train.py

#!/usr/bin/env python
# coding=utf-8
# Copyright 2023 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
# limitations under the License.
"""Fine-tuning script for Stable Diffusion XL for text2image."""

import argparse
import functools
import gc
import logging
import math
import os
import io
import random
import shutil
from pathlib import Path

import accelerate
import datasets
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 datasets import load_dataset, load_from_disk
from huggingface_hub import create_repo, upload_folder
from packaging import version
from torchvision import transforms
from torchvision.transforms.functional import crop
from tqdm.auto import tqdm
from transformers import AutoTokenizer, PretrainedConfig
from PIL import Image

import diffusers
from diffusers import (
    AutoencoderKL,
    DDPMScheduler,
    StableDiffusionXLPipeline,
    UNet2DConditionModel,
)
from diffusers.optimization import get_scheduler
from diffusers.training_utils import EMAModel, compute_snr
from diffusers.utils import check_min_version, is_wandb_available
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("0.25.0.dev0")

logger = get_logger(__name__)


DATASET_NAME_MAPPING = {
    "lambdalabs/pokemon-blip-captions": ("image", "text"),
}


def save_model_card(
    repo_id: str,
    images=None,
    validation_prompt=None,
    base_model=str,
    dataset_name=str,
    repo_folder=None,
    vae_path=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}
dataset: {dataset_name}
tags:
- stable-diffusion-xl
- stable-diffusion-xl-diffusers
- text-to-image
- diffusers
inference: true
---
    """
    model_card = f"""
# Text-to-image finetuning - {repo_id}

This pipeline was finetuned from **{base_model}** on the **{args.dataset_name}** dataset. Below are some example images generated with the finetuned pipeline using the following prompt: {validation_prompt}: \n
{img_str}

Special VAE used for training: {vae_path}.
"""
    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, subfolder: str = "text_encoder"
):
    text_encoder_config = PretrainedConfig.from_pretrained(
        pretrained_model_name_or_path, subfolder=subfolder, revision=revision
    )
    model_class = text_encoder_config.architectures[0]

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

        return CLIPTextModel
    elif model_class == "CLIPTextModelWithProjection":
        from transformers import CLIPTextModelWithProjection

