train_compositional_dreambooth.py

import sys
#sys.path.append('/mnt1/xulu_zhang/git_projects/Diffusion/Text2Image/main')
sys.path.append('/home/xulu/Text2Image/main')
import argparse
import hashlib
import itertools
import json
import logging
import math
import os
import random
import warnings
from pathlib import Path

import numpy as np
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 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, CustomDiffusionXFormersAttnProcessor
from diffusers.optimization import get_scheduler
from diffusers.utils import check_min_version, is_wandb_available
from diffusers.utils.import_utils import is_xformers_available

from model.modeling_clip import CLIPTextModel


# Will error if the minimal version of diffusers is not installed. Remove at your own risks.
check_min_version("0.15.0.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"![img_{i}](./image_{i}.png)\n"

    yaml = f"""
---
license: creativeml-openrail-m
base_model: {base_model}
instance_prompt: {prompt}
tags:
- stable-diffusion
- stable-diffusion-diffusers
- text-to-image
- diffusers
- custom-diffusion
inference: true
---
    """
    model_card = f"""
# DreamBooth - {repo_id}

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

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


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):
    placeholder_index = [example["placeholder_index"] for example in examples]
    input_ids = [example["instance_prompt_ids"] for example in examples]
    reference_input_ids = [example["reference_prompt_ids"] for example in examples]
    pixel_values = [example["instance_images"] 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]

    input_ids = torch.cat(input_ids, dim=0)
    pixel_values = torch.stack(pixel_values)
    placeholder_index = torch.stack(placeholder_index).squeeze(1)
    reference_input_ids = torch.cat(reference_input_ids, dim=0)
    pixel_values = pixel_values.to(memory_format=torch.contiguous_format).float()
    batch = {"input_ids": input_ids, "pixel_values": pixel_values, "reference_prompt_ids": reference_input_ids, "placeholder_index":placeholder_index}
    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 CompositionalDreamboothDataset(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,
        center_crop=False,
        with_prior_preservation=False,
        num_class_images=200,
        hflip=False,
    ):
        self.size = size
        self.center_crop = center_crop
        self.tokenizer = tokenizer
        self.interpolation = Image.BILINEAR

        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"], concept["expected_class"]) 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 = [(x, y) for (x, y) in 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 __getitem__(self, index):
        example = {}
        instance_image, instance_prompt, class_noun = 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")
        example["instance_images"] = self.image_transforms(instance_image)
        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
        for i, w in enumerate(instance_prompt.split(' ')):
            if '<' in w:
                example["placeholder_index"] = torch.tensor([i+1])
                unique_identifier = w
                break

        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_prompt_ids"] = self.tokenizer(
                class_prompt,
                truncation=True,
                padding="max_length",
                max_length=self.tokenizer.model_max_length,
                return_tensors="pt",
            ).input_ids
        example["reference_prompt_ids"] = self.tokenizer(
                instance_prompt.replace(unique_identifier, class_noun),
                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):
    """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]
        torch.save(learned_embeds_dict, f"{output_dir}/{y}.bin")


def parse_args(input_args=None):
    parser = argparse.ArgumentParser(description="Custom Diffusion training script.")
    # for mix up training
    parser.add_argument(
        "--recons_loss_weight",
        type=float,
        default=1.0,
        help="The weight of reconstruction loss.",
    )
    parser.add_argument(
        "--anchor_loss",
        type=str,
        default='none',
        help="The loss function for embedding, choose from ['none', 'l1', 'l2', 'cosine']."
    )
    parser.add_argument(
        "--anchor_loss_weight",
        type=float,
        default=0.0,
        help="The weight of embedding loss."
    )
    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("--lr_power", type=float, default=1.0, help="Power factor of the polynomial scheduler.")
    parser.add_argument(
        "--revision",
        type=str,
        default=None,
        required=False,
        help="Revision of pretrained model identifier from huggingface.co/models.",
    )
    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(
        "--offset_noise",
        action="store_true",
        default=False,
        help=(
            "Fine-tuning against a modified noise"
            " See: https://www.crosslabs.org//blog/diffusion-with-offset-noise for more information."
        ),
    )
    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. Passed as `total_limit` to the `Accelerator` `ProjectConfiguration`."
            " See Accelerator::save_state https://huggingface.co/docs/accelerate/package_reference/accelerator#accelerate.Accelerator.save_state"
            " for more docs"
        ),
    )
    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-6,
        help="Initial learning rate (after the potential warmup period) to use.",
    )
    parser.add_argument(
        "--scale_lr",
        action="store_true",
        default=False,
        help="Scale the learning rate by the number of GPUs, gradient accumulation steps, and batch size.",
    )
    parser.add_argument(
        "--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(
        "--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(
        "--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.",
    )

