train_dreambooth.py 56.3 KB
Newer Older
1
2
#!/usr/bin/env python
# coding=utf-8
Patrick von Platen's avatar
Patrick von Platen committed
3
# Copyright 2023 The HuggingFace Inc. team. All rights reserved.
4
5
6
7
8
9
10
11
12
13
14
15
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and

16
import argparse
17
import copy
18
import gc
19
import hashlib
20
import importlib
21
import itertools
Suraj Patil's avatar
Suraj Patil committed
22
import logging
23
24
import math
import os
25
import shutil
26
import warnings
27
28
from pathlib import Path

29
import numpy as np
30
31
32
import torch
import torch.nn.functional as F
import torch.utils.checkpoint
Suraj Patil's avatar
Suraj Patil committed
33
import transformers
34
35
from accelerate import Accelerator
from accelerate.logging import get_logger
36
from accelerate.utils import ProjectConfiguration, set_seed
Patrick von Platen's avatar
Patrick von Platen committed
37
38
39
40
41
42
43
44
45
46
from huggingface_hub import create_repo, model_info, upload_folder
from packaging import version
from PIL import Image
from PIL.ImageOps import exif_transpose
from torch.utils.data import Dataset
from torchvision import transforms
from tqdm.auto import tqdm
from transformers import AutoTokenizer, PretrainedConfig

import diffusers
47
48
49
50
from diffusers import (
    AutoencoderKL,
    DDPMScheduler,
    DiffusionPipeline,
51
    StableDiffusionPipeline,
52
53
    UNet2DConditionModel,
)
54
from diffusers.optimization import get_scheduler
55
from diffusers.training_utils import compute_snr
56
from diffusers.utils import check_min_version, is_wandb_available
57
58
from diffusers.utils.import_utils import is_xformers_available

59

60
61
62
if is_wandb_available():
    import wandb

63
# Will error if the minimal version of diffusers is not installed. Remove at your own risks.
64
check_min_version("0.22.0.dev0")
65

66
67
68
logger = get_logger(__name__)


69
70
71
72
73
74
75
76
77
def save_model_card(
    repo_id: str,
    images=None,
    base_model=str,
    train_text_encoder=False,
    prompt=str,
    repo_folder=None,
    pipeline: DiffusionPipeline = None,
):
78
79
80
81
82
83
84
85
86
87
88
    img_str = ""
    for i, image in enumerate(images):
        image.save(os.path.join(repo_folder, f"image_{i}.png"))
        img_str += f"![img_{i}](./image_{i}.png)\n"

    yaml = f"""
---
license: creativeml-openrail-m
base_model: {base_model}
instance_prompt: {prompt}
tags:
89
90
- {'stable-diffusion' if isinstance(pipeline, StableDiffusionPipeline) else 'if'}
- {'stable-diffusion-diffusers' if isinstance(pipeline, StableDiffusionPipeline) else 'if-diffusers'}
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
- text-to-image
- diffusers
- dreambooth
inference: true
---
    """
    model_card = f"""
# DreamBooth - {repo_id}

This is a dreambooth model derived from {base_model}. The weights were trained on {prompt} using [DreamBooth](https://dreambooth.github.io/).
You can find some example images in the following. \n
{img_str}

DreamBooth for the text encoder was enabled: {train_text_encoder}.
"""
    with open(os.path.join(repo_folder, "README.md"), "w") as f:
        f.write(yaml + model_card)


110
def log_validation(
111
112
113
114
115
116
117
118
119
120
    text_encoder,
    tokenizer,
    unet,
    vae,
    args,
    accelerator,
    weight_dtype,
    global_step,
    prompt_embeds,
    negative_prompt_embeds,
121
):
122
123
124
125
    logger.info(
        f"Running validation... \n Generating {args.num_validation_images} images with prompt:"
        f" {args.validation_prompt}."
    )
126
127
128
129
130
131

    pipeline_args = {}

    if vae is not None:
        pipeline_args["vae"] = vae

132
133
134
    if text_encoder is not None:
        text_encoder = accelerator.unwrap_model(text_encoder)

135
136
137
138
    # create pipeline (note: unet and vae are loaded again in float32)
    pipeline = DiffusionPipeline.from_pretrained(
        args.pretrained_model_name_or_path,
        tokenizer=tokenizer,
139
        text_encoder=text_encoder,
140
141
142
        unet=accelerator.unwrap_model(unet),
        revision=args.revision,
        torch_dtype=weight_dtype,
143
        **pipeline_args,
144
    )
145
146
147
148
149
150
151
152
153
154
155
156

    # We train on the simplified learning objective. If we were previously predicting a variance, we need the scheduler to ignore it
    scheduler_args = {}

    if "variance_type" in pipeline.scheduler.config:
        variance_type = pipeline.scheduler.config.variance_type

        if variance_type in ["learned", "learned_range"]:
            variance_type = "fixed_small"

        scheduler_args["variance_type"] = variance_type

157
158
159
    module = importlib.import_module("diffusers")
    scheduler_class = getattr(module, args.validation_scheduler)
    pipeline.scheduler = scheduler_class.from_config(pipeline.scheduler.config, **scheduler_args)
160
161
162
    pipeline = pipeline.to(accelerator.device)
    pipeline.set_progress_bar_config(disable=True)

163
164
165
166
167
168
169
170
    if args.pre_compute_text_embeddings:
        pipeline_args = {
            "prompt_embeds": prompt_embeds,
            "negative_prompt_embeds": negative_prompt_embeds,
        }
    else:
        pipeline_args = {"prompt": args.validation_prompt}

171
172
173
    # run inference
    generator = None if args.seed is None else torch.Generator(device=accelerator.device).manual_seed(args.seed)
    images = []
174
175
176
177
178
179
180
181
182
183
    if args.validation_images is None:
        for _ in range(args.num_validation_images):
            with torch.autocast("cuda"):
                image = pipeline(**pipeline_args, num_inference_steps=25, generator=generator).images[0]
            images.append(image)
    else:
        for image in args.validation_images:
            image = Image.open(image)
            image = pipeline(**pipeline_args, image=image, generator=generator).images[0]
            images.append(image)
184
185
186
187

    for tracker in accelerator.trackers:
        if tracker.name == "tensorboard":
            np_images = np.stack([np.asarray(img) for img in images])
188
            tracker.writer.add_images("validation", np_images, global_step, dataformats="NHWC")
189
190
191
192
193
194
195
196
197
198
199
200
        if tracker.name == "wandb":
            tracker.log(
                {
                    "validation": [
                        wandb.Image(image, caption=f"{i}: {args.validation_prompt}") for i, image in enumerate(images)
                    ]
                }
            )

    del pipeline
    torch.cuda.empty_cache()

201
202
    return images

203

204
def import_model_class_from_model_name_or_path(pretrained_model_name_or_path: str, revision: str):
205
206
207
    text_encoder_config = PretrainedConfig.from_pretrained(
        pretrained_model_name_or_path,
        subfolder="text_encoder",
208
        revision=revision,
209
210
211
212
213
214
215
216
217
218
219
    )
    model_class = text_encoder_config.architectures[0]

    if model_class == "CLIPTextModel":
        from transformers import CLIPTextModel

        return CLIPTextModel
    elif model_class == "RobertaSeriesModelWithTransformation":
        from diffusers.pipelines.alt_diffusion.modeling_roberta_series import RobertaSeriesModelWithTransformation

        return RobertaSeriesModelWithTransformation
220
221
222
223
    elif model_class == "T5EncoderModel":
        from transformers import T5EncoderModel

        return T5EncoderModel
224
225
226
227
    else:
        raise ValueError(f"{model_class} is not supported.")