        return CLIPTextModelWithProjection
    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(
        "--pretrained_vae_model_name_or_path",
        type=str,
        default=None,
        help="Path to pretrained VAE model with better numerical stability. More details: https://github.com/huggingface/diffusers/pull/4038.",
    )
    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(
        "--dataset_name",
        type=str,
        default=None,
        help=(
            "The name of the Dataset (from the HuggingFace hub) to train on (could be your own, possibly private,"
            " dataset). It can also be a path pointing to a local copy of a dataset in your filesystem,"
            " or to a folder containing files that 🤗 Datasets can understand."
        ),
    )
    parser.add_argument(
        "--dataset_config_name",
        type=str,
        default=None,
        help="The config of the Dataset, leave as None if there's only one config.",
    )
    parser.add_argument(
        "--train_data_dir",
        type=str,
        default=None,
        help=(
            "A folder containing the training data. Folder contents must follow the structure described in"
            " https://huggingface.co/docs/datasets/image_dataset#imagefolder. In particular, a `metadata.jsonl` file"
            " must exist to provide the captions for the images. Ignored if `dataset_name` is specified."
        ),
    )
    parser.add_argument(
        "--good_image_column",
        type=str,
        default="image",
        help="The column of the dataset containing an good image.",
    )
    parser.add_argument(
        "--bad_image_column",
        type=str,
        default="image",
        help="The column of the dataset containing an bad image.",
    )
    parser.add_argument(
        "--caption_column",
        type=str,
        default="text",
        help="The column of the dataset containing a caption or a list of captions.",
    )
    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=1,
        help=(
            "Run fine-tuning validation every X epochs. The validation process consists of running the prompt"
            " `args.validation_prompt` multiple times: `args.num_validation_images`."
        ),
    )
    parser.add_argument(
        "--output_dir",
        type=str,
        default="sdxl-model-finetuned",
        help="The output directory where the model predictions and checkpoints will be written.",
    )
    parser.add_argument(
        "--cache_dir",
        type=str,
        default=None,
        help="The directory where the downloaded models and datasets will be stored.",
    )
    parser.add_argument(
        "--seed", type=int, default=None, help="A seed for reproducible training."
    )
    parser.add_argument(
        "--resolution",
        type=int,
        default=1024,
        help=(
            "The resolution for input images, all the images in the train/validation dataset will be resized to this"
            " resolution"
        ),
    )
    parser.add_argument(
        "--train_batch_size",
        type=int,
        default=16,
        help="Batch size (per device) for the training dataloader.",
    )
    parser.add_argument("--num_train_epochs", type=int, default=100)
    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`."
        ),
    )
    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(
        "--dpo_beta",
        type=float,
        default=1,
        help="DPO beta.",
    )
    parser.add_argument(
        "--learning_rate",
        type=float,
        default=1e-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(
        "--timestep_bias_strategy",
        type=str,
        default="none",
        choices=["earlier", "later", "range", "none"],
        help=(
            "The timestep bias strategy, which may help direct the model toward learning low or high frequency details."
            " Choices: ['earlier', 'later', 'range', 'none']."
            " The default is 'none', which means no bias is applied, and training proceeds normally."
            " The value of 'later' will increase the frequency of the model's final training timesteps."
        ),
    )
    parser.add_argument(
        "--timestep_bias_multiplier",
        type=float,
        default=1.0,
        help=(
            "The multiplier for the bias. Defaults to 1.0, which means no bias is applied."
            " A value of 2.0 will double the weight of the bias, and a value of 0.5 will halve it."
        ),
    )
    parser.add_argument(
        "--timestep_bias_begin",
        type=int,
        default=0,
        help=(
            "When using `--timestep_bias_strategy=range`, the beginning (inclusive) timestep to bias."
            " Defaults to zero, which equates to having no specific bias."
        ),
    )
    parser.add_argument(
        "--timestep_bias_end",
        type=int,
        default=1000,
        help=(
            "When using `--timestep_bias_strategy=range`, the final timestep (inclusive) to bias."
            " Defaults to 1000, which is the number of timesteps that Stable Diffusion is trained on."
        ),
    )
    parser.add_argument(
        "--timestep_bias_portion",
        type=float,
        default=0.25,
        help=(
            "The portion of timesteps to bias. Defaults to 0.25, which 25% of timesteps will be biased."
            " A value of 0.5 will bias one half of the timesteps. The value provided for `--timestep_bias_strategy` determines"
            " whether the biased portions are in the earlier or later timesteps."
        ),
    )
    parser.add_argument(
        "--snr_gamma",
        type=float,
        default=None,
        help="SNR weighting gamma to be used if rebalancing the loss. Recommended value is 5.0. "
        "More details here: https://arxiv.org/abs/2303.09556.",
    )
    parser.add_argument(
        "--use_ema", action="store_true", help="Whether to use EMA model."
    )
    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(
        "--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."
        ),
    )
    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(
        "--prediction_type",
        type=str,
        default=None,
        help="The prediction_type that shall be used for training. Choose between 'epsilon' or 'v_prediction' or leave `None`. If left to `None` the default prediction type of the scheduler: `noise_scheduler.config.prediciton_type` is chosen.",
    )
    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(
        "--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(
        "--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(
        "--noise_offset", type=float, default=0, help="The scale of noise offset."
    )
    parser.add_argument(
        "--sd_version", type=str, default="1.5", help="sd version"
    )
    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

    # Sanity checks
    if args.dataset_name is None and args.train_data_dir is None:
        raise ValueError("Need either a dataset name or a training folder.")
    
    return args

def generate_timestep_weights(args, num_timesteps):
    weights = torch.ones(num_timesteps)

