train.py

#!/usr/bin/env python
# coding=utf-8
# Copyright 2024 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 Lciense is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and


import argparse
import gc
import itertools
import json
import logging
import math
import os
import random
import shutil
import warnings
from contextlib import nullcontext
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 (
    DistributedDataParallelKwargs,
    ProjectConfiguration,
    set_seed,
)
from huggingface_hub import create_repo, hf_hub_download, upload_folder
from huggingface_hub.utils import insecure_hashlib
from packaging import version
from peft import LoraConfig, set_peft_model_state_dict
from peft.utils import get_peft_model_state_dict
from PIL import Image
from PIL.ImageOps import exif_transpose
from safetensors.torch import load_file, save_file
from torch.utils.data import Dataset
from torchvision import transforms
from torchvision.transforms.functional import crop
from tqdm.auto import tqdm
from transformers import AutoTokenizer, PretrainedConfig

import diffusers
from diffusers import (
    AutoencoderKL,
    DDPMScheduler,
    DPMSolverMultistepScheduler,
    EDMEulerScheduler,
    EulerDiscreteScheduler,
    StableDiffusionXLPipeline,
    UNet2DConditionModel,
)
from diffusers.loaders import LoraLoaderMixin
from diffusers.optimization import get_scheduler
from diffusers.training_utils import (
    _set_state_dict_into_text_encoder,
    cast_training_params,
    compute_snr,
)
from diffusers.utils import (
    check_min_version,
    convert_all_state_dict_to_peft,
    convert_state_dict_to_diffusers,
    convert_state_dict_to_kohya,
    convert_unet_state_dict_to_peft,
    is_wandb_available,
)
from diffusers.utils.hub_utils import load_or_create_model_card, populate_model_card
from diffusers.utils.import_utils import is_xformers_available
from diffusers.utils.torch_utils import is_compiled_module


if is_wandb_available():
    import wandb

# Will error if the minimal version of diffusers is not installed. Remove at your own risks.
# check_min_version("0.28.0.dev0")

logger = get_logger(__name__)

from utils import extend_mask_with_bezier, mask_paint2bbox


def determine_scheduler_type(pretrained_model_name_or_path, revision):
    model_index_filename = "model_index.json"
    if os.path.isdir(pretrained_model_name_or_path):
        model_index = os.path.join(pretrained_model_name_or_path, model_index_filename)
    else:
        model_index = hf_hub_download(
            repo_id=pretrained_model_name_or_path,
            filename=model_index_filename,
            revision=revision,
        )

    with open(model_index, "r") as f:
        scheduler_type = json.load(f)["scheduler"][1]
    return scheduler_type


def parse_args(input_args=None):
    parser = argparse.ArgumentParser(description="Simple example of a training script.")

    # parser.add_argument(
    #     "--ref_image_path",
    #     type=str,
    #     default=None,
    #     required=True,
    #     help="Path to ref_image_path",
    # )

    parser.add_argument(
        "--category",
        type=str,
        default="",
        help="the category of the image, for decompose",
    )
    parser.add_argument(
        "--regular_prob",
        type=float,
        default=0.5,
        help="probability of using regular image",
    )

    parser.add_argument(
        "--regular_dir",
        type=str,
        default=None,
        required=True,
        help="Path to regular img and mask",
    )

    parser.add_argument(
        "--loss_reweight_object",
        type=float,
        default=1,
    )

    parser.add_argument(
        "--loss_reweight_background",
        type=float,
        default=1,
    )

    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) containing the training data of instance images (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(
        "--instance_data_dir",
        type=str,
        default=None,
        help=("A folder containing the training data. "),
    )

    parser.add_argument(
        "--cache_dir",
        type=str,
        default=None,
        help="The directory where the downloaded models and datasets will be stored.",
    )

    parser.add_argument(
        "--image_column",
        type=str,
        default="image",
        help="The column of the dataset containing the target image. By "
        "default, the standard Image Dataset maps out 'file_name' "
        "to 'image'.",
    )
    parser.add_argument(
        "--caption_column",
        type=str,
        default=None,
        help="The column of the dataset containing the instance prompt for each image",
    )

    parser.add_argument(
        "--repeats",
        type=int,
        default=1,
        help="How many times to repeat the training data.",
    )