    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):
    logging_dir = os.path.join(args.output_dir, args.logging_dir)
    accelerator_project_config = ProjectConfiguration(
        total_limit=args.checkpoints_total_limit, 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()

    # 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,
                    )
                    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):
                            hash_image = hashlib.sha1(image.tobytes()).hexdigest()
                            image_filename = (
                                class_images_dir / f"{example['index'][i] + cur_class_images}-{hash_image}.jpg"
                            )
                            image.save(image_filename)

                    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 = CLIPTextModel.from_pretrained(
        args.pretrained_model_name_or_path, subfolder="text_encoder", revision=args.revision
    )
    vae = AutoencoderKL.from_pretrained(args.pretrained_model_name_or_path, subfolder="vae", revision=args.revision)
    unet = UNet2DConditionModel.from_pretrained(
        args.pretrained_model_name_or_path, subfolder="unet", revision=args.revision
    )

    # create custom saving & loading hooks so that `accelerator.save_state(...)` serializes in a nice format
    def save_model_hook(models, weights, output_dir):
        for model in models:
            if isinstance(model, type(accelerator.unwrap_model(text_encoder.text_model.embeddings))):
                continue
            sub_dir = "unet" if isinstance(model, type(accelerator.unwrap_model(unet))) else "text_encoder"
            model.save_pretrained(os.path.join(output_dir, sub_dir))

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

    def load_model_hook(models, input_dir):
        while len(models) > 0:
            
            # pop models so that they are not loaded again
            model = models.pop()
            if isinstance(model, type(accelerator.unwrap_model(text_encoder.text_model.embeddings))):
                continue
            if isinstance(model, type(accelerator.unwrap_model(text_encoder))):
                # load transformers style into model
                load_model = CLIPTextModel.from_pretrained(input_dir, subfolder="text_encoder")
                model.config = load_model.config
            else:
                # load diffusers style into model
                load_model = UNet2DConditionModel.from_pretrained(input_dir, subfolder="unet")
                model.register_to_config(**load_model.config)

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

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


    # 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)
            # 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)

    # 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

    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
        )

    # 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
    optimizer = optimizer_class(
        itertools.chain(text_encoder.get_input_embeddings().parameters(), unet.parameters())
        if args.modifier_token is not None
        else unet.parameters(),
        lr=args.learning_rate,
        betas=(args.adam_beta1, args.adam_beta2),
        weight_decay=args.adam_weight_decay,
        eps=args.adam_epsilon,
    )

    # Dataset and DataLoaders creation:
    train_dataset = CompositionalDreamboothDataset(
        concepts_list=args.concepts_list,
        tokenizer=tokenizer,
        with_prior_preservation=args.with_prior_preservation,
        size=args.resolution,
        center_crop=args.center_crop,
        num_class_images=args.num_class_images,
        hflip=args.hflip
    )

    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),
        num_workers=args.dataloader_num_workers,
    )

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

    lr_scheduler = get_scheduler(
        args.lr_scheduler,
        optimizer=optimizer,
        num_warmup_steps=args.lr_warmup_steps * args.gradient_accumulation_steps,
        num_training_steps=args.max_train_steps * args.gradient_accumulation_steps,
        num_cycles=args.lr_num_cycles,
        power=args.lr_power
    )

    # Prepare everything with our `accelerator`.
    if args.modifier_token is not None:
        unet, text_encoder, optimizer, train_dataloader, lr_scheduler, Embeddings = accelerator.prepare(
            unet, text_encoder, optimizer, train_dataloader, lr_scheduler, text_encoder.text_model.embeddings
        )
    else:
        unet, optimizer, train_dataloader, lr_scheduler, Embeddings = accelerator.prepare(
            unet, optimizer, train_dataloader, lr_scheduler, text_encoder.text_model.embeddings
        )

    # Move unet, vae and text_encoder to device and cast to weight_dtype
    text_encoder.to(accelerator.device, dtype=weight_dtype)
    unet.to(accelerator.device, dtype=weight_dtype)
    vae.to(accelerator.device, dtype=weight_dtype)

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


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


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

    logger.info("***** Running training *****")
    logger.info(f"  Num examples = {len(train_dataset)}")
    logger.info(f"  Num batches each epoch = {len(train_dataloader)}")
    logger.info(f"  Num Epochs = {args.num_train_epochs}")
    logger.info(f"  Instantaneous batch size per device = {args.train_batch_size}")
    logger.info(f"  Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}")
    logger.info(f"  Gradient Accumulation steps = {args.gradient_accumulation_steps}")
    logger.info(f"  Total optimization steps = {args.max_train_steps}")
    logger.info(f"  Reconstruction loss weight = {args.recons_loss_weight}")
    logger.info(f"  Anchor loss weight = {args.anchor_loss_weight}")
    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 most 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]))
            path = dirs[-1] if len(dirs) > 0 else None
 
        if path is None:
            accelerator.print(
                f"Checkpoint '{args.resume_from_checkpoint}' does not exist. Starting a new training run."
            )
            args.resume_from_checkpoint = None
        else:
            accelerator.print(f"Resuming from checkpoint {path}")
            accelerator.load_state(os.path.join(args.output_dir, path))
            global_step = int(path.split("-")[1])

