Spaces:
Paused
Paused
Update train.py
Browse files
train.py
CHANGED
|
@@ -1,1370 +1,106 @@
|
|
| 1 |
-
|
| 2 |
-
# coding=utf-8
|
| 3 |
-
# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
|
| 4 |
-
#
|
| 5 |
-
# Licensed under the Apache License, Version 2.0 (the "License");
|
| 6 |
-
# you may not use this file except in compliance with the License.
|
| 7 |
-
# You may obtain a copy of the License at
|
| 8 |
-
#
|
| 9 |
-
# http://www.apache.org/licenses/LICENSE-2.0
|
| 10 |
-
#
|
| 11 |
-
# Unless required by applicable law or agreed to in writing, software
|
| 12 |
-
# distributed under the License is distributed on an "AS IS" BASIS,
|
| 13 |
-
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
| 14 |
-
# See the License for the specific language governing permissions and
|
| 15 |
-
# limitations under the License.
|
| 16 |
-
"""Fine-tuning script for Stable Diffusion XL for text2image."""
|
| 17 |
-
|
| 18 |
-
import argparse
|
| 19 |
-
import functools
|
| 20 |
-
import gc
|
| 21 |
-
import logging
|
| 22 |
-
import math
|
| 23 |
-
import os
|
| 24 |
-
import random
|
| 25 |
-
import shutil
|
| 26 |
-
from contextlib import nullcontext
|
| 27 |
-
from pathlib import Path
|
| 28 |
-
|
| 29 |
-
import accelerate
|
| 30 |
-
import datasets
|
| 31 |
-
import numpy as np
|
| 32 |
import torch
|
| 33 |
-
import
|
| 34 |
-
import torch.utils.checkpoint
|
| 35 |
-
import transformers
|
| 36 |
-
from accelerate import Accelerator
|
| 37 |
-
from accelerate.logging import get_logger
|
| 38 |
-
from accelerate.utils import DistributedType, ProjectConfiguration, set_seed
|
| 39 |
-
from datasets import concatenate_datasets, load_dataset
|
| 40 |
-
from huggingface_hub import create_repo, upload_folder
|
| 41 |
-
from packaging import version
|
| 42 |
-
from torchvision import transforms
|
| 43 |
-
from torchvision.transforms.functional import crop
|
| 44 |
-
from tqdm.auto import tqdm
|
| 45 |
-
from transformers import AutoTokenizer, PretrainedConfig
|
| 46 |
-
|
| 47 |
-
import diffusers
|
| 48 |
-
from diffusers import AutoencoderKL, DDPMScheduler, StableDiffusionXLPipeline, UNet2DConditionModel
|
| 49 |
-
from diffusers.optimization import get_scheduler
|
| 50 |
-
from diffusers.training_utils import EMAModel, compute_snr
|
| 51 |
-
from diffusers.utils import check_min_version, is_wandb_available
|
| 52 |
-
from diffusers.utils.hub_utils import load_or_create_model_card, populate_model_card
|
| 53 |
-
from diffusers.utils.import_utils import is_torch_npu_available, is_xformers_available
|
| 54 |
-
from diffusers.utils.torch_utils import is_compiled_module
|
| 55 |
-
|
| 56 |
-
|
| 57 |
-
# Will error if the minimal version of diffusers is not installed. Remove at your own risks.
|
| 58 |
-
check_min_version("0.36.0.dev0")
|
| 59 |
-
|
| 60 |
-
logger = get_logger(__name__)
|
| 61 |
-
if is_torch_npu_available():
|
| 62 |
-
import torch_npu
|
| 63 |
-
|
| 64 |
-
torch.npu.config.allow_internal_format = False
|
| 65 |
-
|
| 66 |
-
DATASET_NAME_MAPPING = {
|
| 67 |
-
"lambdalabs/naruto-blip-captions": ("image", "text"),
|
| 68 |
-
}
|
| 69 |
-
|
| 70 |
-
|
| 71 |
-
def save_model_card(
|
| 72 |
-
repo_id: str,
|
| 73 |
-
images: list = None,
|
| 74 |
-
validation_prompt: str = None,
|
| 75 |
-
base_model: str = None,
|
| 76 |
-
dataset_name: str = None,
|
| 77 |
-
repo_folder: str = None,
|
| 78 |
-
vae_path: str = None,
|
| 79 |
-
):
|
| 80 |
-
img_str = ""
|
| 81 |
-
if images is not None:
|
| 82 |
-
for i, image in enumerate(images):
|
| 83 |
-
image.save(os.path.join(repo_folder, f"image_{i}.png"))
|
| 84 |
-
img_str += f"\n"
|
| 85 |
-
|
| 86 |
-
model_description = f"""
|
| 87 |
-
# Text-to-image finetuning - {repo_id}
|
| 88 |
-
|
| 89 |
-
This pipeline was finetuned from **{base_model}** on the **{dataset_name}** dataset. Below are some example images generated with the finetuned pipeline using the following prompt: {validation_prompt}: \n
|
| 90 |
-
{img_str}
|
| 91 |
-
|
| 92 |
-
Special VAE used for training: {vae_path}.
|
| 93 |
-
"""
|
| 94 |
-
|
| 95 |
-
model_card = load_or_create_model_card(
|
| 96 |
-
repo_id_or_path=repo_id,
|
| 97 |
-
from_training=True,
|
| 98 |
-
license="creativeml-openrail-m",
|
| 99 |
-
base_model=base_model,
|
| 100 |
-
model_description=model_description,
|
| 101 |
-
inference=True,
|
| 102 |
-
)
|
| 103 |
-
|
| 104 |
-
tags = [
|
| 105 |
-
"stable-diffusion-xl",
|
| 106 |
-
"stable-diffusion-xl-diffusers",
|
| 107 |
-
"text-to-image",
|
| 108 |
-
"diffusers-training",
|
| 109 |
-
"diffusers",
|
| 110 |
-
]
|
| 111 |
-
model_card = populate_model_card(model_card, tags=tags)
|
| 112 |
-
|
| 113 |
-
model_card.save(os.path.join(repo_folder, "README.md"))
|
| 114 |
-
|
| 115 |
-
|
| 116 |
-
def import_model_class_from_model_name_or_path(
|
| 117 |
-
pretrained_model_name_or_path: str, revision: str, subfolder: str = "text_encoder"
|
| 118 |
-
):
|
| 119 |
-
text_encoder_config = PretrainedConfig.from_pretrained(
|
| 120 |
-
pretrained_model_name_or_path, subfolder=subfolder, revision=revision
|
| 121 |
-
)
|
| 122 |
-
model_class = text_encoder_config.architectures[0]
|
| 123 |
-
|
| 124 |
-
if model_class == "CLIPTextModel":
|
| 125 |
-
from transformers import CLIPTextModel
|
| 126 |
-
|
| 127 |
-
return CLIPTextModel
|
| 128 |
-
elif model_class == "CLIPTextModelWithProjection":
|
| 129 |
-
from transformers import CLIPTextModelWithProjection
|
| 130 |
-
|
| 131 |
-
return CLIPTextModelWithProjection
|
| 132 |
-
else:
|
| 133 |
-
raise ValueError(f"{model_class} is not supported.")
|
| 134 |
-
|
| 135 |
-
|
| 136 |
-
def parse_args(input_args=None):
|
| 137 |
-
parser = argparse.ArgumentParser(description="Simple example of a training script.")
|
| 138 |
-
parser.add_argument(
|
| 139 |
-
"--pretrained_model_name_or_path",
|
| 140 |
-
type=str,
|
| 141 |
-
default=None,
|
| 142 |
-
required=True,
|
| 143 |
-
help="Path to pretrained model or model identifier from huggingface.co/models.",
|
| 144 |
-
)
|
| 145 |
-
parser.add_argument(
|
| 146 |
-
"--pretrained_vae_model_name_or_path",
|
| 147 |
-
type=str,
|
| 148 |
-
default=None,
|
| 149 |
-
help="Path to pretrained VAE model with better numerical stability. More details: https://github.com/huggingface/diffusers/pull/4038.",
|
| 150 |
-
)
|
| 151 |
-
parser.add_argument(
|
| 152 |
-
"--revision",
|
| 153 |
-
type=str,
|
| 154 |
-
default=None,
|
| 155 |
-
required=False,
|
| 156 |
-
help="Revision of pretrained model identifier from huggingface.co/models.",
|
| 157 |
-
)
|
| 158 |
-
parser.add_argument(
|
| 159 |
-
"--variant",
|
| 160 |
-
type=str,
|
| 161 |
-
default=None,
|
| 162 |
-
help="Variant of the model files of the pretrained model identifier from huggingface.co/models, 'e.g.' fp16",
|
| 163 |
-
)
|
| 164 |
-
parser.add_argument(
|
| 165 |
-
"--dataset_name",
|
| 166 |
-
type=str,
|
| 167 |
-
default=None,
|
| 168 |
-
help=(
|
| 169 |
-
"The name of the Dataset (from the HuggingFace hub) to train on (could be your own, possibly private,"
|
| 170 |
-
" dataset). It can also be a path pointing to a local copy of a dataset in your filesystem,"
|
| 171 |
-
" or to a folder containing files that 🤗 Datasets can understand."
|
| 172 |
-
),
|
| 173 |
-
)
|
| 174 |
-
parser.add_argument(
|
| 175 |
-
"--dataset_config_name",
|
| 176 |
-
type=str,
|
| 177 |
-
default=None,
|
| 178 |
-
help="The config of the Dataset, leave as None if there's only one config.",
|
| 179 |
-
)
|
| 180 |
-
parser.add_argument(
|
| 181 |
-
"--train_data_dir",
|
| 182 |
-
type=str,
|
| 183 |
-
default=None,
|
| 184 |
-
help=(
|
| 185 |
-
"A folder containing the training data. Folder contents must follow the structure described in"
|
| 186 |
-
" https://huggingface.co/docs/datasets/image_dataset#imagefolder. In particular, a `metadata.jsonl` file"
|
| 187 |
-
" must exist to provide the captions for the images. Ignored if `dataset_name` is specified."
|
| 188 |
-
),
|
| 189 |
-
)
|
| 190 |
-
parser.add_argument(
|
| 191 |
-
"--image_column", type=str, default="image", help="The column of the dataset containing an image."
