unidisc / model_utils.py
aswerdlow's picture
Initial commit
131da64
import math
import os
import random
import typing
from contextlib import nullcontext
from dataclasses import dataclass
from pathlib import Path
from types import FrameType
from typing import Dict, List, Optional, Tuple, Union
import einops
import hydra
import hydra.utils
import numpy as np
import pandas as pd
import torch
import torch.distributed as dist
import torch.nn as nn
import torch.nn.functional as F
import torch.utils.checkpoint
import torchmetrics
import transformers
from image_utils import Im
from torch import Tensor, nn
from torch.utils.data import DataLoader, Dataset
from tqdm import tqdm
from tqdm.auto import tqdm
import models
import wandb
from decoupled_utils import (Profiler, barrier, dprint, get_rank,
get_slurm_job_id, get_world_size, gprint,
is_local_main_process, is_main_process,
is_torch_cuda_available, is_torch_xla_available,
module_hash, mprint, parameter_hash, print_memory,
rank_zero_fn, rprint, save_memory_profile,
show_memory_usage, try_except, use_dist)
is_xla_available = is_torch_xla_available()
if is_xla_available:
from unidisc.utils.standalone_metrics import MeanMetric, MetricCollection
else:
from torchmetrics import MetricCollection
from torchmetrics.aggregation import MeanMetric
LOG2 = math.log(2)
@try_except(write_error_to_file=True)
def log(*arg, **kwargs):
for key, value in arg[0].items():
if isinstance(value, torch.Tensor):
arg[0][key] = value.detach().cpu().float()
if is_main_process():
wandb.log(*arg, **kwargs)
def replace_nan_dict(x):
return {k: v.nan_to_num(0) for k, v in x.items()}
def ddprint(*args, **kwargs):
mprint(*args, **kwargs)
def empty_device_cache():
if is_torch_cuda_available():
torch.cuda.empty_cache()
else:
dprint("Not using cuda, skipping cache clear")
def update_logs(_logs, _extra_logs):
_logs.update(_extra_logs())
for k, v in _logs.items():
if isinstance(v, torch.Tensor):
_logs[k] = v.detach().cpu().item()
gprint(f"Converting {k} to item: {v}")
log(_logs)
def ema_update(model_dest: nn.Module, model_src: nn.Module, rate):
param_dict_src = dict(model_src.named_parameters())
for p_name, p_dest in model_dest.named_parameters():
if p_name not in param_dict_src:
print(f"Parameter {p_name} not found in src: {param_dict_src}")
p_src = param_dict_src[p_name]
assert p_src is not p_dest
p_dest.data.mul_(rate).add_((1 - rate) * p_src.data)
def identity(x):
return x
def remap_image_torch(image):
image_torch = image * 255
image_torch = torch.clip(image_torch, 0, 255).to(torch.uint8)
return image_torch
def _sample_categorical(categorical_probs):
gumbel_norm = 1e-10 - (torch.rand_like(categorical_probs) + 1e-10).log()
return (categorical_probs / gumbel_norm).argmax(dim=-1)
def wrapped_batch_decode(tokenizer, tokens, disable_mask_after_eos=False, **kwargs):
tokens = tokens.clone()
if (tokenizer.bos_token_id != tokenizer.eos_token_id) and not disable_mask_after_eos:
after_first_five = torch.cumsum(tokens == tokenizer.eos_token_id, dim=1).bool()
tokens[after_first_five.cumsum(dim=1) > 1] = tokenizer.pad_token_id
return tokenizer.batch_decode(tokens, **kwargs)
def _unsqueeze(x, reference):
return x.view(*x.shape, *((1,) * (len(reference.shape) - len(x.shape))))
@dataclass
class Loss:
loss: torch.FloatTensor
img_loss: torch.FloatTensor = None
txt_loss: torch.FloatTensor = None
nlls: torch.FloatTensor = None
token_mask: torch.FloatTensor = None
txt_nlls: torch.FloatTensor = None
img_nlls: torch.FloatTensor = None
extra_losses: dict = None
modality_mask: torch.FloatTensor = None
class NLL(MeanMetric):
pass
class BPD(NLL):
def compute(self) -> Tensor:
"""Computes the bits per dimension.
