train_dreambooth.py 54.3 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 gc
18
import hashlib
19
import itertools
Suraj Patil's avatar
Suraj Patil committed
20
import logging
21
22
import math
import os
23
import shutil
24
import warnings
25
26
from pathlib import Path

27
import numpy as np
28
29
30
import torch
import torch.nn.functional as F
import torch.utils.checkpoint
Suraj Patil's avatar
Suraj Patil committed
31
import transformers
32
33
from accelerate import Accelerator
from accelerate.logging import get_logger
34
from accelerate.utils import ProjectConfiguration, set_seed
Patrick von Platen's avatar
Patrick von Platen committed
35
36
37
38
39
40
41
42
43
44
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
45
46
47
48
49
from diffusers import (
    AutoencoderKL,
    DDPMScheduler,
    DiffusionPipeline,
    DPMSolverMultistepScheduler,
50
    StableDiffusionPipeline,
51
52
    UNet2DConditionModel,
)
53
from diffusers.optimization import get_scheduler
54
from diffusers.utils import check_min_version, is_wandb_available
55
56
from diffusers.utils.import_utils import is_xformers_available

57

58
59
60
if is_wandb_available():
    import wandb

61
# Will error if the minimal version of diffusers is not installed. Remove at your own risks.
62
check_min_version("0.18.0.dev0")
63

64
65
66
logger = get_logger(__name__)


67
68
69
70
71
72
73
74
75
def save_model_card(
    repo_id: str,
    images=None,
    base_model=str,
    train_text_encoder=False,
    prompt=str,
    repo_folder=None,
    pipeline: DiffusionPipeline = None,
):
76
77
78
79
80
81
82
83
84
85
86
    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:
87
88
- {'stable-diffusion' if isinstance(pipeline, StableDiffusionPipeline) else 'if'}
- {'stable-diffusion-diffusers' if isinstance(pipeline, StableDiffusionPipeline) else 'if-diffusers'}
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
- 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)


108
109
110
def log_validation(
    text_encoder, tokenizer, unet, vae, args, accelerator, weight_dtype, epoch, prompt_embeds, negative_prompt_embeds
):
111
112
113
114
    logger.info(
        f"Running validation... \n Generating {args.num_validation_images} images with prompt:"
        f" {args.validation_prompt}."
    )
115
116
117
118
119
120

    pipeline_args = {}

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

121
122
123
    if text_encoder is not None:
        text_encoder = accelerator.unwrap_model(text_encoder)

124
125
126
127
    # create pipeline (note: unet and vae are loaded again in float32)
    pipeline = DiffusionPipeline.from_pretrained(
        args.pretrained_model_name_or_path,
        tokenizer=tokenizer,
128
        text_encoder=text_encoder,
129
130
131
        unet=accelerator.unwrap_model(unet),
        revision=args.revision,
        torch_dtype=weight_dtype,
132
        **pipeline_args,
133
    )
134
135
136
137
138
139
140
141
142
143
144
145
146

    # 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 = DPMSolverMultistepScheduler.from_config(pipeline.scheduler.config, **scheduler_args)
147
148
149
    pipeline = pipeline.to(accelerator.device)
    pipeline.set_progress_bar_config(disable=True)

150
151
152
153
154
155
156
157
    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}

158
159
160
    # run inference
    generator = None if args.seed is None else torch.Generator(device=accelerator.device).manual_seed(args.seed)
    images = []
161
162
163
164
165
166
167
168
169
170
    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)
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187

    for tracker in accelerator.trackers:
        if tracker.name == "tensorboard":
            np_images = np.stack([np.asarray(img) for img in images])
            tracker.writer.add_images("validation", np_images, epoch, dataformats="NHWC")
        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()

188
189
    return images

190

191
def import_model_class_from_model_name_or_path(pretrained_model_name_or_path: str, revision: str):
192
193
194
    text_encoder_config = PretrainedConfig.from_pretrained(
        pretrained_model_name_or_path,
        subfolder="text_encoder",
195
        revision=revision,
196
197
198
199
200
201
202
203
204
205
206
    )
    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
207
208
209
210
    elif model_class == "T5EncoderModel":
        from transformers import T5EncoderModel

        return T5EncoderModel
211
212
213
214
    else:
        raise ValueError(f"{model_class} is not supported.")


