]
+        # take features from the eot embedding (eot_token is the highest number in each sequence)
+        x = x[torch.arange(x.shape[0]), text.argmax(dim=-1)] @ self.text_projection
+        return F.normalize(x, dim=-1) if normalize else x
+
+    def forward(self, image, text):
+        image_features = self.encode_image(image, normalize=True)
+        text_features = self.encode_text(text, normalize=True)
+        return image_features, text_features, self.logit_scale.exp()
+
+
+class CustomCLIP(nn.Module):
+    def __init__(
+            self,
+            embed_dim: int,
+            vision_cfg: CLIPVisionCfg,
+            text_cfg: CLIPTextCfg,
+            quick_gelu: bool = False,
+            cast_dtype: Optional[torch.dtype] = None,
+            itm_task: bool = False,
+    ):
+        super().__init__()
+        self.visual = _build_vision_tower(embed_dim, vision_cfg, quick_gelu, cast_dtype)
+        self.text = _build_text_tower(embed_dim, text_cfg, quick_gelu, cast_dtype)
+        self.logit_scale = nn.Parameter(torch.ones([]) * np.log(1 / 0.07))
+
+    def lock_image_tower(self, unlocked_groups=0, freeze_bn_stats=False):
+        # lock image tower as per LiT - https://arxiv.org/abs/2111.07991
+        self.visual.lock(unlocked_groups=unlocked_groups, freeze_bn_stats=freeze_bn_stats)
+
+    def lock_text_tower(self, unlocked_layers:int=0, freeze_layer_norm:bool=True):
+        self.text.lock(unlocked_layers, freeze_layer_norm)
+
+    @torch.jit.ignore
+    def set_grad_checkpointing(self, enable=True):
+        self.visual.set_grad_checkpointing(enable)
+        self.text.set_grad_checkpointing(enable)
+
+    @torch.jit.ignore
+    def no_weight_decay(self):
+        return {'logit_scale'}
+
+    def encode_image(self, image, normalize: bool = False):
+        features = self.visual(image)
+        return F.normalize(features, dim=-1) if normalize else features
+
+    def encode_text(self, text, normalize: bool = False):
+        features = self.text(text)
+        return F.normalize(features, dim=-1) if normalize else features
+
+    def forward(self, image, text):
+        image_features = self.encode_image(image, normalize=True)
+        text_features = self.encode_text(text, normalize=True)
+        return image_features, text_features, self.logit_scale.exp()
+
+
+def convert_weights_to_lp(model: nn.Module, dtype=torch.float16):
+    """Convert applicable model parameters to low-precision (bf16 or fp16)"""
+
+    def _convert_weights(l):
+        
+        if isinstance(l, (nn.Conv1d, nn.Conv2d, nn.Linear)):
+            l.weight.data = l.weight.data.to(dtype)
+            if l.bias is not None:
+                l.bias.data = l.bias.data.to(dtype)
+
+        if isinstance(l, (nn.MultiheadAttention, Attention)):
+            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, None)
+                if tensor is not None:
+                    tensor.data = tensor.data.to(dtype)
+
+        if isinstance(l, nn.Parameter):
+            l.data = l.data.to(dtype)
+
+        for name in ["text_projection", "proj"]:
+            if hasattr(l, name) and isinstance(l, nn.Parameter):
+                attr = getattr(l, name, None)
+                if attr is not None:
+                    attr.data = attr.data.to(dtype)
+
+    model.apply(_convert_weights)
+
+
+convert_weights_to_fp16 = convert_weights_to_lp  # backwards compat
+
+
+# used to maintain checkpoint compatibility
+def convert_to_custom_text_state_dict(state_dict: dict):
+    if 'text_projection' in state_dict:
+        # old format state_dict, move text tower -> .text
+        new_state_dict = {}
+        for k, v in state_dict.items():
+            if any(k.startswith(p) for p in (
+                'text_projection',
+                'positional_embedding',
+                'token_embedding',
+                'transformer',
+                'ln_final',
+                'logit_scale'
+            )):
+                k = 'text.' + k
+            new_state_dict[k] = v
+        return new_state_dict
+    return state_dict
+
+
+def build_model_from_openai_state_dict(
+        state_dict: dict,
+        quick_gelu=True,
+        cast_dtype=torch.float16,
+):
+    vit = "visual.proj" in state_dict
+
+    if vit:
+        vision_width = state_dict["visual.conv1.weight"].shape[0]
+        vision_layers = len(
+            [k for k in state_dict.keys() if k.startswith("visual.") and k.endswith(".attn.in_proj_weight")])
+        vision_patch_size = state_dict["visual.conv1.weight"].shape[-1]
+        grid_size = round((state_dict["visual.positional_embedding"].shape[0] - 1) ** 0.5)
+        image_size = vision_patch_size * grid_size
+    else:
+        counts: list = [
+            len(set(k.split(".")[2] for k in state_dict if k.startswith(f"visual.layer{b}"))) for b in [1, 2, 3, 4]]
+        vision_layers = tuple(counts)
+        vision_width = state_dict["visual.layer1.0.conv1.weight"].shape[0]
+        output_width = round((state_dict["visual.attnpool.positional_embedding"].shape[0] - 1) ** 0.5)
+        vision_patch_size = None
+        assert output_width ** 2 + 1 == state_dict["visual.attnpool.positional_embedding"].shape[0]
+        image_size = output_width * 32
+
+    embed_dim = state_dict["text_projection"].shape[1]
+    context_length = state_dict["positional_embedding"].shape[0]
+    vocab_size = state_dict["token_embedding.weight"].shape[0]
+    transformer_width = state_dict["ln_final.weight"].shape[0]
+    transformer_heads = transformer_width // 64
+    transformer_layers = len(set(k.split(".")[2] for k in state_dict if k.startswith(f"transformer.resblocks")))
+
+    vision_cfg = CLIPVisionCfg(
+        layers=vision_layers,
+        width=vision_width,
+        patch_size=vision_patch_size,
+        image_size=image_size,
+    )
+    text_cfg = CLIPTextCfg(
+        context_length=context_length,
+        vocab_size=vocab_size,
+        width=transformer_width,
+        heads=transformer_heads,
+        layers=transformer_layers
+    )
+    model = CLIP(
+        embed_dim,
+        vision_cfg=vision_cfg,
+        text_cfg=text_cfg,
+        quick_gelu=quick_gelu,  # OpenAI models were trained with QuickGELU
+        cast_dtype=cast_dtype,
+    )
+
+    for key in ["input_resolution", "context_length", "vocab_size"]:
+        state_dict.pop(key, None)
+
+    convert_weights_to_fp16(model)  # OpenAI state dicts are partially converted to float16
+    model.load_state_dict(state_dict)
+    return model.eval()
+
+
+def trace_model(model, batch_size=256, device=torch.device('cpu')):
+    model.eval()
+    image_size = model.visual.image_size
+    example_images = torch.ones((batch_size, 3, image_size, image_size), 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_images, example_text),
+            encode_text=(example_text,),
+            encode_image=(example_images,)
+        ))
+    model.visual.image_size = image_size
+    return model

eva_clip/model_configs/EVA01-CLIP-B-16.json

@@ -0,0 +1,19 @@
+{
+    "embed_dim": 512,
+    "vision_cfg": {
+        "image_size": 224,
+        "layers": 12,
+        "width": 768,
+        "patch_size": 16,
+        "eva_model_name": "eva-clip-b-16",
+        "ls_init_value": 0.1,
+        "drop_path_rate": 0.0
+    },
+    "text_cfg": {
+        "context_length": 77,
+        "vocab_size": 49408,
+        "width": 512,
+        "heads": 8,
+        "layers": 12
+    }
+}

eva_clip/model_configs/EVA01-CLIP-g-14-plus.json

@@ -0,0 +1,24 @@
+{
+    "embed_dim": 1024,
+    "vision_cfg": {
+        "image_size": 224,
+        "layers": 40,
+        "width": 1408,
+        "head_width": 88,
+        "mlp_ratio": 4.3637,
+        "patch_size": 14,
+        "eva_model_name": "eva-clip-g-14-x",
+        "drop_path_rate": 0,
+        "xattn": true,
+        "fusedLN": true
+    },
+    "text_cfg": {
+        "context_length": 77,
+        "vocab_size": 49408,
+        "width": 1024,
+        "heads": 16,
+        "layers": 24,
+        "xattn": false,
+        "fusedLN": true
+    }
+}

eva_clip/model_configs/EVA01-CLIP-g-14.json

@@ -0,0 +1,24 @@
+{
+    "embed_dim": 1024,
+    "vision_cfg": {
+        "image_size": 224,
+        "layers": 40,
+        "width": 1408,
+        "head_width": 88,
+        "mlp_ratio": 4.3637,
+        "patch_size": 14,
+        "eva_model_name": "eva-clip-g-14-x",
+        "drop_path_rate": 0.4,
+        "xattn": true,
+        "fusedLN": true
+    },
+    "text_cfg": {
+        "context_length": 77,
+        "vocab_size": 49408,
+        "width": 768,
+        "heads": 12,
+        "layers": 12,
+        "xattn": false,
+        "fusedLN": true
+    }
+}

eva_clip/model_configs/EVA02-CLIP-B-16.json

@@ -0,0 +1,29 @@
+{
+    "embed_dim": 512,
+    "vision_cfg": {
+        "image_size": 224,
+        "layers": 12,
+        "width": 768,
+        "head_width": 64,
+        "patch_size": 16,
+        "mlp_ratio": 2.6667,
+        "eva_model_name": "eva-clip-b-16-X",
+        "drop_path_rate": 0.0,
+        "xattn": true,
+        "fusedLN": true,
+        "rope": true,
+        "pt_hw_seq_len": 16,
+        "intp_freq": true,
+        "naiveswiglu": true,
+        "subln": true
+    },
+    "text_cfg": {
+        "context_length": 77,
+        "vocab_size": 49408,
+        "width": 512,
+        "heads": 8,
+        "layers": 12,
+        "xattn": true,
+        "fusedLN": true
+    }
+}

eva_clip/model_configs/EVA02-CLIP-L-14-336.json

@@ -0,0 +1,29 @@
+{
+    "embed_dim": 768,
+    "vision_cfg": {
+        "image_size": 336,
+        "layers": 24,
+        "width": 1024,
+        "drop_path_rate": 0,
+        "head_width": 64,
+        "mlp_ratio": 2.6667,
+        "patch_size": 14,
+        "eva_model_name": "eva-clip-l-14-336",
+        "xattn": true,
+        "fusedLN": true,
+        "rope": true,
+        "pt_hw_seq_len": 16,
+        "intp_freq": true,
+        "naiveswiglu": true,
+        "subln": true
+    },
+    "text_cfg": {
+        "context_length": 77,
+        "vocab_size": 49408,
+        "width": 768,
+        "heads": 12,
+        "layers": 12,
+        "xattn": false,
+        "fusedLN": true
+    }
+}

eva_clip/model_configs/EVA02-CLIP-L-14.json

@@ -0,0 +1,29 @@
+{
+    "embed_dim": 768,
+    "vision_cfg": {
+        "image_size": 224,
+        "layers": 24,
+        "width": 1024,
+        "drop_path_rate": 0,
+        "head_width": 64,
+        "mlp_ratio": 2.6667,
+        "patch_size": 14,
+        "eva_model_name": "eva-clip-l-14",
+        "xattn": true,
+        "fusedLN": true,
+        "rope": true,
+        "pt_hw_seq_len": 16,
+        "intp_freq": true,
+        "naiveswiglu": true,
+        "subln": true
+    },
+    "text_cfg": {
+        "context_length": 77,
+        "vocab_size": 49408,
+        "width": 768,
+        "heads": 12,
+        "layers": 12,
+        "xattn": false,
+        "fusedLN": true
+    }
+}

eva_clip/model_configs/EVA02-CLIP-bigE-14-plus.json

@@ -0,0 +1,25 @@
+{
+    "embed_dim": 1024,
+    "vision_cfg": {
+        "image_size": 224,
+        "layers": 64,
+        "width": 1792,
+        "head_width": 112,
+        "mlp_ratio": 8.571428571428571,
+        "patch_size": 14,
+        "eva_model_name": "eva-clip-4b-14-x",
+        "drop_path_rate": 0,
+        "xattn": true,
+        "postnorm": true,
+        "fusedLN": true
+    },
+    "text_cfg": {
+        "context_length": 77,
+        "vocab_size": 49408,
+        "width": 1280,
+        "heads": 20,
+        "layers": 32,
+        "xattn": false,
+        "fusedLN": true
+    }
+}

eva_clip/model_configs/EVA02-CLIP-bigE-14.json

@@ -0,0 +1,25 @@
+{
+    "embed_dim": 1024,
+    "vision_cfg": {
+        "image_size": 224,
+        "layers": 64,
+        "width": 1792,
+        "head_width": 112,
+        "mlp_ratio": 8.571428571428571,
+        "patch_size": 14,
+        "eva_model_name": "eva-clip-4b-14-x",
+        "drop_path_rate": 0,
+        "xattn": true,
+        "postnorm": true,
+        "fusedLN": true
+    },
+    "text_cfg": {
+        "context_length": 77,
+        "vocab_size": 49408,
+        "width": 1024,
+        "heads": 16,
+        "layers": 24,
+        "xattn": false,
+        "fusedLN": true
+    }
+}

eva_clip/modified_resnet.py

@@ -0,0 +1,181 @@
+from collections import OrderedDict
+
+import torch
+from torch import nn
+from torch.nn import functional as F
+
+from eva_clip.utils import freeze_batch_norm_2d
+
+
+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.act1 = nn.ReLU(inplace=True)
+
+        self.conv2 = nn.Conv2d(planes, planes, 3, padding=1, bias=False)
+        self.bn2 = nn.BatchNorm2d(planes)
+        self.act2 = nn.ReLU(inplace=True)
+
+        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.act3 = 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.act1(self.bn1(self.conv1(x)))
+        out = self.act2(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.act3(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.act1 = nn.ReLU(inplace=True)
+        self.conv2 = nn.Conv2d(width // 2, width // 2, kernel_size=3, padding=1, bias=False)
+        self.bn2 = nn.BatchNorm2d(width // 2)
+        self.act2 = nn.ReLU(inplace=True)
+        self.conv3 = nn.Conv2d(width // 2, width, kernel_size=3, padding=1, bias=False)
+        self.bn3 = nn.BatchNorm2d(width)
+        self.act3 = nn.ReLU(inplace=True)
+        self.avgpool = nn.AvgPool2d(2)
+
+        # 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)
+
+    @torch.jit.ignore
+    def set_grad_checkpointing(self, enable=True):
+        # FIXME support for non-transformer
+        pass
+
+    def stem(self, x):
+        x = self.act1(self.bn1(self.conv1(x)))
+        x = self.act2(self.bn2(self.conv2(x)))
+        x = self.act3(self.bn3(self.conv3(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

eva_clip/openai.py

@@ -0,0 +1,144 @@
+""" 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 List, Optional, Union
+
+import torch
+
+from .model import build_model_from_openai_state_dict, convert_weights_to_lp, get_cast_dtype
+from .pretrained import get_pretrained_url, list_pretrained_models_by_tag, download_pretrained_from_url
+
+__all__ = ["list_openai_models", "load_openai_model"]
+
+
+def list_openai_models() -> List[str]:
+    """Returns the names of available CLIP models"""
+    return list_pretrained_models_by_tag('openai')
+
+
+def load_openai_model(
+        name: str,
+        precision: Optional[str] = None,
+        device: Optional[Union[str, torch.device]] = None,
+        jit: bool = True,
+        cache_dir: Optional[str] = None,
+):
+    """Load a CLIP model
+
+    Parameters
+    ----------
+    name : str
+        A model name listed by `clip.available_models()`, or the path to a model checkpoint containing the state_dict
+    precision: str
+        Model precision, if None defaults to 'fp32' if device == 'cpu' else 'fp16'.
+    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.
+    cache_dir : Optional[str]
+        The directory to cache the downloaded model weights
+
+    Returns
+    -------
+    model : torch.nn.Module
+        The CLIP 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 device is None:
+        device = "cuda" if torch.cuda.is_available() else "cpu"
+    if precision is None:
+        precision = 'fp32' if device == 'cpu' else 'fp16'
+
+    if get_pretrained_url(name, 'openai'):
+        model_path = download_pretrained_from_url(get_pretrained_url(name, 'openai'), cache_dir=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:
+        # Build a non-jit model from the OpenAI jitted model state dict
+        cast_dtype = get_cast_dtype(precision)
+        try:
+            model = build_model_from_openai_state_dict(state_dict or model.state_dict(), cast_dtype=cast_dtype)
+        except KeyError:
+            sd = {k[7:]: v for k, v in state_dict["state_dict"].items()}
+            model = build_model_from_openai_state_dict(sd, cast_dtype=cast_dtype)
+
+        # model from OpenAI state dict is in manually cast fp16 mode, must be converted for AMP/fp32/bf16 use
+        model = model.to(device)
+        if precision.startswith('amp') or precision == 'fp32':
+            model.float()
+        elif precision == 'bf16':
+            convert_weights_to_lp(model, dtype=torch.bfloat16)
+
+        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_image)
+    patch_device(model.encode_text)
+
+    # patch dtype to float32 (typically for CPU)
+    if precision == 'fp32':
+        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_image)
+        patch_float(model.encode_text)
+        model.float()
+
+    # ensure image_size attr available at consistent location for both jit and non-jit
+    model.visual.image_size = model.input_resolution.item()
+    return model