Returns:
bpd
"""
return self.mean_value / self.weight / LOG2
class Perplexity(NLL):
def compute(self) -> Tensor:
"""Computes the Perplexity.
Returns:
Perplexity
"""
return torch.exp(self.mean_value / self.weight)
class Entropy(NLL):
def compute(self) -> Tensor:
"""Computes the Entropy.
Returns:
Entropy
"""
return self.mean_value / self.weight
class MauveScore(NLL):
def compute(self) -> Tensor:
"""Computes the Mauve Score.
Returns:
Mauve Score
"""
return self.mean_value / self.weight
class CIDErScore(NLL):
def compute(self) -> Tensor:
"""Computes the CIDEr Score.
Returns:
CIDEr Score
"""
return self.mean_value / self.weight
class Accuracy(NLL):
def compute(self) -> Tensor:
"""Computes the Accuracy.
Returns:
Accuracy
"""
return self.mean_value / self.weight
def get_coord_plot(self):
from mup.coord_check import get_coord_data, plot_coord_data
def gen(w):
def f():
from copy import deepcopy
from omegaconf import read_write
import models as _models
_conf = deepcopy(self.config)
with read_write(_conf):
_conf.model.hidden_size = _conf.model.n_heads * w
_backbone = _models.dit.DIT(
_conf, vocab_size=self.vocab_size, mask_index=self.mask_index, text_vocab_size=self.text_vocab_size, dtype=self.dtype
)
self.get_base_shapes_for_mup(_backbone)
return _backbone
return f
optimizer = 'adamw'
widths = np.array([2**i for i in range(2, 6)])
models = {int(w) * self.config.model.n_heads: gen(int(w)) for w in widths}
fake_dataloader = []
self.validation_dataloader.num_workers = 0
nsteps = 30
for i, dataloader_batch in enumerate(self.validation_dataloader):
fake_batch = self.update_batch(dataloader_batch)
fake_batch['x0'] = fake_batch["input_ids"]
t = self._sample_t(fake_batch['x0'].shape[0], fake_batch['x0'].device)
sigma, dsigma = self.noise(t)
move_chance = 1 - torch.exp(-sigma[:, None])
xt = self.q_xt(fake_batch['x0'], move_chance)
fake_batch['xt'] = xt
fake_dataloader.append(fake_batch)
if i >= nsteps:
break
def loss_fn(_batch, _logits):
attention_mask = _batch['attention_mask']
model_output = self._subs_parameterization(logits=_logits, xt=_batch['xt'])
log_p_theta = torch.gather(input=model_output, dim=-1, index=_batch['x0'][:, :, None]).squeeze(-1)
std_weighting = (dsigma / torch.expm1(sigma))[:, None]
loss = -log_p_theta * std_weighting
loss = (loss * attention_mask).sum() / attention_mask.sum()
return loss
mup = True
lr = 1e-2
prm = 'μP' if mup else 'SP'
nseeds = 2
with torch.autocast(device_type=self.device.type, dtype=self.dtype):
df = get_coord_data(
models,
fake_dataloader,
lr=lr,
optimizer=optimizer,
nsteps=nsteps,
nseeds=nseeds,
dict_in_out=True,
lossfn=loss_fn,
mup=mup,
)
output_path = Path(__file__).parent / 'output' / f'{prm.lower()}_trsfmr_{optimizer}_coord.png'
output_path.parent.mkdir(parents=True, exist_ok=True)
plot_coord_data(
df,
legend='brief',
save_to=str(output_path.resolve()),
suptitle=f'{prm} Transformer {optimizer} lr={lr} nseeds={nseeds}',
face_color='xkcd:light grey' if not mup else None,
loglog=True
)
rprint(f"Saved coord plot to {output_path.resolve()}")
csv_path = output_path.with_suffix('.csv')
df.to_csv(csv_path, index=False)
rprint(f"DataFrame saved as CSV to {csv_path.resolve()}")
result = df[df['t'] == 1].nsmallest(100, 'l1').sort_values('l1', ascending=True)
with pd.option_context('display.max_rows', None, 'display.max_columns', None):
print(result[['module', 'width', 'l1']])
exit()
def _score_entropy(self, log_score, sigma, xt, x0):
"""Computes the SEDD loss.