Suraj Patil's avatar
Suraj Patil committed
228
def parse_args(input_args=None):
229
230
231
232
233
234
235
236
    parser = argparse.ArgumentParser(description="Simple example of a training script.")
    parser.add_argument(
        "--pretrained_model_name_or_path",
        type=str,
        default=None,
        required=True,
        help="Path to pretrained model or model identifier from huggingface.co/models.",
    )
237
238
239
240
241
    parser.add_argument(
        "--revision",
        type=str,
        default=None,
        required=False,
242
243
244
245
        help=(
            "Revision of pretrained model identifier from huggingface.co/models. Trainable model components should be"
            " float32 precision."
        ),
246
    )
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
    parser.add_argument(
        "--tokenizer_name",
        type=str,
        default=None,
        help="Pretrained tokenizer name or path if not the same as model_name",
    )
    parser.add_argument(
        "--instance_data_dir",
        type=str,
        default=None,
        required=True,
        help="A folder containing the training data of instance images.",
    )
    parser.add_argument(
        "--class_data_dir",
        type=str,
        default=None,
        required=False,
        help="A folder containing the training data of class images.",
    )
    parser.add_argument(
        "--instance_prompt",
        type=str,
        default=None,
271
        required=True,
Yuta Hayashibe's avatar
Yuta Hayashibe committed
272
        help="The prompt with identifier specifying the instance",
273
274
275
276
277
    )
    parser.add_argument(
        "--class_prompt",
        type=str,
        default=None,
Yuta Hayashibe's avatar
Yuta Hayashibe committed
278
        help="The prompt to specify images in the same class as provided instance images.",
279
280
281
282
283
    )
    parser.add_argument(
        "--with_prior_preservation",
        default=False,
        action="store_true",
Yuta Hayashibe's avatar
Yuta Hayashibe committed
284
        help="Flag to add prior preservation loss.",
285
286
287
288
289
290
291
    )
    parser.add_argument("--prior_loss_weight", type=float, default=1.0, help="The weight of prior preservation loss.")
    parser.add_argument(
        "--num_class_images",
        type=int,
        default=100,
        help=(
292
293
            "Minimal class images for prior preservation loss. If there are not enough images already present in"
            " class_data_dir, additional images will be sampled with class_prompt."
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
        ),
    )
    parser.add_argument(
        "--output_dir",
        type=str,
        default="text-inversion-model",
        help="The output directory where the model predictions and checkpoints will be written.",
    )
    parser.add_argument("--seed", type=int, default=None, help="A seed for reproducible training.")
    parser.add_argument(
        "--resolution",
        type=int,
        default=512,
        help=(
            "The resolution for input images, all the images in the train/validation dataset will be resized to this"
            " resolution"
        ),
    )
    parser.add_argument(
313
314
315
        "--center_crop",
        default=False,
        action="store_true",
patil-suraj's avatar
patil-suraj committed
316
317
318
319
        help=(
            "Whether to center crop the input images to the resolution. If not set, the images will be randomly"
            " cropped. The images will be resized to the resolution first before cropping."
        ),
320
    )
321
322
323
324
325
    parser.add_argument(
        "--train_text_encoder",
        action="store_true",
        help="Whether to train the text encoder. If set, the text encoder should be float32 precision.",
    )
326
327
328
329
330
331
332
333
334
335
336
337
338
    parser.add_argument(
        "--train_batch_size", type=int, default=4, help="Batch size (per device) for the training dataloader."
    )
    parser.add_argument(
        "--sample_batch_size", type=int, default=4, help="Batch size (per device) for sampling images."
    )
    parser.add_argument("--num_train_epochs", type=int, default=1)
    parser.add_argument(
        "--max_train_steps",
        type=int,
        default=None,
        help="Total number of training steps to perform.  If provided, overrides num_train_epochs.",
    )
339
340
341
342
343
    parser.add_argument(
        "--checkpointing_steps",
        type=int,
        default=500,
        help=(
344
345
346
347
348
            "Save a checkpoint of the training state every X updates. Checkpoints can be used for resuming training via `--resume_from_checkpoint`. "
            "In the case that the checkpoint is better than the final trained model, the checkpoint can also be used for inference."
            "Using a checkpoint for inference requires separate loading of the original pipeline and the individual checkpointed model components."
            "See https://huggingface.co/docs/diffusers/main/en/training/dreambooth#performing-inference-using-a-saved-checkpoint for step by step"
            "instructions."
349
350
        ),
    )
351
    parser.add_argument(
352
        "--checkpoints_total_limit",
353
354
355
356
357
358
359
360
        type=int,
        default=None,
        help=(
            "Max number of checkpoints to store. Passed as `total_limit` to the `Accelerator` `ProjectConfiguration`."
            " See Accelerator::save_state https://huggingface.co/docs/accelerate/package_reference/accelerator#accelerate.Accelerator.save_state"
            " for more details"
        ),
    )
361
362
363
364
365
366
367
368
369
    parser.add_argument(
        "--resume_from_checkpoint",
        type=str,
        default=None,
        help=(
            "Whether training should be resumed from a previous checkpoint. Use a path saved by"
            ' `--checkpointing_steps`, or `"latest"` to automatically select the last available checkpoint.'
        ),
    )
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
    parser.add_argument(
        "--gradient_accumulation_steps",
        type=int,
        default=1,
        help="Number of updates steps to accumulate before performing a backward/update pass.",
    )
    parser.add_argument(
        "--gradient_checkpointing",
        action="store_true",
        help="Whether or not to use gradient checkpointing to save memory at the expense of slower backward pass.",
    )
    parser.add_argument(
        "--learning_rate",
        type=float,
        default=5e-6,
        help="Initial learning rate (after the potential warmup period) to use.",
    )
    parser.add_argument(
        "--scale_lr",
        action="store_true",
        default=False,
        help="Scale the learning rate by the number of GPUs, gradient accumulation steps, and batch size.",
    )
    parser.add_argument(
        "--lr_scheduler",
        type=str,
        default="constant",
        help=(
            'The scheduler type to use. Choose between ["linear", "cosine", "cosine_with_restarts", "polynomial",'
            ' "constant", "constant_with_warmup"]'
        ),
    )
    parser.add_argument(
        "--lr_warmup_steps", type=int, default=500, help="Number of steps for the warmup in the lr scheduler."
    )
405
406
407
408
409
410
411
    parser.add_argument(
        "--lr_num_cycles",
        type=int,
        default=1,
        help="Number of hard resets of the lr in cosine_with_restarts scheduler.",
    )
    parser.add_argument("--lr_power", type=float, default=1.0, help="Power factor of the polynomial scheduler.")
412
413
414
    parser.add_argument(
        "--use_8bit_adam", action="store_true", help="Whether or not to use 8-bit Adam from bitsandbytes."
    )
415
416
417
418
419
420
421
422
    parser.add_argument(
        "--dataloader_num_workers",
        type=int,
        default=0,
        help=(
            "Number of subprocesses to use for data loading. 0 means that the data will be loaded in the main process."
        ),
    )
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
    parser.add_argument("--adam_beta1", type=float, default=0.9, help="The beta1 parameter for the Adam optimizer.")
    parser.add_argument("--adam_beta2", type=float, default=0.999, help="The beta2 parameter for the Adam optimizer.")
    parser.add_argument("--adam_weight_decay", type=float, default=1e-2, help="Weight decay to use.")
    parser.add_argument("--adam_epsilon", type=float, default=1e-08, help="Epsilon value for the Adam optimizer")
    parser.add_argument("--max_grad_norm", default=1.0, type=float, help="Max gradient norm.")
    parser.add_argument("--push_to_hub", action="store_true", help="Whether or not to push the model to the Hub.")
    parser.add_argument("--hub_token", type=str, default=None, help="The token to use to push to the Model Hub.")
    parser.add_argument(
        "--hub_model_id",
        type=str,
        default=None,
        help="The name of the repository to keep in sync with the local `output_dir`.",
    )
    parser.add_argument(
        "--logging_dir",
        type=str,
        default="logs",
        help=(
            "[TensorBoard](https://www.tensorflow.org/tensorboard) log directory. Will default to"
            " *output_dir/runs/**CURRENT_DATETIME_HOSTNAME***."
        ),
    )
Suraj Patil's avatar
Suraj Patil committed
445
446
447
448
449
450
451
452
453
454
455
456
457
    parser.add_argument(
        "--allow_tf32",
        action="store_true",
        help=(
            "Whether or not to allow TF32 on Ampere GPUs. Can be used to speed up training. For more information, see"
            " https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices"
        ),
    )
    parser.add_argument(
        "--report_to",
        type=str,
        default="tensorboard",
        help=(
458
459
            'The integration to report the results and logs to. Supported platforms are `"tensorboard"`'
            ' (default), `"wandb"` and `"comet_ml"`. Use `"all"` to report to all integrations.'
Suraj Patil's avatar
Suraj Patil committed
460
461
        ),
    )
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
    parser.add_argument(
        "--validation_prompt",
        type=str,
        default=None,
        help="A prompt that is used during validation to verify that the model is learning.",
    )
    parser.add_argument(
        "--num_validation_images",
        type=int,
        default=4,
        help="Number of images that should be generated during validation with `validation_prompt`.",
    )
    parser.add_argument(
        "--validation_steps",
        type=int,
        default=100,
        help=(
            "Run validation every X steps. Validation consists of running the prompt"
            " `args.validation_prompt` multiple times: `args.num_validation_images`"
            " and logging the images."
        ),
    )
484
485
486
    parser.add_argument(
        "--mixed_precision",
        type=str,
487
        default=None,
488
489
        choices=["no", "fp16", "bf16"],
        help=(
490
491
492
            "Whether to use mixed precision. Choose between fp16 and bf16 (bfloat16). Bf16 requires PyTorch >="
            " 1.10.and an Nvidia Ampere GPU.  Default to the value of accelerate config of the current system or the"
            " flag passed with the `accelerate.launch` command. Use this argument to override the accelerate config."
493
494
        ),
    )
495
496
497
498
499
500
501
502
503
504
    parser.add_argument(
        "--prior_generation_precision",
        type=str,
        default=None,
        choices=["no", "fp32", "fp16", "bf16"],
        help=(
            "Choose prior generation precision between fp32, fp16 and bf16 (bfloat16). Bf16 requires PyTorch >="
            " 1.10.and an Nvidia Ampere GPU.  Default to  fp16 if a GPU is available else fp32."
        ),
    )
505
    parser.add_argument("--local_rank", type=int, default=-1, help="For distributed training: local_rank")
506
507
508
    parser.add_argument(
        "--enable_xformers_memory_efficient_attention", action="store_true", help="Whether or not to use xformers."
    )
509
510
511
512
513
514
515
516
517
    parser.add_argument(
        "--set_grads_to_none",
        action="store_true",
        help=(
            "Save more memory by using setting grads to None instead of zero. Be aware, that this changes certain"
            " behaviors, so disable this argument if it causes any problems. More info:"
            " https://pytorch.org/docs/stable/generated/torch.optim.Optimizer.zero_grad.html"
        ),
    )
518

