train_dreambooth_lora.py 57.6 KB
Newer Older
1
2
#!/usr/bin/env python
# coding=utf-8
3
# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
4
5
6
7
8
9
10
11
12
13
14
15
16
#
# 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

import argparse
17
import copy
Will Berman's avatar
Will Berman committed
18
import gc
19
20
21
import logging
import math
import os
22
import shutil
23
24
25
import warnings
from pathlib import Path

26
import numpy as np
27
28
29
30
31
32
import torch
import torch.nn.functional as F
import torch.utils.checkpoint
import transformers
from accelerate import Accelerator
from accelerate.logging import get_logger
33
from accelerate.utils import ProjectConfiguration, set_seed
Patrick von Platen's avatar
Patrick von Platen committed
34
from huggingface_hub import create_repo, upload_folder
35
from huggingface_hub.utils import insecure_hashlib
Patrick von Platen's avatar
Patrick von Platen committed
36
from packaging import version
37
from peft import LoraConfig
38
from peft.utils import get_peft_model_state_dict, set_peft_model_state_dict
Patrick von Platen's avatar
Patrick von Platen committed
39
40
41
42
43
44
45
46
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
51
from diffusers import (
    AutoencoderKL,
    DDPMScheduler,
    DiffusionPipeline,
    DPMSolverMultistepScheduler,
52
    StableDiffusionPipeline,
53
54
    UNet2DConditionModel,
)
55
from diffusers.loaders import StableDiffusionLoraLoaderMixin
56
from diffusers.optimization import get_scheduler
57
58
59
60
61
from diffusers.training_utils import (
    _set_state_dict_into_text_encoder,
    cast_training_params,
    free_memory,
)
62
63
64
65
66
67
from diffusers.utils import (
    check_min_version,
    convert_state_dict_to_diffusers,
    convert_unet_state_dict_to_peft,
    is_wandb_available,
)
68
from diffusers.utils.hub_utils import load_or_create_model_card, populate_model_card
69
from diffusers.utils.import_utils import is_xformers_available
70
from diffusers.utils.torch_utils import is_compiled_module
71
72


73
74
75
if is_wandb_available():
    import wandb

76
# Will error if the minimal version of diffusers is not installed. Remove at your own risks.
Sayak Paul's avatar
Sayak Paul committed
77
check_min_version("0.34.0.dev0")
78
79
80
81

logger = get_logger(__name__)


82
83
84
85
86
87
88
89
90
def save_model_card(
    repo_id: str,
    images=None,
    base_model=str,
    train_text_encoder=False,
    prompt=str,
    repo_folder=None,
    pipeline: DiffusionPipeline = None,
):
Patrick von Platen's avatar
Patrick von Platen committed
91
92
93
94
95
    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"

96
    model_description = f"""
97
# LoRA DreamBooth - {repo_id}
Patrick von Platen's avatar
Patrick von Platen committed
98

hysts's avatar
hysts committed
99
These are LoRA adaption weights for {base_model}. The weights were trained on {prompt} using [DreamBooth](https://dreambooth.github.io/). You can find some example images in the following. \n
Patrick von Platen's avatar
Patrick von Platen committed
100
{img_str}
101
102

LoRA for the text encoder was enabled: {train_text_encoder}.
Patrick von Platen's avatar
Patrick von Platen committed
103
"""
104
105
106
107
108
    model_card = load_or_create_model_card(
        repo_id_or_path=repo_id,
        from_training=True,
        license="creativeml-openrail-m",
        base_model=base_model,
109
        prompt=prompt,
110
111
112
        model_description=model_description,
        inference=True,
    )
113
    tags = ["text-to-image", "diffusers", "lora", "diffusers-training"]
114
115
116
117
118
119
120
    if isinstance(pipeline, StableDiffusionPipeline):
        tags.extend(["stable-diffusion", "stable-diffusion-diffusers"])
    else:
        tags.extend(["if", "if-diffusers"])
    model_card = populate_model_card(model_card, tags=tags)

    model_card.save(os.path.join(repo_folder, "README.md"))
Patrick von Platen's avatar
Patrick von Platen committed
121
122


123
124
125
126
127
128
def log_validation(
    pipeline,
    args,
    accelerator,
    pipeline_args,
    epoch,
129
    torch_dtype,
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
    is_final_validation=False,
):
    logger.info(
        f"Running validation... \n Generating {args.num_validation_images} images with prompt:"
        f" {args.validation_prompt}."
    )
    # 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)

149
    pipeline = pipeline.to(accelerator.device, dtype=torch_dtype)
150
151
152
    pipeline.set_progress_bar_config(disable=True)

