#!/usr/bin/env python
# coding=utf-8
# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
# Copyright 2024 Huawei Technologies Co., Ltd
#
# 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 random
import shutil
import stat
import time
from multiprocessing import Value
from pathlib import Path

import accelerate
import collect_dataset
import datasets
import diffusers
import numpy as np
import pretrain_model
import torch
import torch.nn.functional as F
import torch.utils.checkpoint
import transformers
from accelerate import Accelerator, DistributedType
from accelerate.logging import get_logger
from accelerate.utils import ProjectConfiguration, set_seed
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_torch_npu_available, is_xformers_available
from huggingface_hub import create_repo, upload_folder
from packaging import version
from patch_sdxl import TorchPatcher, config_gc
from torchvision import transforms
from torchvision.transforms.functional import crop
from tqdm.auto import tqdm
from transformers import (
    AutoTokenizer,
    CLIPTextModel,
    CLIPTextModelWithProjection,
    PretrainedConfig,
)

TorchPatcher.apply_patch()
config_gc()

try:
    from torch_npu.utils.profiler import Profile
except ImportError:
    print(
        "Profile not in torch_npu.utils.profiler now.. Auto Profile disabled.",
        flush=True,
    )

    class Profile:
        def __init__(self, *args, **kwargs):
            pass

        def start(self):
            pass

        def end(self):
            pass


# Will error if the minimal version of diffusers is not installed. Remove at your own risks.
check_min_version("0.30.0")

logger = get_logger(__name__)


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

torch.npu.config.allow_internal_format = False


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}.
"""
    file_path = os.path.join(repo_folder, "README.md")
    file_descriptor = os.open(
        file_path, os.O_WRONLY | os.O_CREAT, stat.S_IRUSR | stat.S_IWUSR
    )
    with os.fdopen(file_descriptor, "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, local_files_only=True
    )
    model_class = text_encoder_config.architectures[0]

    if model_class == "CLIPTextModel":
        return CLIPTextModel
    elif model_class == "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",
    )
    parser.add_argument(
        "--pretrained_vae_model_name_or_path",
        type=str,
        default=None,
        help="Path to pretrained VAE model with better numerical stability.",
    )
    parser.add_argument(
        "--revision",
        type=str,
        default=None,
        required=False,
        help="Revision of pretrained model identifier from huggingface.co",
    )
    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"
            " In particular, a `metadata.jsonl` file"
            " must exist to provide the captions for the images. Ignored if `dataset_name` is specified."
        ),
    )
    parser.add_argument(
        "--image_column",
        type=str,
        default="image",
        help="The column of the dataset containing an 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(
        "--max_train_samples",
        type=int,
        default=None,
        help=(
            "For debugging purposes or quicker training, truncate the number of training examples to this "
            "value if set."
        ),
    )
    parser.add_argument(
        "--proportion_empty_prompts",
        type=float,
        default=0,
        help="Proportion of image prompts to be replaced with empty strings. Defaults to 0 (no prompt replacement).",
    )
    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(
        "--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(
        "--random_flip",
        action="store_true",
        help="whether to randomly flip images horizontally",
    )
    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(
        "--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. ",
    )
    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"
        ),
    )
    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]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_npu_flash_attention",
        action="store_true",
        help="Whether or not use npu flash attention",
    )
    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(
        "--max_data_loader_n_workers",
        type=int,
        default=8,
        help="max num workers for DataLoader (lower is less main RAM usage, faster epoch start and slower data loading",
    )
    parser.add_argument("--support_dreambooth", action="store_true")
    parser.add_argument("--support_dropout", action="store_true")
    parser.add_argument(
        "--min_bucket_reso",
        type=int,
        default=512,
        help="minimum resolution for buckets",
    )
    parser.add_argument(
        "--max_bucket_reso",
        type=int,
        default=2048,
        help="maximum resolution for buckets",
    )
    parser.add_argument(
        "--enable_bucket",
        action="store_true",
        help="enable buckets for multi aspect ratio training",
    )
    parser.add_argument(
        "--max_token_length",
        type=int,
        default=None,
        choices=[None, 150, 225],
        help="max token length of text encoder (default for 75, 150 or 225)",
    )
    parser.add_argument(
        "--tokenizer_cache_dir",
        type=str,
        default=None,
        help="directory for caching Tokenizer (for offline training)",
    )
    parser.add_argument(
        "--no_half_vae",
        action="store_true",
        help="do not use fp16/bf16 VAE in mixed precision (use float VAE)",
    )
    parser.add_argument(
        "--half_vae",
        action="store_true",
        help="use fp16/bf16 VAE in mixed precision ",
    )

    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.")