eva_clip/pretrained.py

@@ -0,0 +1,332 @@
+import hashlib
+import os
+import urllib
+import warnings
+from functools import partial
+from typing import Dict, Union
+
+from tqdm import tqdm
+
+try:
+    from huggingface_hub import hf_hub_download
+    _has_hf_hub = True
+except ImportError:
+    hf_hub_download = None
+    _has_hf_hub = False
+
+
+def _pcfg(url='', hf_hub='', filename='', mean=None, std=None):
+    return dict(
+        url=url,
+        hf_hub=hf_hub,
+        mean=mean,
+        std=std,
+    )
+
+_VITB32 = dict(
+    openai=_pcfg(
+        "https://openaipublic.azureedge.net/clip/models/40d365715913c9da98579312b702a82c18be219cc2a73407c4526f58eba950af/ViT-B-32.pt"),
+    laion400m_e31=_pcfg(
+        "https://github.com/mlfoundations/open_clip/releases/download/v0.2-weights/vit_b_32-quickgelu-laion400m_e31-d867053b.pt"),
+    laion400m_e32=_pcfg(
+        "https://github.com/mlfoundations/open_clip/releases/download/v0.2-weights/vit_b_32-quickgelu-laion400m_e32-46683a32.pt"),
+    laion2b_e16=_pcfg(
+        "https://github.com/mlfoundations/open_clip/releases/download/v0.2-weights/vit_b_32-laion2b_e16-af8dbd0c.pth"),
+    laion2b_s34b_b79k=_pcfg(hf_hub='laion/CLIP-ViT-B-32-laion2B-s34B-b79K/')
+)
+
+_VITB32_quickgelu = dict(
+    openai=_pcfg(
+        "https://openaipublic.azureedge.net/clip/models/40d365715913c9da98579312b702a82c18be219cc2a73407c4526f58eba950af/ViT-B-32.pt"),
+    laion400m_e31=_pcfg(
+        "https://github.com/mlfoundations/open_clip/releases/download/v0.2-weights/vit_b_32-quickgelu-laion400m_e31-d867053b.pt"),
+    laion400m_e32=_pcfg(
+        "https://github.com/mlfoundations/open_clip/releases/download/v0.2-weights/vit_b_32-quickgelu-laion400m_e32-46683a32.pt"),
+)
+
+_VITB16 = dict(
+    openai=_pcfg(
+        "https://openaipublic.azureedge.net/clip/models/5806e77cd80f8b59890b7e101eabd078d9fb84e6937f9e85e4ecb61988df416f/ViT-B-16.pt"),
+    laion400m_e31=_pcfg(
+        "https://github.com/mlfoundations/open_clip/releases/download/v0.2-weights/vit_b_16-laion400m_e31-00efa78f.pt"),
+    laion400m_e32=_pcfg(
+        "https://github.com/mlfoundations/open_clip/releases/download/v0.2-weights/vit_b_16-laion400m_e32-55e67d44.pt"),
+    laion2b_s34b_b88k=_pcfg(hf_hub='laion/CLIP-ViT-B-16-laion2B-s34B-b88K/'),
+)
+
+_EVAB16 = dict(
+    eva=_pcfg(hf_hub='QuanSun/EVA-CLIP/EVA02_B_psz14to16.pt'),
+    eva02=_pcfg(hf_hub='QuanSun/EVA-CLIP/EVA02_B_psz14to16.pt'),
+    eva_clip=_pcfg(hf_hub='QuanSun/EVA-CLIP/EVA02_CLIP_B_psz16_s8B.pt'),
+    eva02_clip=_pcfg(hf_hub='QuanSun/EVA-CLIP/EVA02_CLIP_B_psz16_s8B.pt'),
+)
+
+_VITB16_PLUS_240 = dict(
+    laion400m_e31=_pcfg(
+        "https://github.com/mlfoundations/open_clip/releases/download/v0.2-weights/vit_b_16_plus_240-laion400m_e31-8fb26589.pt"),
+    laion400m_e32=_pcfg(
+        "https://github.com/mlfoundations/open_clip/releases/download/v0.2-weights/vit_b_16_plus_240-laion400m_e32-699c4b84.pt"),
+)
+
+_VITL14 = dict(
+    openai=_pcfg(
+        "https://openaipublic.azureedge.net/clip/models/b8cca3fd41ae0c99ba7e8951adf17d267cdb84cd88be6f7c2e0eca1737a03836/ViT-L-14.pt"),
+    laion400m_e31=_pcfg(
+        "https://github.com/mlfoundations/open_clip/releases/download/v0.2-weights/vit_l_14-laion400m_e31-69988bb6.pt"),
+    laion400m_e32=_pcfg(
+        "https://github.com/mlfoundations/open_clip/releases/download/v0.2-weights/vit_l_14-laion400m_e32-3d133497.pt"),
+    laion2b_s32b_b82k=_pcfg(
+        hf_hub='laion/CLIP-ViT-L-14-laion2B-s32B-b82K/',
+        mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5)),
+)
+
+_EVAL14 = dict(
+    eva=_pcfg(hf_hub='QuanSun/EVA-CLIP/EVA02_L_psz14.pt'),
+    eva02=_pcfg(hf_hub='QuanSun/EVA-CLIP/EVA02_L_psz14.pt'),
+    eva_clip=_pcfg(hf_hub='QuanSun/EVA-CLIP/EVA02_CLIP_L_psz14_s4B.pt'),
+    eva02_clip=_pcfg(hf_hub='QuanSun/EVA-CLIP/EVA02_CLIP_L_psz14_s4B.pt'),
+)
+
+_VITL14_336 = dict(
+    openai=_pcfg(
+        "https://openaipublic.azureedge.net/clip/models/3035c92b350959924f9f00213499208652fc7ea050643e8b385c2dac08641f02/ViT-L-14-336px.pt"),
+)
+
+_EVAL14_336 = dict(
+    eva_clip=_pcfg(hf_hub='QuanSun/EVA-CLIP/EVA02_CLIP_L_336_psz14_s6B.pt'),
+    eva02_clip=_pcfg(hf_hub='QuanSun/EVA-CLIP/EVA02_CLIP_L_336_psz14_s6B.pt'),
+    eva_clip_224to336=_pcfg(hf_hub='QuanSun/EVA-CLIP/EVA02_CLIP_L_psz14_224to336.pt'),
+    eva02_clip_224to336=_pcfg(hf_hub='QuanSun/EVA-CLIP/EVA02_CLIP_L_psz14_224to336.pt'),
+)
+
+_VITH14 = dict(
+    laion2b_s32b_b79k=_pcfg(hf_hub='laion/CLIP-ViT-H-14-laion2B-s32B-b79K/'),
+)
+
+_VITg14 = dict(
+    laion2b_s12b_b42k=_pcfg(hf_hub='laion/CLIP-ViT-g-14-laion2B-s12B-b42K/'),
+    laion2b_s34b_b88k=_pcfg(hf_hub='laion/CLIP-ViT-g-14-laion2B-s34B-b88K/'),
+)
+
+_EVAg14 = dict(
+    eva=_pcfg(hf_hub='QuanSun/EVA-CLIP/'),
+    eva01=_pcfg(hf_hub='QuanSun/EVA-CLIP/EVA01_g_psz14.pt'),
+    eva_clip=_pcfg(hf_hub='QuanSun/EVA-CLIP/EVA01_CLIP_g_14_psz14_s11B.pt'),
+    eva01_clip=_pcfg(hf_hub='QuanSun/EVA-CLIP/EVA01_CLIP_g_14_psz14_s11B.pt'),
+)
+
+_EVAg14_PLUS = dict(
+    eva=_pcfg(hf_hub='QuanSun/EVA-CLIP/'),
+    eva01=_pcfg(hf_hub='QuanSun/EVA-CLIP/EVA01_g_psz14.pt'),
+    eva_clip=_pcfg(hf_hub='QuanSun/EVA-CLIP/EVA01_CLIP_g_14_plus_psz14_s11B.pt'),
+    eva01_clip=_pcfg(hf_hub='QuanSun/EVA-CLIP/EVA01_CLIP_g_14_plus_psz14_s11B.pt'),
+)
+
+_VITbigG14 = dict(
+    laion2b_s39b_b160k=_pcfg(hf_hub='laion/CLIP-ViT-bigG-14-laion2B-39B-b160k/'),
+)
+
+_EVAbigE14 = dict(
+    eva=_pcfg(hf_hub='QuanSun/EVA-CLIP/EVA02_E_psz14.pt'),
+    eva02=_pcfg(hf_hub='QuanSun/EVA-CLIP/EVA02_E_psz14.pt'),
+    eva_clip=_pcfg(hf_hub='QuanSun/EVA-CLIP/EVA02_CLIP_E_psz14_s4B.pt'),
+    eva02_clip=_pcfg(hf_hub='QuanSun/EVA-CLIP/EVA02_CLIP_E_psz14_s4B.pt'),
+)
+
+_EVAbigE14_PLUS = dict(
+    eva=_pcfg(hf_hub='QuanSun/EVA-CLIP/EVA02_E_psz14.pt'),
+    eva02=_pcfg(hf_hub='QuanSun/EVA-CLIP/EVA02_E_psz14.pt'),
+    eva_clip=_pcfg(hf_hub='QuanSun/EVA-CLIP/EVA02_CLIP_E_psz14_plus_s9B.pt'),
+    eva02_clip=_pcfg(hf_hub='QuanSun/EVA-CLIP/EVA02_CLIP_E_psz14_plus_s9B.pt'),
+)
+
+
+_PRETRAINED = {
+    # "ViT-B-32": _VITB32,
+    "OpenaiCLIP-B-32": _VITB32,
+    "OpenCLIP-B-32": _VITB32,
+
+    # "ViT-B-32-quickgelu": _VITB32_quickgelu,
+    "OpenaiCLIP-B-32-quickgelu": _VITB32_quickgelu,
+    "OpenCLIP-B-32-quickgelu": _VITB32_quickgelu,
+
+    # "ViT-B-16": _VITB16,
+    "OpenaiCLIP-B-16": _VITB16,
+    "OpenCLIP-B-16": _VITB16,
+
+    "EVA02-B-16": _EVAB16,
+    "EVA02-CLIP-B-16": _EVAB16,
+
+    # "ViT-B-16-plus-240": _VITB16_PLUS_240,
+    "OpenCLIP-B-16-plus-240": _VITB16_PLUS_240,
+
+    # "ViT-L-14": _VITL14,
+    "OpenaiCLIP-L-14": _VITL14,
+    "OpenCLIP-L-14": _VITL14,
+
+    "EVA02-L-14": _EVAL14,
+    "EVA02-CLIP-L-14": _EVAL14,
+
+    # "ViT-L-14-336": _VITL14_336,
+    "OpenaiCLIP-L-14-336": _VITL14_336,
+
+    "EVA02-CLIP-L-14-336": _EVAL14_336,
+
+    # "ViT-H-14": _VITH14,
+    # "ViT-g-14": _VITg14,
+    "OpenCLIP-H-14": _VITH14,
+    "OpenCLIP-g-14": _VITg14,
+
+    "EVA01-CLIP-g-14": _EVAg14,
+    "EVA01-CLIP-g-14-plus": _EVAg14_PLUS,
+
+    # "ViT-bigG-14": _VITbigG14,
+    "OpenCLIP-bigG-14": _VITbigG14,
+
+    "EVA02-CLIP-bigE-14": _EVAbigE14,
+    "EVA02-CLIP-bigE-14-plus": _EVAbigE14_PLUS,
+}
+
+
+def _clean_tag(tag: str):
+    # normalize pretrained tags
+    return tag.lower().replace('-', '_')
+
+
+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_models_by_tag(tag: str):
+    """ return all models having the specified pretrain tag """
+    models = []
+    tag = _clean_tag(tag)
+    for k in _PRETRAINED.keys():
+        if tag in _PRETRAINED[k]:
+            models.append(k)
+    return models
+
+
+def list_pretrained_tags_by_model(model: str):
+    """ return all pretrain tags for the specified model architecture """
+    tags = []
+    if model in _PRETRAINED:
+        tags.extend(_PRETRAINED[model].keys())
+    return tags
+
+
+def is_pretrained_cfg(model: str, tag: str):
+    if model not in _PRETRAINED:
+        return False
+    return _clean_tag(tag) in _PRETRAINED[model]
+
+
+def get_pretrained_cfg(model: str, tag: str):
+    if model not in _PRETRAINED:
+        return {}
+    model_pretrained = _PRETRAINED[model]
+    return model_pretrained.get(_clean_tag(tag), {})
+
+
+def get_pretrained_url(model: str, tag: str):
+    cfg = get_pretrained_cfg(model, _clean_tag(tag))
+    return cfg.get('url', '')
+
+
+def download_pretrained_from_url(
+        url: str,
+        cache_dir: Union[str, None] = None,
+):
+    if not cache_dir:
+        cache_dir = os.path.expanduser("~/.cache/clip")
+    os.makedirs(cache_dir, exist_ok=True)
+    filename = os.path.basename(url)
+
+    if 'openaipublic' in url:
+        expected_sha256 = url.split("/")[-2]
+    elif 'mlfoundations' in url:
+        expected_sha256 = os.path.splitext(filename)[0].split("-")[-1]
+    else:
+        expected_sha256 = ''
+
+    download_target = os.path.join(cache_dir, 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().startswith(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.headers.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 not hashlib.sha256(open(download_target, "rb").read()).hexdigest().startswith(expected_sha256):
+        raise RuntimeError(f"Model has been downloaded but the SHA256 checksum does not not match")
+
+    return download_target
+
+
+def has_hf_hub(necessary=False):
+    if not _has_hf_hub and necessary:
+        # if no HF Hub module installed, and it is necessary to continue, raise error
+        raise RuntimeError(
+            'Hugging Face hub model specified but package not installed. Run `pip install huggingface_hub`.')
+    return _has_hf_hub
+
+
+def download_pretrained_from_hf(
+        model_id: str,
+        filename: str = 'open_clip_pytorch_model.bin',
+        revision=None,
+        cache_dir: Union[str, None] = None,
+):
+    has_hf_hub(True)
+    cached_file = hf_hub_download(model_id, filename, revision=revision, cache_dir=cache_dir)
+    return cached_file
+
+
+def download_pretrained(
+        cfg: Dict,
+        force_hf_hub: bool = False,
+        cache_dir: Union[str, None] = None,
+):
+    target = ''
+    if not cfg:
+        return target
+
+    download_url = cfg.get('url', '')
+    download_hf_hub = cfg.get('hf_hub', '')
+    if download_hf_hub and force_hf_hub:
+        # use HF hub even if url exists
+        download_url = ''
+
+    if download_url:
+        target = download_pretrained_from_url(download_url, cache_dir=cache_dir)
+    elif download_hf_hub:
+        has_hf_hub(True)
+        # we assume the hf_hub entries in pretrained config combine model_id + filename in
+        # 'org/model_name/filename.pt' form. To specify just the model id w/o filename and
+        # use 'open_clip_pytorch_model.bin' default, there must be a trailing slash 'org/model_name/'.
+        model_id, filename = os.path.split(download_hf_hub)
+        if filename:
+            target = download_pretrained_from_hf(model_id, filename=filename, cache_dir=cache_dir)
+        else:
+            target = download_pretrained_from_hf(model_id, cache_dir=cache_dir)
+
+    return target

eva_clip/rope.py

@@ -0,0 +1,137 @@
+from math import pi
+import torch
+from torch import nn
+from einops import rearrange, repeat
+import logging
+
+def broadcat(tensors, dim = -1):
+    num_tensors = len(tensors)
+    shape_lens = set(list(map(lambda t: len(t.shape), tensors)))
+    assert len(shape_lens) == 1, 'tensors must all have the same number of dimensions'
+    shape_len = list(shape_lens)[0]
+    dim = (dim + shape_len) if dim < 0 else dim
+    dims = list(zip(*map(lambda t: list(t.shape), tensors)))
+    expandable_dims = [(i, val) for i, val in enumerate(dims) if i != dim]
+    assert all([*map(lambda t: len(set(t[1])) <= 2, expandable_dims)]), 'invalid dimensions for broadcastable concatentation'
+    max_dims = list(map(lambda t: (t[0], max(t[1])), expandable_dims))
+    expanded_dims = list(map(lambda t: (t[0], (t[1],) * num_tensors), max_dims))
+    expanded_dims.insert(dim, (dim, dims[dim]))
+    expandable_shapes = list(zip(*map(lambda t: t[1], expanded_dims)))
+    tensors = list(map(lambda t: t[0].expand(*t[1]), zip(tensors, expandable_shapes)))
+    return torch.cat(tensors, dim = dim)
+
+def rotate_half(x):
+    x = rearrange(x, '... (d r) -> ... d r', r = 2)
+    x1, x2 = x.unbind(dim = -1)
+    x = torch.stack((-x2, x1), dim = -1)
+    return rearrange(x, '... d r -> ... (d r)')
+
+
+class VisionRotaryEmbedding(nn.Module):
+    def __init__(
+        self,
+        dim,
+        pt_seq_len,
+        ft_seq_len=None,
+        custom_freqs = None,
+        freqs_for = 'lang',
+        theta = 10000,
+        max_freq = 10,
+        num_freqs = 1,
+    ):
+        super().__init__()
+        if custom_freqs:
+            freqs = custom_freqs
+        elif freqs_for == 'lang':
+            freqs = 1. / (theta ** (torch.arange(0, dim, 2)[:(dim // 2)].float() / dim))
+        elif freqs_for == 'pixel':
+            freqs = torch.linspace(1., max_freq / 2, dim // 2) * pi
+        elif freqs_for == 'constant':
+            freqs = torch.ones(num_freqs).float()
+        else:
+            raise ValueError(f'unknown modality {freqs_for}')
+
+        if ft_seq_len is None: ft_seq_len = pt_seq_len
+        t = torch.arange(ft_seq_len) / ft_seq_len * pt_seq_len
+
+        freqs_h = torch.einsum('..., f -> ... f', t, freqs)
+        freqs_h = repeat(freqs_h, '... n -> ... (n r)', r = 2)
+
+        freqs_w = torch.einsum('..., f -> ... f', t, freqs)
+        freqs_w = repeat(freqs_w, '... n -> ... (n r)', r = 2)
+
+        freqs = broadcat((freqs_h[:, None, :], freqs_w[None, :, :]), dim = -1) 
+
+        self.register_buffer("freqs_cos", freqs.cos())
+        self.register_buffer("freqs_sin", freqs.sin())
+
+        logging.info(f'Shape of rope freq: {self.freqs_cos.shape}')
+
+    def forward(self, t, start_index = 0):
+        rot_dim = self.freqs_cos.shape[-1]
+        end_index = start_index + rot_dim
+        assert rot_dim <= t.shape[-1], f'feature dimension {t.shape[-1]} is not of sufficient size to rotate in all the positions {rot_dim}'
+        t_left, t, t_right = t[..., :start_index], t[..., start_index:end_index], t[..., end_index:]
+        t = (t * self.freqs_cos) + (rotate_half(t) * self.freqs_sin)
+
+        return torch.cat((t_left, t, t_right), dim = -1)
+
+class VisionRotaryEmbeddingFast(nn.Module):
+    def __init__(
+        self,
+        dim,
+        pt_seq_len,
+        ft_seq_len=None,
+        custom_freqs = None,
+        freqs_for = 'lang',
+        theta = 10000,
+        max_freq = 10,
+        num_freqs = 1,
+        patch_dropout = 0.
+    ):
+        super().__init__()
+        if custom_freqs:
+            freqs = custom_freqs
+        elif freqs_for == 'lang':
+            freqs = 1. / (theta ** (torch.arange(0, dim, 2)[:(dim // 2)].float() / dim))
+        elif freqs_for == 'pixel':
+            freqs = torch.linspace(1., max_freq / 2, dim // 2) * pi
+        elif freqs_for == 'constant':
+            freqs = torch.ones(num_freqs).float()
+        else:
+            raise ValueError(f'unknown modality {freqs_for}')
+
+        if ft_seq_len is None: ft_seq_len = pt_seq_len
+        t = torch.arange(ft_seq_len) / ft_seq_len * pt_seq_len
+
+        freqs = torch.einsum('..., f -> ... f', t, freqs)
+        freqs = repeat(freqs, '... n -> ... (n r)', r = 2)
+        freqs = broadcat((freqs[:, None, :], freqs[None, :, :]), dim = -1)
+
+        freqs_cos = freqs.cos().view(-1, freqs.shape[-1])
+        freqs_sin = freqs.sin().view(-1, freqs.shape[-1])
+
+        self.patch_dropout = patch_dropout
+
+        self.register_buffer("freqs_cos", freqs_cos)
+        self.register_buffer("freqs_sin", freqs_sin)
+
+        logging.info(f'Shape of rope freq: {self.freqs_cos.shape}')
+
+    def forward(self, t, patch_indices_keep=None):
+        if patch_indices_keep is not None:
+            batch = t.size()[0]
+            batch_indices = torch.arange(batch)
+            batch_indices = batch_indices[..., None]
+
+            freqs_cos = repeat(self.freqs_cos, 'i j -> n i m j', n=t.shape[0], m=t.shape[1])
+            freqs_sin = repeat(self.freqs_sin, 'i j -> n i m j', n=t.shape[0], m=t.shape[1])
+
+            freqs_cos = freqs_cos[batch_indices, patch_indices_keep]
+            freqs_cos = rearrange(freqs_cos, 'n i m j -> n m i j')
+            freqs_sin = freqs_sin[batch_indices, patch_indices_keep]
+            freqs_sin = rearrange(freqs_sin, 'n i m j -> n m i j')
+
+            return  t * freqs_cos + rotate_half(t) * freqs_sin
+
+        return  t * self.freqs_cos + rotate_half(t) * self.freqs_sin

eva_clip/timm_model.py

@@ -0,0 +1,122 @@
+""" timm model adapter
+
+Wraps timm (https://github.com/rwightman/pytorch-image-models) models for use as a vision tower in CLIP model.
+"""
+import logging
+from collections import OrderedDict
+
+import torch
+import torch.nn as nn
+
+try:
+    import timm
+    from timm.models.layers import Mlp, to_2tuple
+    try:
+        # old timm imports < 0.8.1
+        from timm.models.layers.attention_pool2d import RotAttentionPool2d
+        from timm.models.layers.attention_pool2d import AttentionPool2d as AbsAttentionPool2d
+    except ImportError:
+        # new timm imports >= 0.8.1
+        from timm.layers import RotAttentionPool2d
+        from timm.layers 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',
+            proj_bias=False,
+            drop=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, bias=proj_bias)
+        elif proj == 'mlp':
+            head_layers['mlp'] = Mlp(prev_chs, 2 * embed_dim, embed_dim, drop=drop, bias=(True, proj_bias))
+
+        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)
+
+    @torch.jit.ignore
+    def set_grad_checkpointing(self, enable=True):
+        try:
+            self.trunk.set_grad_checkpointing(enable)
+        except Exception as e:
+            logging.warning('grad checkpointing not supported for this timm image tower, continuing without...')
+
+    def forward(self, x):
+        x = self.trunk(x)
+        x = self.head(x)
+        return x

eva_clip/tokenizer.py

@@ -0,0 +1,201 @@
+""" 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
+
+# https://stackoverflow.com/q/62691279
+import os
+os.environ["TOKENIZERS_PARALLELISM"] = "false"
+
+
+@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
+            tokens[-1] = eot_token
+        result[i, :len(tokens)] = torch.tensor(tokens)
+
+    return result
+
+
+class HFTokenizer:
+    "HuggingFace tokenizer wrapper"
+    def __init__(self, tokenizer_name:str):
+        from transformers import AutoTokenizer
+        self.tokenizer = AutoTokenizer.from_pretrained(tokenizer_name)
+
+    def __call__(self, texts:Union[str, List[str]], context_length:int=77) -> torch.Tensor:
+        # same cleaning as for default tokenizer, except lowercasing
+        # adding lower (for case-sensitive tokenizers) will make it more robust but less sensitive to nuance
+        if isinstance(texts, str):
+            texts = [texts]
+        texts = [whitespace_clean(basic_clean(text)) for text in texts]
+        input_ids = self.tokenizer(texts, return_tensors='pt', max_length=context_length, padding='max_length', truncation=True).input_ids
+        return input_ids

eva_clip/transform.py

@@ -0,0 +1,103 @@
+from typing import Optional, Sequence, Tuple
+
+import torch
+import torch.nn as nn
+import torchvision.transforms.functional as F
+
+from torchvision.transforms import Normalize, Compose, RandomResizedCrop, InterpolationMode, ToTensor, Resize, \
+    CenterCrop
+
+from .constants import OPENAI_DATASET_MEAN, OPENAI_DATASET_STD
+
+
+class ResizeMaxSize(nn.Module):
+
+    def __init__(self, max_size, interpolation=InterpolationMode.BICUBIC, fn='max', fill=0):
+        super().__init__()
+        if not isinstance(max_size, int):
+            raise TypeError(f"Size should be int. Got {type(max_size)}")
+        self.max_size = max_size
+        self.interpolation = interpolation
+        self.fn = min if fn == 'min' else min
+        self.fill = fill
+
+    def forward(self, img):
+        if isinstance(img, torch.Tensor):
+            height, width = img.shape[:2]
+        else:
+            width, height = img.size
+        scale = self.max_size / float(max(height, width))
+        if scale != 1.0:
+            new_size = tuple(round(dim * scale) for dim in (height, width))
+            img = F.resize(img, new_size, self.interpolation)
+            pad_h = self.max_size - new_size[0]
+            pad_w = self.max_size - new_size[1]
+            img = F.pad(img, padding=[pad_w//2, pad_h//2, pad_w - pad_w//2, pad_h - pad_h//2], fill=self.fill)
+        return img
+
+
+def _convert_to_rgb(image):
+    return image.convert('RGB')
+
+
+# class CatGen(nn.Module):
+#     def __init__(self, num=4):
+#         self.num = num
+#     def mixgen_batch(image, text):
+#         batch_size = image.shape[0]
+#         index = np.random.permutation(batch_size)
+
+#         cat_images = []
+#         for i in range(batch_size):
+#             # image mixup
+#             image[i,:] = lam * image[i,:] + (1 - lam) * image[index[i],:]
+#             # text concat
+#             text[i] = tokenizer((str(text[i]) + " " + str(text[index[i]])))[0]
+#         text = torch.stack(text)
+#         return image, text
+
+
+def image_transform(
+        image_size: int,
+        is_train: bool,
+        mean: Optional[Tuple[float, ...]] = None,
+        std: Optional[Tuple[float, ...]] = None,
+        resize_longest_max: bool = False,
+        fill_color: int = 0,
+):
+    mean = mean or OPENAI_DATASET_MEAN
+    if not isinstance(mean, (list, tuple)):
+        mean = (mean,) * 3
+
+    std = std or OPENAI_DATASET_STD
+    if not isinstance(std, (list, tuple)):
+        std = (std,) * 3
+
+    if isinstance(image_size, (list, tuple)) and image_size[0] == image_size[1]:
+        # for square size, pass size as int so that Resize() uses aspect preserving shortest edge
+        image_size = image_size[0]
+
+    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:
+        if resize_longest_max:
+            transforms = [
+                ResizeMaxSize(image_size, fill=fill_color)
+            ]
+        else:
+            transforms = [
+                Resize(image_size, interpolation=InterpolationMode.BICUBIC),
+                CenterCrop(image_size),
+            ]
+        transforms.extend([
+            _convert_to_rgb,
+            ToTensor(),
+            normalize,
+        ])
+        return Compose(transforms)