    # Determine the indices to bias
    num_to_bias = int(args.timestep_bias_portion * num_timesteps)

    if args.timestep_bias_strategy == "later":
        bias_indices = slice(-num_to_bias, None)
    elif args.timestep_bias_strategy == "earlier":
        bias_indices = slice(0, num_to_bias)
    elif args.timestep_bias_strategy == "range":
        # Out of the possible 1000 timesteps, we might want to focus on eg. 200-500.
        range_begin = args.timestep_bias_begin
        range_end = args.timestep_bias_end
        if range_begin < 0:
            raise ValueError(
                "When using the range strategy for timestep bias, you must provide a beginning timestep greater or equal to zero."
            )
        if range_end > num_timesteps:
            raise ValueError(
                "When using the range strategy for timestep bias, you must provide an ending timestep smaller than the number of timesteps."
            )
        bias_indices = slice(range_begin, range_end)
    else:  # 'none' or any other string
        return weights
    if args.timestep_bias_multiplier <= 0:
        return ValueError(
            "The parameter --timestep_bias_multiplier is not intended to be used to disable the training of specific timesteps."
            " If it was intended to disable timestep bias, use `--timestep_bias_strategy none` instead."
            " A timestep bias multiplier less than or equal to 0 is not allowed."
        )

    # Apply the bias
    weights[bias_indices] *= args.timestep_bias_multiplier

    # Normalize
    weights /= weights.sum()

    return weights


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

    accelerator_project_config = ProjectConfiguration(
        project_dir=args.output_dir, logging_dir=logging_dir
    )
    # 指定要使用的GPU
    # device_placement = {
    #     "0": "cuda:5",
    #     "1": "cuda:6",
    #     "2": "cuda:7"
    # }
    accelerator = Accelerator(
        gradient_accumulation_steps=args.gradient_accumulation_steps,
        mixed_precision=args.mixed_precision,
        log_with=args.report_to,
        project_config=accelerator_project_config,
        # device_placement=device_placement,
    )

    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

    # 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:
        datasets.utils.logging.set_verbosity_warning()
        transformers.utils.logging.set_verbosity_warning()
        diffusers.utils.logging.set_verbosity_info()
    else:
        datasets.utils.logging.set_verbosity_error()
        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)

    # 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 scheduler and models
    noise_scheduler = DDPMScheduler.from_pretrained(
        args.pretrained_model_name_or_path, subfolder="scheduler"
    )
    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,
    )
    ref_unet = UNet2DConditionModel.from_pretrained(
        args.pretrained_model_name_or_path,
        subfolder="unet",
        revision=args.revision,
        variant=args.variant,
    )

    # Freeze vae and text encoders.
    vae.requires_grad_(False)
    ref_unet.requires_grad_(False)
    # Set unet as trainable.
    unet.train()
    ref_unet.eval()

    # For mixed precision training we cast all non-trainable weigths to half-precision
    # as these weights 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 to device and cast to weight_dtype
    ref_unet.to(accelerator.device, dtype=weight_dtype)

    # Create EMA for the unet.
    if args.use_ema:
        ema_unet = UNet2DConditionModel.from_pretrained(
            args.pretrained_model_name_or_path,
            subfolder="unet",
            revision=args.revision,
            variant=args.variant,
        )
        ema_unet = EMAModel(
            ema_unet.parameters(),
            model_cls=UNet2DConditionModel,
            model_config=ema_unet.config,
        )

    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.warn(
                    "xFormers 0.0.16 cannot be used for training in some GPUs. If you observe problems during training, please update xFormers to at least 0.0.17. See https://huggingface.co/docs/diffusers/main/en/optimization/xformers for more details."
                )
            unet.enable_xformers_memory_efficient_attention()
        else:
            raise ValueError(
                "xformers is not available. Make sure it is installed correctly"
            )

    # `accelerate` 0.16.0 will have better support for customized saving
    if version.parse(accelerate.__version__) >= version.parse("0.16.0"):
        # create custom saving & loading hooks so that `accelerator.save_state(...)` serializes in a nice format
        def save_model_hook(models, weights, output_dir):
            if accelerator.is_main_process:
                if args.use_ema:
                    ema_unet.save_pretrained(os.path.join(output_dir, "unet_ema"))

                for i, model in enumerate(models):
                    print(model)
                    model.save_pretrained(os.path.join(output_dir, "unet"))

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

        def load_model_hook(models, input_dir):
            if args.use_ema:
                load_model = EMAModel.from_pretrained(
                    os.path.join(input_dir, "unet_ema"), UNet2DConditionModel
                )
                ema_unet.load_state_dict(load_model.state_dict())
                ema_unet.to(accelerator.device)
                del load_model

            for i in range(len(models)):
                # pop models so that they are not loaded again
                model = models.pop()

                # 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)

    if args.gradient_checkpointing:
        unet.enable_gradient_checkpointing()

    # 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
    params_to_optimize = unet.parameters()
    optimizer = optimizer_class(
        params_to_optimize,
        lr=args.learning_rate,
        betas=(args.adam_beta1, args.adam_beta2),
        weight_decay=args.adam_weight_decay,
        eps=args.adam_epsilon,
    )

    # Get the datasets: you can either provide your own training and evaluation files (see below)
    # or specify a Dataset from the hub (the dataset will be downloaded automatically from the datasets Hub).