    if args.proportion_empty_prompts < 0 or args.proportion_empty_prompts > 1:
        raise ValueError("`--proportion_empty_prompts` must be in the range [0, 1].")

    return args


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
    )

    accelerator = Accelerator(
        gradient_accumulation_steps=args.gradient_accumulation_steps,
        mixed_precision=args.mixed_precision,
        log_with=args.report_to,
        project_config=accelerator_project_config,
    )

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

    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 the tokenizers
    tokenizer_one = AutoTokenizer.from_pretrained(
        args.pretrained_model_name_or_path,
        subfolder="tokenizer",
        revision=args.revision,
        use_fast=False,
        local_files_only=True,
    )
    tokenizer_two = AutoTokenizer.from_pretrained(
        args.pretrained_model_name_or_path,
        subfolder="tokenizer_2",
        revision=args.revision,
        use_fast=False,
        local_files_only=True,
    )

    # import correct text encoder classes
    text_encoder_cls_one = import_model_class_from_model_name_or_path(
        args.pretrained_model_name_or_path, args.revision
    )
    text_encoder_cls_two = import_model_class_from_model_name_or_path(
        args.pretrained_model_name_or_path, args.revision, subfolder="text_encoder_2"
    )

    # Load scheduler and models
    noise_scheduler = DDPMScheduler.from_pretrained(
        args.pretrained_model_name_or_path, subfolder="scheduler", local_files_only=True
    )
    # Check for terminal SNR in combination with SNR Gamma
    text_encoder_one = text_encoder_cls_one.from_pretrained(
        args.pretrained_model_name_or_path,
        subfolder="text_encoder",
        revision=args.revision,
        variant=args.variant,
        local_files_only=True,
    )
    text_encoder_two = text_encoder_cls_two.from_pretrained(
        args.pretrained_model_name_or_path,
        subfolder="text_encoder_2",
        revision=args.revision,
        variant=args.variant,
        local_files_only=True,
    )
    vae_path = (
        args.pretrained_model_name_or_path
        if args.pretrained_vae_model_name_or_path is None
        else args.pretrained_vae_model_name_or_path
    )
    vae = AutoencoderKL.from_pretrained(
        vae_path,
        subfolder="vae" if args.pretrained_vae_model_name_or_path is None else None,
        revision=args.revision,
        variant=args.variant,
        local_files_only=True,
    )
    unet = UNet2DConditionModel.from_pretrained(
        args.pretrained_model_name_or_path,
        subfolder="unet",
        revision=args.revision,
        variant=args.variant,
        local_files_only=True,
    )

    # Freeze vae and text encoders.
    vae.requires_grad_(False)
    text_encoder_one.requires_grad_(False)
    text_encoder_two.requires_grad_(False)

    # Set unet as trainable.

    # For mixed precision training we cast all non-trainable weights 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, vae and text_encoder to device and cast to weight_dtype
    # The VAE is in float32 to avoid NaN losses.

    if args.half_vae:
        vae.to(accelerator.device, dtype=weight_dtype)
        image_dtype = weight_dtype
    else:
        vae.to(accelerator.device, dtype=torch.float32)
        image_dtype = torch.float32
    text_encoder_one.to(accelerator.device, dtype=weight_dtype)
    text_encoder_two.to(accelerator.device, dtype=weight_dtype)

    if args.enable_npu_flash_attention:
        if is_torch_npu_available():
            logger.info("npu flash attention enabled.")
            unet.enable_npu_flash_attention()
        else:
            raise ValueError(
                "npu flash attention requires torch_npu extensions and is supported only on npu devices"
            )

    # Create EMA for the unet.
    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."
                )
            unet.enable_xformers_memory_efficient_attention()
        else:
            raise ValueError(
                "xformers is not available. Make sure it is installed correctly"
            )

    sdxl_model = pretrain_model.SdxlPretrainModels(
        args,
        unet=unet,
        text_encoder1=text_encoder_one,
        text_encoder2=text_encoder_two,
        weight_dtype=weight_dtype,
    )

    # `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:
                for i, model in enumerate(models):
                    model.save_pretrained(os.path.join(output_dir))

                    # make sure to pop weight so that corresponding model is not saved again
                    if len(weights) > 0:
                        weights.pop()

        def load_model_hook(models, input_dir):
            for _ in range(len(models)):
                # pop models so that they are not loaded again
                model = models.pop()

