#!/usr/bin/env python

# coding=utf-8

# Copyright 2024 Custom Diffusion authors and the HuggingFace Inc. team. All rights reserved.

#

# 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

import itertools

import json

import logging

import math

import os

import random

import shutil

import warnings

from pathlib import Path

import numpy as np

import safetensors

import torch

import torch.nn.functional as F

import torch.utils.checkpoint

import transformers

from accelerate import Accelerator

from accelerate.logging import get_logger

from accelerate.utils import ProjectConfiguration, set_seed

from huggingface_hub import HfApi, create_repo

from huggingface_hub.utils import insecure_hashlib

from packaging import version

from PIL import Image

from torch.utils.data import Dataset

from torchvision import transforms

from tqdm.auto import tqdm

from transformers import AutoTokenizer, PretrainedConfig

import diffusers

from diffusers import (

AutoencoderKL,

DDPMScheduler,

DiffusionPipeline,

DPMSolverMultistepScheduler,

UNet2DConditionModel,

)

from diffusers.loaders import AttnProcsLayers

from diffusers.models.attention_processor import (

CustomDiffusionAttnProcessor,

CustomDiffusionAttnProcessor2_0,

CustomDiffusionXFormersAttnProcessor,

)

from diffusers.optimization import get_scheduler

from diffusers.utils import check_min_version, is_wandb_available

from diffusers.utils.hub_utils import load_or_create_model_card, populate_model_card

from diffusers.utils.import_utils import is_xformers_available

# Will error if the minimal version of diffusers is not installed. Remove at your own risks.

check_min_version("0.28.0.dev0")

logger = get_logger(__name__)

def freeze_params(params):

for param in params:

param.requires_grad = False

def save_model_card(repo_id: str, images=None, base_model=str, prompt=str, repo_folder=None):

img_str = ""

for i, image in enumerate(images):

image.save(os.path.join(repo_folder, f"image_{i}.png"))

img_str += f"\n"

model_description = f"""

# Custom Diffusion - {repo_id}

These are Custom Diffusion adaption weights for {base_model}. The weights were trained on {prompt} using [Custom Diffusion](https://www.cs.cmu.edu/~custom-diffusion). You can find some example images in the following. \n

{img_str}

\nFor more details on the training, please follow [this link](https://github.com/huggingface/diffusers/blob/main/examples/custom_diffusion).

"""

model_card = load_or_create_model_card(

repo_id_or_path=repo_id,

from_training=True,

license="creativeml-openrail-m",

base_model=base_model,

prompt=prompt,

model_description=model_description,

inference=True,

)

tags = [

"text-to-image",

"diffusers",

"stable-diffusion",

"stable-diffusion-diffusers",

"custom-diffusion",

"diffusers-training",

]

model_card = populate_model_card(model_card, tags=tags)

model_card.save(os.path.join(repo_folder, "README.md"))

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

else:

raise ValueError(f"{model_class} is not supported.")

def collate_fn(examples, with_prior_preservation):

input_ids = [example["instance_prompt_ids"] for example in examples]

pixel_values = [example["instance_images"] for example in examples]

mask = [example["mask"] for example in examples]

# 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]

mask += [example["class_mask"] for example in examples]

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

pixel_values = torch.stack(pixel_values)

mask = torch.stack(mask)

pixel_values = pixel_values.to(memory_format=torch.contiguous_format).float()

mask = mask.to(memory_format=torch.contiguous_format).float()

batch = {"input_ids": input_ids, "pixel_values": pixel_values, "mask": mask.unsqueeze(1)}

return batch

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

class CustomDiffusionDataset(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,

concepts_list,

tokenizer,

size=512,

mask_size=64,

center_crop=False,

with_prior_preservation=False,

num_class_images=200,

hflip=False,

aug=True,

):

self.size = size

self.mask_size = mask_size

self.center_crop = center_crop

self.tokenizer = tokenizer

self.interpolation = Image.BILINEAR

self.aug = aug

self.instance_images_path = []

self.class_images_path = []

self.with_prior_preservation = with_prior_preservation

for concept in concepts_list:

inst_img_path = [

(x, concept["instance_prompt"]) for x in Path(concept["instance_data_dir"]).iterdir() if x.is_file()