    # In distributed training, the load_dataset function guarantees that only one local process can concurrently
    # download the dataset.
    dataset = load_from_disk(
            args.dataset_name,
        )
    train_dataset = dataset
    
    gc.collect()
    torch.cuda.empty_cache()

    def collate_fn(examples):
        good_model_input = torch.stack(
            [torch.tensor(example[args.good_image_column + "_model_input"]) for example in examples]
        )
        bad_model_input = torch.stack(
            [torch.tensor(example[args.bad_image_column + "_model_input"]) for example in examples]
        )
        original_sizes = [example[args.good_image_column + "_original_sizes"] for example in examples]
        crop_top_lefts = [example[args.good_image_column + "_crop_top_lefts"] for example in examples]
        prompt_embeds = torch.stack(
            [torch.tensor(example["prompt_embeds"]) for example in examples]
        )
        result = {
            "good_model_input": good_model_input,
            "bad_model_input": bad_model_input,
            "prompt_embeds": prompt_embeds,
            "original_sizes": original_sizes,
            "crop_top_lefts": crop_top_lefts,
        }
        if args.sd_version == "xl":
            pooled_prompt_embeds = torch.stack(
                [torch.tensor(example["pooled_prompt_embeds"]) for example in examples]
            )
            result["pooled_prompt_embeds"] = pooled_prompt_embeds
        return result

    # DataLoaders creation:
    train_dataloader = torch.utils.data.DataLoader(
        train_dataset,
        shuffle=True,
        collate_fn=collate_fn,
        batch_size=args.train_batch_size,
        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,
    )

    # Prepare everything with our `accelerator`.
    unet, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
        unet, optimizer, train_dataloader, lr_scheduler
    )
    # We need to recalculate our total training steps as the size of the training dataloader may have changed.
    num_update_steps_per_epoch = math.ceil(
        len(train_dataloader) / args.gradient_accumulation_steps
    )
    if overrode_max_train_steps:
        args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
    # Afterwards we recalculate our number of training epochs
    args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)

    # We need to initialize the trackers we use, and also store our configuration.
    # The trackers initializes automatically on the main process.
    if accelerator.is_main_process:
        accelerator.init_trackers("text2image-fine-tune-sdxl", 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 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 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
            initial_global_step = 0
        else:
            accelerator.print(f"Resuming from checkpoint {path}")
            accelerator.load_state(os.path.join(args.output_dir, path))
            global_step = int(path.split("-")[1])

            initial_global_step = global_step
            first_epoch = global_step // num_update_steps_per_epoch

    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,
    )

    for epoch in range(first_epoch, args.num_train_epochs):
        train_loss = 0.0
        for step, batch in enumerate(train_dataloader):
            with accelerator.accumulate(unet):
                # Sample noise that we'll add to the latents
                good_model_input = batch["good_model_input"].to(accelerator.device) * vae.config.scaling_factor
                bad_model_input = batch["bad_model_input"].to(accelerator.device) * vae.config.scaling_factor

                good_noise = torch.randn_like(good_model_input)
                bad_noise = torch.randn_like(good_model_input)

                bsz = good_noise.shape[0]

                timesteps = torch.randint(
                    0,
                    noise_scheduler.config.num_train_timesteps,
                    (bsz,),
                    device=good_model_input.device,
                )

                good_noisy_model_input = noise_scheduler.add_noise(
                    good_model_input, good_noise, timesteps
                )
                bad_noisy_model_input = noise_scheduler.add_noise(
                    bad_model_input, bad_noise, timesteps
                )