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

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


    for epoch in range(first_epoch, args.num_train_epochs):
        unet.train()
        if args.modifier_token is not None:
            text_encoder.train()
        for step, batch in enumerate(train_dataloader):
            # Skip steps until we reach the resumed step
            if args.resume_from_checkpoint and epoch == first_epoch and step < resume_step:
                if step % args.gradient_accumulation_steps == 0:
                    progress_bar.update(1)
                continue

            with accelerator.accumulate(unet), accelerator.accumulate(text_encoder):
                # Convert images to latent space
                latents = vae.encode(batch["pixel_values"].to(dtype=weight_dtype)).latent_dist.sample()
                latents = latents * vae.config.scaling_factor

                # Sample noise that we'll add to the latents
                if args.offset_noise:
                    noise = torch.randn_like(latents) + 0.1 * torch.randn(
                        latents.shape[0], latents.shape[1], 1, 1, device=latents.device
                    )
                else:
                    noise = torch.randn_like(latents)
                bsz = latents.shape[0]
                # Sample a random timestep for each image
                timesteps = torch.randint(0, noise_scheduler.config.num_train_timesteps, (bsz,), device=latents.device)
                timesteps = timesteps.long()

                # Add noise to the latents according to the noise magnitude at each timestep
                # (this is the forward diffusion process)
                noisy_latents = noise_scheduler.add_noise(latents, noise, timesteps)

                # Get the text embedding for conditioning
                encoder_hidden_states, input_embeddings = text_encoder(batch["input_ids"])
                encoder_hidden_states = encoder_hidden_states[0]

                # Predict the noise residual
                model_pred = unet(noisy_latents, timesteps, encoder_hidden_states).sample

                # 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":
                    target = noise_scheduler.get_velocity(latents, noise, timesteps)
                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") * args.recons_loss_weight

                    # 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") * args.recons_loss_weight
                
                if args.anchor_loss == 'l2':
                    with torch.no_grad():
                        reference_input_embedings = Embeddings(batch["reference_prompt_ids"])  # token_embeddings + position_embeddings
                    optimized_embeddings, _ = torch.chunk(input_embeddings, 2, dim=0)
                    anchor_loss = F.mse_loss(optimized_embeddings, reference_input_embedings.detach(), reduction="mean")
                    loss = loss + args.anchor_loss_weight * anchor_loss

                accelerator.backward(loss)
                # Zero out the gradients for all token embeddings except the newly added
                # embeddings for the concept, as we only want to optimize the concept embeddings
                if args.modifier_token is not None:
                    if accelerator.num_processes > 1:
                        grads_text_encoder = text_encoder.module.get_input_embeddings().weight.grad
                    else:
                        grads_text_encoder = text_encoder.get_input_embeddings().weight.grad
                    # Get the index for tokens that we want to zero the grads for
                    index_grads_to_zero = torch.arange(len(tokenizer)) != modifier_token_id[0]
                    for i in range(len(modifier_token_id[1:])):
                        index_grads_to_zero = index_grads_to_zero & (
                            torch.arange(len(tokenizer)) != modifier_token_id[i]
                        )
                    grads_text_encoder.data[index_grads_to_zero, :] = grads_text_encoder.data[
                        index_grads_to_zero, :
                    ].fill_(0)

                if accelerator.sync_gradients:
                    params_to_clip = (
                        itertools.chain(text_encoder.parameters(), unet.parameters())
                        if args.modifier_token is not None
                        else unet.parameters()
                    )
                    accelerator.clip_grad_norm_(params_to_clip, args.max_grad_norm)
                optimizer.step()
                lr_scheduler.step()
                optimizer.zero_grad(set_to_none=args.set_grads_to_none)

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

                if global_step % args.checkpointing_steps == 0:
                    if accelerator.is_main_process:
                        save_path = os.path.join(args.output_dir, f"checkpoint-{global_step}")
                        accelerator.save_state(save_path)
                        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

    # Save the custom diffusion layers
    accelerator.wait_for_everyone()
    if accelerator.is_main_process:
        if args.modifier_token:
            save_new_embed(text_encoder, modifier_token_id, accelerator, args, args.output_dir)

        pipeline = DiffusionPipeline.from_pretrained(
            args.pretrained_model_name_or_path,
            unet=accelerator.unwrap_model(unet),
            text_encoder=accelerator.unwrap_model(text_encoder),
            revision=args.revision,
        )
        pipeline.save_pretrained(args.output_dir)

    accelerator.end_training()


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