eva_clip/transformer.py

@@ -0,0 +1,737 @@
+import os
+import logging
+from collections import OrderedDict
+import math
+from typing import Callable, Optional, Sequence
+import numpy as np
+import torch
+from torch import nn
+from torch.nn import functional as F
+
+try:
+    from timm.models.layers import trunc_normal_
+except:
+    from timm.layers import trunc_normal_
+    
+from .rope import VisionRotaryEmbedding, VisionRotaryEmbeddingFast
+from .utils import to_2tuple
+
+if os.getenv('ENV_TYPE') == 'deepspeed':
+    try:
+        import deepspeed
+        from deepspeed.runtime.activation_checkpointing.checkpointing import checkpoint
+    except:
+        print("Please 'pip install deepspeed'")
+        deepspeed = None
+        from torch.utils.checkpoint import checkpoint
+else:
+    from torch.utils.checkpoint import checkpoint
+
+try:
+    import xformers.ops as xops
+except ImportError:
+    xops = None
+    print("Please 'pip install xformers'")
+
+class LayerNormFp32(nn.LayerNorm):
+    """Subclass torch's LayerNorm to handle fp16 (by casting to float32 and back)."""
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+    def forward(self, x: torch.Tensor):
+        output = F.layer_norm(
+            x.float(),
+            self.normalized_shape,
+            self.weight.float() if self.weight is not None else None,
+            self.bias.float() if self.bias is not None else None,
+            self.eps,
+        )
+        return output.type_as(x)
+
+
+class LayerNorm(nn.LayerNorm):
+    """Subclass torch's LayerNorm (with cast back to input dtype)."""
+
+    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 LayerScale(nn.Module):
+    def __init__(self, dim, init_values=1e-5, inplace=False):
+        super().__init__()
+        self.inplace = inplace
+        self.gamma = nn.Parameter(init_values * torch.ones(dim))
+
+    def forward(self, x):
+        return x.mul_(self.gamma) if self.inplace else x * self.gamma
+
+class PatchDropout(nn.Module):
+    """
+    https://arxiv.org/abs/2212.00794
+    """
+
+    def __init__(self, prob, exclude_first_token=True):
+        super().__init__()
+        assert 0 <= prob < 1.
+        self.prob = prob
+        self.exclude_first_token = exclude_first_token  # exclude CLS token
+        logging.info(f"os.getenv('RoPE')={os.getenv('RoPE')}")
+
+    def forward(self, x):
+        if not self.training or self.prob == 0.:
+            return x
+
+        if self.exclude_first_token:
+            cls_tokens, x = x[:, :1], x[:, 1:]
+        else:
+            cls_tokens = torch.jit.annotate(torch.Tensor, x[:, :1])
+
+        batch = x.size()[0]
+        num_tokens = x.size()[1]
+
+        batch_indices = torch.arange(batch)
+        batch_indices = batch_indices[..., None]
+
+        keep_prob = 1 - self.prob
+        num_patches_keep = max(1, int(num_tokens * keep_prob))
+
+        rand = torch.randn(batch, num_tokens)
+        patch_indices_keep = rand.topk(num_patches_keep, dim=-1).indices
+
+        x = x[batch_indices, patch_indices_keep]
+
+        if self.exclude_first_token:
+            x = torch.cat((cls_tokens, x), dim=1)
+
+        if self.training and os.getenv('RoPE') == '1':
+            return x, patch_indices_keep
+
+        return x
+
+
+def _in_projection_packed(
+    q: torch.Tensor,
+    k: torch.Tensor,
+    v: torch.Tensor,
+    w: torch.Tensor,
+    b: Optional[torch.Tensor] = None,
+    ):
+    """
+    https://github.com/pytorch/pytorch/blob/db2a237763eb8693a20788be94f8c192e762baa8/torch/nn/functional.py#L4726
+    """
+    E = q.size(-1)
+    if k is v:
+        if q is k:
+            # self-attention
+            return F.linear(q, w, b).chunk(3, dim=-1)
+        else:
+            # encoder-decoder attention
+            w_q, w_kv = w.split([E, E * 2])
+            if b is None:
+                b_q = b_kv = None
+            else:
+                b_q, b_kv = b.split([E, E * 2])
+            return (F.linear(q, w_q, b_q),) + F.linear(k, w_kv, b_kv).chunk(2, dim=-1)
+    else:
+        w_q, w_k, w_v = w.chunk(3)
+        if b is None:
+            b_q = b_k = b_v = None
+        else:
+            b_q, b_k, b_v = b.chunk(3)
+        return F.linear(q, w_q, b_q), F.linear(k, w_k, b_k), F.linear(v, w_v, b_v)
+
+class Attention(nn.Module):
+    def __init__(
+            self,
+            dim,
+            num_heads=8,
+            qkv_bias=True,
+            scaled_cosine=False,
+            scale_heads=False,
+            logit_scale_max=math.log(1. / 0.01),
+            attn_drop=0.,
+            proj_drop=0.,
+            xattn=False,
+            rope=False
+    ):
+        super().__init__()
+        self.scaled_cosine = scaled_cosine
+        self.scale_heads = scale_heads
+        assert dim % num_heads == 0, 'dim should be divisible by num_heads'
+        self.num_heads = num_heads
+        self.head_dim = dim // num_heads
+        self.scale = self.head_dim ** -0.5
+        self.logit_scale_max = logit_scale_max
+
+        # keeping in_proj in this form (instead of nn.Linear) to match weight scheme of original
+        self.in_proj_weight = nn.Parameter(torch.randn((dim * 3, dim)) * self.scale)
+        if qkv_bias:
+            self.in_proj_bias = nn.Parameter(torch.zeros(dim * 3))
+        else:
+            self.in_proj_bias = None
+
+        if self.scaled_cosine:
+            self.logit_scale = nn.Parameter(torch.log(10 * torch.ones((num_heads, 1, 1))))
+        else:
+            self.logit_scale = None
+        self.attn_drop = nn.Dropout(attn_drop)
+        if self.scale_heads:
+            self.head_scale = nn.Parameter(torch.ones((num_heads, 1, 1)))
+        else:
+            self.head_scale = None
+        self.out_proj = nn.Linear(dim, dim)
+        self.out_drop = nn.Dropout(proj_drop)
+        self.xattn = xattn
+        self.xattn_drop = attn_drop
+        self.rope = rope
+
+    def forward(self, x, attn_mask: Optional[torch.Tensor] = None):
+        L, N, C = x.shape
+        q, k, v = F.linear(x, self.in_proj_weight, self.in_proj_bias).chunk(3, dim=-1)
+        if self.xattn:
+            q = q.contiguous().view(L, N, self.num_heads, -1).transpose(0, 1)
+            k = k.contiguous().view(L, N, self.num_heads, -1).transpose(0, 1)
+            v = v.contiguous().view(L, N, self.num_heads, -1).transpose(0, 1)
+
+            x = xops.memory_efficient_attention(
+                q, k, v,
+                p=self.xattn_drop,
+                scale=self.scale if self.logit_scale is None else None,
+                attn_bias=xops.LowerTriangularMask() if attn_mask is not None else None,
+                )
+        else:
+            q = q.contiguous().view(L, N * self.num_heads, -1).transpose(0, 1)
+            k = k.contiguous().view(L, N * self.num_heads, -1).transpose(0, 1)
+            v = v.contiguous().view(L, N * self.num_heads, -1).transpose(0, 1)
+
+            if self.logit_scale is not None:
+                attn = torch.bmm(F.normalize(q, dim=-1), F.normalize(k, dim=-1).transpose(-1, -2))
+                logit_scale = torch.clamp(self.logit_scale, max=self.logit_scale_max).exp()
+                attn = attn.view(N, self.num_heads, L, L) * logit_scale
+                attn = attn.view(-1, L, L)
+            else:
+                q = q * self.scale
+                attn = torch.bmm(q, k.transpose(-1, -2))
+
+            if attn_mask is not None:
+                if attn_mask.dtype == torch.bool:
+                    new_attn_mask = torch.zeros_like(attn_mask, dtype=q.dtype)
+                    new_attn_mask.masked_fill_(attn_mask, float("-inf"))
+                    attn_mask = new_attn_mask
+                attn += attn_mask
+
+            attn = attn.softmax(dim=-1)
+            attn = self.attn_drop(attn)
+
+            x = torch.bmm(attn, v)
+
+        if self.head_scale is not None:
+            x = x.view(N, self.num_heads, L, C) * self.head_scale
+            x = x.view(-1, L, C)
+        x = x.transpose(0, 1).reshape(L, N, C)
+        x = self.out_proj(x)
+        x = self.out_drop(x)
+        return x
+
+class CustomAttention(nn.Module):
+    def __init__(
+            self,
+            dim,
+            num_heads=8,
+            qkv_bias=True,
+            scaled_cosine=True,
+            scale_heads=False,
+            logit_scale_max=math.log(1. / 0.01),
+            attn_drop=0.,
+            proj_drop=0.,
+            xattn=False
+    ):
+        super().__init__()
+        self.scaled_cosine = scaled_cosine
+        self.scale_heads = scale_heads
+        assert dim % num_heads == 0, 'dim should be divisible by num_heads'
+        self.num_heads = num_heads
+        self.head_dim = dim // num_heads
+        self.scale = self.head_dim ** -0.5
+        self.logit_scale_max = logit_scale_max
+
+        # keeping in_proj in this form (instead of nn.Linear) to match weight scheme of original
+        self.in_proj_weight = nn.Parameter(torch.randn((dim * 3, dim)) * self.scale)
+        if qkv_bias:
+            self.in_proj_bias = nn.Parameter(torch.zeros(dim * 3))
+        else:
+            self.in_proj_bias = None
+
+        if self.scaled_cosine:
+            self.logit_scale = nn.Parameter(torch.log(10 * torch.ones((num_heads, 1, 1))))
+        else:
+            self.logit_scale = None
+        self.attn_drop = nn.Dropout(attn_drop)
+        if self.scale_heads:
+            self.head_scale = nn.Parameter(torch.ones((num_heads, 1, 1)))
+        else:
+            self.head_scale = None
+        self.out_proj = nn.Linear(dim, dim)
+        self.out_drop = nn.Dropout(proj_drop)
+        self.xattn = xattn
+        self.xattn_drop = attn_drop
+
+    def forward(self, query: torch.Tensor, key: torch.Tensor, value: torch.Tensor, attn_mask: Optional[torch.Tensor] = None):
+        q, k, v = _in_projection_packed(query, key, value, self.in_proj_weight, self.in_proj_bias)
+        N_q, B_q, C_q = q.shape
+        N_k, B_k, C_k = k.shape
+        N_v, B_v, C_v = v.shape
+        if self.xattn:
+            # B, N, C -> B, N, num_heads, C
+            q = q.permute(1, 0, 2).reshape(B_q, N_q, self.num_heads, -1)
+            k = k.permute(1, 0, 2).reshape(B_k, N_k, self.num_heads, -1)
+            v = v.permute(1, 0, 2).reshape(B_v, N_v, self.num_heads, -1)
+
+            x = xops.memory_efficient_attention(
+                q, k, v,
+                p=self.xattn_drop,
+                scale=self.scale if self.logit_scale is None else None,
+                attn_bias=xops.LowerTriangularMask() if attn_mask is not None else None
+                )
+        else:
+            # B*H, L, C
+            q = q.contiguous().view(N_q, B_q * self.num_heads, -1).transpose(0, 1)
+            k = k.contiguous().view(N_k, B_k * self.num_heads, -1).transpose(0, 1)
+            v = v.contiguous().view(N_v, B_v * self.num_heads, -1).transpose(0, 1)
+
+            if self.logit_scale is not None:
+                # B*H, N_q, N_k
+                attn = torch.bmm(F.normalize(q, dim=-1), F.normalize(k, dim=-1).transpose(-1, -2))
+                logit_scale = torch.clamp(self.logit_scale, max=self.logit_scale_max).exp()
+                attn = attn.view(B_q, self.num_heads, N_q, N_k) * logit_scale
+                attn = attn.view(-1, N_q, N_k)
+            else:
+                q = q * self.scale
+                attn = torch.bmm(q, k.transpose(-1, -2))
+
+            if attn_mask is not None:
+                if attn_mask.dtype == torch.bool:
+                    new_attn_mask = torch.zeros_like(attn_mask, dtype=q.dtype)
+                    new_attn_mask.masked_fill_(attn_mask, float("-inf"))
+                    attn_mask = new_attn_mask
+                attn += attn_mask
+
+            attn = attn.softmax(dim=-1)
+            attn = self.attn_drop(attn)
+
+            x = torch.bmm(attn, v)
+            
+        if self.head_scale is not None:
+            x = x.view(B_q, self.num_heads, N_q, C_q) * self.head_scale
+            x = x.view(-1, N_q, C_q)
+        x = x.transpose(0, 1).reshape(N_q, B_q, C_q)
+        x = self.out_proj(x)
+        x = self.out_drop(x)
+        return x
+
+class CustomResidualAttentionBlock(nn.Module):
+    def __init__(
+            self,
+            d_model: int,
+            n_head: int,
+            mlp_ratio: float = 4.0,
+            ls_init_value: float = None,
+            act_layer: Callable = nn.GELU,
+            norm_layer: Callable = LayerNorm,
+            scale_cosine_attn: bool = False,
+            scale_heads: bool = False,
+            scale_attn: bool = False,
+            scale_fc: bool = False,
+            cross_attn: bool = False,
+            xattn: bool = False,
+    ):
+        super().__init__()
+
+        self.ln_1 = norm_layer(d_model)
+        self.ln_1_k = norm_layer(d_model) if cross_attn else self.ln_1
+        self.ln_1_v = norm_layer(d_model) if cross_attn else self.ln_1
+        self.attn = CustomAttention(
+            d_model, n_head,
+            qkv_bias=True,
+            attn_drop=0.,
+            proj_drop=0.,
+            scaled_cosine=scale_cosine_attn,
+            scale_heads=scale_heads,
+            xattn=xattn
+        )
+
+        self.ln_attn = norm_layer(d_model) if scale_attn else nn.Identity()
+        self.ls_1 = LayerScale(d_model, ls_init_value) if ls_init_value is not None else nn.Identity()
+
+        self.ln_2 = norm_layer(d_model)
+        mlp_width = int(d_model * mlp_ratio)
+        self.mlp = nn.Sequential(OrderedDict([
+            ("c_fc", nn.Linear(d_model, mlp_width)),
+            ('ln', norm_layer(mlp_width) if scale_fc else nn.Identity()),
+            ("gelu", act_layer()),
+            ("c_proj", nn.Linear(mlp_width, d_model))
+        ]))
+
+        self.ls_2 = LayerScale(d_model, ls_init_value) if ls_init_value is not None else nn.Identity()
+
+    def forward(self, q: torch.Tensor, k: torch.Tensor, v: torch.Tensor, attn_mask: Optional[torch.Tensor] = None):
+        q = q + self.ls_1(self.ln_attn(self.attn(self.ln_1(q), self.ln_1_k(k), self.ln_1_v(v), attn_mask=attn_mask)))
+        q = q + self.ls_2(self.mlp(self.ln_2(q)))
+        return q
+
+class CustomTransformer(nn.Module):
+    def __init__(
+            self,
+            width: int,
+            layers: int,
+            heads: int,
+            mlp_ratio: float = 4.0,
+            ls_init_value: float = None,
+            act_layer: Callable = nn.GELU,
+            norm_layer: Callable = LayerNorm,
+            scale_cosine_attn: bool = True,
+            scale_heads: bool = False,
+            scale_attn: bool = False,
+            scale_fc: bool = False,
+            cross_attn: bool = False,
+            xattn: bool = False,
+    ):
+        super().__init__()
+        self.width = width
+        self.layers = layers
+        self.grad_checkpointing = False
+        self.xattn = xattn
+
+        self.resblocks = nn.ModuleList([
+            CustomResidualAttentionBlock(
+                width,
+                heads,
+                mlp_ratio,
+                ls_init_value=ls_init_value,
+                act_layer=act_layer,
+                norm_layer=norm_layer,
+                scale_cosine_attn=scale_cosine_attn,
+                scale_heads=scale_heads,
+                scale_attn=scale_attn,
+                scale_fc=scale_fc,
+                cross_attn=cross_attn,
+                xattn=xattn)
+            for _ in range(layers)
+        ])
+
+    def get_cast_dtype(self) -> torch.dtype:
+        return self.resblocks[0].mlp.c_fc.weight.dtype 
+
+    def forward(self, q: torch.Tensor, k: torch.Tensor = None, v: torch.Tensor = None, attn_mask: Optional[torch.Tensor] = None):
+        if k is None and v is None:
+            k = v = q
+        for r in self.resblocks:
+            if self.grad_checkpointing and not torch.jit.is_scripting():
+                q = checkpoint(r, q, k, v, attn_mask)
+            else:
+                q = r(q, k, v, attn_mask=attn_mask)
+        return q
+
+
+class ResidualAttentionBlock(nn.Module):
+    def __init__(
+            self,
+            d_model: int,
+            n_head: int,
+            mlp_ratio: float = 4.0,
+            ls_init_value: float = None,
+            act_layer: Callable = nn.GELU,
+            norm_layer: Callable = LayerNorm,
+            xattn: bool = False,
+    ):
+        super().__init__()
+
+        self.ln_1 = norm_layer(d_model)
+        if xattn:
+            self.attn = Attention(d_model, n_head, xattn=True)
+        else:
+            self.attn = nn.MultiheadAttention(d_model, n_head)
+        self.ls_1 = LayerScale(d_model, ls_init_value) if ls_init_value is not None else nn.Identity()
+
+        self.ln_2 = norm_layer(d_model)
+        mlp_width = int(d_model * mlp_ratio)
+        self.mlp = nn.Sequential(OrderedDict([
+            ("c_fc", nn.Linear(d_model, mlp_width)),
+            ("gelu", act_layer()),
+            ("c_proj", nn.Linear(mlp_width, d_model))
+        ]))
+
+        self.ls_2 = LayerScale(d_model, ls_init_value) if ls_init_value is not None else nn.Identity()
+        self.xattn = xattn
+
+    def attention(self, x: torch.Tensor, attn_mask: Optional[torch.Tensor] = None):
+        attn_mask = attn_mask.to(x.dtype) if attn_mask is not None else None
+        if self.xattn:
+            return self.attn(x, attn_mask=attn_mask)
+        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.ls_1(self.attention(self.ln_1(x), attn_mask=attn_mask))
+        x = x + self.ls_2(self.mlp(self.ln_2(x)))
+        return x
+
+class Transformer(nn.Module):
+    def __init__(
+            self,
+            width: int,
+            layers: int,
+            heads: int,
+            mlp_ratio: float = 4.0,
+            ls_init_value: float = None,
+            act_layer: Callable = nn.GELU,
+            norm_layer: Callable = LayerNorm,
+            xattn: bool = False,
+    ):
+        super().__init__()
+        self.width = width
+        self.layers = layers
+        self.grad_checkpointing = False
+
+        self.resblocks = nn.ModuleList([
+            ResidualAttentionBlock(
+                width, heads, mlp_ratio, ls_init_value=ls_init_value, act_layer=act_layer, norm_layer=norm_layer, xattn=xattn)
+            for _ in range(layers)
+        ])
+
+    def get_cast_dtype(self) -> torch.dtype:
+        return self.resblocks[0].mlp.c_fc.weight.dtype
+
+    def forward(self, x: torch.Tensor, attn_mask: Optional[torch.Tensor] = None):
+        for r in self.resblocks:
+            if self.grad_checkpointing and not torch.jit.is_scripting():
+                x = checkpoint(r, x, attn_mask)
+            else:
+                x = r(x, attn_mask=attn_mask)
+        return x
+
+
+class VisionTransformer(nn.Module):
+    def __init__(
+            self,
+            image_size: int,
+            patch_size: int,
+            width: int,
+            layers: int,
+            heads: int,
+            mlp_ratio: float,
+            ls_init_value: float = None,
+            patch_dropout: float = 0.,
+            global_average_pool: bool = False,
+            output_dim: int = 512,
+            act_layer: Callable = nn.GELU,
+            norm_layer: Callable = LayerNorm,
+            xattn: bool = False,
+    ):
+        super().__init__()
+        self.image_size = to_2tuple(image_size)
+        self.patch_size = to_2tuple(patch_size)
+        self.grid_size = (self.image_size[0] // self.patch_size[0], self.image_size[1] // self.patch_size[1])
+        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(self.grid_size[0] * self.grid_size[1] + 1, width))
+
+        # setting a patch_dropout of 0. would mean it is disabled and this function would be the identity fn
+        self.patch_dropout = PatchDropout(patch_dropout) if patch_dropout > 0. else nn.Identity()
+        self.ln_pre = norm_layer(width)
+        
+        self.transformer = Transformer(
+            width,
+            layers,
+            heads,
+            mlp_ratio,
+            ls_init_value=ls_init_value,
+            act_layer=act_layer,
+            norm_layer=norm_layer,
+            xattn=xattn
+        )
+
+        self.global_average_pool = global_average_pool
+        self.ln_post = norm_layer(width)
+        self.proj = nn.Parameter(scale * torch.randn(width, output_dim))
+
+    def lock(self, unlocked_groups=0, freeze_bn_stats=False):
+        for param in self.parameters():
+            param.requires_grad = False
+        
+        if unlocked_groups != 0:
+            groups = [
+                [
+                    self.conv1,
+                    self.class_embedding,
+                    self.positional_embedding,
+                    self.ln_pre,
+                ],
+                *self.transformer.resblocks[:-1],
+                [
+                    self.transformer.resblocks[-1],
+                    self.ln_post,
+                ],
+                self.proj,
+            ]
+
+            def _unlock(x):
+                if isinstance(x, Sequence):
+                    for g in x:
+                        _unlock(g)
+                else:
+                    if isinstance(x, torch.nn.Parameter):
+                        x.requires_grad = True
+                    else:
+                        for p in x.parameters():
+                            p.requires_grad = True
+
+            _unlock(groups[-unlocked_groups:])
+
+    def get_num_layers(self):
+        return self.transformer.layers
+
+    @torch.jit.ignore
+    def set_grad_checkpointing(self, enable=True):
+        self.transformer.grad_checkpointing = enable
+
+    @torch.jit.ignore
+    def no_weight_decay(self):
+        return {'positional_embedding', 'class_embedding'}
+
+    def forward(self, x: torch.Tensor, return_all_features: bool=False):
+        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)
+
+        # a patch_dropout of 0. would mean it is disabled and this function would do nothing but return what was passed in
+        x = self.patch_dropout(x)
+        x = self.ln_pre(x)
+
+        x = x.permute(1, 0, 2)  # NLD -> LND
+        x = self.transformer(x)
+        x = x.permute(1, 0, 2)  # LND -> NLD
+
+        if not return_all_features:
+            if self.global_average_pool:
+                x = x.mean(dim=1) #x = x[:,1:,:].mean(dim=1)
+            else:
+                x = x[:, 0]
+
+            x = self.ln_post(x)
+
+            if self.proj is not None:
+                x = x @ self.proj
+
+        return x
+
+
+class TextTransformer(nn.Module):
+    def __init__(
+            self,
+            context_length: int = 77,
+            vocab_size: int = 49408,
+            width: int = 512,
+            heads: int = 8,
+            layers: int = 12,
+            ls_init_value: float = None,
+            output_dim: int = 512,
+            act_layer: Callable = nn.GELU,
+            norm_layer: Callable = LayerNorm,
+            xattn: bool= False,
+            attn_mask: bool = True
+    ):
+        super().__init__()
+        self.context_length = context_length
+        self.vocab_size = vocab_size
+        self.width = width
+        self.output_dim = output_dim
+
+        self.token_embedding = nn.Embedding(vocab_size, width)
+        self.positional_embedding = nn.Parameter(torch.empty(self.context_length, width))
+        self.transformer = Transformer(
+            width=width,
+            layers=layers,
+            heads=heads,
+            ls_init_value=ls_init_value,
+            act_layer=act_layer,
+            norm_layer=norm_layer,
+            xattn=xattn
+        )
+        
+        self.xattn = xattn
+        self.ln_final = norm_layer(width)
+        self.text_projection = nn.Parameter(torch.empty(width, output_dim))
+
+        if attn_mask:
+            self.register_buffer('attn_mask', self.build_attention_mask(), persistent=False)
+        else:
+            self.attn_mask = None
+
+        self.init_parameters()
+
+    def init_parameters(self):
+        nn.init.normal_(self.token_embedding.weight, std=0.02)
+        nn.init.normal_(self.positional_embedding, std=0.01)
+
+        proj_std = (self.transformer.width ** -0.5) * ((2 * self.transformer.layers) ** -0.5)
+        attn_std = self.transformer.width ** -0.5
+        fc_std = (2 * self.transformer.width) ** -0.5
+        for block in self.transformer.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_projection is not None:
+            nn.init.normal_(self.text_projection, std=self.transformer.width ** -0.5)
+
+    @torch.jit.ignore
+    def set_grad_checkpointing(self, enable=True):
+        self.transformer.grad_checkpointing = enable
+    
+    @torch.jit.ignore
+    def no_weight_decay(self):
+        # return {'positional_embedding', 'token_embedding'}
+        return {'positional_embedding'}
+
+    def get_num_layers(self):
+        return self.transformer.layers
+
+    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 forward(self, text, return_all_features: bool=False):
+        cast_dtype = self.transformer.get_cast_dtype()
+        x = self.token_embedding(text).to(cast_dtype)  # [batch_size, n_ctx, d_model]
+
+        x = x + self.positional_embedding.to(cast_dtype)
+        x = x.permute(1, 0, 2)  # NLD -> LND
+        x = self.transformer(x, attn_mask=self.attn_mask)
+        # x = self.transformer(x) # no attention mask is applied
+        x = x.permute(1, 0, 2)  # LND -> NLD
+        x = self.ln_final(x)
+
+        if not return_all_features:
+            # x.shape = [batch_size, n_ctx, transformer.width]
+            # take features from the eot embedding (eot_token is the highest number in each sequence)
+            x = x[torch.arange(x.shape[0]), text.argmax(dim=-1)] @ self.text_projection
+        return x