    # run inference
153
    generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed is not None else None
154
155
156
157

    if args.validation_images is None:
        images = []
        for _ in range(args.num_validation_images):
158
            with torch.amp.autocast(accelerator.device.type):
159
160
161
162
163
164
                image = pipeline(**pipeline_args, generator=generator).images[0]
                images.append(image)
    else:
        images = []
        for image in args.validation_images:
            image = Image.open(image)
165
            with torch.amp.autocast(accelerator.device.type):
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
                image = pipeline(**pipeline_args, image=image, generator=generator).images[0]
            images.append(image)

    for tracker in accelerator.trackers:
        phase_name = "test" if is_final_validation else "validation"
        if tracker.name == "tensorboard":
            np_images = np.stack([np.asarray(img) for img in images])
            tracker.writer.add_images(phase_name, np_images, epoch, dataformats="NHWC")
        if tracker.name == "wandb":
            tracker.log(
                {
                    phase_name: [
                        wandb.Image(image, caption=f"{i}: {args.validation_prompt}") for i, image in enumerate(images)
                    ]
                }
            )

    del pipeline
184
    free_memory()
185
186
187
188

    return images


189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
def import_model_class_from_model_name_or_path(pretrained_model_name_or_path: str, revision: str):
    text_encoder_config = PretrainedConfig.from_pretrained(
        pretrained_model_name_or_path,
        subfolder="text_encoder",
        revision=revision,
    )
    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
Will Berman's avatar
Will Berman committed
205
206
207
208
    elif model_class == "T5EncoderModel":
        from transformers import T5EncoderModel

        return T5EncoderModel
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
    else:
        raise ValueError(f"{model_class} is not supported.")


def parse_args(input_args=None):
    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.",
    )
    parser.add_argument(
        "--revision",
        type=str,
        default=None,
        required=False,
        help="Revision of pretrained model identifier from huggingface.co/models.",
    )
229
230
231
232
233
234
    parser.add_argument(
        "--variant",
        type=str,
        default=None,
        help="Variant of the model files of the pretrained model identifier from huggingface.co/models, 'e.g.' fp16",
    )
235
236
237
238
239
240
241
242
243
244
245
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
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
    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,
        required=True,
        help="The prompt with identifier specifying the instance",
    )
    parser.add_argument(
        "--class_prompt",
        type=str,
        default=None,
        help="The prompt to specify images in the same class as provided instance images.",
    )
    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_epochs",
        type=int,
        default=50,
        help=(
            "Run dreambooth validation every X epochs. Dreambooth validation consists of running the prompt"
            " `args.validation_prompt` multiple times: `args.num_validation_images`."
        ),
    )
    parser.add_argument(
        "--with_prior_preservation",
        default=False,
        action="store_true",
        help="Flag to add prior preservation loss.",
    )
    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=(
            "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."
        ),
    )
    parser.add_argument(
        "--output_dir",
        type=str,
        default="lora-dreambooth-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(
patil-suraj's avatar
patil-suraj committed
322
323
324
325
326
327
328
        "--center_crop",
        default=False,
        action="store_true",
        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."
        ),
329
    )
330
331
332
333
334
    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.",
    )
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
    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.",
    )
    parser.add_argument(
        "--checkpointing_steps",
        type=int,
        default=500,
        help=(
            "Save a checkpoint of the training state every X updates. These checkpoints can be used both as final"
            " checkpoints in case they are better than the last checkpoint, and are also suitable for resuming"
            " training using `--resume_from_checkpoint`."
        ),
    )
358
    parser.add_argument(
359
        "--checkpoints_total_limit",
360
361
        type=int,
        default=None,
362
        help=("Max number of checkpoints to store."),
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
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
    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.'
        ),
    )
    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-4,
        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."
    )
    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.")
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
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
    parser.add_argument(
        "--use_8bit_adam", action="store_true", help="Whether or not to use 8-bit Adam from bitsandbytes."
    )
    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***."
        ),
    )
    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=(
            '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.'
        ),
    )
    parser.add_argument(
        "--mixed_precision",
        type=str,
        default=None,
        choices=["no", "fp16", "bf16"],
        help=(
            "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."
        ),
    )
    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."
        ),
    )
    parser.add_argument("--local_rank", type=int, default=-1, help="For distributed training: local_rank")
    parser.add_argument(
        "--enable_xformers_memory_efficient_attention", action="store_true", help="Whether or not to use xformers."
    )
Will Berman's avatar
Will Berman committed
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
    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",
    )
508
509
510
511
512
513
514
515
516
517
518
519
520
    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`.",
    )
521
522
523
524
525
526
    parser.add_argument(
        "--rank",
        type=int,
        default=4,
        help=("The dimension of the LoRA update matrices."),
    )
527
528
529
530
531
532
533
534
535
    parser.add_argument(
        "--image_interpolation_mode",
        type=str,
        default="lanczos",
        choices=[
            f.lower() for f in dir(transforms.InterpolationMode) if not f.startswith("__") and not f.endswith("__")
        ],
        help="The image interpolation method to use for resizing images.",
    )
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557

    if input_args is not None:
        args = parser.parse_args(input_args)
    else:
        args = parser.parse_args()

    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.")
    else:
        # logger is not available yet
        if args.class_data_dir is not None:
            warnings.warn("You need not use --class_data_dir without --with_prior_preservation.")
        if args.class_prompt is not None:
            warnings.warn("You need not use --class_prompt without --with_prior_preservation.")