Args:
log_score: float torch.Tensor with shape (batch_size,
diffusion_model_input_length, vocab_size),
log score, output of the denoising network.
xt: int torch.Tensor with shape (batch_size,
diffusion_model_input_length), input.
x0: int torch.Tensor with shape (batch_size,
diffusion_model_input_length), input.
sigma: float torch.Tensor with shape (batch_size, 1).
Returns:
loss with shape (batch_size, diffusion_model_input_length)
"""
masked_indices = xt == self.mask_index
expsig_minus_1 = torch.expm1(sigma).expand_as(xt)
q_ratio = 1 / expsig_minus_1[masked_indices]
words_that_were_masked = x0[masked_indices]
neg_term = q_ratio * torch.gather(log_score[masked_indices], -1, words_that_were_masked[..., None]).squeeze(-1)
score = log_score[masked_indices].exp()
if self.mask_index == self.vocab_size - 1:
pos_term = score[:, :-1].sum(dim=-1)
else:
pos_term = score[:, : self.mask_index].sum(dim=-1) + score[:, self.mask_index + 1 :].sum(dim=-1)
const = q_ratio * (q_ratio.log() - 1)
entropy = torch.zeros(*xt.shape, device=xt.device)
entropy[masked_indices] += pos_term - neg_term + const
return entropy
@torch.no_grad
def sample_subs_guidance(self, n_samples, stride_length, num_strides, dt=0.001):
ones = torch.ones(n_samples, dtype=self.dtype, device=self.device)
num_steps = int(1 / dt)
sampling_steps = 0
intermediate_tokens = []
target = None
for _ in range(num_strides + 1):
p_x0_cache = None
x = self._sample_prior(n_samples, self.config.model.length).to(self.device)
if target is not None:
x[:, :-stride_length] = target
for i in range(num_steps + 1):
p_x0_cache, x_next, nfe_cnt = self._ddpm_caching_update(x=x, t=(1 - i * dt) * ones, dt=dt, p_x0=p_x0_cache)
if not torch.allclose(x_next, x) or self.time_conditioning:
p_x0_cache = None
sampling_steps += 1
x = x_next
x = self.forward(x, 0 * ones).argmax(dim=-1)
intermediate_tokens.append(x[:, :stride_length].cpu().numpy())
target = x[:, stride_length:]
intermediate_tokens.append(target.cpu().numpy())
intermediate_text_samples = []
sequence_lengths = ((np.concatenate(intermediate_tokens, axis=1)[:, 1:] == self.tokenizer.eos_token_id).cumsum(-1) == 0).sum(-1)
for i in range(2, len(intermediate_tokens) + 1):
intermediate_text_samples.append(self.tokenizer.batch_decode(np.concatenate(intermediate_tokens[:i], axis=1)))
return (sampling_steps, intermediate_text_samples, sequence_lengths)
def restore_model_and_semi_ar_sample(self, stride_length, num_strides, dt=0.001):
"""Generate samples from the model."""
# Lightning auto-casting is not working in this method for some reason
if self.ema:
self.ema.store(self.get_params())
self.ema.copy_to(self.get_params())
self.backbone.eval()
(sampling_steps, samples, sequence_lengths) = self.sample_subs_guidance(
n_samples=self.config.loader.eval_batch_size, stride_length=stride_length, num_strides=num_strides, dt=dt
)
if self.ema:
self.ema.restore(self.get_params())
self.backbone.train()
self.noise.train()
return sampling_steps, samples, sequence_lengths
def _reconstruction_loss(self, x0):
t0 = torch.zeros(x0.shape[0], dtype=self.dtype, device=self.device)
assert self.config.noise.type == "loglinear"
# The above assert is for d3pm parameterization
unet_conditioning = self.noise(t0)[0][:, None]
model_output_t0 = self.forward(x0, unet_conditioning)
return -torch.gather(input=model_output_t0, dim=-1, index=x0[:, :, None]).squeeze(-1)
def restore_model_and_sample(self, num_steps, eps=1e-5):
"""Generate samples from the model."""