eva_clip/utils.py

@@ -0,0 +1,326 @@
+from itertools import repeat
+import collections.abc
+import logging
+import math
+import numpy as np
+
+import torch
+from torch import nn as nn
+from torchvision.ops.misc import FrozenBatchNorm2d
+import torch.nn.functional as F
+
+# open CLIP
+def resize_clip_pos_embed(state_dict, model, interpolation: str = 'bicubic', seq_dim=1):
+    # Rescale the grid of position embeddings when loading from state_dict
+    old_pos_embed = state_dict.get('visual.positional_embedding', None)
+    if old_pos_embed is None or not hasattr(model.visual, 'grid_size'):
+        return
+    grid_size = to_2tuple(model.visual.grid_size)
+    extra_tokens = 1  # FIXME detect different token configs (ie no class token, or more)
+    new_seq_len = grid_size[0] * grid_size[1] + extra_tokens
+    if new_seq_len == old_pos_embed.shape[0]:
+        return
+
+    if extra_tokens:
+        pos_emb_tok, pos_emb_img = old_pos_embed[:extra_tokens], old_pos_embed[extra_tokens:]
+    else:
+        pos_emb_tok, pos_emb_img = None, old_pos_embed
+    old_grid_size = to_2tuple(int(math.sqrt(len(pos_emb_img))))
+
+    logging.info('Resizing position embedding grid-size from %s to %s', old_grid_size, grid_size)
+    pos_emb_img = pos_emb_img.reshape(1, old_grid_size[0], old_grid_size[1], -1).permute(0, 3, 1, 2)
+    pos_emb_img = F.interpolate(
+        pos_emb_img,
+        size=grid_size,
+        mode=interpolation,
+        align_corners=True,
+    )
+    pos_emb_img = pos_emb_img.permute(0, 2, 3, 1).reshape(1, grid_size[0] * grid_size[1], -1)[0]
+    if pos_emb_tok is not None:
+        new_pos_embed = torch.cat([pos_emb_tok, pos_emb_img], dim=0)
+    else:
+        new_pos_embed = pos_emb_img
+    state_dict['visual.positional_embedding'] = new_pos_embed
+
+
+def resize_visual_pos_embed(state_dict, model, interpolation: str = 'bicubic', seq_dim=1):
+    # Rescale the grid of position embeddings when loading from state_dict
+    old_pos_embed = state_dict.get('positional_embedding', None)
+    if old_pos_embed is None or not hasattr(model.visual, 'grid_size'):
+        return
+    grid_size = to_2tuple(model.visual.grid_size)
+    extra_tokens = 1  # FIXME detect different token configs (ie no class token, or more)
+    new_seq_len = grid_size[0] * grid_size[1] + extra_tokens
+    if new_seq_len == old_pos_embed.shape[0]:
+        return
+
+    if extra_tokens:
+        pos_emb_tok, pos_emb_img = old_pos_embed[:extra_tokens], old_pos_embed[extra_tokens:]
+    else:
+        pos_emb_tok, pos_emb_img = None, old_pos_embed
+    old_grid_size = to_2tuple(int(math.sqrt(len(pos_emb_img))))
+
+    logging.info('Resizing position embedding grid-size from %s to %s', old_grid_size, grid_size)
+    pos_emb_img = pos_emb_img.reshape(1, old_grid_size[0], old_grid_size[1], -1).permute(0, 3, 1, 2)
+    pos_emb_img = F.interpolate(
+        pos_emb_img,
+        size=grid_size,
+        mode=interpolation,
+        align_corners=True,
+    )
+    pos_emb_img = pos_emb_img.permute(0, 2, 3, 1).reshape(1, grid_size[0] * grid_size[1], -1)[0]
+    if pos_emb_tok is not None:
+        new_pos_embed = torch.cat([pos_emb_tok, pos_emb_img], dim=0)
+    else:
+        new_pos_embed = pos_emb_img
+    state_dict['positional_embedding'] = new_pos_embed
+
+def resize_evaclip_pos_embed(state_dict, model, interpolation: str = 'bicubic', seq_dim=1):
+    all_keys = list(state_dict.keys())
+    # interpolate position embedding
+    if 'visual.pos_embed' in state_dict:
+        pos_embed_checkpoint = state_dict['visual.pos_embed']
+        embedding_size = pos_embed_checkpoint.shape[-1]
+        num_patches = model.visual.patch_embed.num_patches
+        num_extra_tokens = model.visual.pos_embed.shape[-2] - num_patches
+        # height (== width) for the checkpoint position embedding
+        orig_size = int((pos_embed_checkpoint.shape[-2] - num_extra_tokens) ** 0.5)
+        # height (== width) for the new position embedding
+        new_size = int(num_patches ** 0.5)
+        # class_token and dist_token are kept unchanged
+        if orig_size != new_size:
+            print("Position interpolate from %dx%d to %dx%d" % (orig_size, orig_size, new_size, new_size))
+            extra_tokens = pos_embed_checkpoint[:, :num_extra_tokens]
+            # only the position tokens are interpolated
+            pos_tokens = pos_embed_checkpoint[:, num_extra_tokens:]
+            pos_tokens = pos_tokens.reshape(-1, orig_size, orig_size, embedding_size).permute(0, 3, 1, 2)
+            pos_tokens = torch.nn.functional.interpolate(
+                pos_tokens, size=(new_size, new_size), mode='bicubic', align_corners=False)
+            pos_tokens = pos_tokens.permute(0, 2, 3, 1).flatten(1, 2)
+            new_pos_embed = torch.cat((extra_tokens, pos_tokens), dim=1)
+            state_dict['visual.pos_embed'] = new_pos_embed
+
+            patch_embed_proj = state_dict['visual.patch_embed.proj.weight']
+            patch_size = model.visual.patch_embed.patch_size
+            state_dict['visual.patch_embed.proj.weight'] = torch.nn.functional.interpolate(
+                patch_embed_proj.float(), size=patch_size, mode='bicubic', align_corners=False)
+
+
+def resize_eva_pos_embed(state_dict, model, interpolation: str = 'bicubic', seq_dim=1):
+    all_keys = list(state_dict.keys())
+    # interpolate position embedding
+    if 'pos_embed' in state_dict:
+        pos_embed_checkpoint = state_dict['pos_embed']
+        embedding_size = pos_embed_checkpoint.shape[-1]
+        num_patches = model.visual.patch_embed.num_patches
+        num_extra_tokens = model.visual.pos_embed.shape[-2] - num_patches
+        # height (== width) for the checkpoint position embedding
+        orig_size = int((pos_embed_checkpoint.shape[-2] - num_extra_tokens) ** 0.5)
+        # height (== width) for the new position embedding
+        new_size = int(num_patches ** 0.5)
+        # class_token and dist_token are kept unchanged
+        if orig_size != new_size:
+            print("Position interpolate from %dx%d to %dx%d" % (orig_size, orig_size, new_size, new_size))
+            extra_tokens = pos_embed_checkpoint[:, :num_extra_tokens]
+            # only the position tokens are interpolated
+            pos_tokens = pos_embed_checkpoint[:, num_extra_tokens:]
+            pos_tokens = pos_tokens.reshape(-1, orig_size, orig_size, embedding_size).permute(0, 3, 1, 2)
+            pos_tokens = torch.nn.functional.interpolate(
+                pos_tokens, size=(new_size, new_size), mode='bicubic', align_corners=False)
+            pos_tokens = pos_tokens.permute(0, 2, 3, 1).flatten(1, 2)
+            new_pos_embed = torch.cat((extra_tokens, pos_tokens), dim=1)
+            state_dict['pos_embed'] = new_pos_embed
+
+            patch_embed_proj = state_dict['patch_embed.proj.weight']
+            patch_size = model.visual.patch_embed.patch_size
+            state_dict['patch_embed.proj.weight'] = torch.nn.functional.interpolate(
+                patch_embed_proj.float(), size=patch_size, mode='bicubic', align_corners=False)
+                
+
+def resize_rel_pos_embed(state_dict, model, interpolation: str = 'bicubic', seq_dim=1):
+    all_keys = list(state_dict.keys())
+    for key in all_keys:
+        if "relative_position_index" in key:
+            state_dict.pop(key)
+
+        if "relative_position_bias_table" in key:
+            rel_pos_bias = state_dict[key]
+            src_num_pos, num_attn_heads = rel_pos_bias.size()
+            dst_num_pos, _ = model.visual.state_dict()[key].size()
+            dst_patch_shape = model.visual.patch_embed.patch_shape
+            if dst_patch_shape[0] != dst_patch_shape[1]:
+                raise NotImplementedError()
+            num_extra_tokens = dst_num_pos - (dst_patch_shape[0] * 2 - 1) * (dst_patch_shape[1] * 2 - 1)
+            src_size = int((src_num_pos - num_extra_tokens) ** 0.5)
+            dst_size = int((dst_num_pos - num_extra_tokens) ** 0.5)
+            if src_size != dst_size:
+                print("Position interpolate for %s from %dx%d to %dx%d" % (
+                    key, src_size, src_size, dst_size, dst_size))
+                extra_tokens = rel_pos_bias[-num_extra_tokens:, :]
+                rel_pos_bias = rel_pos_bias[:-num_extra_tokens, :]
+
+                def geometric_progression(a, r, n):
+                    return a * (1.0 - r ** n) / (1.0 - r)
+
+                left, right = 1.01, 1.5
+                while right - left > 1e-6:
+                    q = (left + right) / 2.0
+                    gp = geometric_progression(1, q, src_size // 2)
+                    if gp > dst_size // 2:
+                        right = q
+                    else:
+                        left = q
+
+                # if q > 1.090307:
+                #     q = 1.090307
+
+                dis = []
+                cur = 1
+                for i in range(src_size // 2):
+                    dis.append(cur)
+                    cur += q ** (i + 1)
+
+                r_ids = [-_ for _ in reversed(dis)]
+
+                x = r_ids + [0] + dis
+                y = r_ids + [0] + dis
+
+                t = dst_size // 2.0
+                dx = np.arange(-t, t + 0.1, 1.0)
+                dy = np.arange(-t, t + 0.1, 1.0)
+
+                print("Original positions = %s" % str(x))
+                print("Target positions = %s" % str(dx))
+
+                all_rel_pos_bias = []
+
+                for i in range(num_attn_heads):
+                    z = rel_pos_bias[:, i].view(src_size, src_size).float().numpy()
+                    f = F.interpolate.interp2d(x, y, z, kind='cubic')
+                    all_rel_pos_bias.append(
+                        torch.Tensor(f(dx, dy)).contiguous().view(-1, 1).to(rel_pos_bias.device))
+
+                rel_pos_bias = torch.cat(all_rel_pos_bias, dim=-1)
+
+                new_rel_pos_bias = torch.cat((rel_pos_bias, extra_tokens), dim=0)
+                state_dict[key] = new_rel_pos_bias
+
+    # interpolate position embedding
+    if 'pos_embed' in state_dict:
+        pos_embed_checkpoint = state_dict['pos_embed']
+        embedding_size = pos_embed_checkpoint.shape[-1]
+        num_patches = model.visual.patch_embed.num_patches
+        num_extra_tokens = model.visual.pos_embed.shape[-2] - num_patches
+        # height (== width) for the checkpoint position embedding
+        orig_size = int((pos_embed_checkpoint.shape[-2] - num_extra_tokens) ** 0.5)
+        # height (== width) for the new position embedding
+        new_size = int(num_patches ** 0.5)
+        # class_token and dist_token are kept unchanged
+        if orig_size != new_size:
+            print("Position interpolate from %dx%d to %dx%d" % (orig_size, orig_size, new_size, new_size))
+            extra_tokens = pos_embed_checkpoint[:, :num_extra_tokens]
+            # only the position tokens are interpolated
+            pos_tokens = pos_embed_checkpoint[:, num_extra_tokens:]
+            pos_tokens = pos_tokens.reshape(-1, orig_size, orig_size, embedding_size).permute(0, 3, 1, 2)
+            pos_tokens = torch.nn.functional.interpolate(
+                pos_tokens, size=(new_size, new_size), mode='bicubic', align_corners=False)
+            pos_tokens = pos_tokens.permute(0, 2, 3, 1).flatten(1, 2)
+            new_pos_embed = torch.cat((extra_tokens, pos_tokens), dim=1)
+            state_dict['pos_embed'] = new_pos_embed
+
+            patch_embed_proj = state_dict['patch_embed.proj.weight']
+            patch_size = model.visual.patch_embed.patch_size
+            state_dict['patch_embed.proj.weight'] = torch.nn.functional.interpolate(
+                patch_embed_proj.float(), size=patch_size, mode='bicubic', align_corners=False)
+
+
+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
+
+
+# From PyTorch internals
+def _ntuple(n):
+    def parse(x):
+        if isinstance(x, collections.abc.Iterable):
+            return x
+        return tuple(repeat(x, n))
+    return parse
+
+
+to_1tuple = _ntuple(1)
+to_2tuple = _ntuple(2)
+to_3tuple = _ntuple(3)
+to_4tuple = _ntuple(4)
+to_ntuple = lambda n, x: _ntuple(n)(x)
+
+
+def is_logging(args):
+    def is_global_master(args):
+        return args.rank == 0
+
+    def is_local_master(args):
+        return args.local_rank == 0
+
+    def is_master(args, local=False):
+        return is_local_master(args) if local else is_global_master(args)
+    return is_master
+
+
+class AllGather(torch.autograd.Function):
+    """An autograd function that performs allgather on a tensor.
+    Performs all_gather operation on the provided tensors.
+    *** Warning ***: torch.distributed.all_gather has no gradient.
+    """
+
+    @staticmethod
+    def forward(ctx, tensor, rank, world_size):
+        tensors_gather = [torch.empty_like(tensor) for _ in range(world_size)]
+        torch.distributed.all_gather(tensors_gather, tensor)
+        ctx.rank = rank
+        ctx.batch_size = tensor.shape[0]
+        return torch.cat(tensors_gather, 0)
+
+    @staticmethod
+    def backward(ctx, grad_output):
+        return (
+            grad_output[ctx.batch_size * ctx.rank: ctx.batch_size * (ctx.rank + 1)],
+            None,
+            None
+        )
+
+allgather = AllGather.apply

flux/__init__.py

@@ -0,0 +1,11 @@
+try:
+    from ._version import version as __version__  # type: ignore
+    from ._version import version_tuple
+except ImportError:
+    __version__ = "unknown (no version information available)"
+    version_tuple = (0, 0, "unknown", "noinfo")
+
+from pathlib import Path
+
+PACKAGE = __package__.replace("_", "-")
+PACKAGE_ROOT = Path(__file__).parent

flux/image_utils.py

@@ -0,0 +1,210 @@
+from PIL import Image, ImageDraw, ImageFont
+import os
+import torch
+import glob
+import matplotlib.pyplot as plt
+
+def read_images_in_path(path, size = (512,512)):
+    image_paths = []
+    for filename in os.listdir(path):
+        if filename.endswith(".png") or filename.endswith(".jpg") or filename.endswith(".jpeg"):
+            image_path = os.path.join(path, filename)
+            image_paths.append(image_path)
+    image_paths = sorted(image_paths)
+    return [Image.open(image_path).convert("RGB").resize(size) for image_path in image_paths]
+
+def concatenate_images(image_lists, return_list = False):
+    num_rows = len(image_lists[0])
+    num_columns = len(image_lists)
+    image_width = image_lists[0][0].width
+    image_height = image_lists[0][0].height
+
+    grid_width = num_columns * image_width
+    grid_height = num_rows * image_height if not return_list else image_height
+    if not return_list:
+        grid_image = [Image.new('RGB', (grid_width, grid_height))]
+    else:
+        grid_image = [Image.new('RGB', (grid_width, grid_height)) for i in range(num_rows)]
+
+    for i in range(num_rows):
+        row_index = i if return_list else 0
+        for j in range(num_columns):
+            image = image_lists[j][i]
+            x_offset = j * image_width
+            y_offset = i * image_height if not return_list else 0
+            grid_image[row_index].paste(image, (x_offset, y_offset))
+
+    return grid_image if return_list else grid_image[0]
+
+def concatenate_images_single(image_lists):
+    num_columns = len(image_lists)
+    image_width = image_lists[0].width
+    image_height = image_lists[0].height
+
+    grid_width = num_columns * image_width
+    grid_height = image_height
+    grid_image = Image.new('RGB', (grid_width, grid_height))
+
+    for j in range(num_columns):
+        image = image_lists[j]
+        x_offset = j * image_width
+        y_offset = 0
+        grid_image.paste(image, (x_offset, y_offset))
+
+    return grid_image
+
+def get_captions_for_images(images, device):
+    from transformers import Blip2Processor, Blip2ForConditionalGeneration
+
+    processor = Blip2Processor.from_pretrained("Salesforce/blip2-opt-2.7b")
+    model = Blip2ForConditionalGeneration.from_pretrained(
+        "Salesforce/blip2-opt-2.7b", load_in_8bit=True, device_map={"": 0}, torch_dtype=torch.float16
+    )  # doctest: +IGNORE_RESULT
+
+    res = []
+    
+    for image in images:
+        inputs = processor(images=image, return_tensors="pt").to(device, torch.float16)
+
+        generated_ids = model.generate(**inputs)
+        generated_text = processor.batch_decode(generated_ids, skip_special_tokens=True)[0].strip()
+        res.append(generated_text)
+
+    del processor
+    del model
+    
+    return res
+
+def find_and_plot_images(directory, output_file, recursive=True, figsize=(15, 15), image_formats=("*.png", "*.jpg", "*.jpeg", "*.bmp", "*.tiff")):
+    """
+    Finds all images in the specified directory (optionally recursively) 
+    and saves them in a single figure with their filenames.
+
+    Parameters:
+        directory (str): Path to the directory.
+        output_file (str): Path to save the resulting figure (e.g., 'output.png').
+        recursive (bool): Whether to search directories recursively.
+        figsize (tuple): Size of the resulting figure.
+        image_formats (tuple): Image file formats to look for.
+
+    Returns:
+        None
+    """
+    # Gather all image file paths
+    pattern = "**/" if recursive else ""
+    images = []
+    for fmt in image_formats:
+        images.extend(glob.glob(os.path.join(directory, pattern + fmt), recursive=recursive))
+
+    images = [image for image in images if "noise.jpg" not in image and "results.jpg" not in image]  # Filter out noise and result images
+    # move "original" to the front, followed by "reconstruction" and then the rest
+    images = sorted(
+        images,
+        key=lambda x: (not x.endswith("original.jpg"), not x.endswith("reconstruction.jpg"), x)
+    )
+    
+    if not images:
+        print("No images found!")
+        return
+
+    # Create a figure
+    num_images = len(images)
+    cols = num_images  # Max 5 images per row
+    rows = (num_images + cols - 1) // cols  # Calculate number of rows
+    fig, axs = plt.subplots(rows, cols, figsize=figsize)
+    axs = axs.flatten() if num_images > 1 else [axs]  # Flatten axes for single image case
+
+    for i, image_path in enumerate(images):
+        # Open and plot image
+        img = Image.open(image_path)
+        axs[i].imshow(img)
+        axs[i].axis('off')  # Remove axes
+        axs[i].set_title(os.path.basename(image_path), fontsize=8)  # Add filename
+
+    # Hide any remaining empty axes
+    for j in range(i + 1, len(axs)):
+        axs[j].axis('off')
+
+    plt.tight_layout()
+    plt.savefig(output_file, bbox_inches='tight', dpi=300)  # Save the figure to the file
+    plt.close(fig)  # Close the figure to free up memory
+    print(f"Figure saved to {output_file}")
+
+
+def add_label_to_image(image, label):
+    """
+    Adds a label to the lower-right corner of an image.
+
+    Args:
+        image (PIL.Image): Image to add the label to.
+        label (str): Text to add as a label.
+
+    Returns:
+        PIL.Image: Image with the added label.
+    """
+    # Create a drawing context
+    draw = ImageDraw.Draw(image)
+
+
+    # Create a drawing context
+    draw = ImageDraw.Draw(image)
+
+    # Define font and size
+    font_size = int(min(image.size) * 0.05)  # Adjust font size based on image dimensions
+    try:
+        font = ImageFont.truetype("fonts/arial.ttf", font_size)  # Replace with a font path if needed
+    except IOError:
+        font = ImageFont.load_default()  # Fallback to default font if arial.ttf is not found
+
+    # Measure text size using textbbox
+    text_bbox = draw.textbbox((0, 0), label, font=font)  # (left, top, right, bottom)
+    text_width = text_bbox[2] - text_bbox[0]
+    text_height = text_bbox[3] - text_bbox[1]
+
+    # Position the text in the lower-right corner with some padding
+    padding = 10
+    position = (image.width - text_width - padding, image.height - text_height - padding)
+
+    # Add a semi-transparent background for the label
+    draw.rectangle(
+        [
+            (position[0] - padding, position[1] - padding),
+            (position[0] + text_width + padding, position[1] + text_height + padding)
+        ],
+        fill=(0, 0, 0, 150)  # Black with transparency
+    )
+
+    # Draw the label
+    draw.text(position, label, fill="white", font=font)
+
+    return image
+
+def crop_center_square_and_resize(img, size, output_path=None):
+    """
+    Crops the center of an image to make it square.
+    
+    Args:
+        img (PIL.Image): Image to crop.
+        output_path (str, optional): Path to save the cropped image. If None, the cropped image is not saved.
+    
+    Returns:
+        Image: The cropped square image.
+    """
+    width, height = img.size
+    # Determine the shorter side
+    side_length = min(width, height)
+    # Calculate the cropping box
+    left = (width - side_length) // 2
+    top = (height - side_length) // 2
+    right = left + side_length
+    bottom = top + side_length
+    # Crop the image
+    cropped_img = img.crop((left, top, right, bottom))
+    # Resize the image
+    cropped_img = cropped_img.resize(size)
+    
+    # Save the cropped image if output path is specified
+    if output_path:
+        cropped_img.save(output_path)
+    
+    return cropped_img