Will Berman's avatar
Will Berman committed
558
559
560
    if args.train_text_encoder and args.pre_compute_text_embeddings:
        raise ValueError("`--train_text_encoder` cannot be used with `--pre_compute_text_embeddings`")

561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
    return args


class DreamBoothDataset(Dataset):
    """
    A dataset to prepare the instance and class images with the prompts for fine-tuning the model.
    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,
577
        class_num=None,
578
579
        size=512,
        center_crop=False,
Will Berman's avatar
Will Berman committed
580
        encoder_hidden_states=None,
581
        class_prompt_encoder_hidden_states=None,
Will Berman's avatar
Will Berman committed
582
        tokenizer_max_length=None,
583
584
585
586
    ):
        self.size = size
        self.center_crop = center_crop
        self.tokenizer = tokenizer
Will Berman's avatar
Will Berman committed
587
        self.encoder_hidden_states = encoder_hidden_states
588
        self.class_prompt_encoder_hidden_states = class_prompt_encoder_hidden_states
Will Berman's avatar
Will Berman committed
589
        self.tokenizer_max_length = tokenizer_max_length
590
591
592
593
594
595
596
597
598
599
600
601
602
603

        self.instance_data_root = Path(instance_data_root)
        if not self.instance_data_root.exists():
            raise ValueError("Instance images root doesn't exists.")

        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)
            self.class_images_path = list(self.class_data_root.iterdir())
604
605
606
607
            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)
608
609
610
611
612
            self._length = max(self.num_class_images, self.num_instance_images)
            self.class_prompt = class_prompt
        else:
            self.class_data_root = None

613
614
615
616
        interpolation = getattr(transforms.InterpolationMode, args.image_interpolation_mode.upper(), None)
        if interpolation is None:
            raise ValueError(f"Unsupported interpolation mode {interpolation=}.")

617
618
        self.image_transforms = transforms.Compose(
            [
619
                transforms.Resize(size, interpolation=interpolation),
620
621
622
623
624
625
626
627
628
629
630
631
                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])
632
633
        instance_image = exif_transpose(instance_image)

634
635
636
        if not instance_image.mode == "RGB":
            instance_image = instance_image.convert("RGB")
        example["instance_images"] = self.image_transforms(instance_image)
Will Berman's avatar
Will Berman committed
637
638
639
640
641
642
643
644
645

        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
646
647
648

        if self.class_data_root:
            class_image = Image.open(self.class_images_path[index % self.num_class_images])
649
650
            class_image = exif_transpose(class_image)

651
652
653
            if not class_image.mode == "RGB":
                class_image = class_image.convert("RGB")
            example["class_images"] = self.image_transforms(class_image)
Will Berman's avatar
Will Berman committed
654

655
656
            if self.class_prompt_encoder_hidden_states is not None:
                example["class_prompt_ids"] = self.class_prompt_encoder_hidden_states
Will Berman's avatar
Will Berman committed
657
658
659
660
661
662
            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
663
664
665
666
667

        return example


def collate_fn(examples, with_prior_preservation=False):
Will Berman's avatar
Will Berman committed
668
669
    has_attention_mask = "instance_attention_mask" in examples[0]

670
671
672
    input_ids = [example["instance_prompt_ids"] for example in examples]
    pixel_values = [example["instance_images"] for example in examples]

Will Berman's avatar
Will Berman committed
673
674
675
    if has_attention_mask:
        attention_mask = [example["instance_attention_mask"] for example in examples]

676
677
678
679
680
    # 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]
Will Berman's avatar
Will Berman committed
681
682
        if has_attention_mask:
            attention_mask += [example["class_attention_mask"] for example in examples]
683
684
685
686
687
688
689
690
691
692

    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,
    }
Will Berman's avatar
Will Berman committed
693
694
695
696

    if has_attention_mask:
        batch["attention_mask"] = attention_mask

697
698
699
700
    return batch


class PromptDataset(Dataset):
701
    """A simple dataset to prepare the prompts to generate class images on multiple GPUs."""
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716