]

self.instance_images_path.extend(inst_img_path)

if with_prior_preservation:

class_data_root = Path(concept["class_data_dir"])

if os.path.isdir(class_data_root):

class_images_path = list(class_data_root.iterdir())

class_prompt = [concept["class_prompt"] for _ in range(len(class_images_path))]

else:

with open(class_data_root, "r") as f:

class_images_path = f.read().splitlines()

with open(concept["class_prompt"], "r") as f:

class_prompt = f.read().splitlines()

class_img_path = list(zip(class_images_path, class_prompt))

self.class_images_path.extend(class_img_path[:num_class_images])

random.shuffle(self.instance_images_path)

self.num_instance_images = len(self.instance_images_path)

self.num_class_images = len(self.class_images_path)

self._length = max(self.num_class_images, self.num_instance_images)

self.flip = transforms.RandomHorizontalFlip(0.5 * hflip)

self.image_transforms = transforms.Compose(

[

self.flip,

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 preprocess(self, image, scale, resample):

outer, inner = self.size, scale

factor = self.size // self.mask_size

if scale > self.size:

outer, inner = scale, self.size

top, left = np.random.randint(0, outer - inner + 1), np.random.randint(0, outer - inner + 1)

image = image.resize((scale, scale), resample=resample)

image = np.array(image).astype(np.uint8)

image = (image / 127.5 - 1.0).astype(np.float32)

instance_image = np.zeros((self.size, self.size, 3), dtype=np.float32)

mask = np.zeros((self.size // factor, self.size // factor))

if scale > self.size:

instance_image = image[top : top + inner, left : left + inner, :]

mask = np.ones((self.size // factor, self.size // factor))

else:

instance_image[top : top + inner, left : left + inner, :] = image

mask[

top // factor + 1 : (top + scale) // factor - 1, left // factor + 1 : (left + scale) // factor - 1

] = 1.0

return instance_image, mask

def __getitem__(self, index):

example = {}

instance_image, instance_prompt = self.instance_images_path[index % self.num_instance_images]

instance_image = Image.open(instance_image)

if not instance_image.mode == "RGB":

instance_image = instance_image.convert("RGB")

instance_image = self.flip(instance_image)

# apply resize augmentation and create a valid image region mask

random_scale = self.size

if self.aug:

random_scale = (

np.random.randint(self.size // 3, self.size + 1)

if np.random.uniform() < 0.66

else np.random.randint(int(1.2 * self.size), int(1.4 * self.size))

)

instance_image, mask = self.preprocess(instance_image, random_scale, self.interpolation)

if random_scale < 0.6 * self.size:

instance_prompt = np.random.choice(["a far away ", "very small "]) + instance_prompt

elif random_scale > self.size:

instance_prompt = np.random.choice(["zoomed in ", "close up "]) + instance_prompt

example["instance_images"] = torch.from_numpy(instance_image).permute(2, 0, 1)

example["mask"] = torch.from_numpy(mask)

example["instance_prompt_ids"] = self.tokenizer(

instance_prompt,

truncation=True,

padding="max_length",

max_length=self.tokenizer.model_max_length,

return_tensors="pt",

).input_ids

if self.with_prior_preservation:

class_image, class_prompt = self.class_images_path[index % self.num_class_images]

class_image = Image.open(class_image)

if not class_image.mode == "RGB":

class_image = class_image.convert("RGB")

example["class_images"] = self.image_transforms(class_image)

example["class_mask"] = torch.ones_like(example["mask"])

example["class_prompt_ids"] = self.tokenizer(

class_prompt,

truncation=True,

padding="max_length",

max_length=self.tokenizer.model_max_length,

return_tensors="pt",

).input_ids

return example

def save_new_embed(text_encoder, modifier_token_id, accelerator, args, output_dir, safe_serialization=True):

"""Saves the new token embeddings from the text encoder."""