|
| 192 |
-
)
|
| 193 |
-
parser.add_argument(
|
| 194 |
-
"--caption_column",
|
| 195 |
-
type=str,
|
| 196 |
-
default="text",
|
| 197 |
-
help="The column of the dataset containing a caption or a list of captions.",
|
| 198 |
-
)
|
| 199 |
-
parser.add_argument(
|
| 200 |
-
"--validation_prompt",
|
| 201 |
-
type=str,
|
| 202 |
-
default=None,
|
| 203 |
-
help="A prompt that is used during validation to verify that the model is learning.",
|
| 204 |
-
)
|
| 205 |
-
parser.add_argument(
|
| 206 |
-
"--num_validation_images",
|
| 207 |
-
type=int,
|
| 208 |
-
default=4,
|
| 209 |
-
help="Number of images that should be generated during validation with `validation_prompt`.",
|
| 210 |
-
)
|
| 211 |
-
parser.add_argument(
|
| 212 |
-
"--validation_epochs",
|
| 213 |
-
type=int,
|
| 214 |
-
default=1,
|
| 215 |
-
help=(
|
| 216 |
-
"Run fine-tuning validation every X epochs. The validation process consists of running the prompt"
|
| 217 |
-
" `args.validation_prompt` multiple times: `args.num_validation_images`."
|
| 218 |
-
),
|
| 219 |
-
)
|
| 220 |
-
parser.add_argument(
|
| 221 |
-
"--max_train_samples",
|
| 222 |
-
type=int,
|
| 223 |
-
default=None,
|
| 224 |
-
help=(
|
| 225 |
-
"For debugging purposes or quicker training, truncate the number of training examples to this "
|
| 226 |
-
"value if set."
|
| 227 |
-
),
|
| 228 |
-
)
|
| 229 |
-
parser.add_argument(
|
| 230 |
-
"--proportion_empty_prompts",
|
| 231 |
-
type=float,
|
| 232 |
-
default=0,
|
| 233 |
-
help="Proportion of image prompts to be replaced with empty strings. Defaults to 0 (no prompt replacement).",
|
| 234 |
-
)
|
| 235 |
-
parser.add_argument(
|
| 236 |
-
"--output_dir",
|
| 237 |
-
type=str,
|
| 238 |
-
default="sdxl-model-finetuned",
|
| 239 |
-
help="The output directory where the model predictions and checkpoints will be written.",
|
| 240 |
-
)
|
| 241 |
-
parser.add_argument(
|
| 242 |
-
"--cache_dir",
|
| 243 |
-
type=str,
|
| 244 |
-
default=None,
|
| 245 |
-
help="The directory where the downloaded models and datasets will be stored.",
|
| 246 |
-
)
|
| 247 |
-
parser.add_argument("--seed", type=int, default=None, help="A seed for reproducible training.")
|
| 248 |
-
parser.add_argument(
|
| 249 |
-
"--resolution",
|
| 250 |
-
type=int,
|
| 251 |
-
default=1024,
|
| 252 |
-
help=(
|
| 253 |
-
"The resolution for input images, all the images in the train/validation dataset will be resized to this"
|
| 254 |
-
" resolution"
|
| 255 |
-
),
|
| 256 |
-
)
|
| 257 |
-
parser.add_argument(
|
| 258 |
-
"--center_crop",
|
| 259 |
-
default=False,
|
| 260 |
-
action="store_true",
|
| 261 |
-
help=(
|
| 262 |
-
"Whether to center crop the input images to the resolution. If not set, the images will be randomly"
|
| 263 |
-
" cropped. The images will be resized to the resolution first before cropping."
|
| 264 |
-
),
|
| 265 |
-
)
|
| 266 |
-
parser.add_argument(
|
| 267 |
-
"--random_flip",
|
| 268 |
-
action="store_true",
|
| 269 |
-
help="whether to randomly flip images horizontally",
|
| 270 |
-
)
|
| 271 |
-
parser.add_argument(
|
| 272 |
-
"--train_batch_size", type=int, default=16, help="Batch size (per device) for the training dataloader."
|
| 273 |
-
)
|
| 274 |
-
parser.add_argument("--num_train_epochs", type=int, default=100)
|
| 275 |
-
parser.add_argument(
|
| 276 |
-
"--max_train_steps",
|
| 277 |
-
type=int,
|
| 278 |
-
default=None,
|
| 279 |
-
help="Total number of training steps to perform. If provided, overrides num_train_epochs.",
|
| 280 |
-
)
|
| 281 |
-
parser.add_argument(
|
| 282 |
-
"--checkpointing_steps",
|
| 283 |
-
type=int,
|
| 284 |
-
default=500,
|
| 285 |
-
help=(
|
| 286 |
-
"Save a checkpoint of the training state every X updates. These checkpoints can be used both as final"
|
| 287 |
-
" checkpoints in case they are better than the last checkpoint, and are also suitable for resuming"
|
| 288 |
-
" training using `--resume_from_checkpoint`."
|
| 289 |
-
),
|
| 290 |
-
)
|
| 291 |
-
parser.add_argument(
|
| 292 |
-
"--checkpoints_total_limit",
|
| 293 |
-
type=int,
|
| 294 |
-
default=None,
|
| 295 |
-
help=("Max number of checkpoints to store."),
|
| 296 |
-
)
|
| 297 |
-
parser.add_argument(
|
| 298 |
-
"--resume_from_checkpoint",
|
| 299 |
-
type=str,
|
| 300 |
-
default=None,
|
| 301 |
-
help=(
|
| 302 |
-
"Whether training should be resumed from a previous checkpoint. Use a path saved by"
|
| 303 |
-
' `--checkpointing_steps`, or `"latest"` to automatically select the last available checkpoint.'
|
| 304 |
-
),
|
| 305 |
-
)
|
| 306 |
-
parser.add_argument(
|
| 307 |
-
"--gradient_accumulation_steps",
|
| 308 |
-
type=int,
|
| 309 |
-
default=1,
|
| 310 |
-
help="Number of updates steps to accumulate before performing a backward/update pass.",
|
| 311 |
-
)
|
| 312 |
-
parser.add_argument(
|
| 313 |
-
"--gradient_checkpointing",
|
| 314 |
-
action="store_true",
|
| 315 |
-
help="Whether or not to use gradient checkpointing to save memory at the expense of slower backward pass.",
|
| 316 |
-
)
|
| 317 |
-
parser.add_argument(
|
| 318 |
-
"--learning_rate",
|
| 319 |
-
type=float,
|
| 320 |
-
default=1e-4,
|
| 321 |
-
help="Initial learning rate (after the potential warmup period) to use.",
|
| 322 |
-
)
|
| 323 |
-
parser.add_argument(
|
| 324 |
-
"--scale_lr",
|
| 325 |
-
action="store_true",
|
| 326 |
-
default=False,
|
| 327 |
-
help="Scale the learning rate by the number of GPUs, gradient accumulation steps, and batch size.",
|
| 328 |
-
)
|
| 329 |
-
parser.add_argument(
|
| 330 |
-
"--lr_scheduler",
|
| 331 |
-
type=str,
|
| 332 |
-
default="constant",
|
| 333 |
-
help=(
|
| 334 |
-
'The scheduler type to use. Choose between ["linear", "cosine", "cosine_with_restarts", "polynomial",'
|
| 335 |
-
' "constant", "constant_with_warmup"]'
|
| 336 |
-
),
|
| 337 |
-
)
|
| 338 |
-
parser.add_argument(
|
| 339 |
-
"--lr_warmup_steps", type=int, default=500, help="Number of steps for the warmup in the lr scheduler."
|
| 340 |
-
)
|
| 341 |
-
parser.add_argument(
|
| 342 |
-
"--timestep_bias_strategy",
|
| 343 |
-
type=str,
|
| 344 |
-
default="none",
|
| 345 |
-
choices=["earlier", "later", "range", "none"],
|
| 346 |
-
help=(
|
| 347 |
-
"The timestep bias strategy, which may help direct the model toward learning low or high frequency details."
|
| 348 |
-
" Choices: ['earlier', 'later', 'range', 'none']."
|
| 349 |
-
" The default is 'none', which means no bias is applied, and training proceeds normally."
|
| 350 |
-
" The value of 'later' will increase the frequency of the model's final training timesteps."
|
| 351 |
-
),
|
| 352 |
-
)
|
| 353 |
-
parser.add_argument(
|
| 354 |
-
"--timestep_bias_multiplier",
|
| 355 |
-
type=float,
|
| 356 |
-
default=1.0,
|
| 357 |
-
help=(
|
| 358 |
-
"The multiplier for the bias. Defaults to 1.0, which means no bias is applied."
|
| 359 |
-
" A value of 2.0 will double the weight of the bias, and a value of 0.5 will halve it."
|
| 360 |
-
),
|
| 361 |
-
)
|
| 362 |
-
parser.add_argument(
|
| 363 |
-
"--timestep_bias_begin",
|
| 364 |
-
type=int,
|
| 365 |
-
default=0,
|
| 366 |
-
help=(
|
| 367 |
-
"When using `--timestep_bias_strategy=range`, the beginning (inclusive) timestep to bias."
|
| 368 |
-
" Defaults to zero, which equates to having no specific bias."
|
| 369 |
-
),
|
| 370 |
-
)
|
| 371 |
-
parser.add_argument(
|
| 372 |
-
"--timestep_bias_end",
|
| 373 |
-
type=int,
|
| 374 |
-
default=1000,
|
| 375 |
-
help=(
|
| 376 |
-
"When using `--timestep_bias_strategy=range`, the final timestep (inclusive) to bias."
|
| 377 |
-
" Defaults to 1000, which is the number of timesteps that Stable Diffusion is trained on."
|
| 378 |
-
),
|
| 379 |
-
)
|
| 380 |
-
parser.add_argument(
|
| 381 |
-
"--timestep_bias_portion",
|
| 382 |
-
type=float,
|
| 383 |
-
default=0.25,
|
| 384 |
-
help=(
|
| 385 |
-
"The portion of timesteps to bias. Defaults to 0.25, which 25% of timesteps will be biased."
|
| 386 |
-
" A value of 0.5 will bias one half of the timesteps. The value provided for `--timestep_bias_strategy` determines"
|
| 387 |
-
" whether the biased portions are in the earlier or later timesteps."
|
| 388 |
-
),
|
| 389 |
-
)
|
| 390 |
-
parser.add_argument(
|
| 391 |
-
"--snr_gamma",
|
| 392 |
-
type=float,
|
| 393 |
-
default=None,
|
| 394 |
-
help="SNR weighting gamma to be used if rebalancing the loss. Recommended value is 5.0. "
|
| 395 |
-
"More details here: https://huggingface.co/papers/2303.09556.",
|
| 396 |
-
)
|
| 397 |
-
parser.add_argument("--use_ema", action="store_true", help="Whether to use EMA model.")