    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


Will Berman's avatar
Will Berman committed
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
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,
745
        return_dict=False,
Will Berman's avatar
Will Berman committed
746
747
748
749
750
751
    )
    prompt_embeds = prompt_embeds[0]

    return prompt_embeds


752
def main(args):
753
754
755
756
757
758
    if args.report_to == "wandb" and args.hub_token is not None:
        raise ValueError(
            "You cannot use both --report_to=wandb and --hub_token due to a security risk of exposing your token."
            " Please use `huggingface-cli login` to authenticate with the Hub."
        )

759
760
    logging_dir = Path(args.output_dir, args.logging_dir)

761
    accelerator_project_config = ProjectConfiguration(project_dir=args.output_dir, logging_dir=logging_dir)
762

763
764
765
766
    accelerator = Accelerator(
        gradient_accumulation_steps=args.gradient_accumulation_steps,
        mixed_precision=args.mixed_precision,
        log_with=args.report_to,
767
        project_config=accelerator_project_config,
768
769
    )

770
771
772
773
    # Disable AMP for MPS.
    if torch.backends.mps.is_available():
        accelerator.native_amp = False

774
775
776
777
778
779
    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.")

    # 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.
780
781
782
783
784
785
786
    # TODO (sayakpaul): 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."
        )

787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
    # 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.
    if args.seed is not None:
        set_seed(args.seed)

    # Generate class images if prior preservation is enabled.
    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:
813
            torch_dtype = torch.float16 if accelerator.device.type in ("cuda", "xpu") else torch.float32
814
815
816
817
818
819
820
821
822
823
824
            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
            pipeline = DiffusionPipeline.from_pretrained(
                args.pretrained_model_name_or_path,
                torch_dtype=torch_dtype,
                safety_checker=None,
                revision=args.revision,
825
                variant=args.variant,
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
            )
            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
            ):
                images = pipeline(example["prompt"]).images

                for i, image in enumerate(images):
844
                    hash_image = insecure_hashlib.sha1(image.tobytes()).hexdigest()
845
846
847
848
                    image_filename = class_images_dir / f"{example['index'][i] + cur_class_images}-{hash_image}.jpg"
                    image.save(image_filename)

            del pipeline
849
            free_memory()
850
851
852

    # Handle the repository creation
    if accelerator.is_main_process:
853
        if args.output_dir is not None:
854
855
            os.makedirs(args.output_dir, exist_ok=True)

856
857
858
859
860
        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

861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
    # Load the tokenizer
    if args.tokenizer_name:
        tokenizer = AutoTokenizer.from_pretrained(args.tokenizer_name, revision=args.revision, use_fast=False)
    elif args.pretrained_model_name_or_path:
        tokenizer = AutoTokenizer.from_pretrained(
            args.pretrained_model_name_or_path,
            subfolder="tokenizer",
            revision=args.revision,
            use_fast=False,
        )

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

    # Load scheduler and models
    noise_scheduler = DDPMScheduler.from_pretrained(args.pretrained_model_name_or_path, subfolder="scheduler")
    text_encoder = text_encoder_cls.from_pretrained(
878
        args.pretrained_model_name_or_path, subfolder="text_encoder", revision=args.revision, variant=args.variant
879
    )
880
    try:
Will Berman's avatar
Will Berman committed
881
        vae = AutoencoderKL.from_pretrained(
882
            args.pretrained_model_name_or_path, subfolder="vae", revision=args.revision, variant=args.variant
Will Berman's avatar
Will Berman committed
883
        )
884
885
886
    except OSError:
        # IF does not have a VAE so let's just set it to None
        # We don't have to error out here
Will Berman's avatar
Will Berman committed
887
888
        vae = None

889
    unet = UNet2DConditionModel.from_pretrained(
890
        args.pretrained_model_name_or_path, subfolder="unet", revision=args.revision, variant=args.variant
891
892
893
    )

    # We only train the additional adapter LoRA layers
Will Berman's avatar
Will Berman committed
894
895
    if vae is not None:
        vae.requires_grad_(False)
896
897
898
    text_encoder.requires_grad_(False)
    unet.requires_grad_(False)

899
    # For mixed precision training we cast all non-trainable weights (vae, non-lora text_encoder and non-lora unet) to half-precision
900
    # as these weights are only used for inference, keeping weights in full precision is not required.
901
902
903
904
905
906
907
908
    weight_dtype = torch.float32
    if accelerator.mixed_precision == "fp16":
        weight_dtype = torch.float16
    elif accelerator.mixed_precision == "bf16":
        weight_dtype = torch.bfloat16

    # Move unet, vae and text_encoder to device and cast to weight_dtype
    unet.to(accelerator.device, dtype=weight_dtype)
Will Berman's avatar
Will Berman committed
909
910
    if vae is not None:
        vae.to(accelerator.device, dtype=weight_dtype)
911
912
913
914
    text_encoder.to(accelerator.device, dtype=weight_dtype)

    if args.enable_xformers_memory_efficient_attention:
        if is_xformers_available():
915
916
917
918
            import xformers

            xformers_version = version.parse(xformers.__version__)
            if xformers_version == version.parse("0.0.16"):
919
                logger.warning(
920
921
                    "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."
                )
922
923
924
925
            unet.enable_xformers_memory_efficient_attention()
        else:
            raise ValueError("xformers is not available. Make sure it is installed correctly")