    parser.add_argument(
        "--class_data_dir",
        type=str,
        default=None,
        required=False,
        help="A folder containing the training data of class images.",
    )
    parser.add_argument(
        "--instance_prompt",
        type=str,
        default=None,
        required=True,
        help="The prompt with identifier specifying the instance, e.g. 'photo of a TOK dog', 'in the style of TOK'",
    )
    parser.add_argument(
        "--class_prompt",
        type=str,
        default=None,
        help="The prompt to specify images in the same class as provided instance images.",
    )
    parser.add_argument(
        "--validation_prompt",
        type=str,
        default=None,
        help="A prompt that is used during validation to verify that the model is learning.",
    )
    parser.add_argument(
        "--num_validation_images",
        type=int,
        default=4,
        help="Number of images that should be generated during validation with `validation_prompt`.",
    )
    parser.add_argument(
        "--validation_epochs",
        type=int,
        default=50,
        help=(
            "Run dreambooth validation every X epochs. Dreambooth validation consists of running the prompt"
            " `args.validation_prompt` multiple times: `args.num_validation_images`."
        ),
    )
    parser.add_argument(
        "--do_edm_style_training",
        default=False,
        action="store_true",
        help="Flag to conduct training using the EDM formulation as introduced in https://arxiv.org/abs/2206.00364.",
    )
    parser.add_argument(
        "--with_prior_preservation",
        default=False,
        action="store_true",
        help="Flag to add prior preservation loss.",
    )
    parser.add_argument(
        "--prior_loss_weight",
        type=float,
        default=1.0,
        help="The weight of prior preservation loss.",
    )
    parser.add_argument(
        "--num_class_images",
        type=int,
        default=100,
        help=(
            "Minimal class images for prior preservation loss. If there are not enough images already present in"
            " class_data_dir, additional images will be sampled with class_prompt."
        ),
    )
    parser.add_argument(
        "--output_dir",
        type=str,
        default="lora-dreambooth-model",
        help="The output directory where the model predictions and checkpoints will be written.",
    )
    parser.add_argument(
        "--output_kohya_format",
        action="store_true",
        help="Flag to additionally generate final state dict in the Kohya format so that it becomes compatible with A111, Comfy, Kohya, etc.",
    )
    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_text_encoder",
        action="store_true",
        help="Whether to train the text encoder. If set, the text encoder should be float32 precision.",
    )
    parser.add_argument(
        "--train_batch_size",
        type=int,
        default=4,
        help="Batch size (per device) for the training dataloader.",
    )
    parser.add_argument(
        "--sample_batch_size",
        type=int,
        default=4,
        help="Batch size (per device) for sampling images.",
    )
    parser.add_argument("--num_train_epochs", type=int, default=1)
    parser.add_argument(
        "--max_train_steps",
        type=int,
        default=None,
        help="Total number of training steps to perform.  If provided, overrides num_train_epochs.",
    )
    parser.add_argument(
        "--checkpointing_steps",
        type=int,
        default=500,
        help=(
            "Save a checkpoint of the training state every X updates. These checkpoints can be used both as final"
            " checkpoints in case they are better than the last checkpoint, and are also suitable for resuming"
            " training using `--resume_from_checkpoint`."
        ),
    )
    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(
        "--text_encoder_lr",
        type=float,
        default=5e-6,
        help="Text encoder learning rate 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(
        "--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(
        "--lr_warmup_steps",
        type=int,
        default=500,
        help="Number of steps for the warmup in the lr scheduler.",
    )
    parser.add_argument(
        "--lr_num_cycles",
        type=int,
        default=1,
        help="Number of hard resets of the lr in cosine_with_restarts scheduler.",
    )
    parser.add_argument(
        "--lr_power",
        type=float,
        default=1.0,
        help="Power factor of the polynomial scheduler.",
    )
    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(
        "--optimizer",
        type=str,
        default="AdamW",
        help=('The optimizer type to use. Choose between ["AdamW", "prodigy"]'),
    )

    parser.add_argument(
        "--use_8bit_adam",
        action="store_true",
        help="Whether or not to use 8-bit Adam from bitsandbytes. Ignored if optimizer is not set to AdamW",
    )

    parser.add_argument(
        "--adam_beta1",
        type=float,
        default=0.9,
        help="The beta1 parameter for the Adam and Prodigy optimizers.",
    )
    parser.add_argument(
        "--adam_beta2",
        type=float,
        default=0.999,
        help="The beta2 parameter for the Adam and Prodigy optimizers.",
    )
    parser.add_argument(
        "--prodigy_beta3",
        type=float,
        default=None,
        help="coefficients for computing the Prodidy stepsize using running averages. If set to None, "
        "uses the value of square root of beta2. Ignored if optimizer is adamW",
    )
    parser.add_argument(
        "--prodigy_decouple",
        type=bool,
        default=True,
        help="Use AdamW style decoupled weight decay",
    )
    parser.add_argument(
        "--adam_weight_decay",
        type=float,
        default=1e-04,
        help="Weight decay to use for unet params",
    )
    parser.add_argument(
        "--adam_weight_decay_text_encoder",
        type=float,
        default=1e-03,
        help="Weight decay to use for text_encoder",
    )

    parser.add_argument(
        "--adam_epsilon",
        type=float,
        default=1e-08,
        help="Epsilon value for the Adam optimizer and Prodigy optimizers.",
    )