# Lightning auto-casting is not working in this method for some reason
if self.ema is not None:
self.ema.store(self.get_params())
self.ema.copy_to(self.get_params())
self.backbone.eval()
samples = self._sample(num_steps=num_steps, eps=eps)
if self.ema is not None:
self.ema.restore(self.get_params())
self.backbone.train()
return samples
def get_score(self, x, sigma, **kwargs):
model_output = self.forward(x, sigma, **kwargs)
if self.parameterization == "subs":
# score(x, t) = p_t(y) / p_t(x)
# => log score(x, t) = log p_t(y) - log p_t(x)
# case 1: x = masked
# (i) y = unmasked
# log score(x, t) = log p_\theta(x)|_y + log k
# where k = exp(- sigma) / (1 - exp(- sigma))
# (ii) y = masked
# log score(x, t) = 0
# case 2: x = unmasked
# (i) y != masked, y != x
# log score(x_i, t) = - inf
# (ii) y = x
# log score(x_i, t) = 0
# (iii) y = masked token
# log score(x_i, t) = - log k
# where k = exp(- sigma) / (1 - exp(- sigma))
log_k = -torch.log(torch.expm1(sigma)).squeeze(-1)
assert log_k.ndim == 1
masked_score = model_output + log_k[:, None, None]
masked_score[:, :, self.mask_index] = 0
unmasked_score = self.neg_infinity * torch.ones_like(model_output)
unmasked_score = torch.scatter(unmasked_score, -1, x[..., None], torch.zeros_like(unmasked_score[..., :1]))
unmasked_score[:, :, self.mask_index] = -(log_k[:, None] * torch.ones_like(x))
masked_indices = (x == self.mask_index).to(model_output.dtype)[:, :, None]
model_output = masked_score * masked_indices + unmasked_score * (1 - masked_indices)
return model_output.exp()
def _staggered_score(self, score, dsigma):
score = score.clone()
extra_const = (1 - dsigma.exp()) * score.sum(dim=-1)
score *= dsigma.exp()[:, None]
score[..., self.mask_index] += extra_const
return score
def _analytic_update(self, x, t, step_size):
curr_sigma, _ = self.noise(t)
next_sigma, _ = self.noise(t - step_size)
dsigma = curr_sigma - next_sigma
nfe_cnt = 0
score = self.get_score(x, curr_sigma)
nfe_cnt += 1
stag_score = self._staggered_score(score, dsigma)
probs = stag_score * self._transp_transition(x, dsigma)
return _sample_categorical(probs), nfe_cnt
def _denoiser_update(self, x, t):
sigma, _ = self.noise(t)
score = self.get_score(x, sigma)
stag_score = self._staggered_score(score, sigma)
probs = stag_score * self._transp_transition(x, sigma)
probs[..., self.mask_index] = 0
samples = _sample_categorical(probs)
return samples
def _transp_transition(self, i, sigma):
sigma = _unsqueeze(sigma, reference=i[..., None])
edge = torch.exp(-sigma) * F.one_hot(i, num_classes=self.vocab_size)
edge += torch.where(i == self.mask_index, 1 - torch.exp(-sigma).squeeze(-1), 0)[..., None]
return edge
@torch.no_grad()
def eval_retokenize(self, text_samples, max_length):
"""Retokenizes samples for the eval model.
Args:
text_samples: List of sentences generated by the model.