flux/math.py

@@ -0,0 +1,31 @@
+import torch
+from einops import rearrange
+from torch import Tensor
+
+
+def attention(q: Tensor, k: Tensor, v: Tensor, pe: Tensor) -> Tensor:
+    if pe is not None:
+        q, k = apply_rope(q, k, pe)
+
+    x = torch.nn.functional.scaled_dot_product_attention(q, k, v)
+    x = rearrange(x, "B H L D -> B L (H D)")
+
+    return x
+
+
+def rope(pos: Tensor, dim: int, theta: int) -> Tensor:
+    assert dim % 2 == 0
+    scale = torch.arange(0, dim, 2, dtype=torch.float64, device=pos.device) / dim
+    omega = 1.0 / (theta**scale)
+    out = torch.einsum("...n,d->...nd", pos, omega)
+    out = torch.stack([torch.cos(out), -torch.sin(out), torch.sin(out), torch.cos(out)], dim=-1)
+    out = rearrange(out, "b n d (i j) -> b n d i j", i=2, j=2)
+    return out.float()
+
+
+def apply_rope(xq: Tensor, xk: Tensor, freqs_cis: Tensor) -> tuple[Tensor, Tensor]:
+    xq_ = xq.float().reshape(*xq.shape[:-1], -1, 1, 2)
+    xk_ = xk.float().reshape(*xk.shape[:-1], -1, 1, 2)
+    xq_out = freqs_cis[..., 0] * xq_[..., 0] + freqs_cis[..., 1] * xq_[..., 1]
+    xk_out = freqs_cis[..., 0] * xk_[..., 0] + freqs_cis[..., 1] * xk_[..., 1]
+    return xq_out.reshape(*xq.shape).type_as(xq), xk_out.reshape(*xk.shape).type_as(xk)

flux/model.py

@@ -0,0 +1,165 @@
+from dataclasses import dataclass
+
+import torch
+from torch import Tensor, nn
+
+from flux.modules.layers import (
+    DoubleStreamBlock,
+    EmbedND,
+    LastLayer,
+    MLPEmbedder,
+    SingleStreamBlock,
+    timestep_embedding,
+)
+
+DEVICE = torch.device("cuda")
+
+@dataclass
+class FluxParams:
+    in_channels: int
+    vec_in_dim: int
+    context_in_dim: int
+    hidden_size: int
+    mlp_ratio: float
+    num_heads: int
+    depth: int
+    depth_single_blocks: int
+    axes_dim: list[int]
+    theta: int
+    qkv_bias: bool
+    guidance_embed: bool
+
+
+class Flux(nn.Module):
+    """
+    Transformer model for flow matching on sequences.
+    """
+
+    def __init__(self, params: FluxParams):
+        super().__init__()
+
+        self.params = params
+        self.in_channels = params.in_channels
+        self.out_channels = self.in_channels
+        if params.hidden_size % params.num_heads != 0:
+            raise ValueError(
+                f"Hidden size {params.hidden_size} must be divisible by num_heads {params.num_heads}"
+            )
+        pe_dim = params.hidden_size // params.num_heads
+        if sum(params.axes_dim) != pe_dim:
+            raise ValueError(f"Got {params.axes_dim} but expected positional dim {pe_dim}")
+        self.hidden_size = params.hidden_size
+        self.num_heads = params.num_heads
+        self.pe_embedder = EmbedND(dim=pe_dim, theta=params.theta, axes_dim=params.axes_dim)
+        self.img_in = nn.Linear(self.in_channels, self.hidden_size, bias=True)
+        self.time_in = MLPEmbedder(in_dim=256, hidden_dim=self.hidden_size)
+        self.vector_in = MLPEmbedder(params.vec_in_dim, self.hidden_size)
+        self.guidance_in = (
+            MLPEmbedder(in_dim=256, hidden_dim=self.hidden_size) if params.guidance_embed else nn.Identity()
+        )
+        self.txt_in = nn.Linear(params.context_in_dim, self.hidden_size)
+
+        self.double_blocks = nn.ModuleList(
+            [
+                DoubleStreamBlock(
+                    self.hidden_size,
+                    self.num_heads,
+                    mlp_ratio=params.mlp_ratio,
+                    qkv_bias=params.qkv_bias,
+                ).to(torch.bfloat16)
+                for _ in range(params.depth)
+            ]
+        )
+
+        self.single_blocks = nn.ModuleList(
+            [
+                SingleStreamBlock(self.hidden_size, self.num_heads, mlp_ratio=params.mlp_ratio).to(torch.bfloat16)
+                for _ in range(params.depth_single_blocks)
+            ]
+        )
+
+        self.final_layer = LastLayer(self.hidden_size, 1, self.out_channels)
+
+        self.pulid_ca = None
+        self.pulid_double_interval = 2
+        self.pulid_single_interval = 4
+
+    def forward(
+        self,
+        img: Tensor,
+        img_ids: Tensor,
+        txt: Tensor,
+        txt_ids: Tensor,
+        timesteps: Tensor,
+        y: Tensor,
+        guidance: Tensor = None,
+        id: Tensor = None,
+        id_weight: float = 1.0,
+        aggressive_offload: bool = False,
+    ) -> Tensor:
+        if img.ndim != 3 or txt.ndim != 3:
+            raise ValueError("Input img and txt tensors must have 3 dimensions.")
+
+        # running on sequences img
+        img = self.img_in(img)
+        vec = self.time_in(timestep_embedding(timesteps, 256))
+        if self.params.guidance_embed:
+            if guidance is None:
+                raise ValueError("Didn't get guidance strength for guidance distilled model.")
+            vec = vec + self.guidance_in(timestep_embedding(guidance, 256))
+        vec = vec + self.vector_in(y)
+        txt = self.txt_in(txt)
+
+        ids = torch.cat((txt_ids, img_ids), dim=1)
+        pe = self.pe_embedder(ids)
+
+        ca_idx = 0
+        if aggressive_offload:
+            self.double_blocks = self.double_blocks.to(DEVICE)
+        for i, block in enumerate(self.double_blocks):
+            img, txt = block(img=img, txt=txt, vec=vec, pe=pe)
+
+            if i % self.pulid_double_interval == 0 and id is not None:
+                img = img + id_weight * self.pulid_ca[ca_idx](id, img)
+                ca_idx += 1
+        if aggressive_offload:
+            self.double_blocks.cpu()
+
+        img = torch.cat((txt, img), 1)
+        if aggressive_offload:
+            # put half of the single blcoks to gpu
+            for i in range(len(self.single_blocks) // 2):
+                self.single_blocks[i] = self.single_blocks[i].to(DEVICE)
+        for i, block in enumerate(self.single_blocks):
+            if aggressive_offload and i == len(self.single_blocks)//2:
+                # put first half of the single blcoks to cpu and last half to gpu
+                for j in range(len(self.single_blocks) // 2):
+                    self.single_blocks[j].cpu()
+                for j in range(len(self.single_blocks) // 2, len(self.single_blocks)):
+                    self.single_blocks[j] = self.single_blocks[j].to(DEVICE)
+            x = block(img, vec=vec, pe=pe)
+            real_img, txt = x[:, txt.shape[1]:, ...], x[:, :txt.shape[1], ...]
+
+            if i % self.pulid_single_interval == 0 and id is not None:
+                real_img = real_img + id_weight * self.pulid_ca[ca_idx](id, real_img)
+                ca_idx += 1
+
+            img = torch.cat((txt, real_img), 1)
+        if aggressive_offload:
+            self.single_blocks.cpu()
+        img = img[:, txt.shape[1] :, ...]
+
+        img = self.final_layer(img, vec)  # (N, T, patch_size ** 2 * out_channels)
+        return img
+
+    def components_to_gpu(self):
+        # everything but double_blocks, single_blocks
+        self.img_in.to(DEVICE)
+        self.time_in.to(DEVICE)
+        self.guidance_in.to(DEVICE)
+        self.vector_in.to(DEVICE)
+        self.txt_in.to(DEVICE)
+        self.pe_embedder.to(DEVICE)
+        self.final_layer.to(DEVICE)
+        if self.pulid_ca:
+            self.pulid_ca.to(DEVICE)

flux/modules/autoencoder.py

@@ -0,0 +1,317 @@
+from dataclasses import dataclass
+
+import torch
+from einops import rearrange
+from torch import Tensor, nn
+
+
+@dataclass
+class AutoEncoderParams:
+    resolution: int
+    in_channels: int
+    ch: int
+    out_ch: int
+    ch_mult: list[int]
+    num_res_blocks: int
+    z_channels: int
+    scale_factor: float
+    shift_factor: float
+
+
+def swish(x: Tensor) -> Tensor:
+    return x * torch.sigmoid(x)
+
+
+class AttnBlock(nn.Module):
+    def __init__(self, in_channels: int):
+        super().__init__()
+        self.in_channels = in_channels
+
+        self.norm = nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True)
+
+        self.q = nn.Conv2d(in_channels, in_channels, kernel_size=1)
+        self.k = nn.Conv2d(in_channels, in_channels, kernel_size=1)
+        self.v = nn.Conv2d(in_channels, in_channels, kernel_size=1)
+        self.proj_out = nn.Conv2d(in_channels, in_channels, kernel_size=1)
+
+    def attention(self, h_: Tensor) -> Tensor:
+        h_ = self.norm(h_)
+        q = self.q(h_)
+        k = self.k(h_)
+        v = self.v(h_)
+
+        b, c, h, w = q.shape
+        q = rearrange(q, "b c h w -> b 1 (h w) c").contiguous()
+        k = rearrange(k, "b c h w -> b 1 (h w) c").contiguous()
+        v = rearrange(v, "b c h w -> b 1 (h w) c").contiguous()
+        h_ = nn.functional.scaled_dot_product_attention(q, k, v)
+
+        return rearrange(h_, "b 1 (h w) c -> b c h w", h=h, w=w, c=c, b=b)
+
+    def forward(self, x: Tensor) -> Tensor:
+        return x + self.proj_out(self.attention(x))
+
+
+class ResnetBlock(nn.Module):
+    def __init__(self, in_channels: int, out_channels: int):
+        super().__init__()
+        self.in_channels = in_channels
+        out_channels = in_channels if out_channels is None else out_channels
+        self.out_channels = out_channels
+
+        self.norm1 = nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True)
+        self.conv1 = nn.Conv2d(in_channels, out_channels, kernel_size=3, stride=1, padding=1)
+        self.norm2 = nn.GroupNorm(num_groups=32, num_channels=out_channels, eps=1e-6, affine=True)
+        self.conv2 = nn.Conv2d(out_channels, out_channels, kernel_size=3, stride=1, padding=1)
+        if self.in_channels != self.out_channels:
+            self.nin_shortcut = nn.Conv2d(in_channels, out_channels, kernel_size=1, stride=1, padding=0)
+
+    def forward(self, x):
+        h = x
+        h = self.norm1(h)
+        h = swish(h)
+        h = self.conv1(h)
+
+        h = self.norm2(h)
+        h = swish(h)
+        h = self.conv2(h)
+
+        if self.in_channels != self.out_channels:
+            x = self.nin_shortcut(x)
+
+        return x + h
+
+
+class Downsample(nn.Module):
+    def __init__(self, in_channels: int):
+        super().__init__()
+        # no asymmetric padding in torch conv, must do it ourselves
+        self.conv = nn.Conv2d(in_channels, in_channels, kernel_size=3, stride=2, padding=0)
+
+    def forward(self, x: Tensor):
+        pad = (0, 1, 0, 1)
+        x = nn.functional.pad(x, pad, mode="constant", value=0)
+        x = self.conv(x)
+        return x
+
+
+class Upsample(nn.Module):
+    def __init__(self, in_channels: int):
+        super().__init__()
+        self.conv = nn.Conv2d(in_channels, in_channels, kernel_size=3, stride=1, padding=1)
+
+    def forward(self, x: Tensor):
+        x = nn.functional.interpolate(x, scale_factor=2.0, mode="nearest")
+        x = self.conv(x)
+        return x
+
+
+class Encoder(nn.Module):
+    def __init__(
+        self,
+        resolution: int,
+        in_channels: int,
+        ch: int,
+        ch_mult: list[int],
+        num_res_blocks: int,
+        z_channels: int,
+    ):
+        super().__init__()
+        self.ch = ch
+        self.num_resolutions = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        self.resolution = resolution
+        self.in_channels = in_channels
+        # downsampling
+        self.conv_in = nn.Conv2d(in_channels, self.ch, kernel_size=3, stride=1, padding=1)
+
+        curr_res = resolution
+        in_ch_mult = (1,) + tuple(ch_mult)
+        self.in_ch_mult = in_ch_mult
+        self.down = nn.ModuleList()
+        block_in = self.ch
+        for i_level in range(self.num_resolutions):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_in = ch * in_ch_mult[i_level]
+            block_out = ch * ch_mult[i_level]
+            for _ in range(self.num_res_blocks):
+                block.append(ResnetBlock(in_channels=block_in, out_channels=block_out))
+                block_in = block_out
+            down = nn.Module()
+            down.block = block
+            down.attn = attn
+            if i_level != self.num_resolutions - 1:
+                down.downsample = Downsample(block_in)
+                curr_res = curr_res // 2
+            self.down.append(down)
+
+        # middle
+        self.mid = nn.Module()
+        self.mid.block_1 = ResnetBlock(in_channels=block_in, out_channels=block_in)
+        self.mid.attn_1 = AttnBlock(block_in)
+        self.mid.block_2 = ResnetBlock(in_channels=block_in, out_channels=block_in)
+
+        # end
+        self.norm_out = nn.GroupNorm(num_groups=32, num_channels=block_in, eps=1e-6, affine=True)
+        self.conv_out = nn.Conv2d(block_in, 2 * z_channels, kernel_size=3, stride=1, padding=1)
+
+    def forward(self, x: Tensor) -> Tensor:
+        # downsampling
+        hs = [self.conv_in(x)]
+        for i_level in range(self.num_resolutions):
+            for i_block in range(self.num_res_blocks):
+                h = self.down[i_level].block[i_block](hs[-1])
+                if len(self.down[i_level].attn) > 0:
+                    h = self.down[i_level].attn[i_block](h)
+                hs.append(h)
+            if i_level != self.num_resolutions - 1:
+                hs.append(self.down[i_level].downsample(hs[-1]))
+
+        # middle
+        h = hs[-1]
+        h = self.mid.block_1(h)
+        h = self.mid.attn_1(h)
+        h = self.mid.block_2(h)
+        # end
+        h = self.norm_out(h)
+        h = swish(h)
+        h = self.conv_out(h)
+        return h
+
+
+class Decoder(nn.Module):
+    def __init__(
+        self,
+        ch: int,
+        out_ch: int,
+        ch_mult: list[int],
+        num_res_blocks: int,
+        in_channels: int,
+        resolution: int,
+        z_channels: int,
+    ):
+        super().__init__()
+        self.ch = ch
+        self.num_resolutions = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        self.resolution = resolution
+        self.in_channels = in_channels
+        self.ffactor = 2 ** (self.num_resolutions - 1)
+
+        # compute in_ch_mult, block_in and curr_res at lowest res
+        block_in = ch * ch_mult[self.num_resolutions - 1]
+        curr_res = resolution // 2 ** (self.num_resolutions - 1)
+        self.z_shape = (1, z_channels, curr_res, curr_res)
+
+        # z to block_in
+        self.conv_in = nn.Conv2d(z_channels, block_in, kernel_size=3, stride=1, padding=1)
+
+        # middle
+        self.mid = nn.Module()
+        self.mid.block_1 = ResnetBlock(in_channels=block_in, out_channels=block_in)
+        self.mid.attn_1 = AttnBlock(block_in)
+        self.mid.block_2 = ResnetBlock(in_channels=block_in, out_channels=block_in)
+
+        # upsampling
+        self.up = nn.ModuleList()
+        for i_level in reversed(range(self.num_resolutions)):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_out = ch * ch_mult[i_level]
+            for _ in range(self.num_res_blocks + 1):
+                block.append(ResnetBlock(in_channels=block_in, out_channels=block_out))
+                block_in = block_out
+            up = nn.Module()
+            up.block = block
+            up.attn = attn
+            if i_level != 0:
+                up.upsample = Upsample(block_in)
+                curr_res = curr_res * 2
+            self.up.insert(0, up)  # prepend to get consistent order
+
+        # end
+        self.norm_out = nn.GroupNorm(num_groups=32, num_channels=block_in, eps=1e-6, affine=True)
+        self.conv_out = nn.Conv2d(block_in, out_ch, kernel_size=3, stride=1, padding=1)
+
+    def forward(self, z: Tensor) -> Tensor:
+        # z to block_in
+        h = self.conv_in(z)
+
+        # middle
+        h = self.mid.block_1(h)
+        h = self.mid.attn_1(h)
+        h = self.mid.block_2(h)
+
+        # upsampling
+        for i_level in reversed(range(self.num_resolutions)):
+            for i_block in range(self.num_res_blocks + 1):
+                h = self.up[i_level].block[i_block](h)
+                if len(self.up[i_level].attn) > 0:
+                    h = self.up[i_level].attn[i_block](h)
+            if i_level != 0:
+                h = self.up[i_level].upsample(h)
+
+        # end
+        h = self.norm_out(h)
+        h = swish(h)
+        h = self.conv_out(h)
+        return h
+
+
+class DiagonalGaussian(nn.Module):
+    def __init__(self, sample: bool = True, chunk_dim: int = 1):
+        super().__init__()
+        self.sample = sample
+        self.chunk_dim = chunk_dim
+
+    def forward(self, z: Tensor) -> Tensor:
+        mean, logvar = torch.chunk(z, 2, dim=self.chunk_dim)
+        if self.sample:
+            std = torch.exp(0.5 * logvar)
+            return mean + std * torch.randn_like(mean)
+        else:
+            return mean
+
+
+class AutoEncoder(nn.Module):
+    def __init__(self, params: AutoEncoderParams):
+        super().__init__()
+        self.encoder = Encoder(
+            resolution=params.resolution,
+            in_channels=params.in_channels,
+            ch=params.ch,
+            ch_mult=params.ch_mult,
+            num_res_blocks=params.num_res_blocks,
+            z_channels=params.z_channels,
+        )
+        self.decoder = Decoder(
+            resolution=params.resolution,
+            in_channels=params.in_channels,
+            ch=params.ch,
+            out_ch=params.out_ch,
+            ch_mult=params.ch_mult,
+            num_res_blocks=params.num_res_blocks,
+            z_channels=params.z_channels,
+        )
+        self.reg = DiagonalGaussian()
+
+        self.scale_factor = params.scale_factor
+        self.shift_factor = params.shift_factor
+
+    def encode(self, x: Tensor) -> Tensor:
+        z = self.reg(self.encoder(x))
+        z = self.scale_factor * (z - self.shift_factor)
+        return z
+    
+    def encode_no_sampling(self, x: Tensor) -> Tensor:
+        z, _ = torch.chunk(self.encoder(x), 2, dim=1)
+        z = self.scale_factor * (z - self.shift_factor)
+        return z
+
+    def decode(self, z: Tensor) -> Tensor:
+        z = z / self.scale_factor + self.shift_factor
+        return self.decoder(z)
+
+    def forward(self, x: Tensor) -> Tensor:
+        return self.decode(self.encode(x))

flux/modules/conditioner.py

@@ -0,0 +1,37 @@
+from torch import Tensor, nn
+from transformers import CLIPTextModel, CLIPTokenizer, T5EncoderModel, T5Tokenizer
+
+
+class HFEmbedder(nn.Module):
+    def __init__(self, version: str, max_length: int, **hf_kwargs):
+        super().__init__()
+        self.is_clip = version.startswith("openai")
+        self.max_length = max_length
+        self.output_key = "pooler_output" if self.is_clip else "last_hidden_state"
+
+        if self.is_clip:
+            self.tokenizer: CLIPTokenizer = CLIPTokenizer.from_pretrained(version, max_length=max_length)
+            self.hf_module: CLIPTextModel = CLIPTextModel.from_pretrained(version, **hf_kwargs)
+        else:
+            self.tokenizer: T5Tokenizer = T5Tokenizer.from_pretrained(version, max_length=max_length)
+            self.hf_module: T5EncoderModel = T5EncoderModel.from_pretrained(version, **hf_kwargs)
+
+        self.hf_module = self.hf_module.eval().requires_grad_(False)
+
+    def forward(self, text: list[str]) -> Tensor:
+        batch_encoding = self.tokenizer(
+            text,
+            truncation=True,
+            max_length=self.max_length,
+            return_length=False,
+            return_overflowing_tokens=False,
+            padding="max_length",
+            return_tensors="pt",
+        )
+
+        outputs = self.hf_module(
+            input_ids=batch_encoding["input_ids"].to(self.hf_module.device),
+            attention_mask=None,
+            output_hidden_states=False,
+        )
+        return outputs[self.output_key]