    parser.add_argument(
        "--prodigy_use_bias_correction",
        type=bool,
        default=True,
        help="Turn on Adam's bias correction. True by default. Ignored if optimizer is adamW",
    )
    parser.add_argument(
        "--prodigy_safeguard_warmup",
        type=bool,
        default=True,
        help="Remove lr from the denominator of D estimate to avoid issues during warm-up stage. True by default. "
        "Ignored if optimizer is adamW",
    )
    parser.add_argument(
        "--max_grad_norm", default=1.0, type=float, help="Max gradient norm."
    )
    parser.add_argument(
        "--hub_token",
        type=str,
        default=None,
        help="The token to use to push to the Model Hub.",
    )
    parser.add_argument(
        "--hub_model_id",
        type=str,
        default=None,
        help="The name of the repository to keep in sync with the local `output_dir`.",
    )
    parser.add_argument(
        "--logging_dir",
        type=str,
        default="logs",
        help=(
            "[TensorBoard](https://www.tensorflow.org/tensorboard) log directory. Will default to"
            " *output_dir/runs/**CURRENT_DATETIME_HOSTNAME***."
        ),
    )
    parser.add_argument(
        "--allow_tf32",
        action="store_true",
        help=(
            "Whether or not to allow TF32 on Ampere GPUs. Can be used to speed up training. For more information, see"
            " https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices"
        ),
    )
    parser.add_argument(
        "--report_to",
        type=str,
        default="tensorboard",
        help=(
            'The integration to report the results and logs to. Supported platforms are `"tensorboard"`'
            ' (default), `"wandb"` and `"comet_ml"`. Use `"all"` to report to all integrations.'
        ),
    )
    parser.add_argument(
        "--mixed_precision",
        type=str,
        default=None,
        choices=["no", "fp16", "bf16"],
        help=(
            "Whether to use mixed precision. Choose between fp16 and bf16 (bfloat16). Bf16 requires PyTorch >="
            " 1.10.and an Nvidia Ampere GPU.  Default to the value of accelerate config of the current system or the"
            " flag passed with the `accelerate.launch` command. Use this argument to override the accelerate config."
        ),
    )
    parser.add_argument(
        "--prior_generation_precision",
        type=str,
        default=None,
        choices=["no", "fp32", "fp16", "bf16"],
        help=(
            "Choose prior generation precision between fp32, fp16 and bf16 (bfloat16). Bf16 requires PyTorch >="
            " 1.10.and an Nvidia Ampere GPU.  Default to  fp16 if a GPU is available else fp32."
        ),
    )
    parser.add_argument(
        "--local_rank",
        type=int,
        default=-1,
        help="For distributed training: local_rank",
    )
    parser.add_argument(
        "--enable_xformers_memory_efficient_attention",
        action="store_true",
        help="Whether or not to use xformers.",
    )
    parser.add_argument(
        "--rank",
        type=int,
        default=4,
        help=("The dimension of the LoRA update matrices."),
    )
    parser.add_argument(
        "--use_dora",
        action="store_true",
        default=False,
        help=(
            "Wether to train a DoRA as proposed in- DoRA: Weight-Decomposed Low-Rank Adaptation https://arxiv.org/abs/2402.09353. "
            "Note: to use DoRA you need to install peft from main, `pip install git+https://github.com/huggingface/peft.git`"
        ),
    )

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

    if args.dataset_name is None and args.instance_data_dir is None:
        raise ValueError("Specify either `--dataset_name` or `--instance_data_dir`")

    if args.dataset_name is not None and args.instance_data_dir is not None:
        raise ValueError(
            "Specify only one of `--dataset_name` or `--instance_data_dir`"
        )

    env_local_rank = int(os.environ.get("LOCAL_RANK", -1))
    if env_local_rank != -1 and env_local_rank != args.local_rank:
        args.local_rank = env_local_rank

    if args.with_prior_preservation:
        if args.class_data_dir is None:
            raise ValueError("You must specify a data directory for class images.")
        if args.class_prompt is None:
            raise ValueError("You must specify prompt for class images.")
    else:
        # logger is not available yet
        if args.class_data_dir is not None:
            warnings.warn(
                "You need not use --class_data_dir without --with_prior_preservation."
            )
        if args.class_prompt is not None:
            warnings.warn(
                "You need not use --class_prompt without --with_prior_preservation."
            )

    return args


class DreamBoothDataset(Dataset):
    """
    A dataset to prepare the instance and class images with the prompts for fine-tuning the model.
    It pre-processes the images.
    """

    def __init__(
        self,
        instance_data_root,
        instance_prompt,
        class_prompt,
        class_data_root=None,
        class_num=None,
        size=1024,
        repeats=1,
        center_crop=False,
    ):
        self.size = size
        self.center_crop = center_crop

        self.instance_prompt = instance_prompt
        self.custom_instance_prompts = None
        self.class_prompt = class_prompt


        import modules.inpaint_worker as inpaint_worker
        from utils.add_fooocus_inpaint_head_patch import load_inpaint_head

        global pipe
        _device, _dtype = pipe.vae.device, pipe.vae.dtype
        inpaint_head_model = load_inpaint_head(
            "./models/fooocus_inpaint/fooocus_inpaint_head.pth", _device
        ).to(device=_device, dtype=_dtype)

        print("Processing image and mask to latent")

        train_resize = transforms.Resize(
            size, interpolation=transforms.InterpolationMode.BILINEAR
        )
        train_crop = (
            transforms.CenterCrop(size) if center_crop else transforms.RandomCrop(size)
        )
        train_flip = transforms.RandomHorizontalFlip(p=1.0)
        train_transforms = transforms.Compose(
            [
                transforms.ToTensor(),
                transforms.Normalize([0.5], [0.5]),
            ]
        )

        INSTANCE_REPEAT_NUM = 10
        REGULAR_REPEAT_NUM = 3
        USE_BBOX_MASK = False



        # image processing to prepare for using SD-XL micro-conditioning
        self.original_sizes = []
        self.crop_top_lefts = []
        self.pixel_values = []

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

            self.instance_images = []
            self.instance_masks = []
            # self.ref_images = []

            for path in list(Path(instance_data_root).iterdir()):
                image = Image.open(path)
                mask_path = str(path).replace("/image", "/mask").replace(".jpg", ".png")
                if not Path(mask_path).exists():
                    print(f"Warning! mask of image {path} not exist! Ignore..")
                    continue
                mask_img = Image.open(mask_path).resize(image.size)
                self.instance_images.extend(itertools.repeat(image, repeats))
                self.instance_masks.extend(itertools.repeat(mask_img, repeats))
                # break

            with torch.no_grad():  #
                for _ in tqdm(
                    range(INSTANCE_REPEAT_NUM)
                ):  
                    for image, mask in zip(self.instance_images, self.instance_masks):