Suraj Patil's avatar
Suraj Patil committed
215
def parse_args(input_args=None):
216
217
218
219
220
221
222
223
    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.",
    )
224
225
226
227
228
    parser.add_argument(
        "--revision",
        type=str,
        default=None,
        required=False,
229
230
231
232
        help=(
            "Revision of pretrained model identifier from huggingface.co/models. Trainable model components should be"
            " float32 precision."
        ),
233
    )
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
    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,
258
        required=True,
Yuta Hayashibe's avatar
Yuta Hayashibe committed
259
        help="The prompt with identifier specifying the instance",
260
261
262
263
264
    )
    parser.add_argument(
        "--class_prompt",
        type=str,
        default=None,
Yuta Hayashibe's avatar
Yuta Hayashibe committed
265
        help="The prompt to specify images in the same class as provided instance images.",
266
267
268
269
270
    )
    parser.add_argument(
        "--with_prior_preservation",
        default=False,
        action="store_true",
Yuta Hayashibe's avatar
Yuta Hayashibe committed
271
        help="Flag to add prior preservation loss.",
272
273
274
275
276
277
278
    )
    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=(
279
280
            "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."
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
        ),
    )
    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(
300
301
302
        "--center_crop",
        default=False,
        action="store_true",
patil-suraj's avatar
patil-suraj committed
303
304
305
306
        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."
        ),
307
    )
308
309
310
311
312
    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.",
    )
313
314
315
316
317
318
319
320
321
322
323
324
325
    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.",
    )
326
327
328
329
330
    parser.add_argument(
        "--checkpointing_steps",
        type=int,
        default=500,
        help=(
331
332
333
334
335
            "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."
336
337
        ),
    )
338
    parser.add_argument(
339
        "--checkpoints_total_limit",
340
341
342
343
344
345
346
347
        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"
        ),
    )
348
349
350
351
352
353
354
355
356
    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.'
        ),
    )
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
    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."
    )
392
393
394
395
396
397
398
    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.")
399
400
401
    parser.add_argument(
        "--use_8bit_adam", action="store_true", help="Whether or not to use 8-bit Adam from bitsandbytes."
    )
402
403
404
405
406
407
408
409
    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."
        ),
    )
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
    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
432
433
434
435
436
437
438
439
440
441
442
443
444
    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=(
445
446
            '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
447
448
        ),
    )
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
    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."
        ),
    )
471
472
473
    parser.add_argument(
        "--mixed_precision",
        type=str,
474
        default=None,
475
476
        choices=["no", "fp16", "bf16"],
        help=(
477
478
479
            "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."
480
481
        ),
    )
482
483
484
485
486
487
488
489
490
491
    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."
        ),
    )
492
    parser.add_argument("--local_rank", type=int, default=-1, help="For distributed training: local_rank")
493
494
495
    parser.add_argument(
        "--enable_xformers_memory_efficient_attention", action="store_true", help="Whether or not to use xformers."
    )
496
497
498
499
500
501
502
503
504
    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"
        ),
    )
505

506
507
508
509
510
511
512
513
514
    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."
        ),
    )
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
    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"
    )
536
537
538
539
540
541
542
543
544
545
546
547
548
    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`.",
    )
549

550
551
552
553
554
    if input_args is not None:
        args = parser.parse_args(input_args)
    else:
        args = parser.parse_args()

555
556
557
558
559
560
561
562
563
    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.")
564
    else:
565
        # logger is not available yet
566
        if args.class_data_dir is not None:
567
            warnings.warn("You need not use --class_data_dir without --with_prior_preservation.")
568
        if args.class_prompt is not None:
569
            warnings.warn("You need not use --class_prompt without --with_prior_preservation.")
570