519
520
521
522
523
524
525
526
527
    parser.add_argument(
        "--offset_noise",
        action="store_true",
        default=False,
        help=(
            "Fine-tuning against a modified noise"
            " See: https://www.crosslabs.org//blog/diffusion-with-offset-noise for more information."
        ),
    )
528
529
530
531
532
533
534
    parser.add_argument(
        "--snr_gamma",
        type=float,
        default=None,
        help="SNR weighting gamma to be used if rebalancing the loss. Recommended value is 5.0. "
        "More details here: https://arxiv.org/abs/2303.09556.",
    )
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
    parser.add_argument(
        "--pre_compute_text_embeddings",
        action="store_true",
        help="Whether or not to pre-compute text embeddings. If text embeddings are pre-computed, the text encoder will not be kept in memory during training and will leave more GPU memory available for training the rest of the model. This is not compatible with `--train_text_encoder`.",
    )
    parser.add_argument(
        "--tokenizer_max_length",
        type=int,
        default=None,
        required=False,
        help="The maximum length of the tokenizer. If not set, will default to the tokenizer's max length.",
    )
    parser.add_argument(
        "--text_encoder_use_attention_mask",
        action="store_true",
        required=False,
        help="Whether to use attention mask for the text encoder",
    )
    parser.add_argument(
        "--skip_save_text_encoder", action="store_true", required=False, help="Set to not save text encoder"
    )
556
557
558
559
560
561
562
563
564
565
566
567
568
    parser.add_argument(
        "--validation_images",
        required=False,
        default=None,
        nargs="+",
        help="Optional set of images to use for validation. Used when the target pipeline takes an initial image as input such as when training image variation or superresolution.",
    )
    parser.add_argument(
        "--class_labels_conditioning",
        required=False,
        default=None,
        help="The optional `class_label` conditioning to pass to the unet, available values are `timesteps`.",
    )
569
570
571
572
573
574
575
    parser.add_argument(
        "--validation_scheduler",
        type=str,
        default="DPMSolverMultistepScheduler",
        choices=["DPMSolverMultistepScheduler", "DDPMScheduler"],
        help="Select which scheduler to use for validation. DDPMScheduler is recommended for DeepFloyd IF.",
    )
576

577
578
579
580
581
    if input_args is not None:
        args = parser.parse_args(input_args)
    else:
        args = parser.parse_args()