Returns:
samples: Samples re-tokenized for the eval model
attn_mask: Attention mask for the eval model
eval_context_size: Size of the context for the eval model
"""
if "llama2" in self.gen_ppl_eval_model_name_or_path:
tokenizer_kwargs = {
"text_samples": text_samples,
"return_tensors": "pt",
"return_token_type_ids": False,
"return_attention_mask": True,
"truncation": True,
"padding": True,
"max_length": max_length,
}
eval_context_size = 4096
else:
tokenizer_kwargs = {
"return_tensors": "pt",
"return_token_type_ids": False,
"return_attention_mask": True,
"truncation": True,
"padding": True,
"max_length": max_length,
}
eval_context_size = 1024
if getattr(self.config.eval, "force_eval_context_size_match_model", False):
eval_context_size = self.config.model.txt_length
samples = self.eval_model_tokenizer(text_samples, **tokenizer_kwargs)
attn_mask = samples["attention_mask"]
samples = samples["input_ids"]
if "llama2" not in self.gen_ppl_eval_model_name_or_path:
attn_mask = attn_mask.to(self.device)
samples = samples.to(self.device)
return samples, attn_mask, eval_context_size
@try_except(write_error_to_file=True)
@torch.no_grad()
def compute_cider(self, text_samples, gt_text_samples):
"""Compute the CIDEr score for the generated text.
Args:
text_samples: List of sentences generated by the model.
gt_text_samples: List of ground truth sentences.
Returns:
CIDEr score for the generated text.
"""
for text_sample, gt_text_sample in zip(text_samples, gt_text_samples):
self.cider_score_metric += (text_sample, gt_text_sample)
score = self.cider_score_metric.compute_cider() # list of np.float64
avg_score = sum(score) / len(score)
self.cider_score.update(avg_score.item()) # weight=len(text_samples))
def get_anole_data(model, processor, prompt, image, device):
inputs = processor(prompt, [image], padding=True, return_tensors="pt").to(device=device, dtype=dtype)
image_tokens = model.model.get_image_tokens(inputs["pixel_values"])
special_image_mask = inputs["input_ids"] == model.model.vocabulary_mapping.image_token_id
image_tokens = image_tokens.to(inputs["input_ids"].device, inputs["input_ids"].dtype)
inputs["input_ids"] = inputs["input_ids"].masked_scatter(special_image_mask, image_tokens)
inputs.pop("pixel_values")
inputs['input_ids'] = torch.load('save.pth').to(device)
return inputs
@try_except(write_error_to_file=True)
@torch.inference_mode()
def compute_generative_perplexity(self, text_samples: typing.List[str], retokenize: bool = True, max_length: typing.Optional[int] = None, gt: bool = False, return_raw_score: bool = False) -> None:
"""Compute the generative perplexity of the model.
Args:
text_samples: List of sentences generated by the model.
retokenize: Whether to retokenize using eval model's tokenizer
max_length: Maximum sequence length for tokenization
gt: Whether these are ground truth samples
return_raw_score: Whether to return raw NLL scores instead of updating metrics
Returns:
If return_raw_score is True, returns tensor of NLL scores.
Otherwise updates internal perplexity metrics.
"""
os.environ["TOKENIZERS_PARALLELISM"] = "false"
if not getattr(self.config.eval, 'enable_gen_pplx_cleanup', True):
eval_model = self.gen_pplx_eval_model
elif getattr(self.config.eval, 'gen_ppl_use_chameleon', False):
from transformers import (ChameleonForConditionalGeneration,
ChameleonProcessor)
model = ChameleonForConditionalGeneration.from_pretrained("leloy/Anole-7b-v0.1-hf", torch_dtype=torch.bfloat16).to("cuda")
processor = ChameleonProcessor.from_pretrained("leloy/Anole-7b-v0.1-hf")
image = Im(Im("https://cdn.outsideonline.com/wp-content/uploads/2023/03/Funny_Dog_H.jpg").np[50:-150, 550:-900, :]).resize(256, 256).pil
prompt = "A picture of a cat.<image>"
device = "cuda:0"