|
| 398 |
-
parser.add_argument(
|
| 399 |
-
"--allow_tf32",
|
| 400 |
-
action="store_true",
|
| 401 |
-
help=(
|
| 402 |
-
"Whether or not to allow TF32 on Ampere GPUs. Can be used to speed up training. For more information, see"
|
| 403 |
-
" https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices"
|
| 404 |
-
),
|
| 405 |
-
)
|
| 406 |
-
parser.add_argument(
|
| 407 |
-
"--dataloader_num_workers",
|
| 408 |
-
type=int,
|
| 409 |
-
default=0,
|
| 410 |
-
help=(
|
| 411 |
-
"Number of subprocesses to use for data loading. 0 means that the data will be loaded in the main process."
|
| 412 |
-
),
|
| 413 |
-
)
|
| 414 |
-
parser.add_argument(
|
| 415 |
-
"--use_8bit_adam", action="store_true", help="Whether or not to use 8-bit Adam from bitsandbytes."
|
| 416 |
-
)
|
| 417 |
-
parser.add_argument("--adam_beta1", type=float, default=0.9, help="The beta1 parameter for the Adam optimizer.")
|
| 418 |
-
parser.add_argument("--adam_beta2", type=float, default=0.999, help="The beta2 parameter for the Adam optimizer.")
|
| 419 |
-
parser.add_argument("--adam_weight_decay", type=float, default=1e-2, help="Weight decay to use.")
|
| 420 |
-
parser.add_argument("--adam_epsilon", type=float, default=1e-08, help="Epsilon value for the Adam optimizer")
|
| 421 |
-
parser.add_argument("--max_grad_norm", default=1.0, type=float, help="Max gradient norm.")
|
| 422 |
-
parser.add_argument("--push_to_hub", action="store_true", help="Whether or not to push the model to the Hub.")
|
| 423 |
-
parser.add_argument("--hub_token", type=str, default=None, help="The token to use to push to the Model Hub.")
|
| 424 |
-
parser.add_argument(
|
| 425 |
-
"--prediction_type",
|
| 426 |
-
type=str,
|
| 427 |
-
default=None,
|
| 428 |
-
help="The prediction_type that shall be used for training. Choose between 'epsilon' or 'v_prediction' or leave `None`. If left to `None` the default prediction type of the scheduler: `noise_scheduler.config.prediction_type` is chosen.",
|
| 429 |
-
)
|
| 430 |
-
parser.add_argument(
|
| 431 |
-
"--hub_model_id",
|
| 432 |
-
type=str,
|
| 433 |
-
default=None,
|
| 434 |
-
help="The name of the repository to keep in sync with the local `output_dir`.",
|
| 435 |
-
)
|
| 436 |
-
parser.add_argument(
|
| 437 |
-
"--logging_dir",
|
| 438 |
-
type=str,
|
| 439 |
-
default="logs",
|
| 440 |
-
help=(
|
| 441 |
-
"[TensorBoard](https://www.tensorflow.org/tensorboard) log directory. Will default to"
|
| 442 |
-
" *output_dir/runs/**CURRENT_DATETIME_HOSTNAME***."
|
| 443 |
-
),
|
| 444 |
-
)
|
| 445 |
-
parser.add_argument(
|
| 446 |
-
"--report_to",
|
| 447 |
-
type=str,
|
| 448 |
-
default="tensorboard",
|
| 449 |
-
help=(
|
| 450 |
-
'The integration to report the results and logs to. Supported platforms are `"tensorboard"`'
|
| 451 |
-
' (default), `"wandb"` and `"comet_ml"`. Use `"all"` to report to all integrations.'
|
| 452 |
-
),
|
| 453 |
-
)
|
| 454 |
-
parser.add_argument(
|
| 455 |
-
"--mixed_precision",
|
| 456 |
-
type=str,
|
| 457 |
-
default=None,
|
| 458 |
-
choices=["no", "fp16", "bf16"],
|
| 459 |
-
help=(
|
| 460 |
-
"Whether to use mixed precision. Choose between fp16 and bf16 (bfloat16). Bf16 requires PyTorch >="
|
| 461 |
-
" 1.10.and an Nvidia Ampere GPU. Default to the value of accelerate config of the current system or the"
|
| 462 |
-
" flag passed with the `accelerate.launch` command. Use this argument to override the accelerate config."
|
| 463 |
-
),
|
| 464 |
-
)
|
| 465 |
-
parser.add_argument("--local_rank", type=int, default=-1, help="For distributed training: local_rank")
|
| 466 |
-
parser.add_argument(
|
| 467 |
-
"--enable_npu_flash_attention", action="store_true", help="Whether or not to use npu flash attention."
|
| 468 |
-
)
|
| 469 |
-
parser.add_argument(
|
| 470 |
-
"--enable_xformers_memory_efficient_attention", action="store_true", help="Whether or not to use xformers."
|
| 471 |
-
)
|
| 472 |
-
parser.add_argument("--noise_offset", type=float, default=0, help="The scale of noise offset.")
|
| 473 |
-
parser.add_argument(
|
| 474 |
-
"--image_interpolation_mode",
|
| 475 |
-
type=str,
|
| 476 |
-
default="lanczos",
|
| 477 |
-
choices=[
|
| 478 |
-
f.lower() for f in dir(transforms.InterpolationMode) if not f.startswith("__") and not f.endswith("__")
|
| 479 |
-
],
|
| 480 |
-
help="The image interpolation method to use for resizing images.",
|
| 481 |
-
)
|
| 482 |
-
|
| 483 |
-
if input_args is not None:
|
| 484 |
-
args = parser.parse_args(input_args)
|
| 485 |
-
else:
|
| 486 |
-
args = parser.parse_args()
|
| 487 |
-
|
| 488 |
-
env_local_rank = int(os.environ.get("LOCAL_RANK", -1))
|
| 489 |
-
if env_local_rank != -1 and env_local_rank != args.local_rank:
|
| 490 |
-
args.local_rank = env_local_rank
|
| 491 |
-
|
| 492 |
-
# Sanity checks
|
| 493 |
-
if args.dataset_name is None and args.train_data_dir is None:
|
| 494 |
-
raise ValueError("Need either a dataset name or a training folder.")
|
| 495 |
-
if args.proportion_empty_prompts < 0 or args.proportion_empty_prompts > 1:
|
| 496 |
-
raise ValueError("`--proportion_empty_prompts` must be in the range [0, 1].")
|
| 497 |
-
|
| 498 |
-
return args
|
| 499 |
-
|
| 500 |
-
|
| 501 |
-
# Adapted from pipelines.StableDiffusionXLPipeline.encode_prompt
|
| 502 |
-
def encode_prompt(batch, text_encoders, tokenizers, proportion_empty_prompts, caption_column, is_train=True):
|
| 503 |
-
prompt_embeds_list = []
|
| 504 |
-
prompt_batch = batch[caption_column]
|
| 505 |
-
|
| 506 |
-
captions = []
|
| 507 |
-
for caption in prompt_batch:
|
| 508 |
-
if random.random() < proportion_empty_prompts:
|
| 509 |
-
captions.append("")
|
| 510 |
-
elif isinstance(caption, str):
|
| 511 |
-
captions.append(caption)
|
| 512 |
-
elif isinstance(caption, (list, np.ndarray)):
|
| 513 |
-
# take a random caption if there are multiple
|
| 514 |
-
captions.append(random.choice(caption) if is_train else caption[0])
|
| 515 |
-
|
| 516 |
-
with torch.no_grad():
|
| 517 |
-
for tokenizer, text_encoder in zip(tokenizers, text_encoders):
|
| 518 |
-
text_inputs = tokenizer(
|
| 519 |
-
captions,
|
| 520 |
-
padding="max_length",
|
| 521 |
-
max_length=tokenizer.model_max_length,
|
| 522 |
-
truncation=True,
|
| 523 |
-
return_tensors="pt",
|
| 524 |
-
)
|
| 525 |
-
text_input_ids = text_inputs.input_ids
|
| 526 |
-
prompt_embeds = text_encoder(
|
| 527 |
-
text_input_ids.to(text_encoder.device),
|
| 528 |
-
output_hidden_states=True,
|
| 529 |
-
return_dict=False,
|
| 530 |
-
)
|
| 531 |
-
|
| 532 |
-
# We are only ALWAYS interested in the pooled output of the final text encoder
|
| 533 |
-
pooled_prompt_embeds = prompt_embeds[0]
|
| 534 |
-
prompt_embeds = prompt_embeds[-1][-2]
|
| 535 |
-
bs_embed, seq_len, _ = prompt_embeds.shape
|
| 536 |
-
prompt_embeds = prompt_embeds.view(bs_embed, seq_len, -1)
|
| 537 |
-
prompt_embeds_list.append(prompt_embeds)
|
| 538 |
-
|
| 539 |
-
prompt_embeds = torch.concat(prompt_embeds_list, dim=-1)
|
| 540 |
-
pooled_prompt_embeds = pooled_prompt_embeds.view(bs_embed, -1)
|
| 541 |
-
return {"prompt_embeds": prompt_embeds.cpu(), "pooled_prompt_embeds": pooled_prompt_embeds.cpu()}
|
| 542 |
-
|
| 543 |
-
|
| 544 |
-
def compute_vae_encodings(batch, vae):
|
| 545 |
-
images = batch.pop("pixel_values")
|
| 546 |
-
pixel_values = torch.stack(list(images))
|
| 547 |
-
pixel_values = pixel_values.to(memory_format=torch.contiguous_format).float()
|
| 548 |
-
pixel_values = pixel_values.to(vae.device, dtype=vae.dtype)
|
| 549 |
-
|
| 550 |
-
with torch.no_grad():
|
| 551 |
-
model_input = vae.encode(pixel_values).latent_dist.sample()
|
| 552 |
-
model_input = model_input * vae.config.scaling_factor
|
| 553 |
-
|
| 554 |
-
# There might have slightly performance improvement
|
| 555 |
-
# by changing model_input.cpu() to accelerator.gather(model_input)
|
| 556 |
-
return {"model_input": model_input.cpu()}
|
| 557 |
-
|
| 558 |
-
|
| 559 |
-
def generate_timestep_weights(args, num_timesteps):
|
| 560 |
-
weights = torch.ones(num_timesteps)
|
| 561 |
-
|
| 562 |
-
# Determine the indices to bias
|
| 563 |
-
num_to_bias = int(args.timestep_bias_portion * num_timesteps)
|
| 564 |
-
|
| 565 |
-
if args.timestep_bias_strategy == "later":
|
| 566 |
-
bias_indices = slice(-num_to_bias, None)
|
| 567 |
-
elif args.timestep_bias_strategy == "earlier":
|
| 568 |
-
bias_indices = slice(0, num_to_bias)
|
| 569 |
-
elif args.timestep_bias_strategy == "range":
|
| 570 |
-
# Out of the possible 1000 timesteps, we might want to focus on eg. 200-500.
|
| 571 |
-
range_begin = args.timestep_bias_begin
|
| 572 |
-
range_end = args.timestep_bias_end
|
| 573 |
-
if range_begin < 0:
|
| 574 |
-
raise ValueError(
|
| 575 |
-
"When using the range strategy for timestep bias, you must provide a beginning timestep greater or equal to zero."