926
927
928
929
930
    if args.gradient_checkpointing:
        unet.enable_gradient_checkpointing()
        if args.train_text_encoder:
            text_encoder.gradient_checkpointing_enable()

931
    # now we will add new LoRA weights to the attention layers
932
933
    unet_lora_config = LoraConfig(
        r=args.rank,
934
        lora_alpha=args.rank,
935
936
937
938
        init_lora_weights="gaussian",
        target_modules=["to_k", "to_q", "to_v", "to_out.0", "add_k_proj", "add_v_proj"],
    )
    unet.add_adapter(unet_lora_config)
939

940
    # The text encoder comes from 🤗 transformers, we will also attach adapters to it.
941
    if args.train_text_encoder:
942
        text_lora_config = LoraConfig(
943
944
945
946
            r=args.rank,
            lora_alpha=args.rank,
            init_lora_weights="gaussian",
            target_modules=["q_proj", "k_proj", "v_proj", "out_proj"],
947
948
        )
        text_encoder.add_adapter(text_lora_config)
949

950
951
952
953
954
    def unwrap_model(model):
        model = accelerator.unwrap_model(model)
        model = model._orig_mod if is_compiled_module(model) else model
        return model

955
956
    # create custom saving & loading hooks so that `accelerator.save_state(...)` serializes in a nice format
    def save_model_hook(models, weights, output_dir):
957
958
959
960
961
962
963
        if accelerator.is_main_process:
            # there are only two options here. Either are just the unet attn processor layers
            # or there are the unet and text encoder atten layers
            unet_lora_layers_to_save = None
            text_encoder_lora_layers_to_save = None

            for model in models:
964
                if isinstance(model, type(unwrap_model(unet))):
965
                    unet_lora_layers_to_save = convert_state_dict_to_diffusers(get_peft_model_state_dict(model))
966
                elif isinstance(model, type(unwrap_model(text_encoder))):
967
968
969
                    text_encoder_lora_layers_to_save = convert_state_dict_to_diffusers(
                        get_peft_model_state_dict(model)
                    )
970
971
972
973
974
975
                else:
                    raise ValueError(f"unexpected save model: {model.__class__}")

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

976
            StableDiffusionLoraLoaderMixin.save_lora_weights(
977
978
979
980
                output_dir,
                unet_lora_layers=unet_lora_layers_to_save,
                text_encoder_lora_layers=text_encoder_lora_layers_to_save,
            )
981
982

    def load_model_hook(models, input_dir):
Will Berman's avatar
Will Berman committed
983
984
        unet_ = None
        text_encoder_ = None
985

Will Berman's avatar
Will Berman committed
986
987
        while len(models) > 0:
            model = models.pop()
988

989
            if isinstance(model, type(unwrap_model(unet))):
Will Berman's avatar
Will Berman committed
990
                unet_ = model
991
            elif isinstance(model, type(unwrap_model(text_encoder))):
Will Berman's avatar
Will Berman committed
992
993
994
995
                text_encoder_ = model
            else:
                raise ValueError(f"unexpected save model: {model.__class__}")

996
        lora_state_dict, network_alphas = StableDiffusionLoraLoaderMixin.lora_state_dict(input_dir)
997

998
        unet_state_dict = {f"{k.replace('unet.', '')}": v for k, v in lora_state_dict.items() if k.startswith("unet.")}
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
        unet_state_dict = convert_unet_state_dict_to_peft(unet_state_dict)
        incompatible_keys = set_peft_model_state_dict(unet_, unet_state_dict, adapter_name="default")

        if incompatible_keys is not None:
            # check only for unexpected keys
            unexpected_keys = getattr(incompatible_keys, "unexpected_keys", None)
            if unexpected_keys:
                logger.warning(
                    f"Loading adapter weights from state_dict led to unexpected keys not found in the model: "
                    f" {unexpected_keys}. "
                )

        if args.train_text_encoder:
            _set_state_dict_into_text_encoder(lora_state_dict, prefix="text_encoder.", text_encoder=text_encoder_)

        # Make sure the trainable params are in float32. This is again needed since the base models
        # are in `weight_dtype`. More details:
        # https://github.com/huggingface/diffusers/pull/6514#discussion_r1449796804
        if args.mixed_precision == "fp16":
            models = [unet_]
            if args.train_text_encoder:
                models.append(text_encoder_)

            # only upcast trainable parameters (LoRA) into fp32
            cast_training_params(models, dtype=torch.float32)
1024
1025
1026
1027