571
572
573
    if args.train_text_encoder and args.pre_compute_text_embeddings:
        raise ValueError("`--train_text_encoder` cannot be used with `--pre_compute_text_embeddings`")

574
575
576
577
578
    return args


class DreamBoothDataset(Dataset):
    """
Yuta Hayashibe's avatar
Yuta Hayashibe committed
579
    A dataset to prepare the instance and class images with the prompts for fine-tuning the model.
580
581
582
583
584
585
586
587
588
589
    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,
590
        class_num=None,
591
592
        size=512,
        center_crop=False,
593
594
595
        encoder_hidden_states=None,
        instance_prompt_encoder_hidden_states=None,
        tokenizer_max_length=None,
596
597
598
599
    ):
        self.size = size
        self.center_crop = center_crop
        self.tokenizer = tokenizer
600
601
602
        self.encoder_hidden_states = encoder_hidden_states
        self.instance_prompt_encoder_hidden_states = instance_prompt_encoder_hidden_states
        self.tokenizer_max_length = tokenizer_max_length
603
604
605

        self.instance_data_root = Path(instance_data_root)
        if not self.instance_data_root.exists():
606
            raise ValueError(f"Instance {self.instance_data_root} images root doesn't exists.")
607
608
609
610
611
612
613
614
615

        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)
616
            self.class_images_path = list(self.class_data_root.iterdir())
617
618
619
620
            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)
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
            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])
641
642
        instance_image = exif_transpose(instance_image)

643
644
645
        if not instance_image.mode == "RGB":
            instance_image = instance_image.convert("RGB")
        example["instance_images"] = self.image_transforms(instance_image)
646
647
648
649
650
651
652
653
654

        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
655
656
657

        if self.class_data_root:
            class_image = Image.open(self.class_images_path[index % self.num_class_images])
658
659
            class_image = exif_transpose(class_image)

660
661
662
            if not class_image.mode == "RGB":
                class_image = class_image.convert("RGB")
            example["class_images"] = self.image_transforms(class_image)
663
664
665
666
667
668
669
670
671

            if self.instance_prompt_encoder_hidden_states is not None:
                example["class_prompt_ids"] = self.instance_prompt_encoder_hidden_states
            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
672
673
674
675

        return example


676
def collate_fn(examples, with_prior_preservation=False):
677
678
    has_attention_mask = "instance_attention_mask" in examples[0]

679
680
681
    input_ids = [example["instance_prompt_ids"] for example in examples]
    pixel_values = [example["instance_images"] for example in examples]

682
683
684
    if has_attention_mask:
        attention_mask = [example["instance_attention_mask"] for example in examples]

685
686
687
688
689
690
    # 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]

691
692
693
        if has_attention_mask:
            attention_mask += [example["class_attention_mask"] for example in examples]

694
695
696
697
698
699
700
701
702
    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,
    }
703
704

    if has_attention_mask:
705
        attention_mask = torch.cat(attention_mask, dim=0)
706
707
        batch["attention_mask"] = attention_mask

708
709
710
    return batch


711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
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


728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
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


772
def main(args):
773
774
    logging_dir = Path(args.output_dir, args.logging_dir)

775
    accelerator_project_config = ProjectConfiguration(project_dir=args.output_dir, logging_dir=logging_dir)
776

777
778
779
    accelerator = Accelerator(
        gradient_accumulation_steps=args.gradient_accumulation_steps,
        mixed_precision=args.mixed_precision,
Suraj Patil's avatar
Suraj Patil committed
780
        log_with=args.report_to,
781
        project_config=accelerator_project_config,
782
783
    )

784
785
786
787
    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.")