flux/modules/layers.py

@@ -0,0 +1,253 @@
+import math
+from dataclasses import dataclass
+
+import torch
+from einops import rearrange
+from torch import Tensor, nn
+
+from flux.math import attention, rope
+
+
+class EmbedND(nn.Module):
+    def __init__(self, dim: int, theta: int, axes_dim: list[int]):
+        super().__init__()
+        self.dim = dim
+        self.theta = theta
+        self.axes_dim = axes_dim
+
+    def forward(self, ids: Tensor) -> Tensor:
+        n_axes = ids.shape[-1]
+        emb = torch.cat(
+            [rope(ids[..., i], self.axes_dim[i], self.theta) for i in range(n_axes)],
+            dim=-3,
+        )
+
+        return emb.unsqueeze(1)
+
+
+def timestep_embedding(t: Tensor, dim, max_period=10000, time_factor: float = 1000.0):
+    """
+    Create sinusoidal timestep embeddings.
+    :param t: a 1-D Tensor of N indices, one per batch element.
+                      These may be fractional.
+    :param dim: the dimension of the output.
+    :param max_period: controls the minimum frequency of the embeddings.
+    :return: an (N, D) Tensor of positional embeddings.
+    """
+    t = time_factor * t
+    half = dim // 2
+    freqs = torch.exp(-math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32) / half).to(
+        t.device
+    )
+
+    args = t[:, None].float() * freqs[None]
+    embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
+    if dim % 2:
+        embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
+    if torch.is_floating_point(t):
+        embedding = embedding.to(t)
+    return embedding
+
+
+class MLPEmbedder(nn.Module):
+    def __init__(self, in_dim: int, hidden_dim: int):
+        super().__init__()
+        self.in_layer = nn.Linear(in_dim, hidden_dim, bias=True)
+        self.silu = nn.SiLU()
+        self.out_layer = nn.Linear(hidden_dim, hidden_dim, bias=True)
+
+    def forward(self, x: Tensor) -> Tensor:
+        return self.out_layer(self.silu(self.in_layer(x)))
+
+
+class RMSNorm(torch.nn.Module):
+    def __init__(self, dim: int):
+        super().__init__()
+        self.scale = nn.Parameter(torch.ones(dim))
+
+    def forward(self, x: Tensor):
+        x_dtype = x.dtype
+        x = x.float()
+        rrms = torch.rsqrt(torch.mean(x**2, dim=-1, keepdim=True) + 1e-6)
+        return (x * rrms).to(dtype=x_dtype) * self.scale
+
+
+class QKNorm(torch.nn.Module):
+    def __init__(self, dim: int):
+        super().__init__()
+        self.query_norm = RMSNorm(dim)
+        self.key_norm = RMSNorm(dim)
+
+    def forward(self, q: Tensor, k: Tensor, v: Tensor) -> tuple[Tensor, Tensor]:
+        q = self.query_norm(q)
+        k = self.key_norm(k)
+        return q.to(v), k.to(v)
+
+
+class SelfAttention(nn.Module):
+    def __init__(self, dim: int, num_heads: int = 8, qkv_bias: bool = False):
+        super().__init__()
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+
+        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
+        self.norm = QKNorm(head_dim)
+        self.proj = nn.Linear(dim, dim)
+
+    def forward(self, x: Tensor, pe: Tensor) -> Tensor:
+        qkv = self.qkv(x)
+        q, k, v = rearrange(qkv, "B L (K H D) -> K B H L D", K=3, H=self.num_heads)
+        q, k = self.norm(q, k, v)
+        x = attention(q, k, v, pe=pe)
+        x = self.proj(x)
+        return x
+
+
+@dataclass
+class ModulationOut:
+    shift: Tensor
+    scale: Tensor
+    gate: Tensor
+
+
+class Modulation(nn.Module):
+    def __init__(self, dim: int, double: bool):
+        super().__init__()
+        self.is_double = double
+        self.multiplier = 6 if double else 3
+        self.lin = nn.Linear(dim, self.multiplier * dim, bias=True)
+
+    def forward(self, vec: Tensor) -> tuple[ModulationOut, ModulationOut]:
+        out = self.lin(nn.functional.silu(vec))[:, None, :].chunk(self.multiplier, dim=-1)
+
+        return (
+            ModulationOut(*out[:3]),
+            ModulationOut(*out[3:]) if self.is_double else None,
+        )
+
+
+class DoubleStreamBlock(nn.Module):
+    def __init__(self, hidden_size: int, num_heads: int, mlp_ratio: float, qkv_bias: bool = False):
+        super().__init__()
+
+        mlp_hidden_dim = int(hidden_size * mlp_ratio)
+        self.num_heads = num_heads
+        self.hidden_size = hidden_size
+        self.img_mod = Modulation(hidden_size, double=True)
+        self.img_norm1 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
+        self.img_attn = SelfAttention(dim=hidden_size, num_heads=num_heads, qkv_bias=qkv_bias)
+
+        self.img_norm2 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
+        self.img_mlp = nn.Sequential(
+            nn.Linear(hidden_size, mlp_hidden_dim, bias=True),
+            nn.GELU(approximate="tanh"),
+            nn.Linear(mlp_hidden_dim, hidden_size, bias=True),
+        )
+
+        self.txt_mod = Modulation(hidden_size, double=True)
+        self.txt_norm1 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
+        self.txt_attn = SelfAttention(dim=hidden_size, num_heads=num_heads, qkv_bias=qkv_bias)
+
+        self.txt_norm2 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
+        self.txt_mlp = nn.Sequential(
+            nn.Linear(hidden_size, mlp_hidden_dim, bias=True),
+            nn.GELU(approximate="tanh"),
+            nn.Linear(mlp_hidden_dim, hidden_size, bias=True),
+        )
+
+    def forward(self, img: Tensor, txt: Tensor, vec: Tensor, pe: Tensor) -> tuple[Tensor, Tensor]:
+        img_mod1, img_mod2 = self.img_mod(vec)
+        txt_mod1, txt_mod2 = self.txt_mod(vec)
+
+        # prepare image for attention
+        img_modulated = self.img_norm1(img)
+        img_modulated = (1 + img_mod1.scale) * img_modulated + img_mod1.shift
+        img_qkv = self.img_attn.qkv(img_modulated)
+        img_q, img_k, img_v = rearrange(img_qkv, "B L (K H D) -> K B H L D", K=3, H=self.num_heads)
+        img_q, img_k = self.img_attn.norm(img_q, img_k, img_v)
+
+        # prepare txt for attention
+        txt_modulated = self.txt_norm1(txt)
+        txt_modulated = (1 + txt_mod1.scale) * txt_modulated + txt_mod1.shift
+        txt_qkv = self.txt_attn.qkv(txt_modulated)
+        txt_q, txt_k, txt_v = rearrange(txt_qkv, "B L (K H D) -> K B H L D", K=3, H=self.num_heads)
+        txt_q, txt_k = self.txt_attn.norm(txt_q, txt_k, txt_v)
+
+        # run actual attention
+        q = torch.cat((txt_q, img_q), dim=2)
+        k = torch.cat((txt_k, img_k), dim=2)
+        v = torch.cat((txt_v, img_v), dim=2)
+
+        attn = attention(q, k, v, pe=pe)
+        txt_attn, img_attn = attn[:, : txt.shape[1]], attn[:, txt.shape[1] :]
+
+        # calculate the img bloks
+        img = img + img_mod1.gate * self.img_attn.proj(img_attn)
+        img = img + img_mod2.gate * self.img_mlp((1 + img_mod2.scale) * self.img_norm2(img) + img_mod2.shift)
+
+        # calculate the txt bloks
+        txt = txt + txt_mod1.gate * self.txt_attn.proj(txt_attn)
+        txt = txt + txt_mod2.gate * self.txt_mlp((1 + txt_mod2.scale) * self.txt_norm2(txt) + txt_mod2.shift)
+        return img, txt
+
+
+class SingleStreamBlock(nn.Module):
+    """
+    A DiT block with parallel linear layers as described in
+    https://arxiv.org/abs/2302.05442 and adapted modulation interface.
+    """
+
+    def __init__(
+        self,
+        hidden_size: int,
+        num_heads: int,
+        mlp_ratio: float = 4.0,
+        qk_scale: float = None,
+    ):
+        super().__init__()
+        self.hidden_dim = hidden_size
+        self.num_heads = num_heads
+        head_dim = hidden_size // num_heads
+        self.scale = qk_scale or head_dim**-0.5
+
+        self.mlp_hidden_dim = int(hidden_size * mlp_ratio)
+        # qkv and mlp_in
+        self.linear1 = nn.Linear(hidden_size, hidden_size * 3 + self.mlp_hidden_dim)
+        # proj and mlp_out
+        self.linear2 = nn.Linear(hidden_size + self.mlp_hidden_dim, hidden_size)
+
+        self.norm = QKNorm(head_dim)
+
+        self.hidden_size = hidden_size
+        self.pre_norm = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
+
+        self.mlp_act = nn.GELU(approximate="tanh")
+        self.modulation = Modulation(hidden_size, double=False)
+
+    def forward(self, x: Tensor, vec: Tensor, pe: Tensor) -> Tensor:
+        mod, _ = self.modulation(vec)
+        x_mod = (1 + mod.scale) * self.pre_norm(x) + mod.shift
+        qkv, mlp = torch.split(self.linear1(x_mod), [3 * self.hidden_size, self.mlp_hidden_dim], dim=-1)
+
+        q, k, v = rearrange(qkv, "B L (K H D) -> K B H L D", K=3, H=self.num_heads)
+        q, k = self.norm(q, k, v)
+
+        # compute attention
+        attn = attention(q, k, v, pe=pe)
+        # compute activation in mlp stream, cat again and run second linear layer
+        output = self.linear2(torch.cat((attn, self.mlp_act(mlp)), 2))
+        return x + mod.gate * output
+
+
+class LastLayer(nn.Module):
+    def __init__(self, hidden_size: int, patch_size: int, out_channels: int):
+        super().__init__()
+        self.norm_final = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
+        self.linear = nn.Linear(hidden_size, patch_size * patch_size * out_channels, bias=True)
+        self.adaLN_modulation = nn.Sequential(nn.SiLU(), nn.Linear(hidden_size, 2 * hidden_size, bias=True))
+
+    def forward(self, x: Tensor, vec: Tensor) -> Tensor:
+        shift, scale = self.adaLN_modulation(vec).chunk(2, dim=1)
+        x = (1 + scale[:, None, :]) * self.norm_final(x) + shift[:, None, :]
+        x = self.linear(x)
+        return x

flux/sampling.py

@@ -0,0 +1,299 @@
+import math
+from typing import Callable
+
+import torch
+from einops import rearrange, repeat
+from torch import Tensor
+from tqdm import tqdm
+
+from .model import Flux
+from .modules.conditioner import HFEmbedder
+
+
+def get_noise(
+    num_samples: int,
+    height: int,
+    width: int,
+    device: torch.device,
+    dtype: torch.dtype,
+    seed: int,
+):
+    return torch.randn(
+        num_samples,
+        16,
+        # allow for packing
+        2 * math.ceil(height / 16),
+        2 * math.ceil(width / 16),
+        device=device,
+        dtype=dtype,
+        generator=torch.Generator(device=device).manual_seed(seed),
+    )
+
+
+def prepare(t5: HFEmbedder, clip: HFEmbedder, img: Tensor, prompt: str) -> dict[str, Tensor]:
+    bs, c, h, w = img.shape
+    if bs == 1 and not isinstance(prompt, str):
+        bs = len(prompt)
+
+    img = rearrange(img, "b c (h ph) (w pw) -> b (h w) (c ph pw)", ph=2, pw=2)
+    if img.shape[0] == 1 and bs > 1:
+        img = repeat(img, "1 ... -> bs ...", bs=bs)
+
+    img_ids = torch.zeros(h // 2, w // 2, 3)
+    img_ids[..., 1] = img_ids[..., 1] + torch.arange(h // 2)[:, None]
+    img_ids[..., 2] = img_ids[..., 2] + torch.arange(w // 2)[None, :]
+    img_ids = repeat(img_ids, "h w c -> b (h w) c", b=bs)
+    img_ids = img_ids.to(img.dtype)
+    img_ids = img_ids.to(torch.bfloat16)
+
+    if isinstance(prompt, str):
+        prompt = [prompt]
+    txt = t5(prompt)
+    if txt.shape[0] == 1 and bs > 1:
+        txt = repeat(txt, "1 ... -> bs ...", bs=bs)
+    txt_ids = torch.zeros(bs, txt.shape[1], 3).to(txt.dtype)
+    txt_ids = txt_ids.to(torch.bfloat16)
+
+    vec = clip(prompt)
+    if vec.shape[0] == 1 and bs > 1:
+        vec = repeat(vec, "1 ... -> bs ...", bs=bs)
+
+    return {
+        "img": img,
+        "img_ids": img_ids.to(img.device),
+        "txt": txt.to(img.device),
+        "txt_ids": txt_ids.to(img.device),
+        "vec": vec.to(img.device),
+    }
+
+
+def time_shift(mu: float, sigma: float, t: Tensor):
+    return math.exp(mu) / (math.exp(mu) + (1 / t - 1) ** sigma)
+
+
+def get_lin_function(
+    x1: float = 256, y1: float = 0.5, x2: float = 4096, y2: float = 1.15
+) -> Callable[[float], float]:
+    m = (y2 - y1) / (x2 - x1)
+    b = y1 - m * x1
+    return lambda x: m * x + b
+
+
+def get_schedule(
+    num_steps: int,
+    image_seq_len: int,
+    base_shift: float = 0.5,
+    max_shift: float = 1.15,
+    shift: bool = True,
+) -> list[float]:
+    # extra step for zero
+    timesteps = torch.linspace(1, 0, num_steps + 1)
+
+    # shifting the schedule to favor high timesteps for higher signal images
+    if shift:
+        # eastimate mu based on linear estimation between two points
+        mu = get_lin_function(y1=base_shift, y2=max_shift)(image_seq_len)
+        timesteps = time_shift(mu, 1.0, timesteps)
+
+    return timesteps.tolist()
+
+def rf_inversion(
+    model: Flux,
+    img: Tensor,
+    img_ids: Tensor,
+    txt: Tensor,
+    txt_ids: Tensor,
+    vec: Tensor,
+    timesteps: list[float],
+    guidance: float = 4.0,
+    id_weight=1.0,
+    id=None,
+    start_step=0,
+    uncond_id=None,
+    true_cfg=1.0,
+    timestep_to_start_cfg=1,
+    neg_txt=None,
+    neg_txt_ids=None,
+    neg_vec=None,
+    aggressive_offload=False,
+    y_1: Tensor = None,
+    gamma: float = 0.5,
+):
+    # reverse the timesteps
+    timesteps = timesteps[::-1]
+    guidance_vec = torch.full((img.shape[0],), guidance, device=img.device, dtype=img.dtype)
+    use_true_cfg = abs(true_cfg - 1.0) > 1e-2
+    for i in tqdm(range(len(timesteps) - 1), desc="Inverting"):
+        t_i = i / len(timesteps)
+        t_vec = torch.full((img.shape[0],), t_i, dtype=img.dtype, device=img.device)
+        pred = model(
+            img=img,
+            img_ids=img_ids,
+            txt=txt,
+            txt_ids=txt_ids,
+            y=vec,
+            timesteps=t_vec,
+            guidance=guidance_vec,
+            id=id if (len(timesteps) - 1 - i) >= start_step else None,
+            id_weight=id_weight,
+            aggressive_offload=aggressive_offload,
+        )
+
+        if use_true_cfg and i >= timestep_to_start_cfg:
+            neg_pred = model(
+                img=img,
+                img_ids=img_ids,
+                txt=neg_txt,
+                txt_ids=neg_txt_ids,
+                y=neg_vec,
+                timesteps=t_vec,
+                guidance=guidance_vec,
+                id=uncond_id if (len(timesteps) - 1 - i) >= start_step else None,
+                id_weight=id_weight,
+                aggressive_offload=aggressive_offload,
+            )
+            pred = neg_pred + true_cfg * (pred - neg_pred)
+
+        assert (1 - t_i) != 0
+        u_t_i_cond = (y_1 - img) / (1 - t_i)
+        pred = pred + gamma * (u_t_i_cond - pred)
+
+        img = img + (timesteps[i+1] - timesteps[i]) * pred
+
+    return img
+
+def rf_denoise(
+    model: Flux,
+    img: Tensor,
+    img_ids: Tensor,
+    txt: Tensor,
+    txt_ids: Tensor,
+    vec: Tensor,
+    timesteps: list[float],
+    guidance: float = 4.0,
+    id_weight=1.0,
+    id=None,
+    start_step=0,
+    uncond_id=None,
+    true_cfg=1.0,
+    timestep_to_start_cfg=1,
+    neg_txt=None,
+    neg_txt_ids=None,
+    neg_vec=None,
+    aggressive_offload=False,
+    y_0: Tensor = None,
+    eta=0.9,
+    s=0,
+    tau=6,
+):
+    guidance_vec = torch.full((img.shape[0],), guidance, device=img.device, dtype=img.dtype)
+    use_true_cfg = abs(true_cfg - 1.0) > 1e-2
+    for i in tqdm(range(len(timesteps) - 1), desc="Denoising"):
+        t_i = i / len(timesteps)
+        t_vec = torch.full((img.shape[0],), 1-t_i, dtype=img.dtype, device=img.device)
+        pred = model(
+            img=img,
+            img_ids=img_ids,
+            txt=txt,
+            txt_ids=txt_ids,
+            y=vec,
+            timesteps=t_vec,
+            guidance=guidance_vec,
+            id=id if i >= start_step else None,
+            id_weight=id_weight,
+            aggressive_offload=aggressive_offload,
+        )
+
+        if use_true_cfg and i >= timestep_to_start_cfg:
+            neg_pred = model(
+                img=img,
+                img_ids=img_ids,
+                txt=neg_txt,
+                txt_ids=neg_txt_ids,
+                y=neg_vec,
+                timesteps=t_vec,
+                guidance=guidance_vec,
+                id=uncond_id if i >= start_step else None,
+                id_weight=id_weight,
+                aggressive_offload=aggressive_offload,
+            )
+            pred = neg_pred + true_cfg * (pred - neg_pred)
+        pred = -pred
+
+        assert (1 - t_i) != 0
+        v_t_cond = (y_0 - img) / (1 - t_i)
+        eta_t = eta if s <= i < tau else 0
+        pred = pred + eta_t * (v_t_cond - pred)
+        
+        img = img + (timesteps[i] - timesteps[i+1]) * pred
+
+    return img
+
+def denoise(
+    model: Flux,
+    # model input
+    img: Tensor,
+    img_ids: Tensor,
+    txt: Tensor,
+    txt_ids: Tensor,
+    vec: Tensor,
+    timesteps: list[float],
+    guidance: float = 4.0,
+    id_weight=1.0,
+    id=None,
+    start_step=0,
+    uncond_id=None,
+    true_cfg=1.0,
+    timestep_to_start_cfg=1,
+    neg_txt=None,
+    neg_txt_ids=None,
+    neg_vec=None,
+    aggressive_offload=False,
+):
+    # this is ignored for schnell
+    guidance_vec = torch.full((img.shape[0],), guidance, device=img.device, dtype=img.dtype)
+    use_true_cfg = abs(true_cfg - 1.0) > 1e-2
+    for i, (t_curr, t_prev) in enumerate(zip(timesteps[:-1], timesteps[1:])):
+        t_vec = torch.full((img.shape[0],), t_curr, dtype=img.dtype, device=img.device)
+        pred = model(
+            img=img,
+            img_ids=img_ids,
+            txt=txt,
+            txt_ids=txt_ids,
+            y=vec,
+            timesteps=t_vec,
+            guidance=guidance_vec,
+            id=id if i >= start_step else None,
+            id_weight=id_weight,
+            aggressive_offload=aggressive_offload,
+        )
+
+        if use_true_cfg and i >= timestep_to_start_cfg:
+            neg_pred = model(
+                img=img,
+                img_ids=img_ids,
+                txt=neg_txt,
+                txt_ids=neg_txt_ids,
+                y=neg_vec,
+                timesteps=t_vec,
+                guidance=guidance_vec,
+                id=uncond_id if i >= start_step else None,
+                id_weight=id_weight,
+                aggressive_offload=aggressive_offload,
+            )
+            pred = neg_pred + true_cfg * (pred - neg_pred)
+
+        img = img + (t_prev - t_curr) * pred
+
+    return img
+
+
+def unpack(x: Tensor, height: int, width: int) -> Tensor:
+    return rearrange(
+        x,
+        "b (h w) (c ph pw) -> b c (h ph) (w pw)",
+        h=math.ceil(height / 16),
+        w=math.ceil(width / 16),
+        ph=2,
+        pw=2,
+    )