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

1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
    # 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

    if args.scale_lr:
        args.learning_rate = (
            args.learning_rate * args.gradient_accumulation_steps * args.train_batch_size * accelerator.num_processes
        )

1038
1039
1040
1041
1042
1043
1044
1045
1046
    # Make sure the trainable params are in float32.
    if args.mixed_precision == "fp16":
        models = [unet]
        if args.train_text_encoder:
            models.append(text_encoder)

        # only upcast trainable parameters (LoRA) into fp32
        cast_training_params(models, dtype=torch.float32)

1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
    # 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

    # Optimizer creation
1061
1062
1063
1064
    params_to_optimize = list(filter(lambda p: p.requires_grad, unet.parameters()))
    if args.train_text_encoder:
        params_to_optimize = params_to_optimize + list(filter(lambda p: p.requires_grad, text_encoder.parameters()))

1065
    optimizer = optimizer_class(
1066
        params_to_optimize,
1067
1068
1069
1070
1071
1072
        lr=args.learning_rate,
        betas=(args.adam_beta1, args.adam_beta2),
        weight_decay=args.adam_weight_decay,
        eps=args.adam_epsilon,
    )

Will Berman's avatar
Will Berman committed
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
    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

1095
        if args.class_prompt is not None:
1096
            pre_computed_class_prompt_encoder_hidden_states = compute_text_embeddings(args.class_prompt)
Will Berman's avatar
Will Berman committed
1097
        else:
1098
            pre_computed_class_prompt_encoder_hidden_states = None
Will Berman's avatar
Will Berman committed
1099
1100
1101
1102
1103

        text_encoder = None
        tokenizer = None

        gc.collect()
1104
        free_memory()
Will Berman's avatar
Will Berman committed
1105
1106
1107
1108
    else:
        pre_computed_encoder_hidden_states = None
        validation_prompt_encoder_hidden_states = None
        validation_prompt_negative_prompt_embeds = None
1109
        pre_computed_class_prompt_encoder_hidden_states = None
Will Berman's avatar
Will Berman committed
1110

1111
1112
1113
1114
1115
1116
    # Dataset and DataLoaders creation:
    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,
1117
        class_num=args.num_class_images,
1118
1119
1120
        tokenizer=tokenizer,
        size=args.resolution,
        center_crop=args.center_crop,
Will Berman's avatar
Will Berman committed
1121
        encoder_hidden_states=pre_computed_encoder_hidden_states,
1122
        class_prompt_encoder_hidden_states=pre_computed_class_prompt_encoder_hidden_states,
Will Berman's avatar
Will Berman committed
1123
        tokenizer_max_length=args.tokenizer_max_length,
1124
1125
1126
1127
1128
1129
1130
    )

    train_dataloader = torch.utils.data.DataLoader(
        train_dataset,
        batch_size=args.train_batch_size,
        shuffle=True,
        collate_fn=lambda examples: collate_fn(examples, args.with_prior_preservation),
1131
        num_workers=args.dataloader_num_workers,
1132
1133
1134
    )

    # Scheduler and math around the number of training steps.
1135
1136
    # Check the PR https://github.com/huggingface/diffusers/pull/8312 for detailed explanation.
    num_warmup_steps_for_scheduler = args.lr_warmup_steps * accelerator.num_processes
1137
    if args.max_train_steps is None:
1138
1139
1140
1141
1142
1143
1144
        len_train_dataloader_after_sharding = math.ceil(len(train_dataloader) / accelerator.num_processes)
        num_update_steps_per_epoch = math.ceil(len_train_dataloader_after_sharding / args.gradient_accumulation_steps)
        num_training_steps_for_scheduler = (
            args.num_train_epochs * accelerator.num_processes * num_update_steps_per_epoch
        )
    else:
        num_training_steps_for_scheduler = args.max_train_steps * accelerator.num_processes
1145
1146
1147
1148

    lr_scheduler = get_scheduler(
        args.lr_scheduler,
        optimizer=optimizer,
1149
1150
        num_warmup_steps=num_warmup_steps_for_scheduler,
        num_training_steps=num_training_steps_for_scheduler,
1151
1152
1153
1154
1155
        num_cycles=args.lr_num_cycles,
        power=args.lr_power,
    )

    # Prepare everything with our `accelerator`.
1156
    if args.train_text_encoder:
Will Berman's avatar
Will Berman committed
1157
1158
        unet, text_encoder, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
            unet, text_encoder, optimizer, train_dataloader, lr_scheduler
1159
1160
        )
    else:
Will Berman's avatar
Will Berman committed
1161
1162
        unet, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
            unet, optimizer, train_dataloader, lr_scheduler
1163
        )
1164
1165
1166