788
789
790
791
792
793
794
795
796
    # 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
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
    # 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.
812
813
814
    if args.seed is not None:
        set_seed(args.seed)

Suraj Patil's avatar
Suraj Patil committed
815
    # Generate class images if prior preservation is enabled.
816
817
818
819
820
821
822
823
    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
824
825
826
827
828
829
            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
830
            pipeline = DiffusionPipeline.from_pretrained(
831
832
833
834
                args.pretrained_model_name_or_path,
                torch_dtype=torch_dtype,
                safety_checker=None,
                revision=args.revision,
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
            )
            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
            ):
850
                images = pipeline(example["prompt"]).images
851
852

                for i, image in enumerate(images):
853
854
855
                    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)
856
857
858
859
860
861
862

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

    # Handle the repository creation
    if accelerator.is_main_process:
863
        if args.output_dir is not None:
864
865
            os.makedirs(args.output_dir, exist_ok=True)

866
867
868
869
870
        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

871
872
    # Load the tokenizer
    if args.tokenizer_name:
Suraj Patil's avatar
Suraj Patil committed
873
        tokenizer = AutoTokenizer.from_pretrained(args.tokenizer_name, revision=args.revision, use_fast=False)
874
    elif args.pretrained_model_name_or_path:
875
        tokenizer = AutoTokenizer.from_pretrained(
876
877
878
            args.pretrained_model_name_or_path,
            subfolder="tokenizer",
            revision=args.revision,
879
            use_fast=False,
880
        )
881

882
    # import correct text encoder class
883
    text_encoder_cls = import_model_class_from_model_name_or_path(args.pretrained_model_name_or_path, args.revision)
884

Suraj Patil's avatar
Suraj Patil committed
885
886
    # Load scheduler and models
    noise_scheduler = DDPMScheduler.from_pretrained(args.pretrained_model_name_or_path, subfolder="scheduler")
887
    text_encoder = text_encoder_cls.from_pretrained(
Suraj Patil's avatar
Suraj Patil committed
888
        args.pretrained_model_name_or_path, subfolder="text_encoder", revision=args.revision
889
    )
890
891
892
893
894
895
896
897

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

898
    unet = UNet2DConditionModel.from_pretrained(
Suraj Patil's avatar
Suraj Patil committed
899
        args.pretrained_model_name_or_path, subfolder="unet", revision=args.revision
900
    )
901

902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
    # create custom saving & loading hooks so that `accelerator.save_state(...)` serializes in a nice format
    def save_model_hook(models, weights, output_dir):
        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))

            # make sure to pop weight so that corresponding model is not saved again
            weights.pop()

    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)
930

931
932
933
    if vae is not None:
        vae.requires_grad_(False)

Suraj Patil's avatar
Suraj Patil committed
934
935
936
    if not args.train_text_encoder:
        text_encoder.requires_grad_(False)

937
938
    if args.enable_xformers_memory_efficient_attention:
        if is_xformers_available():
939
940
941
942
943
944
945
            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."
                )
946
            unet.enable_xformers_memory_efficient_attention()
947
948
        else:
            raise ValueError("xformers is not available. Make sure it is installed correctly")
949

950
951
    if args.gradient_checkpointing:
        unet.enable_gradient_checkpointing()
952
953
        if args.train_text_encoder:
            text_encoder.gradient_checkpointing_enable()
954

955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
    # 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
972
973
974
975
976
    # 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

977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
    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
995
    # Optimizer creation
996
997
998
    params_to_optimize = (
        itertools.chain(unet.parameters(), text_encoder.parameters()) if args.train_text_encoder else unet.parameters()
    )
999
    optimizer = optimizer_class(
1000
        params_to_optimize,
1001
1002
1003
1004
1005
1006
        lr=args.learning_rate,
        betas=(args.adam_beta1, args.adam_beta2),
        weight_decay=args.adam_weight_decay,
        eps=args.adam_epsilon,
    )

1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
    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

        if args.instance_prompt is not None:
            pre_computed_instance_prompt_encoder_hidden_states = compute_text_embeddings(args.instance_prompt)
        else:
            pre_computed_instance_prompt_encoder_hidden_states = None