flux/util.py

@@ -0,0 +1,249 @@
+import json
+import os
+from dataclasses import dataclass
+
+import torch
+from huggingface_hub import hf_hub_download
+from safetensors.torch import load_file as load_sft
+
+from flux.model import Flux, FluxParams
+from flux.modules.autoencoder import AutoEncoder, AutoEncoderParams
+from flux.modules.conditioner import HFEmbedder
+
+
+@dataclass
+class SamplingOptions:
+    prompt: str
+    width: int
+    height: int
+    num_steps: int
+    guidance: float
+    seed: int
+
+
+@dataclass
+class ModelSpec:
+    params: FluxParams
+    ae_params: AutoEncoderParams
+    ckpt_path: str
+    ae_path: str
+    repo_id: str
+    repo_flow: str
+    repo_ae: str
+
+
+configs = {
+    "flux-dev": ModelSpec(
+        repo_id="black-forest-labs/FLUX.1-dev",
+        repo_flow="flux1-dev.safetensors",
+        repo_ae="ae.safetensors",
+        ckpt_path='models/flux1-dev.safetensors',
+        params=FluxParams(
+            in_channels=64,
+            vec_in_dim=768,
+            context_in_dim=4096,
+            hidden_size=3072,
+            mlp_ratio=4.0,
+            num_heads=24,
+            depth=19,
+            depth_single_blocks=38,
+            axes_dim=[16, 56, 56],
+            theta=10_000,
+            qkv_bias=True,
+            guidance_embed=True,
+        ),
+        ae_path='models/ae.safetensors',
+        ae_params=AutoEncoderParams(
+            resolution=256,
+            in_channels=3,
+            ch=128,
+            out_ch=3,
+            ch_mult=[1, 2, 4, 4],
+            num_res_blocks=2,
+            z_channels=16,
+            scale_factor=0.3611,
+            shift_factor=0.1159,
+        ),
+    ),
+    "flux-schnell": ModelSpec(
+        repo_id="black-forest-labs/FLUX.1-schnell",
+        repo_flow="flux1-schnell.safetensors",
+        repo_ae="ae.safetensors",
+        ckpt_path=os.getenv("FLUX_SCHNELL"),
+        params=FluxParams(
+            in_channels=64,
+            vec_in_dim=768,
+            context_in_dim=4096,
+            hidden_size=3072,
+            mlp_ratio=4.0,
+            num_heads=24,
+            depth=19,
+            depth_single_blocks=38,
+            axes_dim=[16, 56, 56],
+            theta=10_000,
+            qkv_bias=True,
+            guidance_embed=False,
+        ),
+        ae_path=os.getenv("AE"),
+        ae_params=AutoEncoderParams(
+            resolution=256,
+            in_channels=3,
+            ch=128,
+            out_ch=3,
+            ch_mult=[1, 2, 4, 4],
+            num_res_blocks=2,
+            z_channels=16,
+            scale_factor=0.3611,
+            shift_factor=0.1159,
+        ),
+    ),
+}
+
+
+def print_load_warning(missing: list[str], unexpected: list[str]) -> None:
+    if len(missing) > 0 and len(unexpected) > 0:
+        print(f"Got {len(missing)} missing keys:\n\t" + "\n\t".join(missing))
+        print("\n" + "-" * 79 + "\n")
+        print(f"Got {len(unexpected)} unexpected keys:\n\t" + "\n\t".join(unexpected))
+    elif len(missing) > 0:
+        print(f"Got {len(missing)} missing keys:\n\t" + "\n\t".join(missing))
+    elif len(unexpected) > 0:
+        print(f"Got {len(unexpected)} unexpected keys:\n\t" + "\n\t".join(unexpected))
+
+
+def load_flow_model(name: str, device: str = "cuda", hf_download: bool = True):
+    # Loading Flux
+    print("Init model")
+    ckpt_path = configs[name].ckpt_path
+    if (
+        not os.path.exists(ckpt_path)
+        and configs[name].repo_id is not None
+        and configs[name].repo_flow is not None
+        and hf_download
+    ):
+        ckpt_path = hf_hub_download(configs[name].repo_id, configs[name].repo_flow, local_dir='models')
+
+# Initialize the model on the 'meta' device, which doesn't allocate real memory
+    with torch.device('meta'):
+        model = Flux(configs[name].params)
+    model = model.to_empty(device=device)
+
+    if ckpt_path is not None:
+        print("Loading checkpoint")
+        # Load the state dictionary directly to the desired device
+        sd = load_sft(ckpt_path, device=str(device))
+        # Load the state dictionary into the model
+        missing, unexpected = model.load_state_dict(sd, strict=False)
+        print_load_warning(missing, unexpected)
+    model.to(torch.bfloat16)
+    return model
+
+# from XLabs-AI https://github.com/XLabs-AI/x-flux/blob/1f8ef54972105ad9062be69fe6b7f841bce02a08/src/flux/util.py#L330
+def load_flow_model_quintized(name: str, device: str = "cuda", hf_download: bool = True):
+    # Loading Flux
+    print("Init model")
+    ckpt_path = 'models/flux-dev-fp8.safetensors'
+    if (
+        not os.path.exists(ckpt_path)
+        and hf_download
+    ):
+        print("Downloading model")
+        ckpt_path = hf_hub_download("XLabs-AI/flux-dev-fp8", "flux-dev-fp8.safetensors")
+        print("Model downloaded to", ckpt_path)
+    json_path = hf_hub_download("XLabs-AI/flux-dev-fp8", 'flux_dev_quantization_map.json')
+
+    model = Flux(configs[name].params).to(torch.bfloat16)
+def load_flow_model_quintized(
+    name: str,
+    device: str = "cuda",
+    hf_download: bool = True,
+    cache_path: str = None,
+):
+    """
+    Loads (or downloads) a FLUX-fp8 checkpoint, performs quantization once,
+    and caches the quantized model to disk. Future calls load from cache.
+    
+    :param name: model name key in configs (e.g. "flux-dev-fp8")
+    :param device: Torch device string ("cuda" or "cpu")
+    :param hf_download: Whether to download from HF if local ckpt is missing
+    :param cache_path: Filepath for cached quantized model
+    :return: A quantized FLUX model on the specified device.
+    """
+    if cache_path is None:
+        cache_path = os.path.join(os.path.expanduser("~"), ".cache/flux_dev_fp8_quantized_model.pth")
+
+
+
+    # 1) Check if we already have a cached, quantized model
+    if os.path.exists(cache_path):
+        print(f"Loading cached quantized model from '{cache_path}'...")
+        model = torch.load(cache_path, map_location=device)
+        return model.to(device)
+
+    # 2) If no cache, build and quantize for the first time.
+    print("No cached model found. Initializing + quantizing from scratch.")
+
+    # (A) Download or specify checkpoint paths
+    ckpt_path = "models/flux-dev-fp8.safetensors"
+    if not os.path.exists(ckpt_path) and hf_download:
+        print("Downloading model checkpoint from HF...")
+        ckpt_path = hf_hub_download("XLabs-AI/flux-dev-fp8", "flux-dev-fp8.safetensors")
+        print("Model downloaded to:", ckpt_path)
+
+    json_path = hf_hub_download("XLabs-AI/flux-dev-fp8", "flux_dev_quantization_map.json")
+
+    # (B) Build the unquantized model
+    print("Initializing model in bfloat16...")
+    model = Flux(configs[name].params).to(torch.bfloat16)
+
+    # (C) Load the unquantized weights
+    print("Loading unquantized checkpoint to CPU...")
+    sd = load_sft(ckpt_path, device="cpu")  # CPU load
+
+    # (D) Load quantization map
+    with open(json_path, "r") as f:
+        quantization_map = json.load(f)
+
+    # (E) Quantize
+    print("Starting quantization process...")
+    from optimum.quanto import requantize
+    requantize(model, sd, quantization_map, device=device)
+    print("Quantization complete.")
+
+    # (F) Cache the fully quantized model to disk
+    print(f"Saving the quantized model to '{cache_path}'...")
+    torch.save(model, cache_path)
+    print("Model saved. Future runs will load from cache.")
+
+    return model.to(device)
+
+
+def load_t5(device: str = "cuda", max_length: int = 512) -> HFEmbedder:
+    # max length 64, 128, 256 and 512 should work (if your sequence is short enough)
+    return HFEmbedder("xlabs-ai/xflux_text_encoders", max_length=max_length, torch_dtype=torch.bfloat16).to(device)
+
+
+def load_clip(device: str = "cuda") -> HFEmbedder:
+    return HFEmbedder("openai/clip-vit-large-patch14", max_length=77, torch_dtype=torch.bfloat16).to(device)
+
+
+def load_ae(name: str, device: str = "cuda", hf_download: bool = True) -> AutoEncoder:
+    ckpt_path = configs[name].ae_path
+    if (
+        not os.path.exists(ckpt_path)
+        and configs[name].repo_id is not None
+        and configs[name].repo_ae is not None
+        and hf_download
+    ):
+        ckpt_path = hf_hub_download(configs[name].repo_id, configs[name].repo_ae, local_dir='models')
+
+    # Loading the autoencoder
+    print("Init AE")
+    with torch.device(device):
+        ae = AutoEncoder(configs[name].ae_params)
+
+    if ckpt_path is not None:
+        sd = load_sft(ckpt_path, device=str(device))
+        missing, unexpected = ae.load_state_dict(sd, strict=False)
+        print_load_warning(missing, unexpected)
+    return ae

pulid/attention_processor.py

@@ -0,0 +1,422 @@
+# modified from https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+NUM_ZERO = 0
+ORTHO = False
+ORTHO_v2 = False
+
+
+class AttnProcessor(nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    def __call__(
+        self,
+        attn,
+        hidden_states,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        temb=None,
+        id_embedding=None,
+        id_scale=1.0,
+    ):
+        residual = hidden_states
+
+        if attn.spatial_norm is not None:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+        attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        elif attn.norm_cross:
+            encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        query = attn.head_to_batch_dim(query)
+        key = attn.head_to_batch_dim(key)
+        value = attn.head_to_batch_dim(value)
+
+        attention_probs = attn.get_attention_scores(query, key, attention_mask)
+        hidden_states = torch.bmm(attention_probs, value)
+        hidden_states = attn.batch_to_head_dim(hidden_states)
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states
+
+
+class IDAttnProcessor(nn.Module):
+    r"""
+    Attention processor for ID-Adapater.
+    Args:
+        hidden_size (`int`):
+            The hidden size of the attention layer.
+        cross_attention_dim (`int`):
+            The number of channels in the `encoder_hidden_states`.
+        scale (`float`, defaults to 1.0):
+            the weight scale of image prompt.
+    """
+
+    def __init__(self, hidden_size, cross_attention_dim=None):
+        super().__init__()
+        self.id_to_k = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
+        self.id_to_v = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
+
+    def __call__(
+        self,
+        attn,
+        hidden_states,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        temb=None,
+        id_embedding=None,
+        id_scale=1.0,
+    ):
+        residual = hidden_states
+
+        if attn.spatial_norm is not None:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+        attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        elif attn.norm_cross:
+            encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        query = attn.head_to_batch_dim(query)
+        key = attn.head_to_batch_dim(key)
+        value = attn.head_to_batch_dim(value)
+
+        attention_probs = attn.get_attention_scores(query, key, attention_mask)
+        hidden_states = torch.bmm(attention_probs, value)
+        hidden_states = attn.batch_to_head_dim(hidden_states)
+
+        # for id-adapter
+        if id_embedding is not None:
+            if NUM_ZERO == 0:
+                id_key = self.id_to_k(id_embedding)
+                id_value = self.id_to_v(id_embedding)
+            else:
+                zero_tensor = torch.zeros(
+                    (id_embedding.size(0), NUM_ZERO, id_embedding.size(-1)),
+                    dtype=id_embedding.dtype,
+                    device=id_embedding.device,
+                )
+                id_key = self.id_to_k(torch.cat((id_embedding, zero_tensor), dim=1))
+                id_value = self.id_to_v(torch.cat((id_embedding, zero_tensor), dim=1))
+
+            id_key = attn.head_to_batch_dim(id_key).to(query.dtype)
+            id_value = attn.head_to_batch_dim(id_value).to(query.dtype)
+
+            id_attention_probs = attn.get_attention_scores(query, id_key, None)
+            id_hidden_states = torch.bmm(id_attention_probs, id_value)
+            id_hidden_states = attn.batch_to_head_dim(id_hidden_states)
+
+            if not ORTHO:
+                hidden_states = hidden_states + id_scale * id_hidden_states
+            else:
+                projection = (
+                    torch.sum((hidden_states * id_hidden_states), dim=-2, keepdim=True)
+                    / torch.sum((hidden_states * hidden_states), dim=-2, keepdim=True)
+                    * hidden_states
+                )
+                orthogonal = id_hidden_states - projection
+                hidden_states = hidden_states + id_scale * orthogonal
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states
+
+
+class AttnProcessor2_0(nn.Module):
+    r"""
+    Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0).
+    """
+
+    def __init__(self):
+        super().__init__()
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
+
+    def __call__(
+        self,
+        attn,
+        hidden_states,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        temb=None,
+        id_embedding=None,
+        id_scale=1.0,
+    ):
+        residual = hidden_states
+
+        if attn.spatial_norm is not None:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+
+        if attention_mask is not None:
+            attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+            # scaled_dot_product_attention expects attention_mask shape to be
+            # (batch, heads, source_length, target_length)
+            attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        elif attn.norm_cross:
+            encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        inner_dim = key.shape[-1]
+        head_dim = inner_dim // attn.heads
+
+        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        # the output of sdp = (batch, num_heads, seq_len, head_dim)
+        hidden_states = F.scaled_dot_product_attention(
+            query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
+        )
+
+        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+        hidden_states = hidden_states.to(query.dtype)
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states
+
+
+class IDAttnProcessor2_0(torch.nn.Module):
+    r"""
+    Attention processor for ID-Adapater for PyTorch 2.0.
+    Args:
+        hidden_size (`int`):
+            The hidden size of the attention layer.
+        cross_attention_dim (`int`):
+            The number of channels in the `encoder_hidden_states`.
+    """
+
+    def __init__(self, hidden_size, cross_attention_dim=None):
+        super().__init__()
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
+
+        self.id_to_k = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
+        self.id_to_v = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
+
+    def __call__(
+        self,
+        attn,
+        hidden_states,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        temb=None,
+        id_embedding=None,
+        id_scale=1.0,
+    ):
+        residual = hidden_states
+
+        if attn.spatial_norm is not None:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+
+        if attention_mask is not None:
+            attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+            # scaled_dot_product_attention expects attention_mask shape to be
+            # (batch, heads, source_length, target_length)
+            attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        elif attn.norm_cross:
+            encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        inner_dim = key.shape[-1]
+        head_dim = inner_dim // attn.heads
+
+        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        # the output of sdp = (batch, num_heads, seq_len, head_dim)
+        hidden_states = F.scaled_dot_product_attention(
+            query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
+        )
+
+        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+        hidden_states = hidden_states.to(query.dtype)
+
+        # for id embedding
+        if id_embedding is not None:
+            if NUM_ZERO == 0:
+                id_key = self.id_to_k(id_embedding).to(query.dtype)
+                id_value = self.id_to_v(id_embedding).to(query.dtype)
+            else:
+                zero_tensor = torch.zeros(
+                    (id_embedding.size(0), NUM_ZERO, id_embedding.size(-1)),
+                    dtype=id_embedding.dtype,
+                    device=id_embedding.device,
+                )
+                id_key = self.id_to_k(torch.cat((id_embedding, zero_tensor), dim=1)).to(query.dtype)
+                id_value = self.id_to_v(torch.cat((id_embedding, zero_tensor), dim=1)).to(query.dtype)
+
+            id_key = id_key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+            id_value = id_value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+            # the output of sdp = (batch, num_heads, seq_len, head_dim)
+            id_hidden_states = F.scaled_dot_product_attention(
+                query, id_key, id_value, attn_mask=None, dropout_p=0.0, is_causal=False
+            )
+
+            id_hidden_states = id_hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+            id_hidden_states = id_hidden_states.to(query.dtype)
+
+            if not ORTHO and not ORTHO_v2:
+                hidden_states = hidden_states + id_scale * id_hidden_states
+            elif ORTHO_v2:
+                orig_dtype = hidden_states.dtype
+                hidden_states = hidden_states.to(torch.float32)
+                id_hidden_states = id_hidden_states.to(torch.float32)
+                attn_map = query @ id_key.transpose(-2, -1)
+                attn_mean = attn_map.softmax(dim=-1).mean(dim=1)
+                attn_mean = attn_mean[:, :, :5].sum(dim=-1, keepdim=True)
+                projection = (
+                    torch.sum((hidden_states * id_hidden_states), dim=-2, keepdim=True)
+                    / torch.sum((hidden_states * hidden_states), dim=-2, keepdim=True)
+                    * hidden_states
+                )
+                orthogonal = id_hidden_states + (attn_mean - 1) * projection
+                hidden_states = hidden_states + id_scale * orthogonal
+                hidden_states = hidden_states.to(orig_dtype)
+            else:
+                orig_dtype = hidden_states.dtype
+                hidden_states = hidden_states.to(torch.float32)
+                id_hidden_states = id_hidden_states.to(torch.float32)
+                projection = (
+                    torch.sum((hidden_states * id_hidden_states), dim=-2, keepdim=True)
+                    / torch.sum((hidden_states * hidden_states), dim=-2, keepdim=True)
+                    * hidden_states
+                )
+                orthogonal = id_hidden_states - projection
+                hidden_states = hidden_states + id_scale * orthogonal
+                hidden_states = hidden_states.to(orig_dtype)
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states

pulid/encoders.py

@@ -0,0 +1,64 @@
+import torch
+import torch.nn as nn
+
+
+class IDEncoder(nn.Module):
+    def __init__(self, width=1280, context_dim=2048, num_token=5):
+        super().__init__()
+        self.num_token = num_token
+        self.context_dim = context_dim
+        h1 = min((context_dim * num_token) // 4, 1024)
+        h2 = min((context_dim * num_token) // 2, 1024)
+        self.body = nn.Sequential(
+            nn.Linear(width, h1),
+            nn.LayerNorm(h1),
+            nn.LeakyReLU(),
+            nn.Linear(h1, h2),
+            nn.LayerNorm(h2),
+            nn.LeakyReLU(),
+            nn.Linear(h2, context_dim * num_token),
+        )
+
+        for i in range(5):
+            setattr(
+                self,
+                f'mapping_{i}',
+                nn.Sequential(
+                    nn.Linear(1024, 1024),
+                    nn.LayerNorm(1024),
+                    nn.LeakyReLU(),
+                    nn.Linear(1024, 1024),
+                    nn.LayerNorm(1024),
+                    nn.LeakyReLU(),
+                    nn.Linear(1024, context_dim),
+                ),
+            )
+
+            setattr(
+                self,
+                f'mapping_patch_{i}',
+                nn.Sequential(
+                    nn.Linear(1024, 1024),
+                    nn.LayerNorm(1024),
+                    nn.LeakyReLU(),
+                    nn.Linear(1024, 1024),
+                    nn.LayerNorm(1024),
+                    nn.LeakyReLU(),
+                    nn.Linear(1024, context_dim),
+                ),
+            )
+
+    def forward(self, x, y):
+        # x shape [N, C]
+        x = self.body(x)
+        x = x.reshape(-1, self.num_token, self.context_dim)
+
+        hidden_states = ()
+        for i, emb in enumerate(y):
+            hidden_state = getattr(self, f'mapping_{i}')(emb[:, :1]) + getattr(self, f'mapping_patch_{i}')(
+                emb[:, 1:]
+            ).mean(dim=1, keepdim=True)
+            hidden_states += (hidden_state,)
+        hidden_states = torch.cat(hidden_states, dim=1)
+
+        return torch.cat([x, hidden_states], dim=1)

pulid/encoders_transformer.py

@@ -0,0 +1,209 @@
+import math
+
+import torch
+import torch.nn as nn
+
+
+# FFN
+def FeedForward(dim, mult=4):
+    inner_dim = int(dim * mult)
+    return nn.Sequential(
+        nn.LayerNorm(dim),
+        nn.Linear(dim, inner_dim, bias=False),
+        nn.GELU(),
+        nn.Linear(inner_dim, dim, bias=False),
+    )
+
+
+def reshape_tensor(x, heads):
+    bs, length, width = x.shape
+    # (bs, length, width) --> (bs, length, n_heads, dim_per_head)
+    x = x.view(bs, length, heads, -1)
+    # (bs, length, n_heads, dim_per_head) --> (bs, n_heads, length, dim_per_head)
+    x = x.transpose(1, 2)
+    # (bs, n_heads, length, dim_per_head) --> (bs*n_heads, length, dim_per_head)
+    x = x.reshape(bs, heads, length, -1)
+    return x
+
+
+class PerceiverAttentionCA(nn.Module):
+    def __init__(self, *, dim=3072, dim_head=128, heads=16, kv_dim=2048):
+        super().__init__()
+        self.scale = dim_head ** -0.5
+        self.dim_head = dim_head
+        self.heads = heads
+        inner_dim = dim_head * heads
+
+        self.norm1 = nn.LayerNorm(dim if kv_dim is None else kv_dim)
+        self.norm2 = nn.LayerNorm(dim)
+
+        self.to_q = nn.Linear(dim, inner_dim, bias=False)
+        self.to_kv = nn.Linear(dim if kv_dim is None else kv_dim, inner_dim * 2, bias=False)
+        self.to_out = nn.Linear(inner_dim, dim, bias=False)
+
+    def forward(self, x, latents):
+        """
+        Args:
+            x (torch.Tensor): image features
+                shape (b, n1, D)
+            latent (torch.Tensor): latent features
+                shape (b, n2, D)
+        """
+        x = self.norm1(x)
+        latents = self.norm2(latents)
+
+        b, seq_len, _ = latents.shape
+
+        q = self.to_q(latents)
+        k, v = self.to_kv(x).chunk(2, dim=-1)
+
+        q = reshape_tensor(q, self.heads)
+        k = reshape_tensor(k, self.heads)
+        v = reshape_tensor(v, self.heads)
+
+        # attention
+        scale = 1 / math.sqrt(math.sqrt(self.dim_head))
+        weight = (q * scale) @ (k * scale).transpose(-2, -1)  # More stable with f16 than dividing afterwards
+        weight = torch.softmax(weight.float(), dim=-1).type(weight.dtype)
+        out = weight @ v
+
+        out = out.permute(0, 2, 1, 3).reshape(b, seq_len, -1)
+
+        return self.to_out(out)
+
+
+class PerceiverAttention(nn.Module):
+    def __init__(self, *, dim, dim_head=64, heads=8, kv_dim=None):
+        super().__init__()
+        self.scale = dim_head ** -0.5
+        self.dim_head = dim_head
+        self.heads = heads
+        inner_dim = dim_head * heads
+
+        self.norm1 = nn.LayerNorm(dim if kv_dim is None else kv_dim)
+        self.norm2 = nn.LayerNorm(dim)
+
+        self.to_q = nn.Linear(dim, inner_dim, bias=False)
+        self.to_kv = nn.Linear(dim if kv_dim is None else kv_dim, inner_dim * 2, bias=False)
+        self.to_out = nn.Linear(inner_dim, dim, bias=False)
+
+    def forward(self, x, latents):
+        """
+        Args:
+            x (torch.Tensor): image features
+                shape (b, n1, D)
+            latent (torch.Tensor): latent features
+                shape (b, n2, D)
+        """
+        x = self.norm1(x)
+        latents = self.norm2(latents)
+
+        b, seq_len, _ = latents.shape
+
+        q = self.to_q(latents)
+        kv_input = torch.cat((x, latents), dim=-2)
+        k, v = self.to_kv(kv_input).chunk(2, dim=-1)
+
+        q = reshape_tensor(q, self.heads)
+        k = reshape_tensor(k, self.heads)
+        v = reshape_tensor(v, self.heads)
+
+        # attention
+        scale = 1 / math.sqrt(math.sqrt(self.dim_head))
+        weight = (q * scale) @ (k * scale).transpose(-2, -1)  # More stable with f16 than dividing afterwards
+        weight = torch.softmax(weight.float(), dim=-1).type(weight.dtype)
+        out = weight @ v
+
+        out = out.permute(0, 2, 1, 3).reshape(b, seq_len, -1)
+
+        return self.to_out(out)
+
+
+class IDFormer(nn.Module):
+    """
+    - perceiver resampler like arch (compared with previous MLP-like arch)
+    - we concat id embedding (generated by arcface) and query tokens as latents
+    - latents will attend each other and interact with vit features through cross-attention
+    - vit features are multi-scaled and inserted into IDFormer in order, currently, each scale corresponds to two
+      IDFormer layers
+    """
+    def __init__(
+            self,
+            dim=1024,
+            depth=10,
+            dim_head=64,
+            heads=16,
+            num_id_token=5,
+            num_queries=32,
+            output_dim=2048,
+            ff_mult=4,
+    ):
+        super().__init__()
+
+        self.num_id_token = num_id_token
+        self.dim = dim
+        self.num_queries = num_queries
+        assert depth % 5 == 0
+        self.depth = depth // 5
+        scale = dim ** -0.5
+
+        self.latents = nn.Parameter(torch.randn(1, num_queries, dim) * scale)
+        self.proj_out = nn.Parameter(scale * torch.randn(dim, output_dim))
+
+        self.layers = nn.ModuleList([])
+        for _ in range(depth):
+            self.layers.append(
+                nn.ModuleList(
+                    [
+                        PerceiverAttention(dim=dim, dim_head=dim_head, heads=heads),
+                        FeedForward(dim=dim, mult=ff_mult),
+                    ]
+                )
+            )
+
+        for i in range(5):
+            setattr(
+                self,
+                f'mapping_{i}',
+                nn.Sequential(
+                    nn.Linear(1024, 1024),
+                    nn.LayerNorm(1024),
+                    nn.LeakyReLU(),
+                    nn.Linear(1024, 1024),
+                    nn.LayerNorm(1024),
+                    nn.LeakyReLU(),
+                    nn.Linear(1024, dim),
+                ),
+            )
+
+        self.id_embedding_mapping = nn.Sequential(
+            nn.Linear(1280, 1024),
+            nn.LayerNorm(1024),
+            nn.LeakyReLU(),
+            nn.Linear(1024, 1024),
+            nn.LayerNorm(1024),
+            nn.LeakyReLU(),
+            nn.Linear(1024, dim * num_id_token),
+        )
+
+    def forward(self, x, y):
+
+        latents = self.latents.repeat(x.size(0), 1, 1)
+
+        num_duotu = x.shape[1] if x.ndim == 3 else 1
+
+        x = self.id_embedding_mapping(x)
+        x = x.reshape(-1, self.num_id_token * num_duotu, self.dim)
+
+        latents = torch.cat((latents, x), dim=1)
+
+        for i in range(5):
+            vit_feature = getattr(self, f'mapping_{i}')(y[i])
+            ctx_feature = torch.cat((x, vit_feature), dim=1)
+            for attn, ff in self.layers[i * self.depth: (i + 1) * self.depth]:
+                latents = attn(ctx_feature, latents) + latents
+                latents = ff(latents) + latents
+
+        latents = latents[:, :self.num_queries]
+        latents = latents @ self.proj_out
+        return latents