                # load diffusers style into model
                unet_load_model = UNet2DConditionModel.from_pretrained(
                    input_dir, subfolder="unet", local_files_only=True
                )
                tex1_load_model = CLIPTextModel.from_pretrained(
                    input_dir, subfolder="text_encoder", local_files_only=True
                )
                tex2_load_model = CLIPTextModelWithProjection.from_pretrained(
                    input_dir, subfolder="text_encoder_2", local_files_only=True
                )

                model.unet.register_to_config(**unet_load_model.config)
                model.text_encoder1.register_to_config(**tex1_load_model.config)
                model.text_encoder2.register_to_config(**tex2_load_model.config)

                model.unte.load_state_dict(unet_load_model.state_dict())
                model.text_encoder1.load_state_dict(tex1_load_model.state_dict())
                model.text_encoder2.load_state_dict(tex2_load_model.state_dict())

                del unet_load_model, tex1_load_model, tex2_load_model

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

    if args.gradient_checkpointing:
        sdxl_model.unet.enable_gradient_checkpointing()

    # Enable TF32 for faster training on Ampere GPUs,
    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
        )

    optimizer_class = torch.optim.AdamW

    # Optimizer creation
    params_to_optimize = sdxl_model.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.
    if args.dataset_name is not None:
        user_config = {
            "datasets": [
                {
                    "subsets": collect_dataset.generate_dreambooth_subsets_config_by_subdirs(
                        args.dataset_name,
                    )
                }
            ]
        }
        blueprint_generator = collect_dataset.BlueprintGenerator(
            collect_dataset.ConfigSanitizer(True, True, False, True)
        )
        blueprint = blueprint_generator.generate(
            user_config, args, tokenizer=[tokenizer_one, tokenizer_two]
        )
        train_dataset_group = collect_dataset.generate_dataset_group_by_blueprint(
            blueprint.dataset_group
        )
    else:
        raise NotImplementedError("Dataset is only support Laion now")

    current_epoch = Value("i", 0)
    current_step = Value("i", 0)
    ds_for_collator = (
        train_dataset_group if args.max_data_loader_n_workers == 0 else None
    )
    collator = collect_dataset.collator_class(
        current_epoch, current_step, ds_for_collator
    )

    train_dataset_group.verify_bucket_reso_steps(32)
    # Preprocessing the datasets.

    n_workers = min(args.max_data_loader_n_workers, os.cpu_count() - 1)
    # DataLoaders creation:
    train_dataloader = torch.utils.data.DataLoader(
        train_dataset_group,
        shuffle=True,
        collate_fn=collator,
        batch_size=1,
        num_workers=n_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`.
    sdxl_model, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
        sdxl_model, 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-pretrain-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("  Num examples = %d", len(train_dataset_group))
    logger.info("  Num Epochs = %d", args.num_train_epochs)
    logger.info("  Instantaneous batch size per device = %d", args.train_batch_size)
    logger.info(
        "  Total train batch size (w. parallel, distributed & accumulation) = %d",
        total_batch_size,
    )
    logger.info("  Gradient Accumulation steps = %d", args.gradient_accumulation_steps)
    logger.info("  Total optimization steps = %d", 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,
    )

    profile = Profile(
        start_step=int(os.getenv("PROFILE_START_STEP", 10)),
        profile_type=os.getenv("PROFILE_TYPE"),
    )
    for epoch in range(first_epoch, args.num_train_epochs):
        train_loss = 0.0
        sdxl_model.train()
        current_epoch.value = epoch + 1
        step_end_time = time.time()

        for step, batch in enumerate(train_dataloader):
            step_data_time = time.time() - step_end_time
            current_step.value = global_step
            profile.start()
            with accelerator.accumulate(sdxl_model):
                # Sample noise that we'll add to the latents
                if "latents" in batch and batch["latents"] is not None:
                    latents = (
                        batch["latents"].to(accelerator.device).to(dtype=weight_dtype)
                    )
                else:
                    with torch.no_grad():
                        # latentに変換
                        latents = (
                            vae.encode(batch["images"].to(image_dtype))
                            .latent_dist.sample()
                            .to(weight_dtype)
                        )

                        # NaNが含まれていれば警告を表示し0に置き換える
                        if torch.any(torch.isnan(latents)):
                            accelerator.print(
                                "NaN found in latents, replacing with zeros"
                            )
                            latents = torch.nan_to_num(latents, 0, out=latents)
                latents = latents * vae.config.scaling_factor
                input_ids1 = batch["input_ids"]
                input_ids2 = batch["input_ids2"]
                with torch.set_grad_enabled(True):
                    # Get the text embedding for conditioning
                    input_ids1 = input_ids1.to(accelerator.device)
                    input_ids2 = input_ids2.to(accelerator.device)
                    # unwrap_model is fine for models not wrapped by accelerator
                    encoder_hidden_states1, encoder_hidden_states2, pool2 = (
                        pretrain_model.get_hidden_states_sdxl(
                            args.max_token_length,
                            input_ids1,
                            input_ids2,
                            tokenizer_one,
                            tokenizer_two,
                            sdxl_model.text_encoder1,
                            sdxl_model.text_encoder2,
                            None,
                        )
                    )
                noise_pred, noise, timesteps = sdxl_model(
                    batch,
                    accelerator,
                    noise_scheduler,
                    latents,
                    epoch,
                    step,
                    encoder_hidden_states1,
                    encoder_hidden_states2,
                    pool2,
                )