        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
        pre_computed_instance_prompt_encoder_hidden_states = None

Suraj Patil's avatar
Suraj Patil committed
1045
    # Dataset and DataLoaders creation:
1046
1047
1048
1049
1050
    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,
1051
        class_num=args.num_class_images,
1052
1053
1054
        tokenizer=tokenizer,
        size=args.resolution,
        center_crop=args.center_crop,
1055
1056
1057
        encoder_hidden_states=pre_computed_encoder_hidden_states,
        instance_prompt_encoder_hidden_states=pre_computed_instance_prompt_encoder_hidden_states,
        tokenizer_max_length=args.tokenizer_max_length,
1058
1059
1060
    )

    train_dataloader = torch.utils.data.DataLoader(
1061
1062
1063
1064
        train_dataset,
        batch_size=args.train_batch_size,
        shuffle=True,
        collate_fn=lambda examples: collate_fn(examples, args.with_prior_preservation),
1065
        num_workers=args.dataloader_num_workers,
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
    )

    # 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,
        num_warmup_steps=args.lr_warmup_steps * args.gradient_accumulation_steps,
        num_training_steps=args.max_train_steps * args.gradient_accumulation_steps,
1080
1081
        num_cycles=args.lr_num_cycles,
        power=args.lr_power,
1082
1083
    )

Suraj Patil's avatar
Suraj Patil committed
1084
    # Prepare everything with our `accelerator`.
1085
1086
1087
1088
1089
1090
1091
1092
    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
        )
1093

1094
1095
    # 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.
1096
    weight_dtype = torch.float32
1097
    if accelerator.mixed_precision == "fp16":
1098
        weight_dtype = torch.float16
1099
    elif accelerator.mixed_precision == "bf16":
1100
1101
        weight_dtype = torch.bfloat16

Suraj Patil's avatar
Suraj Patil committed
1102
    # Move vae and text_encoder to device and cast to weight_dtype
1103
1104
1105
1106
    if vae is not None:
        vae.to(accelerator.device, dtype=weight_dtype)

    if not args.train_text_encoder and text_encoder is not None:
1107
        text_encoder.to(accelerator.device, dtype=weight_dtype)
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131

    # 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:
        accelerator.init_trackers("dreambooth", config=vars(args))

    # 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}")
1132
1133
1134
    global_step = 0
    first_epoch = 0

Suraj Patil's avatar
Suraj Patil committed
1135
    # Potentially load in the weights and states from a previous save
1136
1137
1138
1139
1140
1141
1142
1143
    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]))
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
            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)
1159

1160
    # Only show the progress bar once on each machine.
1161
    progress_bar = tqdm(range(global_step, args.max_train_steps), disable=not accelerator.is_local_main_process)
1162
1163
    progress_bar.set_description("Steps")

1164
    for epoch in range(first_epoch, args.num_train_epochs):
1165
        unet.train()
1166
1167
        if args.train_text_encoder:
            text_encoder.train()
1168
        for step, batch in enumerate(train_dataloader):
1169
1170
1171
1172
1173
1174
            # 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

1175
            with accelerator.accumulate(unet):
1176
                pixel_values = batch["pixel_values"].to(dtype=weight_dtype)
1177

1178
1179
1180
1181
1182
1183
1184
1185
                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
1186
                if args.offset_noise:
1187
1188
                    noise = torch.randn_like(model_input) + 0.1 * torch.randn(
                        model_input.shape[0], model_input.shape[1], 1, 1, device=model_input.device
1189
1190
                    )
                else:
1191
                    noise = torch.randn_like(model_input)
1192
                bsz, channels, height, width = model_input.shape
1193
                # Sample a random timestep for each image
1194
1195
1196
                timesteps = torch.randint(
                    0, noise_scheduler.config.num_train_timesteps, (bsz,), device=model_input.device
                )
1197
1198
                timesteps = timesteps.long()