    # 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)
1167
    if args.max_train_steps is None:
1168
        args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
1169
1170
1171
1172
1173
1174
1175
        if num_training_steps_for_scheduler != args.max_train_steps:
            logger.warning(
                f"The length of the 'train_dataloader' after 'accelerator.prepare' ({len(train_dataloader)}) does not match "
                f"the expected length ({len_train_dataloader_after_sharding}) when the learning rate scheduler was created. "
                f"This inconsistency may result in the learning rate scheduler not functioning properly."
            )

1176
1177
1178
1179
1180
1181
    # 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:
1182
        tracker_config = vars(copy.deepcopy(args))
1183
1184
        tracker_config.pop("validation_images")
        accelerator.init_trackers("dreambooth-lora", config=tracker_config)
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208

    # 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}")
    global_step = 0
    first_epoch = 0

    # Potentially load in the weights and states from a previous save
    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]))
1209
1210
1211
1212
1213
1214
1215
            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
1216
            initial_global_step = 0
1217
1218
1219
1220
1221
        else:
            accelerator.print(f"Resuming from checkpoint {path}")
            accelerator.load_state(os.path.join(args.output_dir, path))
            global_step = int(path.split("-")[1])

1222
            initial_global_step = global_step
1223
            first_epoch = global_step // num_update_steps_per_epoch
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
    else:
        initial_global_step = 0

    progress_bar = tqdm(
        range(0, args.max_train_steps),
        initial=initial_global_step,
        desc="Steps",
        # Only show the progress bar once on each machine.
        disable=not accelerator.is_local_main_process,
    )
1234
1235
1236

    for epoch in range(first_epoch, args.num_train_epochs):
        unet.train()
1237
1238
        if args.train_text_encoder:
            text_encoder.train()
1239
1240
        for step, batch in enumerate(train_dataloader):
            with accelerator.accumulate(unet):
Will Berman's avatar
Will Berman committed
1241
1242
1243
1244
1245
1246
1247
1248
                pixel_values = batch["pixel_values"].to(dtype=weight_dtype)

                if vae is not None:
                    # Convert images to latent space
                    model_input = vae.encode(pixel_values).latent_dist.sample()
                    model_input = model_input * vae.config.scaling_factor
                else:
                    model_input = pixel_values
1249
1250

                # Sample noise that we'll add to the latents
Will Berman's avatar
Will Berman committed
1251
                noise = torch.randn_like(model_input)
1252
                bsz, channels, height, width = model_input.shape
1253
                # Sample a random timestep for each image
Will Berman's avatar
Will Berman committed
1254
1255
1256
                timesteps = torch.randint(
                    0, noise_scheduler.config.num_train_timesteps, (bsz,), device=model_input.device
                )
1257
1258
                timesteps = timesteps.long()

Will Berman's avatar
Will Berman committed
1259
                # Add noise to the model input according to the noise magnitude at each timestep
1260
                # (this is the forward diffusion process)
Will Berman's avatar
Will Berman committed
1261
                noisy_model_input = noise_scheduler.add_noise(model_input, noise, timesteps)
1262
1263

                # Get the text embedding for conditioning
Will Berman's avatar
Will Berman committed
1264
1265
1266
1267
1268
1269
1270
1271
1272
                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,
                    )
1273

1274
                if unwrap_model(unet).config.in_channels == channels * 2:
1275
                    noisy_model_input = torch.cat([noisy_model_input, noisy_model_input], dim=1)
1276
1277
1278
1279
1280
1281

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

1282
                # Predict the noise residual
1283
                model_pred = unet(
1284
1285
1286
1287
1288
1289
                    noisy_model_input,
                    timesteps,
                    encoder_hidden_states,
                    class_labels=class_labels,
                    return_dict=False,
                )[0]
Will Berman's avatar
Will Berman committed
1290
1291
1292
1293
1294
1295

                # if model predicts variance, throw away the prediction. we will only train on the
                # simplified training objective. This means that all schedulers using the fine tuned
                # model must be configured to use one of the fixed variance variance types.
                if model_pred.shape[1] == 6:
                    model_pred, _ = torch.chunk(model_pred, 2, dim=1)
1296
1297
1298
1299
1300

                # 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":
Will Berman's avatar
Will Berman committed
1301
                    target = noise_scheduler.get_velocity(model_input, noise, timesteps)
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
                else:
                    raise ValueError(f"Unknown prediction type {noise_scheduler.config.prediction_type}")

                if args.with_prior_preservation:
                    # 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)

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

                    # Compute prior loss
                    prior_loss = F.mse_loss(model_pred_prior.float(), target_prior.float(), reduction="mean")

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

                accelerator.backward(loss)
                if accelerator.sync_gradients:
1323
                    accelerator.clip_grad_norm_(params_to_optimize, args.max_grad_norm)
1324
1325
1326
1327
1328
1329
1330
1331
1332
                optimizer.step()
                lr_scheduler.step()
                optimizer.zero_grad()