inputs = get_anole_data(model, processor, prompt, image, self.dtype, device)
output = model(input_ids=inputs['input_ids'].to(device))
attention_mask = torch.ones_like(inputs["input_ids"])
logits = output.logits
logits = logits.transpose(-1, -2)
sample_chunk = inputs["input_ids"]
nlls = F.cross_entropy(logits[..., :-1].to(device), sample_chunk[..., 1:].to(device), reduction="none")
nlls = nlls * attention_mask[..., 1:].to(nlls.dtype)
nlls = nlls.sum(-1) / attention_mask[..., 1:].sum(-1)
print(torch.exp(nlls))
else:
eval_model = transformers.AutoModelForCausalLM.from_pretrained(self.gen_ppl_eval_model_name_or_path).eval()
if max_length is None:
max_length = self.config.model.txt_length
if "llama2" not in self.gen_ppl_eval_model_name_or_path:
eval_model = eval_model.to(self.device)
# Re-tokenize using eval model's tokenizer
if retokenize:
(samples, attn_mask, eval_context_size) = self.eval_retokenize(text_samples, max_length=max_length)
else:
samples = text_samples
attn_mask = torch.ones(samples.shape).to(self.device)
eval_context_size = samples.shape[-1]
batch_size = min(self.config.eval.perplexity_batch_size, samples.shape[0])
num_batches = (samples.shape[0] + batch_size - 1) // batch_size
all_nlls = []
all_valid_mask = []
for i in range(num_batches):
batch_samples = samples[i * batch_size : (i + 1) * batch_size]
batch_attn_mask = attn_mask[i * batch_size : (i + 1) * batch_size]
with torch.nn.attention.sdpa_kernel([torch.nn.attention.SDPBackend.FLASH_ATTENTION, torch.nn.attention.SDPBackend.EFFICIENT_ATTENTION]):
logits = eval_model(batch_samples, attention_mask=batch_attn_mask)[0]
logits = logits.transpose(-1, -2)
nlls = F.cross_entropy(logits[..., :-1], batch_samples[..., 1:], reduction="none")
# Only consider tokens up to first EOS or padding
first_eos = (batch_samples == self.eval_model_tokenizer.eos_token_id).cumsum(-1) <= 1
token_mask = batch_attn_mask[..., 1:] > 0
valid_mask = first_eos[..., 1:] * token_mask
if not return_raw_score:
if gt:
self.gt_gen_ppl_metric.update(nlls, valid_mask)
else:
self.gen_ppl_metric.update(nlls, valid_mask)
else:
all_nlls.append(nlls)
all_valid_mask.append(valid_mask)
if getattr(self.config.eval, 'enable_gen_pplx_cleanup', True):
eval_model.to(torch.device('cpu'))
del eval_model
if return_raw_score:
all_nlls = torch.cat(all_nlls)
all_valid_mask = torch.cat(all_valid_mask)
# Compute mean NLL per sequence, ignoring padding/post-EOS tokens
nll = (all_nlls * all_valid_mask).sum(-1) / all_valid_mask.sum(-1)
return nll
def _d3pm_loss(self, model_output, xt, x0, t):
dt = 1 / self.T
if torch.is_tensor(t):
t = t[:, None]
assert t.ndim == 2
t = t.clamp(0.0, 1.0 - 1e-4)
alpha_t = 1 - t + torch.zeros_like(xt)
alpha_s = 1 - (t - dt) + torch.zeros_like(xt)
log_x_theta_at_x0 = torch.gather(model_output, -1, x0[:, :, None]).squeeze(-1)
log_x_theta_at_m = model_output[:, :, self.mask_index]
x_theta_at_m = log_x_theta_at_m.exp()
term_1_coef = dt / t
term_1_log_nr = torch.log(alpha_t * x_theta_at_m / t + 1)
term_1_log_dr = log_x_theta_at_x0
term_2_coef = 1 - dt / t
term_2_log_nr = term_1_log_nr
term_2_log_dr = torch.log(alpha_s * x_theta_at_m / (t - dt) + 1)
L_vb_masked = term_1_coef * (term_1_log_nr - term_1_log_dr) + term_2_coef * (term_2_log_nr - term_2_log_dr)
L_vb = L_vb_masked * (xt == self.mask_index)
return self.T * L_vb
def _d3pm_parameterization(self, logits):
if self.subs_masking:
logits[:, :, self.mask_index] += self.neg_infinity
logits = logits - torch.logsumexp(logits, dim=-1, keepdim=True)
return logits
def _sedd_parameterization(self, logits, xt, sigma):
esigm1_log = torch.where(sigma < 0.5, torch.expm1(sigma), sigma.exp() - 1).log().to(logits.dtype)
# logits shape
# (batch_size, diffusion_model_input_length, vocab_size)
logits = logits - esigm1_log[:, None, None] - np.log(logits.shape[-1] - 1)