582
583
584
585
586
587
588
589
590
    env_local_rank = int(os.environ.get("LOCAL_RANK", -1))
    if env_local_rank != -1 and env_local_rank != args.local_rank:
        args.local_rank = env_local_rank

    if args.with_prior_preservation:
        if args.class_data_dir is None:
            raise ValueError("You must specify a data directory for class images.")
        if args.class_prompt is None:
            raise ValueError("You must specify prompt for class images.")
591
    else:
592
        # logger is not available yet
593
        if args.class_data_dir is not None:
594
            warnings.warn("You need not use --class_data_dir without --with_prior_preservation.")
595
        if args.class_prompt is not None:
596
            warnings.warn("You need not use --class_prompt without --with_prior_preservation.")
597

598
599
600
    if args.train_text_encoder and args.pre_compute_text_embeddings:
        raise ValueError("`--train_text_encoder` cannot be used with `--pre_compute_text_embeddings`")

601
602
603
604
605
    return args


class DreamBoothDataset(Dataset):
    """
Yuta Hayashibe's avatar
Yuta Hayashibe committed
606
    A dataset to prepare the instance and class images with the prompts for fine-tuning the model.
607
608
609
610
611
612
613
614
615
616
    It pre-processes the images and the tokenizes prompts.
    """

    def __init__(
        self,
        instance_data_root,
        instance_prompt,
        tokenizer,
        class_data_root=None,
        class_prompt=None,
617
        class_num=None,
618
619
        size=512,
        center_crop=False,
620
        encoder_hidden_states=None,
621
        class_prompt_encoder_hidden_states=None,
622
        tokenizer_max_length=None,
623
624
625
626
    ):
        self.size = size
        self.center_crop = center_crop
        self.tokenizer = tokenizer
627
        self.encoder_hidden_states = encoder_hidden_states
628
        self.class_prompt_encoder_hidden_states = class_prompt_encoder_hidden_states
629
        self.tokenizer_max_length = tokenizer_max_length
630
631
632

        self.instance_data_root = Path(instance_data_root)
        if not self.instance_data_root.exists():
633
            raise ValueError(f"Instance {self.instance_data_root} images root doesn't exists.")
634
635
636
637
638
639
640
641
642

        self.instance_images_path = list(Path(instance_data_root).iterdir())
        self.num_instance_images = len(self.instance_images_path)
        self.instance_prompt = instance_prompt
        self._length = self.num_instance_images

        if class_data_root is not None:
            self.class_data_root = Path(class_data_root)
            self.class_data_root.mkdir(parents=True, exist_ok=True)
643
            self.class_images_path = list(self.class_data_root.iterdir())
644
645
646
647
            if class_num is not None:
                self.num_class_images = min(len(self.class_images_path), class_num)
            else:
                self.num_class_images = len(self.class_images_path)
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
            self._length = max(self.num_class_images, self.num_instance_images)
            self.class_prompt = class_prompt
        else:
            self.class_data_root = None

        self.image_transforms = transforms.Compose(
            [
                transforms.Resize(size, interpolation=transforms.InterpolationMode.BILINEAR),
                transforms.CenterCrop(size) if center_crop else transforms.RandomCrop(size),
                transforms.ToTensor(),
                transforms.Normalize([0.5], [0.5]),
            ]
        )

    def __len__(self):
        return self._length

    def __getitem__(self, index):
        example = {}
        instance_image = Image.open(self.instance_images_path[index % self.num_instance_images])
668
669
        instance_image = exif_transpose(instance_image)

670
671
672
        if not instance_image.mode == "RGB":
            instance_image = instance_image.convert("RGB")
        example["instance_images"] = self.image_transforms(instance_image)
673
674
675
676
677
678
679
680
681

        if self.encoder_hidden_states is not None:
            example["instance_prompt_ids"] = self.encoder_hidden_states
        else:
            text_inputs = tokenize_prompt(
                self.tokenizer, self.instance_prompt, tokenizer_max_length=self.tokenizer_max_length
            )
            example["instance_prompt_ids"] = text_inputs.input_ids
            example["instance_attention_mask"] = text_inputs.attention_mask
682
683
684

        if self.class_data_root:
            class_image = Image.open(self.class_images_path[index % self.num_class_images])
685
686
            class_image = exif_transpose(class_image)

687
688
689
            if not class_image.mode == "RGB":
                class_image = class_image.convert("RGB")
            example["class_images"] = self.image_transforms(class_image)
690

691
692
            if self.class_prompt_encoder_hidden_states is not None:
                example["class_prompt_ids"] = self.class_prompt_encoder_hidden_states
693
694
695
696
697
698
            else:
                class_text_inputs = tokenize_prompt(
                    self.tokenizer, self.class_prompt, tokenizer_max_length=self.tokenizer_max_length
                )
                example["class_prompt_ids"] = class_text_inputs.input_ids
                example["class_attention_mask"] = class_text_inputs.attention_mask
699
700
701
702

        return example


703
def collate_fn(examples, with_prior_preservation=False):
704
705
    has_attention_mask = "instance_attention_mask" in examples[0]

706
707
708
    input_ids = [example["instance_prompt_ids"] for example in examples]
    pixel_values = [example["instance_images"] for example in examples]

709
710
711
    if has_attention_mask:
        attention_mask = [example["instance_attention_mask"] for example in examples]

712
713
714
715
716
717
    # Concat class and instance examples for prior preservation.
    # We do this to avoid doing two forward passes.
    if with_prior_preservation:
        input_ids += [example["class_prompt_ids"] for example in examples]
        pixel_values += [example["class_images"] for example in examples]

718
719
720
        if has_attention_mask:
            attention_mask += [example["class_attention_mask"] for example in examples]

721
722
723
724
725
726
727
728
729
    pixel_values = torch.stack(pixel_values)
    pixel_values = pixel_values.to(memory_format=torch.contiguous_format).float()

    input_ids = torch.cat(input_ids, dim=0)

    batch = {
        "input_ids": input_ids,
        "pixel_values": pixel_values,
    }
730
731

    if has_attention_mask:
732
        attention_mask = torch.cat(attention_mask, dim=0)
733
734
        batch["attention_mask"] = attention_mask

735
736
737
    return batch


738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
class PromptDataset(Dataset):
    "A simple dataset to prepare the prompts to generate class images on multiple GPUs."

    def __init__(self, prompt, num_samples):
        self.prompt = prompt
        self.num_samples = num_samples

    def __len__(self):
        return self.num_samples

    def __getitem__(self, index):
        example = {}
        example["prompt"] = self.prompt
        example["index"] = index
        return example