pulid/pipeline.py

@@ -0,0 +1,228 @@
+import gc
+
+import cv2
+import insightface
+import torch
+import torch.nn as nn
+from diffusers import (
+    DPMSolverMultistepScheduler,
+    StableDiffusionXLPipeline,
+    UNet2DConditionModel,
+)
+from facexlib.parsing import init_parsing_model
+from facexlib.utils.face_restoration_helper import FaceRestoreHelper
+from huggingface_hub import hf_hub_download, snapshot_download
+from insightface.app import FaceAnalysis
+from safetensors.torch import load_file
+from torchvision.transforms import InterpolationMode
+from torchvision.transforms.functional import normalize, resize
+
+from eva_clip import create_model_and_transforms
+from eva_clip.constants import OPENAI_DATASET_MEAN, OPENAI_DATASET_STD
+from pulid.encoders import IDEncoder
+from pulid.utils import img2tensor, is_torch2_available, tensor2img
+
+if is_torch2_available():
+    from pulid.attention_processor import AttnProcessor2_0 as AttnProcessor
+    from pulid.attention_processor import IDAttnProcessor2_0 as IDAttnProcessor
+else:
+    from pulid.attention_processor import AttnProcessor, IDAttnProcessor
+
+
+class PuLIDPipeline:
+    def __init__(self, *args, **kwargs):
+        super().__init__()
+        self.device = 'cuda'
+        sdxl_base_repo = 'stabilityai/stable-diffusion-xl-base-1.0'
+        sdxl_lightning_repo = 'ByteDance/SDXL-Lightning'
+        self.sdxl_base_repo = sdxl_base_repo
+
+        # load base model
+        unet = UNet2DConditionModel.from_config(sdxl_base_repo, subfolder='unet').to(self.device, torch.float16)
+        unet.load_state_dict(
+            load_file(
+                hf_hub_download(sdxl_lightning_repo, 'sdxl_lightning_4step_unet.safetensors'), device=self.device
+            )
+        )
+        self.hack_unet_attn_layers(unet)
+        self.pipe = StableDiffusionXLPipeline.from_pretrained(
+            sdxl_base_repo, unet=unet, torch_dtype=torch.float16, variant="fp16"
+        ).to(self.device)
+        self.pipe.watermark = None
+
+        # scheduler
+        self.pipe.scheduler = DPMSolverMultistepScheduler.from_config(
+            self.pipe.scheduler.config, timestep_spacing="trailing"
+        )
+
+        # ID adapters
+        self.id_adapter = IDEncoder().to(self.device)
+
+        # preprocessors
+        # face align and parsing
+        self.face_helper = FaceRestoreHelper(
+            upscale_factor=1,
+            face_size=512,
+            crop_ratio=(1, 1),
+            det_model='retinaface_resnet50',
+            save_ext='png',
+            device=self.device,
+        )
+        self.face_helper.face_parse = None
+        self.face_helper.face_parse = init_parsing_model(model_name='bisenet', device=self.device)
+        # clip-vit backbone
+        model, _, _ = create_model_and_transforms('EVA02-CLIP-L-14-336', 'eva_clip', force_custom_clip=True)
+        model = model.visual
+        self.clip_vision_model = model.to(self.device)
+        eva_transform_mean = getattr(self.clip_vision_model, 'image_mean', OPENAI_DATASET_MEAN)
+        eva_transform_std = getattr(self.clip_vision_model, 'image_std', OPENAI_DATASET_STD)
+        if not isinstance(eva_transform_mean, (list, tuple)):
+            eva_transform_mean = (eva_transform_mean,) * 3
+        if not isinstance(eva_transform_std, (list, tuple)):
+            eva_transform_std = (eva_transform_std,) * 3
+        self.eva_transform_mean = eva_transform_mean
+        self.eva_transform_std = eva_transform_std
+        # antelopev2
+        snapshot_download('DIAMONIK7777/antelopev2', local_dir='models/antelopev2')
+        self.app = FaceAnalysis(
+            name='antelopev2', root='.', providers=['CUDAExecutionProvider', 'CPUExecutionProvider']
+        )
+        self.app.prepare(ctx_id=0, det_size=(640, 640))
+        self.handler_ante = insightface.model_zoo.get_model('models/antelopev2/glintr100.onnx')
+        self.handler_ante.prepare(ctx_id=0)
+
+        gc.collect()
+        torch.cuda.empty_cache()
+
+        self.load_pretrain()
+
+        # other configs
+        self.debug_img_list = []
+
+    def hack_unet_attn_layers(self, unet):
+        id_adapter_attn_procs = {}
+        for name, _ in unet.attn_processors.items():
+            cross_attention_dim = None if name.endswith("attn1.processor") else unet.config.cross_attention_dim
+            if name.startswith("mid_block"):
+                hidden_size = unet.config.block_out_channels[-1]
+            elif name.startswith("up_blocks"):
+                block_id = int(name[len("up_blocks.")])
+                hidden_size = list(reversed(unet.config.block_out_channels))[block_id]
+            elif name.startswith("down_blocks"):
+                block_id = int(name[len("down_blocks.")])
+                hidden_size = unet.config.block_out_channels[block_id]
+            if cross_attention_dim is not None:
+                id_adapter_attn_procs[name] = IDAttnProcessor(
+                    hidden_size=hidden_size,
+                    cross_attention_dim=cross_attention_dim,
+                ).to(unet.device)
+            else:
+                id_adapter_attn_procs[name] = AttnProcessor()
+        unet.set_attn_processor(id_adapter_attn_procs)
+        self.id_adapter_attn_layers = nn.ModuleList(unet.attn_processors.values())
+
+    def load_pretrain(self):
+        hf_hub_download('guozinan/PuLID', 'pulid_v1.bin', local_dir='models')
+        ckpt_path = 'models/pulid_v1.bin'
+        state_dict = torch.load(ckpt_path, map_location='cpu')
+        state_dict_dict = {}
+        for k, v in state_dict.items():
+            module = k.split('.')[0]
+            state_dict_dict.setdefault(module, {})
+            new_k = k[len(module) + 1 :]
+            state_dict_dict[module][new_k] = v
+
+        for module in state_dict_dict:
+            print(f'loading from {module}')
+            getattr(self, module).load_state_dict(state_dict_dict[module], strict=True)
+
+    def to_gray(self, img):
+        x = 0.299 * img[:, 0:1] + 0.587 * img[:, 1:2] + 0.114 * img[:, 2:3]
+        x = x.repeat(1, 3, 1, 1)
+        return x
+
+    def get_id_embedding(self, image):
+        """
+        Args:
+            image: numpy rgb image, range [0, 255]
+        """
+        self.face_helper.clean_all()
+        image_bgr = cv2.cvtColor(image, cv2.COLOR_RGB2BGR)
+        # get antelopev2 embedding
+        face_info = self.app.get(image_bgr)
+        if len(face_info) > 0:
+            face_info = sorted(face_info, key=lambda x: (x['bbox'][2] - x['bbox'][0]) * (x['bbox'][3] - x['bbox'][1]))[
+                -1
+            ]  # only use the maximum face
+            id_ante_embedding = face_info['embedding']
+            self.debug_img_list.append(
+                image[
+                    int(face_info['bbox'][1]) : int(face_info['bbox'][3]),
+                    int(face_info['bbox'][0]) : int(face_info['bbox'][2]),
+                ]
+            )
+        else:
+            id_ante_embedding = None
+
+        # using facexlib to detect and align face
+        self.face_helper.read_image(image_bgr)
+        self.face_helper.get_face_landmarks_5(only_center_face=True)
+        self.face_helper.align_warp_face()
+        if len(self.face_helper.cropped_faces) == 0:
+            raise RuntimeError('facexlib align face fail')
+        align_face = self.face_helper.cropped_faces[0]
+        # incase insightface didn't detect face
+        if id_ante_embedding is None:
+            print('fail to detect face using insightface, extract embedding on align face')
+            id_ante_embedding = self.handler_ante.get_feat(align_face)
+
+        id_ante_embedding = torch.from_numpy(id_ante_embedding).to(self.device)
+        if id_ante_embedding.ndim == 1:
+            id_ante_embedding = id_ante_embedding.unsqueeze(0)
+
+        # parsing
+        input = img2tensor(align_face, bgr2rgb=True).unsqueeze(0) / 255.0
+        input = input.to(self.device)
+        parsing_out = self.face_helper.face_parse(normalize(input, [0.485, 0.456, 0.406], [0.229, 0.224, 0.225]))[0]
+        parsing_out = parsing_out.argmax(dim=1, keepdim=True)
+        bg_label = [0, 16, 18, 7, 8, 9, 14, 15]
+        bg = sum(parsing_out == i for i in bg_label).bool()
+        white_image = torch.ones_like(input)
+        # only keep the face features
+        face_features_image = torch.where(bg, white_image, self.to_gray(input))
+        self.debug_img_list.append(tensor2img(face_features_image, rgb2bgr=False))
+
+        # transform img before sending to eva-clip-vit
+        face_features_image = resize(face_features_image, self.clip_vision_model.image_size, InterpolationMode.BICUBIC)
+        face_features_image = normalize(face_features_image, self.eva_transform_mean, self.eva_transform_std)
+        id_cond_vit, id_vit_hidden = self.clip_vision_model(
+            face_features_image, return_all_features=False, return_hidden=True, shuffle=False
+        )
+        id_cond_vit_norm = torch.norm(id_cond_vit, 2, 1, True)
+        id_cond_vit = torch.div(id_cond_vit, id_cond_vit_norm)
+
+        id_cond = torch.cat([id_ante_embedding, id_cond_vit], dim=-1)
+        id_uncond = torch.zeros_like(id_cond)
+        id_vit_hidden_uncond = []
+        for layer_idx in range(0, len(id_vit_hidden)):
+            id_vit_hidden_uncond.append(torch.zeros_like(id_vit_hidden[layer_idx]))
+
+        id_embedding = self.id_adapter(id_cond, id_vit_hidden)
+        uncond_id_embedding = self.id_adapter(id_uncond, id_vit_hidden_uncond)
+
+        # return id_embedding
+        return torch.cat((uncond_id_embedding, id_embedding), dim=0)
+
+    def inference(self, prompt, size, prompt_n='', image_embedding=None, id_scale=1.0, guidance_scale=1.2, steps=4):
+        images = self.pipe(
+            prompt=prompt,
+            negative_prompt=prompt_n,
+            num_images_per_prompt=size[0],
+            height=size[1],
+            width=size[2],
+            num_inference_steps=steps,
+            guidance_scale=guidance_scale,
+            cross_attention_kwargs={'id_embedding': image_embedding, 'id_scale': id_scale},
+        ).images
+
+        return images

pulid/pipeline_flux.py

@@ -0,0 +1,194 @@
+import gc
+
+import cv2
+import insightface
+import torch
+import torch.nn as nn
+from facexlib.parsing import init_parsing_model
+from facexlib.utils.face_restoration_helper import FaceRestoreHelper
+from huggingface_hub import hf_hub_download, snapshot_download
+from insightface.app import FaceAnalysis
+from safetensors.torch import load_file
+from torchvision.transforms import InterpolationMode
+from torchvision.transforms.functional import normalize, resize
+
+from eva_clip import create_model_and_transforms
+from eva_clip.constants import OPENAI_DATASET_MEAN, OPENAI_DATASET_STD
+from pulid.encoders_transformer import IDFormer, PerceiverAttentionCA
+from pulid.utils import img2tensor, tensor2img
+
+
+class PuLIDPipeline(nn.Module):
+    def __init__(self, dit, device, weight_dtype=torch.bfloat16, onnx_provider='gpu', *args, **kwargs):
+        super().__init__()
+        self.device = device
+        self.weight_dtype = weight_dtype
+        double_interval = 2
+        single_interval = 4
+
+        # init encoder
+        self.pulid_encoder = IDFormer().to(self.device, self.weight_dtype)
+
+        num_ca = 19 // double_interval + 38 // single_interval
+        if 19 % double_interval != 0:
+            num_ca += 1
+        if 38 % single_interval != 0:
+            num_ca += 1
+        self.pulid_ca = nn.ModuleList([
+            PerceiverAttentionCA().to(self.device, self.weight_dtype) for _ in range(num_ca)
+        ])
+
+        dit.pulid_ca = self.pulid_ca
+        dit.pulid_double_interval = double_interval
+        dit.pulid_single_interval = single_interval
+
+        # preprocessors
+        # face align and parsing
+        self.face_helper = FaceRestoreHelper(
+            upscale_factor=1,
+            face_size=512,
+            crop_ratio=(1, 1),
+            det_model='retinaface_resnet50',
+            save_ext='png',
+            device=self.device,
+        )
+        self.face_helper.face_parse = None
+        self.face_helper.face_parse = init_parsing_model(model_name='bisenet', device=self.device)
+        # clip-vit backbone
+        model, _, _ = create_model_and_transforms('EVA02-CLIP-L-14-336', 'eva_clip', force_custom_clip=True)
+        model = model.visual
+        self.clip_vision_model = model.to(self.device, dtype=self.weight_dtype)
+        eva_transform_mean = getattr(self.clip_vision_model, 'image_mean', OPENAI_DATASET_MEAN)
+        eva_transform_std = getattr(self.clip_vision_model, 'image_std', OPENAI_DATASET_STD)
+        if not isinstance(eva_transform_mean, (list, tuple)):
+            eva_transform_mean = (eva_transform_mean,) * 3
+        if not isinstance(eva_transform_std, (list, tuple)):
+            eva_transform_std = (eva_transform_std,) * 3
+        self.eva_transform_mean = eva_transform_mean
+        self.eva_transform_std = eva_transform_std
+        # antelopev2
+        snapshot_download('DIAMONIK7777/antelopev2', local_dir='models/antelopev2')
+        providers = ['CPUExecutionProvider'] if onnx_provider == 'cpu' \
+            else ['CUDAExecutionProvider', 'CPUExecutionProvider']
+        self.app = FaceAnalysis(name='antelopev2', root='.', providers=providers)
+        self.app.prepare(ctx_id=0, det_size=(640, 640))
+        self.handler_ante = insightface.model_zoo.get_model('models/antelopev2/glintr100.onnx',
+                                                            providers=providers)
+        self.handler_ante.prepare(ctx_id=0)
+
+        gc.collect()
+        torch.cuda.empty_cache()
+
+        # self.load_pretrain()
+
+        # other configs
+        self.debug_img_list = []
+
+    def components_to_device(self, device):
+        # everything but pulid_ca
+        self.face_helper.face_det = self.face_helper.face_det.to(device)
+        self.face_helper.face_parse = self.face_helper.face_parse.to(device)
+        self.clip_vision_model = self.clip_vision_model.to(device)
+        self.pulid_encoder = self.pulid_encoder.to(device)
+
+    def load_pretrain(self, pretrain_path=None, version='v0.9.0'):
+        hf_hub_download('guozinan/PuLID', f'pulid_flux_{version}.safetensors', local_dir='models')
+        ckpt_path = f'models/pulid_flux_{version}.safetensors'
+        if pretrain_path is not None:
+            ckpt_path = pretrain_path
+        state_dict = load_file(ckpt_path)
+        state_dict_dict = {}
+        for k, v in state_dict.items():
+            module = k.split('.')[0]
+            state_dict_dict.setdefault(module, {})
+            new_k = k[len(module) + 1:]
+            state_dict_dict[module][new_k] = v
+
+        for module in state_dict_dict:
+            print(f'loading from {module}')
+            getattr(self, module).load_state_dict(state_dict_dict[module], strict=True)
+
+        del state_dict
+        del state_dict_dict
+
+    def to_gray(self, img):
+        x = 0.299 * img[:, 0:1] + 0.587 * img[:, 1:2] + 0.114 * img[:, 2:3]
+        x = x.repeat(1, 3, 1, 1)
+        return x
+
+    @torch.no_grad()
+    def get_id_embedding(self, image, cal_uncond=False):
+        """
+        Args:
+            image: numpy rgb image, range [0, 255]
+        """
+        self.face_helper.clean_all()
+        self.debug_img_list = []
+        image_bgr = cv2.cvtColor(image, cv2.COLOR_RGB2BGR)
+        # get antelopev2 embedding
+        face_info = self.app.get(image_bgr)
+        if len(face_info) > 0:
+            face_info = sorted(face_info, key=lambda x: (x['bbox'][2] - x['bbox'][0]) * (x['bbox'][3] - x['bbox'][1]))[
+                -1
+            ]  # only use the maximum face
+            id_ante_embedding = face_info['embedding']
+            self.debug_img_list.append(
+                image[
+                    int(face_info['bbox'][1]) : int(face_info['bbox'][3]),
+                    int(face_info['bbox'][0]) : int(face_info['bbox'][2]),
+                ]
+            )
+        else:
+            id_ante_embedding = None
+
+        # using facexlib to detect and align face
+        self.face_helper.read_image(image_bgr)
+        self.face_helper.get_face_landmarks_5(only_center_face=True)
+        self.face_helper.align_warp_face()
+        if len(self.face_helper.cropped_faces) == 0:
+            raise RuntimeError('facexlib align face fail')
+        align_face = self.face_helper.cropped_faces[0]
+        # incase insightface didn't detect face
+        if id_ante_embedding is None:
+            print('fail to detect face using insightface, extract embedding on align face')
+            id_ante_embedding = self.handler_ante.get_feat(align_face)
+
+        id_ante_embedding = torch.from_numpy(id_ante_embedding).to(self.device, self.weight_dtype)
+        if id_ante_embedding.ndim == 1:
+            id_ante_embedding = id_ante_embedding.unsqueeze(0)
+
+        # parsing
+        input = img2tensor(align_face, bgr2rgb=True).unsqueeze(0) / 255.0
+        input = input.to(self.device)
+        parsing_out = self.face_helper.face_parse(normalize(input, [0.485, 0.456, 0.406], [0.229, 0.224, 0.225]))[0]
+        parsing_out = parsing_out.argmax(dim=1, keepdim=True)
+        bg_label = [0, 16, 18, 7, 8, 9, 14, 15]
+        bg = sum(parsing_out == i for i in bg_label).bool()
+        white_image = torch.ones_like(input)
+        # only keep the face features
+        face_features_image = torch.where(bg, white_image, self.to_gray(input))
+        self.debug_img_list.append(tensor2img(face_features_image, rgb2bgr=False))
+
+        # transform img before sending to eva-clip-vit
+        face_features_image = resize(face_features_image, self.clip_vision_model.image_size, InterpolationMode.BICUBIC)
+        face_features_image = normalize(face_features_image, self.eva_transform_mean, self.eva_transform_std)
+        id_cond_vit, id_vit_hidden = self.clip_vision_model(
+            face_features_image.to(self.weight_dtype), return_all_features=False, return_hidden=True, shuffle=False
+        )
+        id_cond_vit_norm = torch.norm(id_cond_vit, 2, 1, True)
+        id_cond_vit = torch.div(id_cond_vit, id_cond_vit_norm)
+
+        id_cond = torch.cat([id_ante_embedding, id_cond_vit], dim=-1)
+
+        id_embedding = self.pulid_encoder(id_cond, id_vit_hidden)
+
+        if not cal_uncond:
+            return id_embedding, None
+
+        id_uncond = torch.zeros_like(id_cond)
+        id_vit_hidden_uncond = []
+        for layer_idx in range(0, len(id_vit_hidden)):
+            id_vit_hidden_uncond.append(torch.zeros_like(id_vit_hidden[layer_idx]))
+        uncond_id_embedding = self.pulid_encoder(id_uncond, id_vit_hidden_uncond)
+
+        return id_embedding, uncond_id_embedding