# The below scatter operation sets the log score
# for the input word to 0.
logits = torch.scatter(logits, -1, xt[..., None], torch.zeros_like(logits[..., :1]))
return logits
def get_base_shapes_for_mup(self, _model):
from copy import deepcopy
from mup import set_base_shapes
from omegaconf import read_write
base_config = deepcopy(self.config)
with read_write(base_config):
base_config.model.hidden_size = base_config.model.n_heads # We need at least n_heads dim
delta_config = deepcopy(base_config)
with read_write(delta_config):
delta_config.model.hidden_size = base_config.model.n_heads * 2
base_model = models.dit.DIT(
base_config, vocab_size=self.vocab_size, mask_index=self.mask_index, text_vocab_size=self.text_vocab_size, dtype=self.dtype
)
delta_model = models.dit.DIT(
delta_config, vocab_size=self.vocab_size, mask_index=self.mask_index, text_vocab_size=self.text_vocab_size, dtype=self.dtype
)
set_base_shapes(_model, base_model, delta=delta_model)
def update_histogram(histogram, timesteps: torch.Tensor, losses: torch.Tensor):
for t, l in zip(timesteps, losses):
if t.item() in histogram:
histogram[t.item()].append(l.item())
else:
histogram[t.item()] = [l.item()]
def _maybe_sub_sample(self, x0, attention_mask):
seqlen = x0.shape[1]
if seqlen > self.config.model.length:
if not getattr(self.config.eval, 'big_seq_len_eval', False):
assert seqlen == 2 * self.config.model.length
# cropping is needed for text8-crop dataset
# try the same starting point for now
start = np.random.choice(self.config.model.length)
end = start + self.config.model.length
input_tokens = x0[:, start:end]
output_tokens = x0[:, start + 1 : end + 1]
new_attention_mask = attention_mask[:, start:end]
# Helps with validation PPL, since the val
# examples will all start and end with BOS/EOS
input_tokens[:, 0] = self.tokenizer.bos_token_id
output_tokens[:, -1] = self.tokenizer.eos_token_id
else:
input_tokens = x0
output_tokens = None
new_attention_mask = attention_mask
return input_tokens, output_tokens, new_attention_mask
from unidisc.tokenizers.image_tokenizers import decode_latents
def viz_images_from_dataloader(self):
_iter = iter(self.train_dataloader)
random_elements = [next(_iter) for _ in range(10)]
# random_elements[0]['input_ids'] - self.text_vocab_size
out = decode_latents(self.config, self.get_vae(), torch.cat([torch.zeros_like(random_elements[0]['input_ids'][:, :1]), (random_elements[0]['input_ids'] - self.text_vocab_size)], dim=-1))
from image_utils import Im
print(Im(out[:16]).save())
breakpoint()
return random_elements
try:
from torch.nn.attention.flex_attention import create_block_mask
except:
pass
def _attn_mask(txt_batch_dropout, img_batch_dropout, txt_length):
def mask_mod(b, h, q_idx, kv_idx):
txt_sees_txt = (q_idx < txt_length) & (kv_idx < txt_length)
img_sees_img_and_txt = (q_idx >= txt_length)
txt_dropout_case = ~txt_batch_dropout[b] | (txt_sees_txt | img_sees_img_and_txt)
img_sees_img = ((q_idx >= txt_length) & (kv_idx >= txt_length))
txt_sees_txt_and_img = (q_idx < txt_length)
img_dropout_case = ~img_batch_dropout[b] | (img_sees_img | txt_sees_txt_and_img)
return txt_dropout_case & img_dropout_case
return mask_mod
def get_block_mask(txt_batch_attn_dropout, img_batch_attn_dropout, txt_length, batch_size, seq_len, device):
return create_block_mask(
_attn_mask(txt_batch_attn_dropout, img_batch_attn_dropout, txt_length),