755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
def model_has_vae(args):
    config_file_name = os.path.join("vae", AutoencoderKL.config_name)
    if os.path.isdir(args.pretrained_model_name_or_path):
        config_file_name = os.path.join(args.pretrained_model_name_or_path, config_file_name)
        return os.path.isfile(config_file_name)
    else:
        files_in_repo = model_info(args.pretrained_model_name_or_path, revision=args.revision).siblings
        return any(file.rfilename == config_file_name for file in files_in_repo)


def tokenize_prompt(tokenizer, prompt, tokenizer_max_length=None):
    if tokenizer_max_length is not None:
        max_length = tokenizer_max_length
    else:
        max_length = tokenizer.model_max_length

    text_inputs = tokenizer(
        prompt,
        truncation=True,
        padding="max_length",
        max_length=max_length,
        return_tensors="pt",
    )

    return text_inputs


def encode_prompt(text_encoder, input_ids, attention_mask, text_encoder_use_attention_mask=None):
    text_input_ids = input_ids.to(text_encoder.device)

    if text_encoder_use_attention_mask:
        attention_mask = attention_mask.to(text_encoder.device)
    else:
        attention_mask = None

    prompt_embeds = text_encoder(
        text_input_ids,
        attention_mask=attention_mask,
    )
    prompt_embeds = prompt_embeds[0]

    return prompt_embeds


799
def main(args):
800
801
    logging_dir = Path(args.output_dir, args.logging_dir)

802
    accelerator_project_config = ProjectConfiguration(project_dir=args.output_dir, logging_dir=logging_dir)
803

804
805
806
    accelerator = Accelerator(
        gradient_accumulation_steps=args.gradient_accumulation_steps,
        mixed_precision=args.mixed_precision,
Suraj Patil's avatar
Suraj Patil committed
807
        log_with=args.report_to,
808
        project_config=accelerator_project_config,
809
810
    )

811
812
813
814
    if args.report_to == "wandb":
        if not is_wandb_available():
            raise ImportError("Make sure to install wandb if you want to use it for logging during training.")

815
816
817
818
819
820
821
822
823
    # Currently, it's not possible to do gradient accumulation when training two models with accelerate.accumulate
    # This will be enabled soon in accelerate. For now, we don't allow gradient accumulation when training two models.
    # TODO (patil-suraj): Remove this check when gradient accumulation with two models is enabled in accelerate.
    if args.train_text_encoder and args.gradient_accumulation_steps > 1 and accelerator.num_processes > 1:
        raise ValueError(
            "Gradient accumulation is not supported when training the text encoder in distributed training. "
            "Please set gradient_accumulation_steps to 1. This feature will be supported in the future."
        )

Suraj Patil's avatar
Suraj Patil committed
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
    # Make one log on every process with the configuration for debugging.
    logging.basicConfig(
        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
        datefmt="%m/%d/%Y %H:%M:%S",
        level=logging.INFO,
    )
    logger.info(accelerator.state, main_process_only=False)
    if accelerator.is_local_main_process:
        transformers.utils.logging.set_verbosity_warning()
        diffusers.utils.logging.set_verbosity_info()
    else:
        transformers.utils.logging.set_verbosity_error()
        diffusers.utils.logging.set_verbosity_error()

    # If passed along, set the training seed now.
839
840
841
    if args.seed is not None:
        set_seed(args.seed)

Suraj Patil's avatar
Suraj Patil committed
842
    # Generate class images if prior preservation is enabled.
843
844
845
846
847
848
849
850
    if args.with_prior_preservation:
        class_images_dir = Path(args.class_data_dir)
        if not class_images_dir.exists():
            class_images_dir.mkdir(parents=True)
        cur_class_images = len(list(class_images_dir.iterdir()))

        if cur_class_images < args.num_class_images:
            torch_dtype = torch.float16 if accelerator.device.type == "cuda" else torch.float32
851
852
853
854
855
856
            if args.prior_generation_precision == "fp32":
                torch_dtype = torch.float32
            elif args.prior_generation_precision == "fp16":
                torch_dtype = torch.float16
            elif args.prior_generation_precision == "bf16":
                torch_dtype = torch.bfloat16
857
            pipeline = DiffusionPipeline.from_pretrained(
858
859
860
861
                args.pretrained_model_name_or_path,
                torch_dtype=torch_dtype,
                safety_checker=None,
                revision=args.revision,
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
            )
            pipeline.set_progress_bar_config(disable=True)

            num_new_images = args.num_class_images - cur_class_images
            logger.info(f"Number of class images to sample: {num_new_images}.")

            sample_dataset = PromptDataset(args.class_prompt, num_new_images)
            sample_dataloader = torch.utils.data.DataLoader(sample_dataset, batch_size=args.sample_batch_size)

            sample_dataloader = accelerator.prepare(sample_dataloader)
            pipeline.to(accelerator.device)

            for example in tqdm(
                sample_dataloader, desc="Generating class images", disable=not accelerator.is_local_main_process
            ):
877
                images = pipeline(example["prompt"]).images
878
879

                for i, image in enumerate(images):
880
881
882
                    hash_image = hashlib.sha1(image.tobytes()).hexdigest()
                    image_filename = class_images_dir / f"{example['index'][i] + cur_class_images}-{hash_image}.jpg"
                    image.save(image_filename)
883
884
885
886
887
888
889

            del pipeline
            if torch.cuda.is_available():
                torch.cuda.empty_cache()

    # Handle the repository creation
    if accelerator.is_main_process:
890
        if args.output_dir is not None:
891
892
            os.makedirs(args.output_dir, exist_ok=True)

893
894
895
896
897
        if args.push_to_hub:
            repo_id = create_repo(
                repo_id=args.hub_model_id or Path(args.output_dir).name, exist_ok=True, token=args.hub_token
            ).repo_id

898
899
    # Load the tokenizer
    if args.tokenizer_name:
Suraj Patil's avatar
Suraj Patil committed
900
        tokenizer = AutoTokenizer.from_pretrained(args.tokenizer_name, revision=args.revision, use_fast=False)
901
    elif args.pretrained_model_name_or_path:
902
        tokenizer = AutoTokenizer.from_pretrained(
903
904
905
            args.pretrained_model_name_or_path,
            subfolder="tokenizer",
            revision=args.revision,
906
            use_fast=False,
907
        )
908

909
    # import correct text encoder class
910
    text_encoder_cls = import_model_class_from_model_name_or_path(args.pretrained_model_name_or_path, args.revision)
911

Suraj Patil's avatar
Suraj Patil committed
912
913
    # Load scheduler and models
    noise_scheduler = DDPMScheduler.from_pretrained(args.pretrained_model_name_or_path, subfolder="scheduler")
914
    text_encoder = text_encoder_cls.from_pretrained(
Suraj Patil's avatar
Suraj Patil committed
915
        args.pretrained_model_name_or_path, subfolder="text_encoder", revision=args.revision
916
    )
917
918
919
920
921
922
923
924

    if model_has_vae(args):
        vae = AutoencoderKL.from_pretrained(
            args.pretrained_model_name_or_path, subfolder="vae", revision=args.revision
        )
    else:
        vae = None

925
    unet = UNet2DConditionModel.from_pretrained(
Suraj Patil's avatar
Suraj Patil committed
926
        args.pretrained_model_name_or_path, subfolder="unet", revision=args.revision
927
    )
928

929
930
    # create custom saving & loading hooks so that `accelerator.save_state(...)` serializes in a nice format
    def save_model_hook(models, weights, output_dir):
931
932
933
934
        if accelerator.is_main_process:
            for model in models:
                sub_dir = "unet" if isinstance(model, type(accelerator.unwrap_model(unet))) else "text_encoder"
                model.save_pretrained(os.path.join(output_dir, sub_dir))
935