                        image = exif_transpose(image)
                        mask = exif_transpose(mask)
                        if not image.mode == "RGB":
                            image = image.convert("RGB")

                        self.original_sizes.append((image.height, image.width))
                        image = train_resize(image)
                        mask = train_resize(mask)
                        if args.random_flip and random.random() < 0.5:
                            # flip
                            image = train_flip(image)
                            mask = train_flip(mask)
                        if args.center_crop:
                            y1 = max(
                                0, int(round((image.height - args.resolution) / 2.0))
                            )
                            x1 = max(
                                0, int(round((image.width - args.resolution) / 2.0))
                            )
                            image = train_crop(image)
                            mask = train_crop(mask)
                        else:
                            y1, x1, h, w = train_crop.get_params(
                                image, (args.resolution, args.resolution)
                            )
                            image = crop(image, y1, x1, h, w)
                            mask = crop(mask, y1, x1, h, w)
                        crop_top_left = (y1, x1)
                        self.crop_top_lefts.append(crop_top_left)

                        mask = extend_mask_with_bezier(
                            (
                                np.asarray(mask)[:, :, 0]
                                if mask.mode == "RGB"
                                else np.asarray(mask)
                            ),
                            extend_ratio=random.randint(1, 10) / 10.0,
                            random_width=random.randint(1, 20),
                        )
                        
                        if USE_BBOX_MASK:
                            mask = mask_paint2bbox(mask)


                        mask = Image.fromarray(mask)
                    
                        for interesting_are in [1]:
                            inpaint_work = inpaint_worker.InpaintWorker(
                                image=np.asarray(image),
                                mask=(
                                    np.asarray(mask)[:, :, 0]
                                    if mask.mode == "RGB"
                                    else np.asarray(mask)
                                ),
                                use_fill=True,
                                k=interesting_are,
                                path_upscale_models="./models/upscale_models/fooocus_upscaler_s409985e5.bin",
                            )

                            pixel_fill = pipe.image_processor.preprocess(
                                Image.fromarray(inpaint_work.interested_fill)
                            )
                            pixel_image = pipe.image_processor.preprocess(
                                Image.fromarray(inpaint_work.interested_image)
                            )
                            pixel_mask = pipe.mask_processor.preprocess(
                                Image.fromarray(inpaint_work.interested_mask)
                            )

                            pixel_fill = pixel_fill.to(device=_device, dtype=_dtype)
                            pixel_image = pixel_image.to(device=_device, dtype=_dtype)
                            pixel_mask = pixel_mask.to(
                                device=_device, dtype=torch.float32
                            )

                            masked_image = pixel_image * (pixel_mask < 0.5)
                            latent_fill = pipe._encode_vae_image(
                                image=pixel_fill, generator=None
                            )
                            latent_masked_image = pipe._encode_vae_image(
                                image=masked_image, generator=None
                            )
                            latent_image = pipe._encode_vae_image(
                                image=pixel_image, generator=None
                            )
                            B, C, H, W = latent_fill.shape
                            latent_mask = torch.nn.functional.interpolate(
                                pixel_mask, size=(H * 8, W * 8), mode="bilinear"
                            ).round()
                            latent_mask = (
                                torch.nn.functional.max_pool2d(latent_mask, (8, 8))
                                .round()
                                .to(latent_fill.device)
                            )

                            feed = torch.cat(
                                [latent_mask, latent_masked_image],
                                dim=1,
                            )
                            inpaint_head_feature = inpaint_head_model(feed).to("cpu")

                            # self.pixel_values.append((inpaint_head_feature, pixel_image))
                            self.pixel_values.append(
                                (
                                    inpaint_head_feature,
                                    latent_image,
                                    latent_mask,
                                    "instance",
                                )
                            )

            self.num_instance_images = len(self.pixel_values)

            # ! get regular image
            global regular_prompts_list
            print("Processing regular image and mask to latent")
            regular_data_root = args.regular_dir
            regular_prompts_list = [name for name in os.listdir(regular_data_root)]

            if args.regular_prob > 0:
                with torch.no_grad():  
                    # ! load regular image and mask
                    for prompt in tqdm(regular_prompts_list):
                        all_images_and_masks = os.listdir(
                            os.path.join(regular_data_root, prompt)
                        )
                        files_name = set(
                            [name.split("-")[0] for name in all_images_and_masks]
                        )

                        for file_name in tqdm(files_name):
                            _image = Image.open(
                                f"{regular_data_root}/{prompt}/{file_name}-image.png"
                            )
                            _mask = Image.open(
                                f"{regular_data_root}/{prompt}/{file_name}-mask.png"
                            )
                            for _ in range(REGULAR_REPEAT_NUM):
                                image = exif_transpose(_image)
                                mask = exif_transpose(_mask)
                                if not image.mode == "RGB":
                                    image = image.convert("RGB")

                                self.original_sizes.append((image.height, image.width))
                                image = train_resize(image)
                                mask = train_resize(mask)
                                if args.random_flip and random.random() < 0.5:
                                    # flip
                                    image = train_flip(image)
                                    mask = train_flip(mask)
                                if args.center_crop:
                                    y1 = max(
                                        0,
                                        int(
                                            round(
                                                (image.height - args.resolution) / 2.0
                                            )
                                        ),
                                    )
                                    x1 = max(
                                        0,
                                        int(
                                            round((image.width - args.resolution) / 2.0)
                                        ),
                                    )
                                    image = train_crop(image)
                                    mask = train_crop(mask)
                                else:
                                    y1, x1, h, w = train_crop.get_params(
                                        image, (args.resolution, args.resolution)
                                    )
                                    image = crop(image, y1, x1, h, w)
                                    mask = crop(mask, y1, x1, h, w)
                                crop_top_left = (y1, x1)
                                self.crop_top_lefts.append(crop_top_left)