logger.info("Saving embeddings")

learned_embeds = accelerator.unwrap_model(text_encoder).get_input_embeddings().weight

for x, y in zip(modifier_token_id, args.modifier_token):

learned_embeds_dict = {}

learned_embeds_dict[y] = learned_embeds[x]

filename = f"{output_dir}/{y}.bin"

if safe_serialization:

safetensors.torch.save_file(learned_embeds_dict, filename, metadata={"format": "pt"})

else:

torch.save(learned_embeds_dict, filename)

def parse_args(input_args=None):

parser = argparse.ArgumentParser(description="Custom Diffusion 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.",

)

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",

)

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,

help="A folder containing the training data of instance images.",

)

parser.add_argument(

"--class_data_dir",

type=str,

default=None,

help="A folder containing the training data of class images.",

)

parser.add_argument(

"--instance_prompt",

type=str,

default=None,

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=2,

help="Number of images that should be generated during validation with `validation_prompt`.",

)

parser.add_argument(

"--validation_steps",

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(

"--real_prior",

default=False,

action="store_true",

help="real images as prior.",

)

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=200,

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="custom-diffusion-model",

help="The output directory where the model predictions and checkpoints will be written.",

)

parser.add_argument("--seed", type=int, default=42, 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(

"--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."

),

)

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=250,

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`."

),

)

parser.add_argument(

"--checkpoints_total_limit",

type=int,

default=None,

help=("Max number of checkpoints to store."),

)

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=1e-5,

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(

"--dataloader_num_workers",

type=int,

default=2,

help=(

"Number of subprocesses to use for data loading. 0 means that the data will be loaded in the main process."

),

)

parser.add_argument(

"--freeze_model",

type=str,

default="crossattn_kv",

choices=["crossattn_kv", "crossattn"],

help="crossattn to enable fine-tuning of all params in the cross attention",

)

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(

"--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(

"--concepts_list",

type=str,

default=None,

help="Path to json containing multiple concepts, will overwrite parameters like instance_prompt, class_prompt, etc.",

)

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."

)

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"

),

)

parser.add_argument(

"--modifier_token",

type=str,

default=None,

help="A token to use as a modifier for the concept.",

)

parser.add_argument(

"--initializer_token", type=str, default="ktn+pll+ucd", help="A token to use as initializer word."

)

parser.add_argument("--hflip", action="store_true", help="Apply horizontal flip data augmentation.")

parser.add_argument(

"--noaug",

action="store_true",

help="Dont apply augmentation during data augmentation when this flag is enabled.",

)

parser.add_argument(

"--no_safe_serialization",

action="store_true",

help="If specified save the checkpoint not in `safetensors` format, but in original PyTorch format instead.",

)

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.concepts_list is None:

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.")

return args

def main(args):

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."

)

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

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

accelerator = Accelerator(

gradient_accumulation_steps=args.gradient_accumulation_steps,

mixed_precision=args.mixed_precision,

log_with=args.report_to,

project_config=accelerator_project_config,

)

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.")

import wandb

# 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.

# 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()

# 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("custom-diffusion", config=vars(args))

# If passed along, set the training seed now.

if args.seed is not None:

set_seed(args.seed)

if args.concepts_list is None:

args.concepts_list = [

{

"instance_prompt": args.instance_prompt,

"class_prompt": args.class_prompt,

"instance_data_dir": args.instance_data_dir,

"class_data_dir": args.class_data_dir,

}

]

else:

with open(args.concepts_list, "r") as f:

args.concepts_list = json.load(f)

# Generate class images if prior preservation is enabled.

if args.with_prior_preservation:

for i, concept in enumerate(args.concepts_list):

class_images_dir = Path(concept["class_data_dir"])

if not class_images_dir.exists():

class_images_dir.mkdir(parents=True, exist_ok=True)

if args.real_prior:

assert (

class_images_dir / "images"