936
937
                # make sure to pop weight so that corresponding model is not saved again
                weights.pop()
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957

    def load_model_hook(models, input_dir):
        while len(models) > 0:
            # pop models so that they are not loaded again
            model = models.pop()

            if isinstance(model, type(accelerator.unwrap_model(text_encoder))):
                # load transformers style into model
                load_model = text_encoder_cls.from_pretrained(input_dir, subfolder="text_encoder")
                model.config = load_model.config
            else:
                # load diffusers style into model
                load_model = UNet2DConditionModel.from_pretrained(input_dir, subfolder="unet")
                model.register_to_config(**load_model.config)

            model.load_state_dict(load_model.state_dict())
            del load_model

    accelerator.register_save_state_pre_hook(save_model_hook)
    accelerator.register_load_state_pre_hook(load_model_hook)
958

959
960
961
    if vae is not None:
        vae.requires_grad_(False)

Suraj Patil's avatar
Suraj Patil committed
962
963
964
    if not args.train_text_encoder:
        text_encoder.requires_grad_(False)

965
966
    if args.enable_xformers_memory_efficient_attention:
        if is_xformers_available():
967
968
969
970
971
972
973
            import xformers

            xformers_version = version.parse(xformers.__version__)
            if xformers_version == version.parse("0.0.16"):
                logger.warn(
                    "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."
                )
974
            unet.enable_xformers_memory_efficient_attention()
975
976
        else:
            raise ValueError("xformers is not available. Make sure it is installed correctly")
977

978
979
    if args.gradient_checkpointing:
        unet.enable_gradient_checkpointing()
980
981
        if args.train_text_encoder:
            text_encoder.gradient_checkpointing_enable()
982

983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
    # Check that all trainable models are in full precision
    low_precision_error_string = (
        "Please make sure to always have all model weights in full float32 precision when starting training - even if"
        " doing mixed precision training. copy of the weights should still be float32."
    )

    if accelerator.unwrap_model(unet).dtype != torch.float32:
        raise ValueError(
            f"Unet loaded as datatype {accelerator.unwrap_model(unet).dtype}. {low_precision_error_string}"
        )

    if args.train_text_encoder and accelerator.unwrap_model(text_encoder).dtype != torch.float32:
        raise ValueError(
            f"Text encoder loaded as datatype {accelerator.unwrap_model(text_encoder).dtype}."
            f" {low_precision_error_string}"
        )

Suraj Patil's avatar
Suraj Patil committed
1000
1001
1002
1003
1004
    # Enable TF32 for faster training on Ampere GPUs,
    # cf https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices
    if args.allow_tf32:
        torch.backends.cuda.matmul.allow_tf32 = True

1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
    if args.scale_lr:
        args.learning_rate = (
            args.learning_rate * args.gradient_accumulation_steps * args.train_batch_size * accelerator.num_processes
        )

    # Use 8-bit Adam for lower memory usage or to fine-tune the model in 16GB GPUs
    if args.use_8bit_adam:
        try:
            import bitsandbytes as bnb
        except ImportError:
            raise ImportError(
                "To use 8-bit Adam, please install the bitsandbytes library: `pip install bitsandbytes`."
            )

        optimizer_class = bnb.optim.AdamW8bit
    else:
        optimizer_class = torch.optim.AdamW

Suraj Patil's avatar
Suraj Patil committed
1023
    # Optimizer creation
1024
1025
1026
    params_to_optimize = (
        itertools.chain(unet.parameters(), text_encoder.parameters()) if args.train_text_encoder else unet.parameters()
    )
1027
    optimizer = optimizer_class(
1028
        params_to_optimize,
1029
1030
1031
1032
1033
1034
        lr=args.learning_rate,
        betas=(args.adam_beta1, args.adam_beta2),
        weight_decay=args.adam_weight_decay,
        eps=args.adam_epsilon,
    )

1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
    if args.pre_compute_text_embeddings:

        def compute_text_embeddings(prompt):
            with torch.no_grad():
                text_inputs = tokenize_prompt(tokenizer, prompt, tokenizer_max_length=args.tokenizer_max_length)
                prompt_embeds = encode_prompt(
                    text_encoder,
                    text_inputs.input_ids,
                    text_inputs.attention_mask,
                    text_encoder_use_attention_mask=args.text_encoder_use_attention_mask,
                )

            return prompt_embeds

        pre_computed_encoder_hidden_states = compute_text_embeddings(args.instance_prompt)
        validation_prompt_negative_prompt_embeds = compute_text_embeddings("")

        if args.validation_prompt is not None:
            validation_prompt_encoder_hidden_states = compute_text_embeddings(args.validation_prompt)
        else:
            validation_prompt_encoder_hidden_states = None

1057
1058
        if args.class_prompt is not None:
            pre_computed_class_prompt_encoder_hidden_states = compute_text_embeddings(args.class_prompt)
1059
        else:
1060
            pre_computed_class_prompt_encoder_hidden_states = None
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070

        text_encoder = None
        tokenizer = None

        gc.collect()
        torch.cuda.empty_cache()
    else:
        pre_computed_encoder_hidden_states = None
        validation_prompt_encoder_hidden_states = None
        validation_prompt_negative_prompt_embeds = None
1071
        pre_computed_class_prompt_encoder_hidden_states = None
1072

Suraj Patil's avatar
Suraj Patil committed
1073
    # Dataset and DataLoaders creation:
1074
1075
1076
1077
1078
    train_dataset = DreamBoothDataset(
        instance_data_root=args.instance_data_dir,
        instance_prompt=args.instance_prompt,
        class_data_root=args.class_data_dir if args.with_prior_preservation else None,
        class_prompt=args.class_prompt,
1079
        class_num=args.num_class_images,
1080
1081
1082
        tokenizer=tokenizer,
        size=args.resolution,
        center_crop=args.center_crop,
1083
        encoder_hidden_states=pre_computed_encoder_hidden_states,
1084
        class_prompt_encoder_hidden_states=pre_computed_class_prompt_encoder_hidden_states,
1085
        tokenizer_max_length=args.tokenizer_max_length,
1086
1087
1088
    )

    train_dataloader = torch.utils.data.DataLoader(
1089
1090
1091
1092
        train_dataset,
        batch_size=args.train_batch_size,
        shuffle=True,
        collate_fn=lambda examples: collate_fn(examples, args.with_prior_preservation),
1093
        num_workers=args.dataloader_num_workers,
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
    )

    # Scheduler and math around the number of training steps.
    overrode_max_train_steps = False
    num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
    if args.max_train_steps is None:
        args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
        overrode_max_train_steps = True

    lr_scheduler = get_scheduler(
        args.lr_scheduler,
        optimizer=optimizer,
1106
1107
        num_warmup_steps=args.lr_warmup_steps * accelerator.num_processes,
        num_training_steps=args.max_train_steps * accelerator.num_processes,
1108
1109
        num_cycles=args.lr_num_cycles,
        power=args.lr_power,
1110
1111
    )