                                mask = extend_mask_with_bezier(
                                    (
                                        np.asarray(mask)[:, :, 0]
                                        if mask.mode == "RGB"
                                        else np.asarray(mask)
                                    ),
                                    extend_ratio=random.randint(1, 10) / 10.0,
                                    random_width=random.randint(1, 20),
                                )
                                
                                if USE_BBOX_MASK:
                                    mask = mask_paint2bbox(mask)
                                

                                mask = Image.fromarray(mask)
                                
                                interesting_are = 1
                                inpaint_work = inpaint_worker.InpaintWorker(
                                    image=np.asarray(image),
                                    mask=(
                                        np.asarray(mask)[:, :, 0]
                                        if mask.mode == "RGB"
                                        else np.asarray(mask)
                                    ),
                                    use_fill=True,
                                    k=interesting_are,
                                    path_upscale_models="./models/upscale_models/fooocus_upscaler_s409985e5.bin",
                                )

                                pixel_fill = pipe.image_processor.preprocess(
                                    Image.fromarray(inpaint_work.interested_fill)
                                )
                                pixel_image = pipe.image_processor.preprocess(
                                    Image.fromarray(inpaint_work.interested_image)
                                )
                                pixel_mask = pipe.mask_processor.preprocess(
                                    Image.fromarray(inpaint_work.interested_mask)
                                )

                                pixel_fill = pixel_fill.to(device=_device, dtype=_dtype)
                                pixel_image = pixel_image.to(
                                    device=_device, dtype=_dtype
                                )
                                pixel_mask = pixel_mask.to(
                                    device=_device, dtype=torch.float32
                                )

                                masked_image = pixel_image * (pixel_mask < 0.5)
                                latent_fill = pipe._encode_vae_image(
                                    image=pixel_fill, generator=None
                                )
                                latent_masked_image = pipe._encode_vae_image(
                                    image=masked_image, generator=None
                                )
                                latent_image = pipe._encode_vae_image(
                                    image=pixel_image, generator=None
                                )
                                B, C, H, W = latent_fill.shape
                                latent_mask = torch.nn.functional.interpolate(
                                    pixel_mask, size=(H * 8, W * 8), mode="bilinear"
                                ).round()
                                latent_mask = (
                                    torch.nn.functional.max_pool2d(latent_mask, (8, 8))
                                    .round()
                                    .to(latent_fill.device)
                                )

                                feed = torch.cat(
                                    [latent_mask, latent_masked_image],
                                    dim=1,
                                )
                                inpaint_head_feature = inpaint_head_model(feed).to(
                                    "cpu"
                                )

                                self.pixel_values.append(
                                    (
                                        inpaint_head_feature,
                                        latent_image,
                                        latent_mask,
                                        prompt,
                                    )
                                )

        inpaint_head_model.to("cpu")
        del inpaint_head_model
        if torch.cuda.is_available():
            torch.cuda.empty_cache()

        self.num_all_images = len(self.pixel_values)
        self.num_regular_images = self.num_all_images - self.num_instance_images
        self._length = self.num_all_images

        self.class_data_root = None

    def __len__(self):
        return self._length

    def __getitem__(self, index):
        example = {}

        self.num_regular_images
        self.num_all_images
        self.num_instance_images

        regular_prob = args.regular_prob
        if random.random() < regular_prob and self.num_regular_images > 0:
            img_id = self.num_instance_images + index % self.num_regular_images
        else:
            img_id = index % self.num_instance_images

        instance_image = self.pixel_values[img_id]
        original_size = self.original_sizes[img_id]
        crop_top_left = self.crop_top_lefts[img_id]

        example["instance_images"] = instance_image
        example["original_size"] = original_size
        example["crop_top_left"] = crop_top_left

        example["instance_prompt"] = self.instance_prompt

        return example


def collate_fn(examples, with_prior_preservation=False):
    pixel_values = [example["instance_images"] for example in examples]
    prompts = [example["instance_prompt"] for example in examples]
    original_sizes = [example["original_size"] for example in examples]
    crop_top_lefts = [example["crop_top_left"] 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:
        pixel_values += [example["class_images"] for example in examples]
        prompts += [example["class_prompt"] for example in examples]
        original_sizes += [example["original_size"] for example in examples]
        crop_top_lefts += [example["crop_top_left"] for example in examples]

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

    batch = {
        "pixel_values": pixel_values,
        "prompts": prompts,
        "original_sizes": original_sizes,
        "crop_top_lefts": crop_top_lefts,
    }
    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


def tokenize_prompt(tokenizer, prompt):
    text_inputs = tokenizer(
        prompt,
        padding="max_length",
        max_length=tokenizer.model_max_length,
        truncation=True,
        return_tensors="pt",
    )
    text_input_ids = text_inputs.input_ids
    return text_input_ids


# Adapted from pipelines.StableDiffusionXLPipeline.encode_prompt
def encode_prompt(text_encoders, tokenizers, prompt, text_input_ids_list=None):
    prompt_embeds_list = []

    for i, text_encoder in enumerate(text_encoders):
        if tokenizers is not None:
            tokenizer = tokenizers[i]
            text_input_ids = tokenize_prompt(tokenizer, prompt)
        else:
            assert text_input_ids_list is not None
            text_input_ids = text_input_ids_list[i]

        prompt_embeds = text_encoder(
            text_input_ids.to(text_encoder.device),
            output_hidden_states=True,
            return_dict=False,
        )