|
| 576 |
-
)
|
| 577 |
-
if range_end > num_timesteps:
|
| 578 |
-
raise ValueError(
|
| 579 |
-
"When using the range strategy for timestep bias, you must provide an ending timestep smaller than the number of timesteps."
|
| 580 |
-
)
|
| 581 |
-
bias_indices = slice(range_begin, range_end)
|
| 582 |
-
else: # 'none' or any other string
|
| 583 |
-
return weights
|
| 584 |
-
if args.timestep_bias_multiplier <= 0:
|
| 585 |
-
return ValueError(
|
| 586 |
-
"The parameter --timestep_bias_multiplier is not intended to be used to disable the training of specific timesteps."
|
| 587 |
-
" If it was intended to disable timestep bias, use `--timestep_bias_strategy none` instead."
|
| 588 |
-
" A timestep bias multiplier less than or equal to 0 is not allowed."
|
| 589 |
-
)
|
| 590 |
-
|
| 591 |
-
# Apply the bias
|
| 592 |
-
weights[bias_indices] *= args.timestep_bias_multiplier
|
| 593 |
|
| 594 |
-
# Normalize
|
| 595 |
-
weights /= weights.sum()
|
| 596 |
|
| 597 |
-
|
| 598 |
-
|
| 599 |
-
|
| 600 |
-
|
| 601 |
-
|
| 602 |
-
|
| 603 |
-
"You cannot use both --report_to=wandb and --hub_token due to a security risk of exposing your token."
|
| 604 |
-
" Please use `hf auth login` to authenticate with the Hub."
|
| 605 |
-
)
|
| 606 |
-
|
| 607 |
-
logging_dir = Path(args.output_dir, args.logging_dir)
|
| 608 |
-
|
| 609 |
-
accelerator_project_config = ProjectConfiguration(project_dir=args.output_dir, logging_dir=logging_dir)
|
| 610 |
-
|
| 611 |
-
if torch.backends.mps.is_available() and args.mixed_precision == "bf16":
|
| 612 |
-
# due to pytorch#99272, MPS does not yet support bfloat16.
|
| 613 |
-
raise ValueError(
|
| 614 |
-
"Mixed precision training with bfloat16 is not supported on MPS. Please use fp16 (recommended) or fp32 instead."
|
| 615 |
-
)
|
| 616 |
-
|
| 617 |
-
accelerator = Accelerator(
|
| 618 |
-
gradient_accumulation_steps=args.gradient_accumulation_steps,
|
| 619 |
-
mixed_precision=args.mixed_precision,
|
| 620 |
-
log_with=args.report_to,
|
| 621 |
-
project_config=accelerator_project_config,
|
| 622 |
-
)
|
| 623 |
-
|
| 624 |
-
# Disable AMP for MPS.
|
| 625 |
-
if torch.backends.mps.is_available():
|
| 626 |
-
accelerator.native_amp = False
|
| 627 |
-
|
| 628 |
-
if args.report_to == "wandb":
|
| 629 |
-
if not is_wandb_available():
|
| 630 |
-
raise ImportError("Make sure to install wandb if you want to use it for logging during training.")
|
| 631 |
-
import wandb
|
| 632 |
-
|
| 633 |
-
# Make one log on every process with the configuration for debugging.
|
| 634 |
-
logging.basicConfig(
|
| 635 |
-
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
|
| 636 |
-
datefmt="%m/%d/%Y %H:%M:%S",
|
| 637 |
-
level=logging.INFO,
|
| 638 |
-
)
|
| 639 |
-
logger.info(accelerator.state, main_process_only=False)
|
| 640 |
-
if accelerator.is_local_main_process:
|
| 641 |
-
datasets.utils.logging.set_verbosity_warning()
|
| 642 |
-
transformers.utils.logging.set_verbosity_warning()
|
| 643 |
-
diffusers.utils.logging.set_verbosity_info()
|
| 644 |
-
else:
|
| 645 |
-
datasets.utils.logging.set_verbosity_error()
|
| 646 |
-
transformers.utils.logging.set_verbosity_error()
|
| 647 |
-
diffusers.utils.logging.set_verbosity_error()
|
| 648 |
-
|
| 649 |
-
# If passed along, set the training seed now.
|
| 650 |
-
if args.seed is not None:
|
| 651 |
-
set_seed(args.seed)
|
| 652 |
-
|
| 653 |
-
# Handle the repository creation
|
| 654 |
-
if accelerator.is_main_process:
|
| 655 |
-
if args.output_dir is not None:
|
| 656 |
-
os.makedirs(args.output_dir, exist_ok=True)
|
| 657 |
-
|
| 658 |
-
if args.push_to_hub:
|
| 659 |
-
repo_id = create_repo(
|
| 660 |
-
repo_id=args.hub_model_id or Path(args.output_dir).name, exist_ok=True, token=args.hub_token
|
| 661 |
-
).repo_id
|
| 662 |
-
|
| 663 |
-
# Load the tokenizers
|
| 664 |
-
tokenizer_one = AutoTokenizer.from_pretrained(
|
| 665 |
-
args.pretrained_model_name_or_path,
|
| 666 |
-
subfolder="tokenizer",
|
| 667 |
-
revision=args.revision,
|
| 668 |
-
use_fast=False,
|
| 669 |
-
)
|
| 670 |
-
tokenizer_two = AutoTokenizer.from_pretrained(
|
| 671 |
-
args.pretrained_model_name_or_path,
|
| 672 |
-
subfolder="tokenizer_2",
|
| 673 |
-
revision=args.revision,
|
| 674 |
-
use_fast=False,
|
| 675 |
-
)
|
| 676 |
-
|
| 677 |
-
# import correct text encoder classes
|
| 678 |
-
text_encoder_cls_one = import_model_class_from_model_name_or_path(
|
| 679 |
-
args.pretrained_model_name_or_path, args.revision
|
| 680 |
-
)
|
| 681 |
-
text_encoder_cls_two = import_model_class_from_model_name_or_path(
|
| 682 |
-
args.pretrained_model_name_or_path, args.revision, subfolder="text_encoder_2"
|
| 683 |
-
)
|
| 684 |
-
|
| 685 |
-
# Load scheduler and models
|
| 686 |
-
noise_scheduler = DDPMScheduler.from_pretrained(args.pretrained_model_name_or_path, subfolder="scheduler")
|
| 687 |
-
# Check for terminal SNR in combination with SNR Gamma
|
| 688 |
-
text_encoder_one = text_encoder_cls_one.from_pretrained(
|
| 689 |
-
args.pretrained_model_name_or_path, subfolder="text_encoder", revision=args.revision, variant=args.variant
|
| 690 |
-
)
|
| 691 |
-
text_encoder_two = text_encoder_cls_two.from_pretrained(
|
| 692 |
-
args.pretrained_model_name_or_path, subfolder="text_encoder_2", revision=args.revision, variant=args.variant
|
| 693 |
-
)
|
| 694 |
-
vae_path = (
|
| 695 |
-
args.pretrained_model_name_or_path
|
| 696 |
-
if args.pretrained_vae_model_name_or_path is None
|
| 697 |
-
else args.pretrained_vae_model_name_or_path
|
| 698 |
-
)
|
| 699 |
-
vae = AutoencoderKL.from_pretrained(
|
| 700 |
-
vae_path,
|
| 701 |
-
subfolder="vae" if args.pretrained_vae_model_name_or_path is None else None,
|
| 702 |
-
revision=args.revision,
|
| 703 |
-
variant=args.variant,
|
| 704 |
)
|
| 705 |
-
unet = UNet2DConditionModel.from_pretrained(
|
| 706 |
-
args.pretrained_model_name_or_path, subfolder="unet", revision=args.revision, variant=args.variant
|
| 707 |
-
)
|
| 708 |
-
|
| 709 |
-
# Freeze vae and text encoders.
|
| 710 |
-
vae.requires_grad_(False)
|
| 711 |
-
text_encoder_one.requires_grad_(False)
|
| 712 |
-
text_encoder_two.requires_grad_(False)
|
| 713 |
-
# Set unet as trainable.
|
| 714 |
-
unet.train()
|
| 715 |
-
|
| 716 |
-
# For mixed precision training we cast all non-trainable weights to half-precision
|
| 717 |
-
# as these weights are only used for inference, keeping weights in full precision is not required.
|
| 718 |
-
weight_dtype = torch.float32
|
| 719 |
-
if accelerator.mixed_precision == "fp16":
|
| 720 |
-
weight_dtype = torch.float16
|
| 721 |
-
elif accelerator.mixed_precision == "bf16":
|
| 722 |
-
weight_dtype = torch.bfloat16
|
| 723 |
-
|
| 724 |
-
# Move unet, vae and text_encoder to device and cast to weight_dtype
|
| 725 |
-
# The VAE is in float32 to avoid NaN losses.
|
| 726 |
-
vae.to(accelerator.device, dtype=torch.float32)
|
| 727 |
-
text_encoder_one.to(accelerator.device, dtype=weight_dtype)
|
| 728 |
-
text_encoder_two.to(accelerator.device, dtype=weight_dtype)
|
| 729 |
-
|
| 730 |
-
# Create EMA for the unet.
|
| 731 |
-
if args.use_ema:
|
| 732 |
-
ema_unet = UNet2DConditionModel.from_pretrained(
|
| 733 |
-
args.pretrained_model_name_or_path, subfolder="unet", revision=args.revision, variant=args.variant
|
| 734 |
-
)
|
| 735 |
-
ema_unet = EMAModel(ema_unet.parameters(), model_cls=UNet2DConditionModel, model_config=ema_unet.config)
|
| 736 |
-
if args.enable_npu_flash_attention:
|
| 737 |
-
if is_torch_npu_available():
|
| 738 |
-
logger.info("npu flash attention enabled.")