Suraj Patil's avatar
Suraj Patil committed
1112
    # Prepare everything with our `accelerator`.
1113
1114
1115
1116
1117
1118
1119
1120
    if args.train_text_encoder:
        unet, text_encoder, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
            unet, text_encoder, optimizer, train_dataloader, lr_scheduler
        )
    else:
        unet, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
            unet, optimizer, train_dataloader, lr_scheduler
        )
1121

1122
1123
    # For mixed precision training we cast all non-trainable weigths (vae, non-lora text_encoder and non-lora unet) to half-precision
    # as these weights are only used for inference, keeping weights in full precision is not required.
1124
    weight_dtype = torch.float32
1125
    if accelerator.mixed_precision == "fp16":
1126
        weight_dtype = torch.float16
1127
    elif accelerator.mixed_precision == "bf16":
1128
1129
        weight_dtype = torch.bfloat16

Suraj Patil's avatar
Suraj Patil committed
1130
    # Move vae and text_encoder to device and cast to weight_dtype
1131
1132
1133
1134
    if vae is not None:
        vae.to(accelerator.device, dtype=weight_dtype)

    if not args.train_text_encoder and text_encoder is not None:
1135
        text_encoder.to(accelerator.device, dtype=weight_dtype)
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146

    # We need to recalculate our total training steps as the size of the training dataloader may have changed.
    num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
    if overrode_max_train_steps:
        args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
    # Afterwards we recalculate our number of training epochs
    args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)

    # We need to initialize the trackers we use, and also store our configuration.
    # The trackers initializes automatically on the main process.
    if accelerator.is_main_process:
1147
        tracker_config = vars(copy.deepcopy(args))
1148
1149
        tracker_config.pop("validation_images")
        accelerator.init_trackers("dreambooth", config=tracker_config)
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161

    # Train!
    total_batch_size = args.train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps

    logger.info("***** Running training *****")
    logger.info(f"  Num examples = {len(train_dataset)}")
    logger.info(f"  Num batches each epoch = {len(train_dataloader)}")
    logger.info(f"  Num Epochs = {args.num_train_epochs}")
    logger.info(f"  Instantaneous batch size per device = {args.train_batch_size}")
    logger.info(f"  Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}")
    logger.info(f"  Gradient Accumulation steps = {args.gradient_accumulation_steps}")
    logger.info(f"  Total optimization steps = {args.max_train_steps}")
1162
1163
1164
    global_step = 0
    first_epoch = 0

Suraj Patil's avatar
Suraj Patil committed
1165
    # Potentially load in the weights and states from a previous save
1166
1167
1168
1169
1170
1171
1172
1173
    if args.resume_from_checkpoint:
        if args.resume_from_checkpoint != "latest":
            path = os.path.basename(args.resume_from_checkpoint)
        else:
            # Get the mos recent checkpoint
            dirs = os.listdir(args.output_dir)
            dirs = [d for d in dirs if d.startswith("checkpoint")]
            dirs = sorted(dirs, key=lambda x: int(x.split("-")[1]))
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
            path = dirs[-1] if len(dirs) > 0 else None

        if path is None:
            accelerator.print(
                f"Checkpoint '{args.resume_from_checkpoint}' does not exist. Starting a new training run."
            )
            args.resume_from_checkpoint = None
        else:
            accelerator.print(f"Resuming from checkpoint {path}")
            accelerator.load_state(os.path.join(args.output_dir, path))
            global_step = int(path.split("-")[1])

            resume_global_step = global_step * args.gradient_accumulation_steps
            first_epoch = global_step // num_update_steps_per_epoch
            resume_step = resume_global_step % (num_update_steps_per_epoch * args.gradient_accumulation_steps)
1189

1190
    # Only show the progress bar once on each machine.
1191
    progress_bar = tqdm(range(global_step, args.max_train_steps), disable=not accelerator.is_local_main_process)
1192
1193
    progress_bar.set_description("Steps")

1194
    for epoch in range(first_epoch, args.num_train_epochs):
1195
        unet.train()
1196
1197
        if args.train_text_encoder:
            text_encoder.train()
1198
        for step, batch in enumerate(train_dataloader):
1199
1200
1201
1202
1203
1204
            # Skip steps until we reach the resumed step
            if args.resume_from_checkpoint and epoch == first_epoch and step < resume_step:
                if step % args.gradient_accumulation_steps == 0:
                    progress_bar.update(1)
                continue

1205
            with accelerator.accumulate(unet):
1206
                pixel_values = batch["pixel_values"].to(dtype=weight_dtype)
1207

1208
1209
1210
1211
1212
1213
1214
1215
                if vae is not None:
                    # Convert images to latent space
                    model_input = vae.encode(batch["pixel_values"].to(dtype=weight_dtype)).latent_dist.sample()
                    model_input = model_input * vae.config.scaling_factor
                else:
                    model_input = pixel_values

                # Sample noise that we'll add to the model input
1216
                if args.offset_noise:
1217
1218
                    noise = torch.randn_like(model_input) + 0.1 * torch.randn(
                        model_input.shape[0], model_input.shape[1], 1, 1, device=model_input.device
1219
1220
                    )
                else:
1221
                    noise = torch.randn_like(model_input)
1222
                bsz, channels, height, width = model_input.shape
1223
                # Sample a random timestep for each image
1224
1225
1226
                timesteps = torch.randint(
                    0, noise_scheduler.config.num_train_timesteps, (bsz,), device=model_input.device
                )
1227
1228
                timesteps = timesteps.long()

1229
                # Add noise to the model input according to the noise magnitude at each timestep
1230
                # (this is the forward diffusion process)
1231
                noisy_model_input = noise_scheduler.add_noise(model_input, noise, timesteps)
1232
1233

                # Get the text embedding for conditioning
1234
1235
1236
1237
1238
1239
1240
1241
1242
                if args.pre_compute_text_embeddings:
                    encoder_hidden_states = batch["input_ids"]
                else:
                    encoder_hidden_states = encode_prompt(
                        text_encoder,
                        batch["input_ids"],
                        batch["attention_mask"],
                        text_encoder_use_attention_mask=args.text_encoder_use_attention_mask,
                    )
1243

1244
                if accelerator.unwrap_model(unet).config.in_channels == channels * 2:
1245
                    noisy_model_input = torch.cat([noisy_model_input, noisy_model_input], dim=1)
1246
1247
1248
1249
1250
1251

                if args.class_labels_conditioning == "timesteps":
                    class_labels = timesteps
                else:
                    class_labels = None

1252
                # Predict the noise residual
1253
1254
1255
                model_pred = unet(
                    noisy_model_input, timesteps, encoder_hidden_states, class_labels=class_labels
                ).sample
1256
1257
1258

                if model_pred.shape[1] == 6:
                    model_pred, _ = torch.chunk(model_pred, 2, dim=1)
1259
1260
1261
1262
1263