        # We are only ALWAYS interested in the pooled output of the final text encoder
        pooled_prompt_embeds = prompt_embeds[0]
        prompt_embeds = prompt_embeds[-1][-2]
        bs_embed, seq_len, _ = prompt_embeds.shape
        prompt_embeds = prompt_embeds.view(bs_embed, seq_len, -1)
        prompt_embeds_list.append(prompt_embeds)

    prompt_embeds = torch.concat(prompt_embeds_list, dim=-1)
    pooled_prompt_embeds = pooled_prompt_embeds.view(bs_embed, -1)
    return prompt_embeds, pooled_prompt_embeds


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

    if args.do_edm_style_training and args.snr_gamma is not None:
        raise ValueError(
            "Min-SNR formulation is not supported when conducting EDM-style training."
        )

    if torch.backends.mps.is_available() and args.mixed_precision == "bf16":
        # due to pytorch#99272, MPS does not yet support bfloat16.
        raise ValueError(
            "Mixed precision training with bfloat16 is not supported on MPS. Please use fp16 (recommended) or fp32 instead."
        )

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

    accelerator_project_config = ProjectConfiguration(
        project_dir=args.output_dir, logging_dir=logging_dir
    )
    kwargs = DistributedDataParallelKwargs(find_unused_parameters=True)
    accelerator = Accelerator(
        gradient_accumulation_steps=args.gradient_accumulation_steps,
        mixed_precision=args.mixed_precision,
        log_with=args.report_to,
        project_config=accelerator_project_config,
        kwargs_handlers=[kwargs],
    )

    # 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."
            )

    # 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)

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

    print("Loading sdxl pipeline")
    from FooocusSDXLInpaintAllInOnePipeline import FooocusSDXLInpaintPipeline

    global pipe
    pipe = FooocusSDXLInpaintPipeline.from_pretrained(
        "frankjoshua/juggernautXL_v8Rundiffusion",
        # torch_dtype=torch.float16,
        # variant="fp16",
        use_safetensors=True,
    ).to("cuda")
    # pipe.text_encoder.to("cpu") 
    # pipe.text_encoder_2.to("cpu")

    print("Loading fooocus unet")

    pipe.only_load_fooocus_unet_and_cover_pipe_unet_for_train(
        fooocus_model_path="./models/fooocus_inpaint/inpaint_v26.fooocus.patch",
    )

    # Load the tokenizers
    tokenizer_one = pipe.tokenizer
    tokenizer_two = pipe.tokenizer_2

    # Load scheduler and models
    scheduler_type = determine_scheduler_type(
        args.pretrained_model_name_or_path, args.revision
    )
    if "EDM" in scheduler_type:
        args.do_edm_style_training = True
        noise_scheduler = EDMEulerScheduler.from_pretrained(
            args.pretrained_model_name_or_path, subfolder="scheduler"
        )
        logger.info("Performing EDM-style training!")
    elif args.do_edm_style_training:
        noise_scheduler = EulerDiscreteScheduler.from_pretrained(
            args.pretrained_model_name_or_path, subfolder="scheduler"
        )
        logger.info("Performing EDM-style training!")
    else:
        noise_scheduler = DDPMScheduler.from_pretrained(
            args.pretrained_model_name_or_path, subfolder="scheduler"
        )

    text_encoder_one = pipe.text_encoder

    text_encoder_two = pipe.text_encoder_2

    vae = pipe.vae

    latents_mean = latents_std = None
    if hasattr(vae.config, "latents_mean") and vae.config.latents_mean is not None:
        latents_mean = torch.tensor(vae.config.latents_mean).view(1, 4, 1, 1)
    if hasattr(vae.config, "latents_std") and vae.config.latents_std is not None:
        latents_std = torch.tensor(vae.config.latents_std).view(1, 4, 1, 1)

    unet = pipe.unet 

    # We only train the additional adapter LoRA layers
    vae.requires_grad_(False)
    text_encoder_one.requires_grad_(False)
    text_encoder_two.requires_grad_(False)
    unet.requires_grad_(False)

    # For mixed precision training we cast all non-trainable weights (vae, non-lora text_encoder and non-lora unet) 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

    if torch.backends.mps.is_available() and weight_dtype == torch.bfloat16:
        # due to pytorch#99272, MPS does not yet support bfloat16.
        raise ValueError(
            "Mixed precision training with bfloat16 is not supported on MPS. Please use fp16 (recommended) or fp32 instead."
        )

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

    # The VAE is always in float32 to avoid NaN losses.
    vae.to(accelerator.device, dtype=torch.float32)

    text_encoder_one.to(accelerator.device, dtype=weight_dtype)
    text_encoder_two.to(accelerator.device, dtype=weight_dtype)

    if args.enable_xformers_memory_efficient_attention:
        if is_xformers_available():
            import xformers

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

    if args.gradient_checkpointing:
        unet.enable_gradient_checkpointing()
        if args.train_text_encoder:
            text_encoder_one.gradient_checkpointing_enable()
            text_encoder_two.gradient_checkpointing_enable()