                # time ids
                def compute_time_ids(original_size, crops_coords_top_left):
                    # Adapted from pipeline.StableDiffusionXLPipeline._get_add_time_ids
                    target_size = (args.resolution, args.resolution)
                    add_time_ids = list(
                        tuple(original_size) + tuple(crops_coords_top_left) + target_size
                    )
                    add_time_ids = torch.tensor([add_time_ids])
                    add_time_ids = add_time_ids.to(
                        accelerator.device, dtype=weight_dtype
                    )
                    return add_time_ids

                add_time_ids = torch.cat(
                    [
                        compute_time_ids(s, c)
                        for s, c in zip(
                            batch["original_sizes"], batch["crop_top_lefts"]
                        )
                    ]
                )

                # Predict the noise residual
                unet_added_conditions = {"time_ids": add_time_ids}
                prompt_embeds = batch["prompt_embeds"].to(accelerator.device)
                if args.sd_version == "xl":
                    pooled_prompt_embeds = batch["pooled_prompt_embeds"].to(
                        accelerator.device
                    )
                    unet_added_conditions.update({"text_embeds": pooled_prompt_embeds})
                good_model_pred = unet(
                    good_noisy_model_input,
                    timesteps,
                    prompt_embeds,
                    added_cond_kwargs=unet_added_conditions,
                ).sample
                bad_model_pred = unet(
                    bad_noisy_model_input,
                    timesteps,
                    prompt_embeds,
                    added_cond_kwargs=unet_added_conditions,
                ).sample
                good_model_ref = ref_unet(
                    good_noisy_model_input.half(),
                    timesteps,
                    prompt_embeds,
                    added_cond_kwargs=unet_added_conditions,
                ).sample
                bad_model_ref = ref_unet(
                    bad_noisy_model_input.half(),
                    timesteps,
                    prompt_embeds,
                    added_cond_kwargs=unet_added_conditions,
                ).sample
                # Get the target for loss depending on the prediction type
                if args.prediction_type is not None:
                    # set prediction_type of scheduler if defined
                    noise_scheduler.register_to_config(
                        prediction_type=args.prediction_type
                    )

                if noise_scheduler.config.prediction_type == "epsilon":
                    good_target = good_noise
                    bad_target = bad_noise
                elif noise_scheduler.config.prediction_type == "v_prediction":
                    good_target = noise_scheduler.get_velocity(
                        good_model_input, good_noise, timesteps
                    )
                    bad_target = noise_scheduler.get_velocity(
                        bad_model_input, bad_noise, timesteps
                    )
                elif noise_scheduler.config.prediction_type == "sample":
                    good_target = good_model_input
                    good_model_pred = good_model_pred - good_noise
                    bad_target = bad_model_input
                    bad_model_pred = bad_model_pred - bad_noise
                else:
                    raise ValueError(
                        f"Unknown prediction type {noise_scheduler.config.prediction_type}"
                    )

                # Compute DPO Loss. A bit different from paper for better stability.
                loss = -F.logsigmoid(
                    -args.dpo_beta
                    * (
                        torch.norm(good_model_pred.float() - good_target.float())
                        - torch.norm(good_target.float() - good_model_ref.float())
                        - (torch.norm(bad_model_pred.float() - bad_target.float())
                        - torch.norm(bad_target.float() - bad_model_ref.float()))
                    )
                ).mean()
                # Gather the losses across all processes for logging (if we use distributed training).
                avg_loss = accelerator.gather(loss.repeat(args.train_batch_size)).mean()
                train_loss += avg_loss.item() / args.gradient_accumulation_steps

                # Backpropagate
                accelerator.backward(loss)
                if accelerator.sync_gradients:
                    params_to_clip = unet.parameters()
                    accelerator.clip_grad_norm_(params_to_clip, args.max_grad_norm)
                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
                accelerator.log({"train_loss": train_loss}, step=global_step)
                train_loss = 0.0

                if accelerator.is_main_process:
                    if global_step % args.checkpointing_steps == 0:
                        # _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)

                        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 = {
                "step_loss": loss.detach().item(),
                "lr": lr_scheduler.get_last_lr()[0],
            }
            progress_bar.set_postfix(**logs)

            if global_step >= args.max_train_steps:
                break
    accelerator.end_training()


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