).exists(), f"Please run: python retrieve.py --class_prompt \"{concept['class_prompt']}\" --class_data_dir {class_images_dir} --num_class_images {args.num_class_images}"

assert (

len(list((class_images_dir / "images").iterdir())) == args.num_class_images

), f"Please run: python retrieve.py --class_prompt \"{concept['class_prompt']}\" --class_data_dir {class_images_dir} --num_class_images {args.num_class_images}"

assert (

class_images_dir / "caption.txt"

).exists(), f"Please run: python retrieve.py --class_prompt \"{concept['class_prompt']}\" --class_data_dir {class_images_dir} --num_class_images {args.num_class_images}"

assert (

class_images_dir / "images.txt"

).exists(), f"Please run: python retrieve.py --class_prompt \"{concept['class_prompt']}\" --class_data_dir {class_images_dir} --num_class_images {args.num_class_images}"

concept["class_prompt"] = os.path.join(class_images_dir, "caption.txt")

concept["class_data_dir"] = os.path.join(class_images_dir, "images.txt")

args.concepts_list[i] = concept

accelerator.wait_for_everyone()

else:

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

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,

variant=args.variant,

)

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(concept["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):

hash_image = insecure_hashlib.sha1(image.tobytes()).hexdigest()

image_filename = (

class_images_dir / f"{example['index'][i] + cur_class_images}-{hash_image}.jpg"

)

image.save(image_filename)

del pipeline

if torch.cuda.is_available():

torch.cuda.empty_cache()

# Handle the repository creation

if accelerator.is_main_process:

if args.output_dir is not None:

os.makedirs(args.output_dir, exist_ok=True)

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

# 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(

args.pretrained_model_name_or_path, subfolder="text_encoder", revision=args.revision, variant=args.variant

)

vae = AutoencoderKL.from_pretrained(

args.pretrained_model_name_or_path, subfolder="vae", revision=args.revision, variant=args.variant

)

unet = UNet2DConditionModel.from_pretrained(

args.pretrained_model_name_or_path, subfolder="unet", revision=args.revision, variant=args.variant

)

# Adding a modifier token which is optimized ####

# Code taken from https://github.com/huggingface/diffusers/blob/main/examples/textual_inversion/textual_inversion.py

modifier_token_id = []

initializer_token_id = []

if args.modifier_token is not None:

args.modifier_token = args.modifier_token.split("+")

args.initializer_token = args.initializer_token.split("+")

if len(args.modifier_token) > len(args.initializer_token):

raise ValueError("You must specify + separated initializer token for each modifier token.")

for modifier_token, initializer_token in zip(

args.modifier_token, args.initializer_token[: len(args.modifier_token)]

):

# Add the placeholder token in tokenizer

num_added_tokens = tokenizer.add_tokens(modifier_token)

if num_added_tokens == 0:

raise ValueError(

f"The tokenizer already contains the token {modifier_token}. Please pass a different"

" `modifier_token` that is not already in the tokenizer."

)

# Convert the initializer_token, placeholder_token to ids

token_ids = tokenizer.encode([initializer_token], add_special_tokens=False)

print(token_ids)

# Check if initializer_token is a single token or a sequence of tokens

if len(token_ids) > 1:

raise ValueError("The initializer token must be a single token.")

initializer_token_id.append(token_ids[0])

modifier_token_id.append(tokenizer.convert_tokens_to_ids(modifier_token))

# Resize the token embeddings as we are adding new special tokens to the tokenizer

text_encoder.resize_token_embeddings(len(tokenizer))

# Initialise the newly added placeholder token with the embeddings of the initializer token

token_embeds = text_encoder.get_input_embeddings().weight.data

for x, y in zip(modifier_token_id, initializer_token_id):

token_embeds[x] = token_embeds[y]

# Freeze all parameters except for the token embeddings in text encoder

params_to_freeze = itertools.chain(

text_encoder.text_model.encoder.parameters(),

text_encoder.text_model.final_layer_norm.parameters(),

text_encoder.text_model.embeddings.position_embedding.parameters(),

)

freeze_params(params_to_freeze)

########################################################

########################################################

vae.requires_grad_(False)

if args.modifier_token is None:

text_encoder.requires_grad_(False)

unet.requires_grad_(False)