    # now we will add new LoRA weights to the attention layers
    unet_lora_config = LoraConfig(
        r=args.rank,
        use_dora=args.use_dora,
        lora_alpha=args.rank,
        init_lora_weights="gaussian",
        target_modules=[
            "to_k",
            # "to_q",
            "to_v",
            "to_out.0",
            # "conv1",
            # "conv2",
            # "attn1.to_k", "attn1.to_q", "attn1.to_v",
            # "attn1.to_out.0"
        ],
    )
    unet.add_adapter(unet_lora_config)

    # The text encoder comes from 🤗 transformers, so we cannot directly modify it.
    # So, instead, we monkey-patch the forward calls of its attention-blocks.

    def unwrap_model(model):
        model = accelerator.unwrap_model(model)
        model = model._orig_mod if is_compiled_module(model) else model
        return model

    # 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:
            # there are only two options here. Either are just the unet attn processor layers
            # or there are the unet and text encoder atten layers
            unet_lora_layers_to_save = None
            text_encoder_one_lora_layers_to_save = None
            text_encoder_two_lora_layers_to_save = None

            for model in models:
                if isinstance(model, type(unwrap_model(unet))):
                    unet_lora_layers_to_save = convert_state_dict_to_diffusers(
                        get_peft_model_state_dict(model)
                    )
                elif isinstance(model, type(unwrap_model(text_encoder_one))):
                    text_encoder_one_lora_layers_to_save = (
                        convert_state_dict_to_diffusers(
                            get_peft_model_state_dict(model)
                        )
                    )
                elif isinstance(model, type(unwrap_model(text_encoder_two))):
                    text_encoder_two_lora_layers_to_save = (
                        convert_state_dict_to_diffusers(
                            get_peft_model_state_dict(model)
                        )
                    )
                else:
                    raise ValueError(f"unexpected save model: {model.__class__}")

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

            StableDiffusionXLPipeline.save_lora_weights(
                output_dir,
                unet_lora_layers=unet_lora_layers_to_save,
                text_encoder_lora_layers=text_encoder_one_lora_layers_to_save,
                text_encoder_2_lora_layers=text_encoder_two_lora_layers_to_save,
            )

    def load_model_hook(models, input_dir):
        unet_ = None
        text_encoder_one_ = None
        text_encoder_two_ = None

        while len(models) > 0:
            model = models.pop()

            if isinstance(model, type(unwrap_model(unet))):
                unet_ = model
            elif isinstance(model, type(unwrap_model(text_encoder_one))):
                text_encoder_one_ = model
            elif isinstance(model, type(unwrap_model(text_encoder_two))):
                text_encoder_two_ = model
            else:
                raise ValueError(f"unexpected save model: {model.__class__}")

        lora_state_dict, network_alphas = LoraLoaderMixin.lora_state_dict(input_dir)

        unet_state_dict = {
            f'{k.replace("unet.", "")}': v
            for k, v in lora_state_dict.items()
            if k.startswith("unet.")
        }
        unet_state_dict = convert_unet_state_dict_to_peft(unet_state_dict)
        incompatible_keys = set_peft_model_state_dict(
            unet_, unet_state_dict, adapter_name="default"
        )
        if incompatible_keys is not None:
            # check only for unexpected keys
            unexpected_keys = getattr(incompatible_keys, "unexpected_keys", None)
            if unexpected_keys:
                logger.warning(
                    f"Loading adapter weights from state_dict led to unexpected keys not found in the model: "
                    f" {unexpected_keys}. "
                )

        if args.train_text_encoder:
            # Do we need to call `scale_lora_layers()` here?
            _set_state_dict_into_text_encoder(
                lora_state_dict, prefix="text_encoder.", text_encoder=text_encoder_one_
            )

            _set_state_dict_into_text_encoder(
                lora_state_dict,
                prefix="text_encoder_2.",
                text_encoder=text_encoder_two_,
            )

        # Make sure the trainable params are in float32. This is again needed since the base models
        # are in `weight_dtype`. More details:
        # https://github.com/huggingface/diffusers/pull/6514#discussion_r1449796804
        if args.mixed_precision == "fp16":
            models = [unet_]
            if args.train_text_encoder:
                models.extend([text_encoder_one_, text_encoder_two_])
                # only upcast trainable parameters (LoRA) into fp32
                cast_training_params(models)

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

    # 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 and torch.cuda.is_available():
        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
        )

    # Make sure the trainable params are in float32.
    if args.mixed_precision == "fp16":
        models = [unet]
        if args.train_text_encoder:
            models.extend([text_encoder_one, text_encoder_two])

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

    unet_lora_parameters = list(filter(lambda p: p.requires_grad, unet.parameters()))

    if args.train_text_encoder:
        text_lora_parameters_one = list(
            filter(lambda p: p.requires_grad, text_encoder_one.parameters())
        )
        text_lora_parameters_two = list(
            filter(lambda p: p.requires_grad, text_encoder_two.parameters())
        )

    # Optimization parameters
    unet_lora_parameters_with_lr = {
        "params": unet_lora_parameters,
        "lr": args.learning_rate,
    }
    if args.train_text_encoder:
        # different learning rate for text encoder and unet
        text_lora_parameters_one_with_lr = {
            "params": text_lora_parameters_one,
            "weight_decay": args.adam_weight_decay_text_encoder,
            "lr": args.text_encoder_lr if args.text_encoder_lr else args.learning_rate,
        }
        text_lora_parameters_two_with_lr = {
            "params": text_lora_parameters_two,
            "weight_decay": args.adam_weight_decay_text_encoder,
            "lr": args.text_encoder_lr if args.text_encoder_lr else args.learning_rate,
        }
        params_to_optimize = [
            unet_lora_parameters_with_lr,
            text_lora_parameters_one_with_lr,
            text_lora_parameters_two_with_lr,
        ]
    else:
        params_to_optimize = [unet_lora_parameters_with_lr]