|
| 739 |
-
unet.enable_npu_flash_attention()
|
| 740 |
-
else:
|
| 741 |
-
raise ValueError("npu flash attention requires torch_npu extensions and is supported only on npu devices.")
|
| 742 |
-
if args.enable_xformers_memory_efficient_attention:
|
| 743 |
-
if is_xformers_available():
|
| 744 |
-
import xformers
|
| 745 |
-
|
| 746 |
-
xformers_version = version.parse(xformers.__version__)
|
| 747 |
-
if xformers_version == version.parse("0.0.16"):
|
| 748 |
-
logger.warning(
|
| 749 |
-
"xFormers 0.0.16 cannot be used for training in some GPUs. If you observe problems during training, please update xFormers to at least 0.0.17. See https://huggingface.co/docs/diffusers/main/en/optimization/xformers for more details."
|
| 750 |
-
)
|
| 751 |
-
unet.enable_xformers_memory_efficient_attention()
|
| 752 |
-
else:
|
| 753 |
-
raise ValueError("xformers is not available. Make sure it is installed correctly")
|
| 754 |
-
|
| 755 |
-
# `accelerate` 0.16.0 will have better support for customized saving
|
| 756 |
-
if version.parse(accelerate.__version__) >= version.parse("0.16.0"):
|
| 757 |
-
# create custom saving & loading hooks so that `accelerator.save_state(...)` serializes in a nice format
|
| 758 |
-
def save_model_hook(models, weights, output_dir):
|
| 759 |
-
if accelerator.is_main_process:
|
| 760 |
-
if args.use_ema:
|
| 761 |
-
ema_unet.save_pretrained(os.path.join(output_dir, "unet_ema"))
|
| 762 |
-
|
| 763 |
-
for i, model in enumerate(models):
|
| 764 |
-
model.save_pretrained(os.path.join(output_dir, "unet"))
|
| 765 |
-
|
| 766 |
-
# make sure to pop weight so that corresponding model is not saved again
|
| 767 |
-
if weights:
|
| 768 |
-
weights.pop()
|
| 769 |
-
|
| 770 |
-
def load_model_hook(models, input_dir):
|
| 771 |
-
if args.use_ema:
|
| 772 |
-
load_model = EMAModel.from_pretrained(os.path.join(input_dir, "unet_ema"), UNet2DConditionModel)
|
| 773 |
-
ema_unet.load_state_dict(load_model.state_dict())
|
| 774 |
-
ema_unet.to(accelerator.device)
|
| 775 |
-
del load_model
|
| 776 |
-
|
| 777 |
-
for _ in range(len(models)):
|
| 778 |
-
# pop models so that they are not loaded again
|
| 779 |
-
model = models.pop()
|
| 780 |
-
|
| 781 |
-
# load diffusers style into model
|
| 782 |
-
load_model = UNet2DConditionModel.from_pretrained(input_dir, subfolder="unet")
|
| 783 |
-
model.register_to_config(**load_model.config)
|
| 784 |
|
| 785 |
-
|
| 786 |
-
del load_model
|
| 787 |
|
| 788 |
-
|
| 789 |
-
|
| 790 |
-
|
| 791 |
-
|
| 792 |
-
|
| 793 |
-
|
| 794 |
-
# Enable TF32 for faster training on Ampere GPUs,
|
| 795 |
-
# cf https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices
|
| 796 |
-
if args.allow_tf32:
|
| 797 |
-
torch.backends.cuda.matmul.allow_tf32 = True
|
| 798 |
-
|
| 799 |
-
if args.scale_lr:
|
| 800 |
-
args.learning_rate = (
|
| 801 |
-
args.learning_rate * args.gradient_accumulation_steps * args.train_batch_size * accelerator.num_processes
|
| 802 |
-
)
|
| 803 |
-
|
| 804 |
-
# Use 8-bit Adam for lower memory usage or to fine-tune the model in 16GB GPUs
|
| 805 |
-
if args.use_8bit_adam:
|
| 806 |
-
try:
|
| 807 |
-
import bitsandbytes as bnb
|
| 808 |
-
except ImportError:
|
| 809 |
-
raise ImportError(
|
| 810 |
-
"To use 8-bit Adam, please install the bitsandbytes library: `pip install bitsandbytes`."
|
| 811 |
-
)
|
| 812 |
-
|
| 813 |
-
optimizer_class = bnb.optim.AdamW8bit
|
| 814 |
-
else:
|
| 815 |
-
optimizer_class = torch.optim.AdamW
|
| 816 |
-
|
| 817 |
-
# Optimizer creation
|
| 818 |
-
params_to_optimize = unet.parameters()
|
| 819 |
-
optimizer = optimizer_class(
|
| 820 |
-
params_to_optimize,
|
| 821 |
-
lr=args.learning_rate,
|
| 822 |
-
betas=(args.adam_beta1, args.adam_beta2),
|
| 823 |
-
weight_decay=args.adam_weight_decay,
|
| 824 |
-
eps=args.adam_epsilon,
|
| 825 |
)
|
| 826 |
|
| 827 |
-
|
| 828 |
-
# or specify a Dataset from the hub (the dataset will be downloaded automatically from the datasets Hub).
|
| 829 |
-
|
| 830 |
-
# In distributed training, the load_dataset function guarantees that only one local process can concurrently
|
| 831 |
-
# download the dataset.
|
| 832 |
-
if args.dataset_name is not None:
|
| 833 |
-
# Downloading and loading a dataset from the hub.
|
| 834 |
-
dataset = load_dataset(
|
| 835 |
-
args.dataset_name, args.dataset_config_name, cache_dir=args.cache_dir, data_dir=args.train_data_dir
|
| 836 |
-
)
|
| 837 |
-
else:
|
| 838 |
-
data_files = {}
|
| 839 |
-
if args.train_data_dir is not None:
|
| 840 |
-
data_files["train"] = os.path.join(args.train_data_dir, "**")
|
| 841 |
-
dataset = load_dataset(
|
| 842 |
-
"imagefolder",
|
| 843 |
-
data_files=data_files,
|
| 844 |
-
cache_dir=args.cache_dir,
|
| 845 |
-
)
|
| 846 |
-
# See more about loading custom images at
|
| 847 |
-
# https://huggingface.co/docs/datasets/v2.4.0/en/image_load#imagefolder
|
| 848 |
-
|
| 849 |
-
# Preprocessing the datasets.
|
| 850 |
-
# We need to tokenize inputs and targets.
|
| 851 |
-
column_names = dataset["train"].column_names
|
| 852 |
-
|
| 853 |
-
# 6. Get the column names for input/target.
|
| 854 |
-
dataset_columns = DATASET_NAME_MAPPING.get(args.dataset_name, None)
|
| 855 |
-
if args.image_column is None:
|
| 856 |
-
image_column = dataset_columns[0] if dataset_columns is not None else column_names[0]
|
| 857 |
-
else:
|
| 858 |
-
image_column = args.image_column
|
| 859 |
-
if image_column not in column_names:
|
| 860 |
-
raise ValueError(
|
| 861 |
-
f"--image_column' value '{args.image_column}' needs to be one of: {', '.join(column_names)}"
|
| 862 |
-
)
|
| 863 |
-
if args.caption_column is None:
|
| 864 |
-
caption_column = dataset_columns[1] if dataset_columns is not None else column_names[1]
|
| 865 |
-
else:
|
| 866 |
-
caption_column = args.caption_column
|
| 867 |
-
if caption_column not in column_names:
|
| 868 |
-
raise ValueError(
|
| 869 |
-
f"--caption_column' value '{args.caption_column}' needs to be one of: {', '.join(column_names)}"
|
| 870 |
-
)
|
| 871 |
-
|
| 872 |
-
# Preprocessing the datasets.
|
| 873 |
-
interpolation = getattr(transforms.InterpolationMode, args.image_interpolation_mode.upper(), None)
|
| 874 |
-
if interpolation is None:
|
| 875 |
-
raise ValueError(f"Unsupported interpolation mode {interpolation=}.")