B = batch_size, H = None, Q_LEN = seq_len, KV_LEN = seq_len, device = device
)
def _interleaved_attn_mask(interleaved_sample_ids):
def mask_mod(b, h, q_idx, kv_idx):
return (interleaved_sample_ids[b, q_idx] == interleaved_sample_ids[b, kv_idx]) & (interleaved_sample_ids[b, q_idx] != -1)
return mask_mod
def visualize_flex_attention(mask_mod, B, SEQ_LEN, H=16, HEAD_DIM=64, device="cuda"):
from models.archived.utils import visualize_attention_scores
def make_tensor():
return torch.ones(B, H, SEQ_LEN, HEAD_DIM, device=device)
query, key = make_tensor(), make_tensor()
visualize_attention_scores(
query,
key,
mask_mod=mask_mod,
device=device,
name="interleaved_attn_mask",
)
def get_interleaved_block_mask(interleaved_sample_ids, batch_size, seq_len, device, visualize=False):
# Uncomment this to visualize the mask
if visualize:
visualize_flex_attention(_interleaved_attn_mask(interleaved_sample_ids), batch_size, seq_len, device=device)
if (interleaved_sample_ids == -1).all(dim=-1).any():
gprint(f"WARNING: Found all -1s in interleaved_sample_ids, setting one to 0")
interleaved_sample_ids = interleaved_sample_ids.clone()
interleaved_sample_ids[(interleaved_sample_ids == -1).all(dim=-1), 0] = 0
return create_block_mask(
_interleaved_attn_mask(interleaved_sample_ids),
B = batch_size, H = None, Q_LEN = seq_len, KV_LEN = seq_len, device = device
)
def calculate_clip_score(
image_paths: List[str],
captions_mapping: Dict[str, str],
device: torch.device = "cuda",
seed: Optional[int] = 42,
batch_size: int = 128,
dataloader_workers: int = 16,
verbose: bool = True,
):
import clip
from T2IBenchmark.feature_extractors import (BaseFeatureExtractor,
InceptionV3FE)
from T2IBenchmark.loaders import CaptionImageDataset
from T2IBenchmark.model_wrapper import (ModelWrapperDataloader,
T2IModelWrapper)
from T2IBenchmark.utils import dprint, set_all_seeds
if seed:
set_all_seeds(seed)
model, preprocess = clip.load("ViT-B/32", device=device)
dataset = CaptionImageDataset(
images_paths=image_paths,
captions=list(map(lambda x: captions_mapping[x], image_paths)),
preprocess_fn=preprocess,
)
dataloader = DataLoader(
dataset,
batch_size=batch_size,
shuffle=False,
drop_last=False,
num_workers=dataloader_workers,
)
score_acc = 0.0
num_samples = 0.0
for image, caption in tqdm(dataloader):
image_embedding = model.encode_image(image.to(device))
caption_embedding = model.encode_text(clip.tokenize(caption, truncate=True).to(device))
image_features = image_embedding / image_embedding.norm(dim=1, keepdim=True).to(
torch.float32
)
caption_features = caption_embedding / caption_embedding.norm(
dim=1, keepdim=True
).to(torch.float32)
score = (image_features * caption_features).sum()
score_acc += score
num_samples += image.shape[0]
clip_score = score_acc / num_samples
dprint(verbose, f"CLIP score is {clip_score}")
return clip_score
def get_chameleon_txt_indices(vae, include_special_tokens=True):
image_indices = set(vae.chameleon_ori_translation.bpe2img.keys())
if include_special_tokens:
h_grids, w_grids = 32, 32
image_start_token = vae.token2id(vae.image_start_token)
n_grids_token = vae.token2id(vae.get_n_grids_token(h_grids))
image_end_token = vae.token2id(vae.image_end_token)
image_indices.add(image_start_token)
image_indices.add(n_grids_token)
image_indices.add(image_end_token)
image_indices.add(-100)
image_indices.add(1)
image_indices.update(range(8192, 8820 + 1))
return image_indices