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

1333
1334
                if accelerator.is_main_process:
                    if global_step % args.checkpointing_steps == 0:
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
                        # _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)

1355
                        save_path = os.path.join(args.output_dir, f"checkpoint-{global_step}")
1356
                        accelerator.save_state(save_path)
1357
1358
1359
1360
1361
1362
1363
1364
1365
                        logger.info(f"Saved state to {save_path}")

            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

1366
1367
1368
1369
1370
        if accelerator.is_main_process:
            if args.validation_prompt is not None and epoch % args.validation_epochs == 0:
                # create pipeline
                pipeline = DiffusionPipeline.from_pretrained(
                    args.pretrained_model_name_or_path,
1371
1372
                    unet=unwrap_model(unet),
                    text_encoder=None if args.pre_compute_text_embeddings else unwrap_model(text_encoder),
1373
                    revision=args.revision,
1374
                    variant=args.variant,
1375
1376
                    torch_dtype=weight_dtype,
                )
Will Berman's avatar
Will Berman committed
1377
1378
1379
1380
1381
1382
1383
1384

                if args.pre_compute_text_embeddings:
                    pipeline_args = {
                        "prompt_embeds": validation_prompt_encoder_hidden_states,
                        "negative_prompt_embeds": validation_prompt_negative_prompt_embeds,
                    }
                else:
                    pipeline_args = {"prompt": args.validation_prompt}
1385

1386
1387
1388
1389
1390
1391
                images = log_validation(
                    pipeline,
                    args,
                    accelerator,
                    pipeline_args,
                    epoch,
1392
                    torch_dtype=weight_dtype,
1393
                )
1394
1395
1396
1397

    # Save the lora layers
    accelerator.wait_for_everyone()
    if accelerator.is_main_process:
1398
        unet = unwrap_model(unet)
1399
        unet = unet.to(torch.float32)
1400

1401
        unet_lora_state_dict = convert_state_dict_to_diffusers(get_peft_model_state_dict(unet))
1402
1403

        if args.train_text_encoder:
1404
            text_encoder = unwrap_model(text_encoder)
1405
            text_encoder_state_dict = convert_state_dict_to_diffusers(get_peft_model_state_dict(text_encoder))
Will Berman's avatar
Will Berman committed
1406
        else:
1407
            text_encoder_state_dict = None
1408

1409
        StableDiffusionLoraLoaderMixin.save_lora_weights(
1410
            save_directory=args.output_dir,
1411
1412
            unet_lora_layers=unet_lora_state_dict,
            text_encoder_lora_layers=text_encoder_state_dict,
1413
        )
1414

Patrick von Platen's avatar
Patrick von Platen committed
1415
1416
1417
        # Final inference
        # Load previous pipeline
        pipeline = DiffusionPipeline.from_pretrained(
1418
            args.pretrained_model_name_or_path, revision=args.revision, variant=args.variant, torch_dtype=weight_dtype
Patrick von Platen's avatar
Patrick von Platen committed
1419
        )
Will Berman's avatar
Will Berman committed
1420

Patrick von Platen's avatar
Patrick von Platen committed
1421
        # load attention processors
1422
        pipeline.load_lora_weights(args.output_dir, weight_name="pytorch_lora_weights.safetensors")
Patrick von Platen's avatar
Patrick von Platen committed
1423
1424

        # run inference
1425
        images = []
1426
        if args.validation_prompt and args.num_validation_images > 0:
1427
1428
1429
1430
1431
1432
1433
1434
            pipeline_args = {"prompt": args.validation_prompt, "num_inference_steps": 25}
            images = log_validation(
                pipeline,
                args,
                accelerator,
                pipeline_args,
                epoch,
                is_final_validation=True,
1435
                torch_dtype=weight_dtype,
1436
            )
1437

Patrick von Platen's avatar
Patrick von Platen committed
1438
1439
        if args.push_to_hub:
            save_model_card(
1440
                repo_id,
Patrick von Platen's avatar
Patrick von Platen committed
1441
1442
                images=images,
                base_model=args.pretrained_model_name_or_path,
1443
                train_text_encoder=args.train_text_encoder,
Patrick von Platen's avatar
Patrick von Platen committed
1444
1445
                prompt=args.instance_prompt,
                repo_folder=args.output_dir,
1446
                pipeline=pipeline,
1447
            )
1448
1449
1450
1451
1452
1453
            upload_folder(
                repo_id=repo_id,
                folder_path=args.output_dir,
                commit_message="End of training",
                ignore_patterns=["step_*", "epoch_*"],
            )
1454
1455
1456
1457
1458
1459
1460

    accelerator.end_training()


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