                # 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":
                    target = noise
                elif noise_scheduler.config.prediction_type == "v_prediction":
                    target = noise_scheduler.get_velocity(latents, noise, timesteps)
                elif noise_scheduler.config.prediction_type == "sample":
                    # We set the target to latents here, but the model_pred will return the noise sample prediction.
                    target = latents
                    # We will have to subtract the noise residual from the prediction to get the target sample.
                    model_pred = noise_pred - noise
                else:
                    raise ValueError(
                        f"Unknown prediction type {noise_scheduler.config.prediction_type}"
                    )

                if args.snr_gamma is None:
                    loss = F.mse_loss(
                        noise_pred.float(), target.float(), reduction="mean"
                    )
                else:
                    # Compute loss-weights
                    # Since we predict the noise instead of x_0, the original formulation is slightly changed.
                    # This is discussed in Section 4.2 of the same paper.
                    snr = compute_snr(noise_scheduler, timesteps)
                    if noise_scheduler.config.prediction_type == "v_prediction":
                        # Velocity objective requires that we add one to SNR values before we divide by them.
                        snr = snr + 1
                    mse_loss_weights = (
                        torch.stack(
                            [snr, args.snr_gamma * torch.ones_like(timesteps)], dim=1
                        ).min(dim=1)[0]
                        / snr
                    )

                    loss = F.mse_loss(
                        model_pred.float(), target.float(), reduction="none"
                    )
                    loss = (
                        loss.mean(dim=list(range(1, len(loss.shape))))
                        * mse_loss_weights
                    ).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 = sdxl_model.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
                train_loss = 0.0

                if accelerator.is_main_process or accelerator.distributed_type == DistributedType.DEEPSPEED:
                    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(
                                    "%d checkpoints already exist, removing %d checkpoints",
                                    (len(checkpoints), len(removing_checkpoints)),
                                )
                                logger.info(
                                    "removing checkpoints: %s",
                                    ", ".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)
                        accelerator.wait_for_everyone()
                        logger.info("Saved state to %s", save_path)

            profile.end()
            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

            step_total_time = time.time() - step_end_time
            accelerator.print(f"step_train_time: {step_total_time}")
            accelerator.print(f"step_data_time: {step_data_time}")
            accelerator.print(
                f"FPS: {args.train_batch_size * accelerator.num_processes / (step_total_time - step_data_time)}"
            )
            step_end_time = time.time()

    accelerator.wait_for_everyone()
    if accelerator.is_main_process:
        unet = accelerator.unwrap_model(sdxl_model.unet)
        text_encoder = accelerator.unwrap_model(sdxl_model.text_encoder1)
        text_encoder_bak = CLIPTextModel.from_pretrained(
            args.pretrained_model_name_or_path, subfolder="text_encoder", local_files_only=True
        )
        text_encoder_bak.load_state_dict(text_encoder.state_dict(), strict=False)

        text_encoder_2 = accelerator.unwrap_model(sdxl_model.text_encoder2)
        text_encoder_2_bak = CLIPTextModelWithProjection.from_pretrained(
            args.pretrained_model_name_or_path, subfolder="text_encoder_2", local_files_only=True
        )
        text_encoder_2_bak.load_state_dict(text_encoder_2.state_dict(), strict=False)

        # Serialize pipeline.
        vae = AutoencoderKL.from_pretrained(
            vae_path,
            subfolder="vae" if args.pretrained_vae_model_name_or_path is None else None,
            revision=args.revision,
            variant=args.variant,
            torch_dtype=weight_dtype,
            local_files_only=True,
        )
        pipeline = StableDiffusionXLPipeline.from_pretrained(
            args.pretrained_model_name_or_path,
            unet=unet,
            vae=vae,
            text_encoder=text_encoder_bak,
            text_encoder_2=text_encoder_2_bak,
            revision=args.revision,
            variant=args.variant,
            torch_dtype=weight_dtype,
            local_files_only=True,
        )
        if args.prediction_type is not None:
            scheduler_args = {"prediction_type": args.prediction_type}
            pipeline.scheduler = pipeline.scheduler.from_config(
                pipeline.scheduler.config, **scheduler_args
            )
        pipeline.save_pretrained(args.output_dir)

        del pipeline

    accelerator.end_training()


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