1199
                # Add noise to the model input according to the noise magnitude at each timestep
1200
                # (this is the forward diffusion process)
1201
                noisy_model_input = noise_scheduler.add_noise(model_input, noise, timesteps)
1202
1203

                # Get the text embedding for conditioning
1204
1205
1206
1207
1208
1209
1210
1211
1212
                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,
                    )
1213

1214
                if accelerator.unwrap_model(unet).config.in_channels == channels * 2:
1215
                    noisy_model_input = torch.cat([noisy_model_input, noisy_model_input], dim=1)
1216
1217
1218
1219
1220
1221

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

1222
                # Predict the noise residual
1223
1224
1225
                model_pred = unet(
                    noisy_model_input, timesteps, encoder_hidden_states, class_labels=class_labels
                ).sample
1226
1227
1228

                if model_pred.shape[1] == 6:
                    model_pred, _ = torch.chunk(model_pred, 2, dim=1)
1229
1230
1231
1232
1233

                # 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":
1234
                    target = noise_scheduler.get_velocity(model_input, noise, timesteps)
1235
1236
                else:
                    raise ValueError(f"Unknown prediction type {noise_scheduler.config.prediction_type}")
1237
1238

                if args.with_prior_preservation:
1239
1240
1241
                    # 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)
1242
1243

                    # Compute instance loss
1244
                    loss = F.mse_loss(model_pred.float(), target.float(), reduction="mean")
1245
1246

                    # Compute prior loss
1247
                    prior_loss = F.mse_loss(model_pred_prior.float(), target_prior.float(), reduction="mean")
1248
1249
1250
1251

                    # Add the prior loss to the instance loss.
                    loss = loss + args.prior_loss_weight * prior_loss
                else:
1252
                    loss = F.mse_loss(model_pred.float(), target.float(), reduction="mean")
1253
1254

                accelerator.backward(loss)
1255
                if accelerator.sync_gradients:
1256
1257
1258
1259
1260
1261
                    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)
1262
1263
                optimizer.step()
                lr_scheduler.step()
1264
                optimizer.zero_grad(set_to_none=args.set_grads_to_none)
1265
1266
1267
1268
1269
1270

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

1271
1272
                if accelerator.is_main_process:
                    if global_step % args.checkpointing_steps == 0:
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
                        # _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)

1293
                        save_path = os.path.join(args.output_dir, f"checkpoint-{global_step}")
1294
1295
                        accelerator.save_state(save_path)
                        logger.info(f"Saved state to {save_path}")
1296

1297
1298
                    images = []

1299
                    if args.validation_prompt is not None and global_step % args.validation_steps == 0:
1300
                        images = log_validation(
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
                            text_encoder,
                            tokenizer,
                            unet,
                            vae,
                            args,
                            accelerator,
                            weight_dtype,
                            epoch,
                            validation_prompt_encoder_hidden_states,
                            validation_prompt_negative_prompt_embeds,
1311
                        )
1312

1313
1314
1315
1316
1317
1318
1319
1320
            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
1321
    accelerator.wait_for_everyone()
1322
    if accelerator.is_main_process:
1323
1324
1325
1326
1327
1328
1329
1330
        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

1331
        pipeline = DiffusionPipeline.from_pretrained(
1332
1333
            args.pretrained_model_name_or_path,
            unet=accelerator.unwrap_model(unet),
1334
            revision=args.revision,
1335
            **pipeline_args,
1336
        )
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350

        # 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)

1351
1352
1353
        pipeline.save_pretrained(args.output_dir)

        if args.push_to_hub:
1354
1355
1356
1357
1358
1359
1360
            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,
1361
                pipeline=pipeline,
1362
            )
1363
1364
1365
1366
1367
1368
            upload_folder(
                repo_id=repo_id,
                folder_path=args.output_dir,
                commit_message="End of training",
                ignore_patterns=["step_*", "epoch_*"],
            )
1369
1370
1371
1372
1373

    accelerator.end_training()


if __name__ == "__main__":
1374
1375
    args = parse_args()
    main(args)