|
| 876 |
-
train_resize = transforms.Resize(args.resolution, interpolation=interpolation)
|
| 877 |
-
train_crop = transforms.CenterCrop(args.resolution) if args.center_crop else transforms.RandomCrop(args.resolution)
|
| 878 |
-
train_flip = transforms.RandomHorizontalFlip(p=1.0)
|
| 879 |
-
train_transforms = transforms.Compose([transforms.ToTensor(), transforms.Normalize([0.5], [0.5])])
|
| 880 |
-
|
| 881 |
-
def preprocess_train(examples):
|
| 882 |
-
images = [image.convert("RGB") for image in examples[image_column]]
|
| 883 |
-
# image aug
|
| 884 |
-
original_sizes = []
|
| 885 |
-
all_images = []
|
| 886 |
-
crop_top_lefts = []
|
| 887 |
-
for image in images:
|
| 888 |
-
original_sizes.append((image.height, image.width))
|
| 889 |
-
image = train_resize(image)
|
| 890 |
-
if args.random_flip and random.random() < 0.5:
|
| 891 |
-
# flip
|
| 892 |
-
image = train_flip(image)
|
| 893 |
-
if args.center_crop:
|
| 894 |
-
y1 = max(0, int(round((image.height - args.resolution) / 2.0)))
|
| 895 |
-
x1 = max(0, int(round((image.width - args.resolution) / 2.0)))
|
| 896 |
-
image = train_crop(image)
|
| 897 |
-
else:
|
| 898 |
-
y1, x1, h, w = train_crop.get_params(image, (args.resolution, args.resolution))
|
| 899 |
-
image = crop(image, y1, x1, h, w)
|
| 900 |
-
crop_top_left = (y1, x1)
|
| 901 |
-
crop_top_lefts.append(crop_top_left)
|
| 902 |
-
image = train_transforms(image)
|
| 903 |
-
all_images.append(image)
|
| 904 |
-
|
| 905 |
-
examples["original_sizes"] = original_sizes
|
| 906 |
-
examples["crop_top_lefts"] = crop_top_lefts
|
| 907 |
-
examples["pixel_values"] = all_images
|
| 908 |
-
return examples
|
| 909 |
|
| 910 |
-
|
| 911 |
-
|
| 912 |
-
|
| 913 |
-
# Set the training transforms
|
| 914 |
-
train_dataset = dataset["train"].with_transform(preprocess_train)
|
| 915 |
|
| 916 |
-
|
| 917 |
-
|
| 918 |
-
|
| 919 |
-
|
| 920 |
-
|
| 921 |
-
|
| 922 |
-
|
| 923 |
-
|
| 924 |
-
|
| 925 |
-
|
| 926 |
-
)
|
| 927 |
-
compute_vae_encodings_fn = functools.partial(compute_vae_encodings, vae=vae)
|
| 928 |
-
with accelerator.main_process_first():
|
| 929 |
-
from datasets.fingerprint import Hasher
|
| 930 |
-
|
| 931 |
-
# fingerprint used by the cache for the other processes to load the result
|
| 932 |
-
# details: https://github.com/huggingface/diffusers/pull/4038#discussion_r1266078401
|
| 933 |
-
new_fingerprint = Hasher.hash(args)
|
| 934 |
-
new_fingerprint_for_vae = Hasher.hash((vae_path, args))
|
| 935 |
-
train_dataset_with_embeddings = train_dataset.map(
|
| 936 |
-
compute_embeddings_fn, batched=True, new_fingerprint=new_fingerprint
|
| 937 |
-
)
|
| 938 |
-
train_dataset_with_vae = train_dataset.map(
|
| 939 |
-
compute_vae_encodings_fn,
|
| 940 |
-
batched=True,
|
| 941 |
-
batch_size=args.train_batch_size,
|
| 942 |
-
new_fingerprint=new_fingerprint_for_vae,
|
| 943 |
-
)
|
| 944 |
-
precomputed_dataset = concatenate_datasets(
|
| 945 |
-
[train_dataset_with_embeddings, train_dataset_with_vae.remove_columns(["image", "text"])], axis=1
|
| 946 |
-
)
|
| 947 |
-
precomputed_dataset = precomputed_dataset.with_transform(preprocess_train)
|
| 948 |
-
|
| 949 |
-
del compute_vae_encodings_fn, compute_embeddings_fn, text_encoder_one, text_encoder_two
|
| 950 |
-
del text_encoders, tokenizers, vae
|
| 951 |
-
gc.collect()
|
| 952 |
-
if is_torch_npu_available():
|
| 953 |
-
torch_npu.npu.empty_cache()
|
| 954 |
-
elif torch.cuda.is_available():
|
| 955 |
-
torch.cuda.empty_cache()
|
| 956 |
-
|
| 957 |
-
def collate_fn(examples):
|
| 958 |
-
model_input = torch.stack([torch.tensor(example["model_input"]) for example in examples])
|
| 959 |
-
original_sizes = [example["original_sizes"] for example in examples]
|
| 960 |
-
crop_top_lefts = [example["crop_top_lefts"] for example in examples]
|
| 961 |
-
prompt_embeds = torch.stack([torch.tensor(example["prompt_embeds"]) for example in examples])
|
| 962 |
-
pooled_prompt_embeds = torch.stack([torch.tensor(example["pooled_prompt_embeds"]) for example in examples])
|
| 963 |
-
|
| 964 |
-
return {
|
| 965 |
-
"model_input": model_input,
|
| 966 |
-
"prompt_embeds": prompt_embeds,
|
| 967 |
-
"pooled_prompt_embeds": pooled_prompt_embeds,
|
| 968 |
-
"original_sizes": original_sizes,
|
| 969 |
-
"crop_top_lefts": crop_top_lefts,
|
| 970 |
}
|
|
|
|
|
|
|
| 971 |
|
| 972 |
-
|
| 973 |
-
|
| 974 |
-
|
| 975 |
-
|
| 976 |
-
|
| 977 |
-
|
| 978 |
-
|
| 979 |
-
)
|
| 980 |
-
|
| 981 |
-
# Scheduler and math around the number of training steps.
|
| 982 |
-
overrode_max_train_steps = False
|
| 983 |
-
num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
|
| 984 |
-
if args.max_train_steps is None:
|
| 985 |
-
args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
|
| 986 |
-
overrode_max_train_steps = True
|
| 987 |
-
|
| 988 |
-
lr_scheduler = get_scheduler(
|
| 989 |
-
args.lr_scheduler,
|
| 990 |
-
optimizer=optimizer,
|
| 991 |
-
num_warmup_steps=args.lr_warmup_steps * args.gradient_accumulation_steps,
|
| 992 |
-
num_training_steps=args.max_train_steps * args.gradient_accumulation_steps,
|
| 993 |
-
)
|
| 994 |
|
| 995 |
-
|
| 996 |
-
|
| 997 |
-
unet, optimizer, train_dataloader, lr_scheduler
|
| 998 |
-
)
|
| 999 |
|
| 1000 |
-
|
| 1001 |
-
|
| 1002 |
|
| 1003 |
-
|
| 1004 |
-
num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
|
| 1005 |
-
if overrode_max_train_steps:
|
| 1006 |
-
args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
|
| 1007 |
-
# Afterwards we recalculate our number of training epochs
|
| 1008 |
-
args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)
|
| 1009 |
|
| 1010 |
-
|
| 1011 |
-
|
| 1012 |
-
|
| 1013 |
-
|
|
|
|
| 1014 |
|
| 1015 |
-
# Function for unwrapping if torch.compile() was used in accelerate.
|
| 1016 |
-
def unwrap_model(model):
|
| 1017 |
-
model = accelerator.unwrap_model(model)
|
| 1018 |
-
model = model._orig_mod if is_compiled_module(model) else model
|
| 1019 |
-
return model
|
| 1020 |
|
| 1021 |
-
|
| 1022 |
-
|
| 1023 |
-
else:
|
| 1024 |
-
autocast_ctx = torch.autocast(accelerator.device.type)
|
| 1025 |
|
| 1026 |
-
|
| 1027 |
-
total_batch_size = args.train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps
|
| 1028 |
|
| 1029 |
-
|
| 1030 |
-
|
| 1031 |
-
|
| 1032 |
-
|
| 1033 |
-
logger.info(f" Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}")
|
| 1034 |
-
logger.info(f" Gradient Accumulation steps = {args.gradient_accumulation_steps}")
|
| 1035 |
-
logger.info(f" Total optimization steps = {args.max_train_steps}")
|
| 1036 |
-
global_step = 0
|
| 1037 |
-
first_epoch = 0
|
| 1038 |
|
| 1039 |
-
|
| 1040 |
-
|
| 1041 |
-
if args.resume_from_checkpoint != "latest":
|
| 1042 |
-
path = os.path.basename(args.resume_from_checkpoint)
|
| 1043 |
-
else:
|
| 1044 |
-
# Get the most recent checkpoint
|
| 1045 |
-
dirs = os.listdir(args.output_dir)
|
| 1046 |
-
dirs = [d for d in dirs if d.startswith("checkpoint")]
|
| 1047 |
-
dirs = sorted(dirs, key=lambda x: int(x.split("-")[1]))
|
| 1048 |
-
path = dirs[-1] if len(dirs) > 0 else None
|
| 1049 |
|
| 1050 |
-
|
| 1051 |
-
|
| 1052 |
-
f"Checkpoint '{args.resume_from_checkpoint}' does not exist. Starting a new training run."
|
| 1053 |
-
)
|
| 1054 |
-
args.resume_from_checkpoint = None
|
| 1055 |
-
initial_global_step = 0
|
| 1056 |
-
else:
|
| 1057 |
-
accelerator.print(f"Resuming from checkpoint {path}")
|
| 1058 |
-
accelerator.load_state(os.path.join(args.output_dir, path))
|
| 1059 |
-
global_step = int(path.split("-")[1])
|
| 1060 |
|
| 1061 |
-
initial_global_step = global_step
|
| 1062 |
-
first_epoch = global_step // num_update_steps_per_epoch
|
| 1063 |
|
| 1064 |
-
|
| 1065 |
-
|
|
|
|
|
|
|
| 1066 |
|
| 1067 |
-
|
| 1068 |
-
|
| 1069 |
-
|
| 1070 |
-
|
| 1071 |
-
|
| 1072 |
-
disable=not accelerator.is_local_main_process,
|
| 1073 |
)
|
| 1074 |
|
| 1075 |
-
for epoch in range(first_epoch, args.num_train_epochs):
|
| 1076 |
-
train_loss = 0.0
|
| 1077 |
-
for step, batch in enumerate(train_dataloader):
|
| 1078 |
-
with accelerator.accumulate(unet):
|
| 1079 |
-
# Sample noise that we'll add to the latents
|
| 1080 |
-
model_input = batch["model_input"].to(accelerator.device)
|
| 1081 |
-
noise = torch.randn_like(model_input)
|
| 1082 |
-
if args.noise_offset:
|
| 1083 |
-
# https://www.crosslabs.org//blog/diffusion-with-offset-noise
|
| 1084 |
-
noise += args.noise_offset * torch.randn(
|
| 1085 |
-
(model_input.shape[0], model_input.shape[1], 1, 1), device=model_input.device
|
| 1086 |
-
)
|
| 1087 |
-
|
| 1088 |
-
bsz = model_input.shape[0]
|
| 1089 |
-
if args.timestep_bias_strategy == "none":
|
| 1090 |
-
# Sample a random timestep for each image without bias.
|
| 1091 |
-
timesteps = torch.randint(
|
| 1092 |
-
0, noise_scheduler.config.num_train_timesteps, (bsz,), device=model_input.device
|
| 1093 |
-
)
|
| 1094 |
-
else:
|
| 1095 |
-
# Sample a random timestep for each image, potentially biased by the timestep weights.
|
| 1096 |
-
# Biasing the timestep weights allows us to spend less time training irrelevant timesteps.
|
| 1097 |
-
weights = generate_timestep_weights(args, noise_scheduler.config.num_train_timesteps).to(
|
| 1098 |
-
model_input.device
|
| 1099 |
-
)
|
| 1100 |
-
timesteps = torch.multinomial(weights, bsz, replacement=True).long()
|
| 1101 |
-
|
| 1102 |
-
# Add noise to the model input according to the noise magnitude at each timestep
|
| 1103 |
-
# (this is the forward diffusion process)
|
| 1104 |
-
noisy_model_input = noise_scheduler.add_noise(model_input, noise, timesteps).to(dtype=weight_dtype)
|
| 1105 |
-
|
| 1106 |
-
# time ids
|
| 1107 |
-
def compute_time_ids(original_size, crops_coords_top_left):
|
| 1108 |
-
# Adapted from pipeline.StableDiffusionXLPipeline._get_add_time_ids
|
| 1109 |
-
target_size = (args.resolution, args.resolution)
|
| 1110 |
-
add_time_ids = list(original_size + crops_coords_top_left + target_size)
|
| 1111 |
-
add_time_ids = torch.tensor([add_time_ids], device=accelerator.device, dtype=weight_dtype)
|
| 1112 |
-
return add_time_ids
|
| 1113 |
-
|
| 1114 |
-
add_time_ids = torch.cat(
|
| 1115 |
-
[compute_time_ids(s, c) for s, c in zip(batch["original_sizes"], batch["crop_top_lefts"])]
|
| 1116 |
-
)
|
| 1117 |
-
|
| 1118 |
-
# Predict the noise residual
|
| 1119 |
-
unet_added_conditions = {"time_ids": add_time_ids}
|
| 1120 |
-
prompt_embeds = batch["prompt_embeds"].to(accelerator.device, dtype=weight_dtype)
|
| 1121 |
-
pooled_prompt_embeds = batch["pooled_prompt_embeds"].to(accelerator.device)
|
| 1122 |
-
unet_added_conditions.update({"text_embeds": pooled_prompt_embeds})
|
| 1123 |
-
model_pred = unet(
|
| 1124 |
-
noisy_model_input,
|
| 1125 |
-
timesteps,
|
| 1126 |
-
prompt_embeds,
|
| 1127 |
-
added_cond_kwargs=unet_added_conditions,
|
| 1128 |
-
return_dict=False,
|
| 1129 |
-
)[0]
|
| 1130 |
-
|
| 1131 |
-
# Get the target for loss depending on the prediction type
|
| 1132 |
-
if args.prediction_type is not None:
|
| 1133 |
-
# set prediction_type of scheduler if defined
|
| 1134 |
-
noise_scheduler.register_to_config(prediction_type=args.prediction_type)
|
| 1135 |
-
|
| 1136 |
-
if noise_scheduler.config.prediction_type == "epsilon":
|
| 1137 |
-
target = noise
|
| 1138 |
-
elif noise_scheduler.config.prediction_type == "v_prediction":
|
| 1139 |
-
target = noise_scheduler.get_velocity(model_input, noise, timesteps)
|
| 1140 |
-
elif noise_scheduler.config.prediction_type == "sample":
|
| 1141 |
-
# We set the target to latents here, but the model_pred will return the noise sample prediction.
|
| 1142 |
-
target = model_input
|
| 1143 |
-
# We will have to subtract the noise residual from the prediction to get the target sample.
|
| 1144 |
-
model_pred = model_pred - noise
|
| 1145 |
-
else:
|
| 1146 |
-
raise ValueError(f"Unknown prediction type {noise_scheduler.config.prediction_type}")
|
| 1147 |
-
|
| 1148 |
-
if args.snr_gamma is None:
|
| 1149 |
-
loss = F.mse_loss(model_pred.float(), target.float(), reduction="mean")
|
| 1150 |
-
else:
|
| 1151 |
-
# Compute loss-weights as per Section 3.4 of https://huggingface.co/papers/2303.09556.
|
| 1152 |
-
# Since we predict the noise instead of x_0, the original formulation is slightly changed.
|
| 1153 |
-
# This is discussed in Section 4.2 of the same paper.
|
| 1154 |
-
snr = compute_snr(noise_scheduler, timesteps)
|
| 1155 |
-
mse_loss_weights = torch.stack([snr, args.snr_gamma * torch.ones_like(timesteps)], dim=1).min(
|
| 1156 |
-
dim=1
|
| 1157 |
-
)[0]
|
| 1158 |
-
if noise_scheduler.config.prediction_type == "epsilon":
|
| 1159 |
-
mse_loss_weights = mse_loss_weights / snr
|
| 1160 |
-
elif noise_scheduler.config.prediction_type == "v_prediction":
|
| 1161 |
-
mse_loss_weights = mse_loss_weights / (snr + 1)
|
| 1162 |
-
|
| 1163 |
-
loss = F.mse_loss(model_pred.float(), target.float(), reduction="none")
|
| 1164 |
-
loss = loss.mean(dim=list(range(1, len(loss.shape)))) * mse_loss_weights
|
| 1165 |
-
loss = loss.mean()
|
| 1166 |
-
|
| 1167 |
-
# Gather the losses across all processes for logging (if we use distributed training).
|
| 1168 |
-
avg_loss = accelerator.gather(loss.repeat(args.train_batch_size)).mean()
|
| 1169 |
-
train_loss += avg_loss.item() / args.gradient_accumulation_steps
|
| 1170 |
-
|
| 1171 |
-
# Backpropagate
|
| 1172 |
-
accelerator.backward(loss)
|
| 1173 |
-
if accelerator.sync_gradients:
|
| 1174 |
-
params_to_clip = unet.parameters()
|
| 1175 |
-
accelerator.clip_grad_norm_(params_to_clip, args.max_grad_norm)
|
| 1176 |
-
optimizer.step()
|
| 1177 |
-
lr_scheduler.step()
|
| 1178 |
-
optimizer.zero_grad()
|
| 1179 |
|
| 1180 |
-
|
| 1181 |
-
|
| 1182 |
-
|
| 1183 |
-
|
| 1184 |
-
|
| 1185 |
-
|
| 1186 |
-
accelerator.log({"train_loss": train_loss}, step=global_step)
|
| 1187 |
-
train_loss = 0.0
|
| 1188 |
-
|
| 1189 |
-
# DeepSpeed requires saving weights on every device; saving weights only on the main process would cause issues.
|
| 1190 |
-
if accelerator.distributed_type == DistributedType.DEEPSPEED or accelerator.is_main_process:
|
| 1191 |
-
if global_step % args.checkpointing_steps == 0:
|
| 1192 |
-
# _before_ saving state, check if this save would set us over the `checkpoints_total_limit`
|
| 1193 |
-
if args.checkpoints_total_limit is not None:
|
| 1194 |
-
checkpoints = os.listdir(args.output_dir)
|
| 1195 |
-
checkpoints = [d for d in checkpoints if d.startswith("checkpoint")]
|
| 1196 |
-
checkpoints = sorted(checkpoints, key=lambda x: int(x.split("-")[1]))
|
| 1197 |
-
|
| 1198 |
-
# before we save the new checkpoint, we need to have at _most_ `checkpoints_total_limit - 1` checkpoints
|
| 1199 |
-
if len(checkpoints) >= args.checkpoints_total_limit:
|
| 1200 |
-
num_to_remove = len(checkpoints) - args.checkpoints_total_limit + 1
|
| 1201 |
-
removing_checkpoints = checkpoints[0:num_to_remove]
|
| 1202 |
-
|
| 1203 |
-
logger.info(
|
| 1204 |
-
f"{len(checkpoints)} checkpoints already exist, removing {len(removing_checkpoints)} checkpoints"
|
| 1205 |
-
)
|
| 1206 |
-
logger.info(f"removing checkpoints: {', '.join(removing_checkpoints)}")
|
| 1207 |
-
|
| 1208 |
-
for removing_checkpoint in removing_checkpoints:
|
| 1209 |
-
removing_checkpoint = os.path.join(args.output_dir, removing_checkpoint)
|
| 1210 |
-
shutil.rmtree(removing_checkpoint)
|
| 1211 |
-
|
| 1212 |
-
save_path = os.path.join(args.output_dir, f"checkpoint-{global_step}")
|
| 1213 |
-
accelerator.save_state(save_path)
|
| 1214 |
-
logger.info(f"Saved state to {save_path}")
|
| 1215 |
-
|
| 1216 |
-
logs = {"step_loss": loss.detach().item(), "lr": lr_scheduler.get_last_lr()[0]}
|
| 1217 |
-
progress_bar.set_postfix(**logs)
|
| 1218 |
-
|
| 1219 |
-
if global_step >= args.max_train_steps:
|
| 1220 |
-
break
|
| 1221 |
-
|
| 1222 |
-
if accelerator.is_main_process:
|
| 1223 |
-
if args.validation_prompt is not None and epoch % args.validation_epochs == 0:
|
| 1224 |
-
logger.info(
|
| 1225 |
-
f"Running validation... \n Generating {args.num_validation_images} images with prompt:"
|
| 1226 |
-
f" {args.validation_prompt}."
|
| 1227 |
-
)
|
| 1228 |
-
if args.use_ema:
|
| 1229 |
-
# Store the UNet parameters temporarily and load the EMA parameters to perform inference.
|
| 1230 |
-
ema_unet.store(unet.parameters())
|
| 1231 |
-
ema_unet.copy_to(unet.parameters())
|
| 1232 |
-
|
| 1233 |
-
# create pipeline
|
| 1234 |
-
vae = AutoencoderKL.from_pretrained(
|
| 1235 |
-
vae_path,
|
| 1236 |
-
subfolder="vae" if args.pretrained_vae_model_name_or_path is None else None,
|
| 1237 |
-
revision=args.revision,
|
| 1238 |
-
variant=args.variant,
|
| 1239 |
-
)
|
| 1240 |
-
pipeline = StableDiffusionXLPipeline.from_pretrained(
|
| 1241 |
-
args.pretrained_model_name_or_path,
|
| 1242 |
-
vae=vae,
|
| 1243 |
-
unet=accelerator.unwrap_model(unet),
|
| 1244 |
-
revision=args.revision,
|
| 1245 |
-
variant=args.variant,
|
| 1246 |
-
torch_dtype=weight_dtype,
|
| 1247 |
-
)
|
| 1248 |
-
if args.prediction_type is not None:
|
| 1249 |
-
scheduler_args = {"prediction_type": args.prediction_type}
|
| 1250 |
-
pipeline.scheduler = pipeline.scheduler.from_config(pipeline.scheduler.config, **scheduler_args)
|
| 1251 |
-
|
| 1252 |
-
pipeline = pipeline.to(accelerator.device)
|
| 1253 |
-
pipeline.set_progress_bar_config(disable=True)
|
| 1254 |
-
|
| 1255 |
-
# run inference
|
| 1256 |
-
generator = (
|
| 1257 |
-
torch.Generator(device=accelerator.device).manual_seed(args.seed)
|
| 1258 |
-
if args.seed is not None
|
| 1259 |
-
else None
|
| 1260 |
-
)
|
| 1261 |
-
pipeline_args = {"prompt": args.validation_prompt}
|
| 1262 |
-
|
| 1263 |
-
with autocast_ctx:
|
| 1264 |
-
images = [
|
| 1265 |
-
pipeline(**pipeline_args, generator=generator, num_inference_steps=25).images[0]
|
| 1266 |
-
for _ in range(args.num_validation_images)
|
| 1267 |
-
]
|
| 1268 |
-
|
| 1269 |
-
for tracker in accelerator.trackers:
|
| 1270 |
-
if tracker.name == "tensorboard":
|
| 1271 |
-
np_images = np.stack([np.asarray(img) for img in images])
|
| 1272 |
-
tracker.writer.add_images("validation", np_images, epoch, dataformats="NHWC")
|
| 1273 |
-
if tracker.name == "wandb":
|
| 1274 |
-
tracker.log(
|
| 1275 |
-
{
|
| 1276 |
-
"validation": [
|
| 1277 |
-
wandb.Image(image, caption=f"{i}: {args.validation_prompt}")
|
| 1278 |
-
for i, image in enumerate(images)
|
| 1279 |
-
]
|
| 1280 |
-
}
|
| 1281 |
-
)
|
| 1282 |
-
|
| 1283 |
-
del pipeline
|
| 1284 |
-
if is_torch_npu_available():
|
| 1285 |
-
torch_npu.npu.empty_cache()
|
| 1286 |
-
elif torch.cuda.is_available():
|
| 1287 |
-
torch.cuda.empty_cache()
|
| 1288 |
-
|
| 1289 |
-
if args.use_ema:
|
| 1290 |
-
# Switch back to the original UNet parameters.
|
| 1291 |
-
ema_unet.restore(unet.parameters())
|
| 1292 |
-
|
| 1293 |
-
accelerator.wait_for_everyone()
|
| 1294 |
-
if accelerator.is_main_process:
|
| 1295 |
-
unet = unwrap_model(unet)
|
| 1296 |
-
if args.use_ema:
|
| 1297 |
-
ema_unet.copy_to(unet.parameters())
|
| 1298 |
-
|
| 1299 |
-
# Serialize pipeline.
|
| 1300 |
-
vae = AutoencoderKL.from_pretrained(
|
| 1301 |
-
vae_path,
|
| 1302 |
-
subfolder="vae" if args.pretrained_vae_model_name_or_path is None else None,
|
| 1303 |
-
revision=args.revision,
|
| 1304 |
-
variant=args.variant,
|
| 1305 |
-
torch_dtype=weight_dtype,
|
| 1306 |
-
)
|
| 1307 |
-
pipeline = StableDiffusionXLPipeline.from_pretrained(
|
| 1308 |
-
args.pretrained_model_name_or_path,
|
| 1309 |
-
unet=unet,
|
| 1310 |
-
vae=vae,
|
| 1311 |
-
revision=args.revision,
|
| 1312 |
-
variant=args.variant,
|
| 1313 |
-
torch_dtype=weight_dtype,
|
| 1314 |
-
)
|
| 1315 |
-
if args.prediction_type is not None:
|
| 1316 |
-
scheduler_args = {"prediction_type": args.prediction_type}
|
| 1317 |
-
pipeline.scheduler = pipeline.scheduler.from_config(pipeline.scheduler.config, **scheduler_args)
|
| 1318 |
-
pipeline.save_pretrained(args.output_dir)
|
| 1319 |
-
|
| 1320 |
-
# run inference
|
| 1321 |
-
images = []
|
| 1322 |
-
if args.validation_prompt and args.num_validation_images > 0:
|
| 1323 |
-
pipeline = pipeline.to(accelerator.device)
|
| 1324 |
-
generator = (
|
| 1325 |
-
torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed is not None else None
|
| 1326 |
-
)
|
| 1327 |
-
|
| 1328 |
-
with autocast_ctx:
|
| 1329 |
-
images = [
|
| 1330 |
-
pipeline(args.validation_prompt, num_inference_steps=25, generator=generator).images[0]
|
| 1331 |
-
for _ in range(args.num_validation_images)
|
| 1332 |
-
]
|
| 1333 |
-
|
| 1334 |
-
for tracker in accelerator.trackers:
|
| 1335 |
-
if tracker.name == "tensorboard":
|
| 1336 |
-
np_images = np.stack([np.asarray(img) for img in images])
|
| 1337 |
-
tracker.writer.add_images("test", np_images, epoch, dataformats="NHWC")
|
| 1338 |
-
if tracker.name == "wandb":
|
| 1339 |
-
tracker.log(
|
| 1340 |
-
{
|
| 1341 |
-
"test": [
|
| 1342 |
-
wandb.Image(image, caption=f"{i}: {args.validation_prompt}")
|
| 1343 |
-
for i, image in enumerate(images)
|
| 1344 |
-
]
|
| 1345 |
-
}
|
| 1346 |
-
)
|
| 1347 |
-
|
| 1348 |
-
if args.push_to_hub:
|
| 1349 |
-
save_model_card(
|
| 1350 |
-
repo_id=repo_id,
|
| 1351 |
-
images=images,
|
| 1352 |
-
validation_prompt=args.validation_prompt,
|
| 1353 |
-
base_model=args.pretrained_model_name_or_path,
|
| 1354 |
-
dataset_name=args.dataset_name,
|
| 1355 |
-
repo_folder=args.output_dir,
|
| 1356 |
-
vae_path=args.pretrained_vae_model_name_or_path,
|
| 1357 |
-
)
|
| 1358 |
-
upload_folder(
|
| 1359 |
-
repo_id=repo_id,
|
| 1360 |
-
folder_path=args.output_dir,
|
| 1361 |
-
commit_message="End of training",
|
| 1362 |
-
ignore_patterns=["step_*", "epoch_*"],
|
| 1363 |
-
)
|
| 1364 |
-
|
| 1365 |
-
accelerator.end_training()
|
| 1366 |
|
|
|
|
| 1367 |
|
| 1368 |
-
|
| 1369 |
-
|
| 1370 |
-
main(args)
|
|
|
|
| 1 |
+
# %%
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 2 |
import torch
|
| 3 |
+
from transformers import AutoProcessor, AutoModelForVision2Seq, BitsAndBytesConfig
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 4 |
|
|
|
|
|
|
|
| 5 |
|
| 6 |
+
def load_model(model_name="datalab-to/chandra", device_id=0):
|
| 7 |
+
bnb_config = BitsAndBytesConfig(
|
| 8 |
+
load_in_4bit=True,
|
| 9 |
+
bnb_4bit_compute_dtype=torch.bfloat16,
|
| 10 |
+
bnb_4bit_quant_type="nf4",
|
| 11 |
+
bnb_4bit_use_double_quant=True,
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 12 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 13 |
|
| 14 |
+
processor = AutoProcessor.from_pretrained(model_name)
|
|
|
|
| 15 |
|
| 16 |
+
model = AutoModelForVision2Seq.from_pretrained(
|
| 17 |
+
model_name,
|
| 18 |
+
quantization_config=bnb_config,
|
| 19 |
+
dtype=torch.bfloat16,
|
| 20 |
+
device_map={"": device_id},
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 21 |
)
|
| 22 |
|
| 23 |
+
return processor, model
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 24 |
|
| 25 |
+
# %%
|
| 26 |
+
def caption_batch(batch, processor, model):
|
| 27 |
+
images = batch["image"]
|
|
|
|
|
|
|
| 28 |
|
| 29 |
+
messages = [
|
| 30 |
+
{
|
| 31 |
+
"role": "user",
|
| 32 |
+
"content": [
|
| 33 |
+
{"type": "image", "image": image},
|
| 34 |
+
{
|
| 35 |
+
"type": "text",
|
| 36 |
+
"text": "Describe the image, and skip mentioning that it's illustrated or from anime.",
|
| 37 |
+
},
|
| 38 |
+
],
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 39 |
}
|
| 40 |
+
for image in images
|
| 41 |
+
]
|
| 42 |
|
| 43 |
+
inputs = processor.apply_chat_template(
|
| 44 |
+
messages,
|
| 45 |
+
tokenize=True,
|
| 46 |
+
add_generation_prompt=True,
|
| 47 |
+
return_dict=True,
|
| 48 |
+
return_tensors="pt",
|
| 49 |
+
).to(model.device)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 50 |
|
| 51 |
+
with torch.no_grad():
|
| 52 |
+
generated = model.generate(**inputs)
|
|
|
|
|
|
|
| 53 |
|
| 54 |
+
decoded = processor.batch_decode(generated)
|
| 55 |
+
captions = [d.split("<|im_start|>assistant\n")[-1] for d in decoded]
|
| 56 |
|
| 57 |
+
return {"image": images, "text": captions}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 58 |
|
| 59 |
+
# %%
|
| 60 |
+
import datasets
|
| 61 |
+
from datasets import Dataset
|
| 62 |
+
from typing import cast
|
| 63 |
+
from concurrent.futures import ThreadPoolExecutor
|
| 64 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 65 |
|
| 66 |
+
input_dataset = "none-yet/anime-captions"
|
| 67 |
+
output_dataset = "nroggendorff/anime-captions"
|
|
|
|
|
|
|
| 68 |
|
| 69 |
+
loaded = datasets.load_dataset(input_dataset, split="train")
|
|
|
|
| 70 |
|
| 71 |
+
if isinstance(loaded, datasets.DatasetDict):
|
| 72 |
+
ds = cast(Dataset, loaded["train"])
|
| 73 |
+
else:
|
| 74 |
+
ds = cast(Dataset, loaded)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 75 |
|
| 76 |
+
num_gpus = torch.cuda.device_count()
|
| 77 |
+
models = [load_model(device_id=i) for i in range(num_gpus)]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 78 |
|
| 79 |
+
batch_size = 8
|
| 80 |
+
shard_size = len(ds) // num_gpus
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 81 |
|
|
|
|
|
|
|
| 82 |
|
| 83 |
+
def process_shard(shard_idx, processor, model):
|
| 84 |
+
start = shard_idx * shard_size
|
| 85 |
+
end = start + shard_size if shard_idx < num_gpus - 1 else len(ds)
|
| 86 |
+
shard = ds.select(range(start, end))
|
| 87 |
|
| 88 |
+
return shard.map(
|
| 89 |
+
lambda batch: caption_batch(batch, processor, model),
|
| 90 |
+
batched=True,
|
| 91 |
+
batch_size=batch_size,
|
| 92 |
+
remove_columns=shard.column_names,
|
|
|
|
| 93 |
)
|
| 94 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 95 |
|
| 96 |
+
with ThreadPoolExecutor(max_workers=num_gpus) as executor:
|
| 97 |
+
futures = [
|
| 98 |
+
executor.submit(process_shard, i, proc, model)
|
| 99 |
+
for i, (proc, model) in enumerate(models)
|
| 100 |
+
]
|
| 101 |
+
shards = [f.result() for f in futures]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 102 |
|
| 103 |
+
ds = datasets.concatenate_datasets(shards)
|
| 104 |
|
| 105 |
+
# %%
|
| 106 |
+
ds.push_to_hub(output_dataset, create_pr=True)
|
|
|