                # Get the target for loss depending on the prediction type
                if noise_scheduler.config.prediction_type == "epsilon":
                    target = noise
                elif noise_scheduler.config.prediction_type == "v_prediction":
1264
                    target = noise_scheduler.get_velocity(model_input, noise, timesteps)
1265
1266
                else:
                    raise ValueError(f"Unknown prediction type {noise_scheduler.config.prediction_type}")
1267
1268

                if args.with_prior_preservation:
1269
1270
1271
                    # Chunk the noise and model_pred into two parts and compute the loss on each part separately.
                    model_pred, model_pred_prior = torch.chunk(model_pred, 2, dim=0)
                    target, target_prior = torch.chunk(target, 2, dim=0)
1272
1273
                    # Compute prior loss
                    prior_loss = F.mse_loss(model_pred_prior.float(), target_prior.float(), reduction="mean")
1274

1275
1276
                # Compute instance loss
                if args.snr_gamma is None:
1277
                    loss = F.mse_loss(model_pred.float(), target.float(), reduction="mean")
1278
1279
1280
1281
                else:
                    # Compute loss-weights as per Section 3.4 of https://arxiv.org/abs/2303.09556.
                    # Since we predict the noise instead of x_0, the original formulation is slightly changed.
                    # This is discussed in Section 4.2 of the same paper.
1282
                    snr = compute_snr(noise_scheduler, timesteps)
1283
1284
1285
                    base_weight = (
                        torch.stack([snr, args.snr_gamma * torch.ones_like(timesteps)], dim=1).min(dim=1)[0] / snr
                    )
1286

1287
1288
1289
1290
1291
1292
1293
1294
1295
                    if noise_scheduler.config.prediction_type == "v_prediction":
                        # Velocity objective needs to be floored to an SNR weight of one.
                        mse_loss_weights = base_weight + 1
                    else:
                        # Epsilon and sample both use the same loss weights.
                        mse_loss_weights = base_weight
                    loss = F.mse_loss(model_pred.float(), target.float(), reduction="none")
                    loss = loss.mean(dim=list(range(1, len(loss.shape)))) * mse_loss_weights
                    loss = loss.mean()
1296

1297
                if args.with_prior_preservation:
1298
1299
1300
1301
                    # Add the prior loss to the instance loss.
                    loss = loss + args.prior_loss_weight * prior_loss

                accelerator.backward(loss)
1302
                if accelerator.sync_gradients:
1303
1304
1305
1306
1307
1308
                    params_to_clip = (
                        itertools.chain(unet.parameters(), text_encoder.parameters())
                        if args.train_text_encoder
                        else unet.parameters()
                    )
                    accelerator.clip_grad_norm_(params_to_clip, args.max_grad_norm)
1309
1310
                optimizer.step()
                lr_scheduler.step()
1311
                optimizer.zero_grad(set_to_none=args.set_grads_to_none)
1312
1313
1314
1315
1316
1317

            # Checks if the accelerator has performed an optimization step behind the scenes
            if accelerator.sync_gradients:
                progress_bar.update(1)
                global_step += 1

1318
1319
                if accelerator.is_main_process:
                    if global_step % args.checkpointing_steps == 0:
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
                        # _before_ saving state, check if this save would set us over the `checkpoints_total_limit`
                        if args.checkpoints_total_limit is not None:
                            checkpoints = os.listdir(args.output_dir)
                            checkpoints = [d for d in checkpoints if d.startswith("checkpoint")]
                            checkpoints = sorted(checkpoints, key=lambda x: int(x.split("-")[1]))

                            # before we save the new checkpoint, we need to have at _most_ `checkpoints_total_limit - 1` checkpoints
                            if len(checkpoints) >= args.checkpoints_total_limit:
                                num_to_remove = len(checkpoints) - args.checkpoints_total_limit + 1
                                removing_checkpoints = checkpoints[0:num_to_remove]

                                logger.info(
                                    f"{len(checkpoints)} checkpoints already exist, removing {len(removing_checkpoints)} checkpoints"
                                )
                                logger.info(f"removing checkpoints: {', '.join(removing_checkpoints)}")

                                for removing_checkpoint in removing_checkpoints:
                                    removing_checkpoint = os.path.join(args.output_dir, removing_checkpoint)
                                    shutil.rmtree(removing_checkpoint)

1340
                        save_path = os.path.join(args.output_dir, f"checkpoint-{global_step}")
1341
1342
                        accelerator.save_state(save_path)
                        logger.info(f"Saved state to {save_path}")
1343

1344
1345
                    images = []

1346
                    if args.validation_prompt is not None and global_step % args.validation_steps == 0:
1347
                        images = log_validation(
1348
1349
1350
1351
1352
1353
1354
                            text_encoder,
                            tokenizer,
                            unet,
                            vae,
                            args,
                            accelerator,
                            weight_dtype,
1355
                            global_step,
1356
1357
                            validation_prompt_encoder_hidden_states,
                            validation_prompt_negative_prompt_embeds,
1358
                        )
1359

1360
1361
1362
1363
1364
1365
1366
1367
            logs = {"loss": loss.detach().item(), "lr": lr_scheduler.get_last_lr()[0]}
            progress_bar.set_postfix(**logs)
            accelerator.log(logs, step=global_step)

            if global_step >= args.max_train_steps:
                break

    # Create the pipeline using using the trained modules and save it.
Suraj Patil's avatar
Suraj Patil committed
1368
    accelerator.wait_for_everyone()
1369
    if accelerator.is_main_process:
1370
1371
1372
1373
1374
1375
1376
1377
        pipeline_args = {}

        if text_encoder is not None:
            pipeline_args["text_encoder"] = accelerator.unwrap_model(text_encoder)

        if args.skip_save_text_encoder:
            pipeline_args["text_encoder"] = None

1378
        pipeline = DiffusionPipeline.from_pretrained(
1379
1380
            args.pretrained_model_name_or_path,
            unet=accelerator.unwrap_model(unet),
1381
            revision=args.revision,
1382
            **pipeline_args,
1383
        )
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397

        # We train on the simplified learning objective. If we were previously predicting a variance, we need the scheduler to ignore it
        scheduler_args = {}

        if "variance_type" in pipeline.scheduler.config:
            variance_type = pipeline.scheduler.config.variance_type

            if variance_type in ["learned", "learned_range"]:
                variance_type = "fixed_small"

            scheduler_args["variance_type"] = variance_type

        pipeline.scheduler = pipeline.scheduler.from_config(pipeline.scheduler.config, **scheduler_args)

1398
1399
1400
        pipeline.save_pretrained(args.output_dir)

        if args.push_to_hub:
1401
1402
1403
1404
1405
1406
1407
            save_model_card(
                repo_id,
                images=images,
                base_model=args.pretrained_model_name_or_path,
                train_text_encoder=args.train_text_encoder,
                prompt=args.instance_prompt,
                repo_folder=args.output_dir,
1408
                pipeline=pipeline,
1409
            )
1410
1411
1412
1413
1414
1415
            upload_folder(
                repo_id=repo_id,
                folder_path=args.output_dir,
                commit_message="End of training",
                ignore_patterns=["step_*", "epoch_*"],
            )
1416
1417
1418
1419
1420

    accelerator.end_training()


if __name__ == "__main__":
1421
1422
    args = parse_args()
    main(args)