# For mixed precision training we cast the text_encoder and vae weights to half-precision

# as these models are only used for inference, keeping weights in full precision is not required.

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

if accelerator.mixed_precision != "fp16" and args.modifier_token is not None:

text_encoder.to(accelerator.device, dtype=weight_dtype)

unet.to(accelerator.device, dtype=weight_dtype)

vae.to(accelerator.device, dtype=weight_dtype)

attention_class = (

CustomDiffusionAttnProcessor2_0 if hasattr(F, "scaled_dot_product_attention") else CustomDiffusionAttnProcessor

)

if args.enable_xformers_memory_efficient_attention:

if is_xformers_available():

import xformers

xformers_version = version.parse(xformers.__version__)

if xformers_version == version.parse("0.0.16"):

logger.warning(

"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."

)

attention_class = CustomDiffusionXFormersAttnProcessor

else:

raise ValueError("xformers is not available. Make sure it is installed correctly")

# now we will add new Custom Diffusion weights to the attention layers

# It's important to realize here how many attention weights will be added and of which sizes

# The sizes of the attention layers consist only of two different variables:

# 1) - the "hidden_size", which is increased according to `unet.config.block_out_channels`.

# 2) - the "cross attention size", which is set to `unet.config.cross_attention_dim`.

# Let's first see how many attention processors we will have to set.

# For Stable Diffusion, it should be equal to:

# - down blocks (2x attention layers) * (2x transformer layers) * (3x down blocks) = 12

# - mid blocks (2x attention layers) * (1x transformer layers) * (1x mid blocks) = 2

# - up blocks (2x attention layers) * (3x transformer layers) * (3x down blocks) = 18

# => 32 layers

# Only train key, value projection layers if freeze_model = 'crossattn_kv' else train all params in the cross attention layer

train_kv = True

train_q_out = False if args.freeze_model == "crossattn_kv" else True

custom_diffusion_attn_procs = {}

st = unet.state_dict()

for name, _ in unet.attn_processors.items():

cross_attention_dim = None if name.endswith("attn1.processor") else unet.config.cross_attention_dim

if name.startswith("mid_block"):

hidden_size = unet.config.block_out_channels[-1]

elif name.startswith("up_blocks"):

block_id = int(name[len("up_blocks.")])

hidden_size = list(reversed(unet.config.block_out_channels))[block_id]

elif name.startswith("down_blocks"):

block_id = int(name[len("down_blocks.")])

hidden_size = unet.config.block_out_channels[block_id]

layer_name = name.split(".processor")[0]

weights = {

"to_k_custom_diffusion.weight": st[layer_name + ".to_k.weight"],

"to_v_custom_diffusion.weight": st[layer_name + ".to_v.weight"],

}

if train_q_out:

weights["to_q_custom_diffusion.weight"] = st[layer_name + ".to_q.weight"]

weights["to_out_custom_diffusion.0.weight"] = st[layer_name + ".to_out.0.weight"]

weights["to_out_custom_diffusion.0.bias"] = st[layer_name + ".to_out.0.bias"]

if cross_attention_dim is not None:

custom_diffusion_attn_procs[name] = attention_class(

train_kv=train_kv,

train_q_out=train_q_out,

hidden_size=hidden_size,

cross_attention_dim=cross_attention_dim,

).to(unet.device)

custom_diffusion_attn_procs[name].load_state_dict(weights)

else:

custom_diffusion_attn_procs[name] = attention_class(

train_kv=False,

train_q_out=False,

hidden_size=hidden_size,

cross_attention_dim=cross_attention_dim,

)

del st

unet.set_attn_processor(custom_diffusion_attn_procs)

custom_diffusion_layers = AttnProcsLayers(unet.attn_processors)

accelerator.register_for_checkpointing(custom_diffusion_layers)

if args.gradient_checkpointing:

unet.enable_gradient_checkpointing()

if args.modifier_token is not None:

text_encoder.gradient_checkpointing_enable()

# 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

)

if args.with_prior_preservation:

args.learning_rate = args.learning_rate * 2.0

# 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