    # Optimizer creation
    if not (args.optimizer.lower() == "prodigy" or args.optimizer.lower() == "adamw"):
        logger.warning(
            f"Unsupported choice of optimizer: {args.optimizer}.Supported optimizers include [adamW, prodigy]."
            "Defaulting to adamW"
        )
        args.optimizer = "adamw"

    if args.use_8bit_adam and not args.optimizer.lower() == "adamw":
        logger.warning(
            f"use_8bit_adam is ignored when optimizer is not set to 'AdamW'. Optimizer was "
            f"set to {args.optimizer.lower()}"
        )

    if args.optimizer.lower() == "adamw":
        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 = optimizer_class(
            params_to_optimize,
            betas=(args.adam_beta1, args.adam_beta2),
            weight_decay=args.adam_weight_decay,
            eps=args.adam_epsilon,
        )

    if args.optimizer.lower() == "prodigy":
        try:
            import prodigyopt
        except ImportError:
            raise ImportError(
                "To use Prodigy, please install the prodigyopt library: `pip install prodigyopt`"
            )

        optimizer_class = prodigyopt.Prodigy

        if args.learning_rate <= 0.1:
            logger.warning(
                "Learning rate is too low. When using prodigy, it's generally better to set learning rate around 1.0"
            )
        if args.train_text_encoder and args.text_encoder_lr:
            logger.warning(
                f"Learning rates were provided both for the unet and the text encoder- e.g. text_encoder_lr:"
                f" {args.text_encoder_lr} and learning_rate: {args.learning_rate}. "
                f"When using prodigy only learning_rate is used as the initial learning rate."
            )
            # changes the learning rate of text_encoder_parameters_one and text_encoder_parameters_two to be
            # --learning_rate
            params_to_optimize[1]["lr"] = args.learning_rate
            params_to_optimize[2]["lr"] = args.learning_rate

        optimizer = optimizer_class(
            params_to_optimize,
            lr=args.learning_rate,
            betas=(args.adam_beta1, args.adam_beta2),
            beta3=args.prodigy_beta3,
            weight_decay=args.adam_weight_decay,
            eps=args.adam_epsilon,
            decouple=args.prodigy_decouple,
            use_bias_correction=args.prodigy_use_bias_correction,
            safeguard_warmup=args.prodigy_safeguard_warmup,
        )

    # Dataset and DataLoaders creation:
    train_dataset = DreamBoothDataset(
        instance_data_root=args.instance_data_dir,
        instance_prompt=args.instance_prompt,
        class_prompt=args.class_prompt,
        class_data_root=args.class_data_dir if args.with_prior_preservation else None,
        class_num=args.num_class_images,
        size=args.resolution,
        repeats=args.repeats,
        center_crop=args.center_crop,
    )

    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,
    )

    del vae
    del pipe.vae
    gc.collect()
    if torch.cuda.is_available():
        torch.cuda.empty_cache()
    # Computes additional embeddings/ids required by the SDXL UNet.
    # regular text embeddings (when `train_text_encoder` is not True)
    # pooled text embeddings
    # 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(original_size + 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

    if not args.train_text_encoder:
        tokenizers = [tokenizer_one, tokenizer_two]
        text_encoders = [text_encoder_one, text_encoder_two]

        def compute_text_embeddings(prompt, text_encoders, tokenizers):
            with torch.no_grad():
                prompt_embeds, pooled_prompt_embeds = encode_prompt(
                    text_encoders, tokenizers, prompt
                )
                prompt_embeds = prompt_embeds.to(accelerator.device)
                pooled_prompt_embeds = pooled_prompt_embeds.to(accelerator.device)
            return prompt_embeds, pooled_prompt_embeds

    # If no type of tuning is done on the text_encoder and custom instance prompts are NOT
    # provided (i.e. the --instance_prompt is used for all images), we encode the instance prompt once to avoid
    # the redundant encoding.
    if not args.train_text_encoder and not train_dataset.custom_instance_prompts:
        instance_prompt_hidden_states, instance_pooled_prompt_embeds = (
            compute_text_embeddings(args.instance_prompt, text_encoders, tokenizers)
        )

        global prompt_embeds_dict, unet_add_text_embeds_dict
        prompt_embeds_dict = {"instance": instance_prompt_hidden_states.to("cpu")}
        unet_add_text_embeds_dict = {
            "instance": instance_pooled_prompt_embeds.to("cpu")
        }

        print("Encoding regular prompts")
        global regular_prompts_list

        for _prompt in tqdm(regular_prompts_list):
            prompt = _prompt.replace("_", " ")
            prompt_embeds, pooled_prompt_embeds = compute_text_embeddings(
                prompt, text_encoders, tokenizers
            )

            prompt_embeds_dict[_prompt] = prompt_embeds.to("cpu")
            unet_add_text_embeds_dict[_prompt] = pooled_prompt_embeds.to("cpu")

    # Handle class prompt for prior-preservation.
    if args.with_prior_preservation:
        if not args.train_text_encoder:
            class_prompt_hidden_states, class_pooled_prompt_embeds = (
                compute_text_embeddings(args.class_prompt, text_encoders, tokenizers)
            )

    # Clear the memory here
    if not args.train_text_encoder and not train_dataset.custom_instance_prompts:
        del tokenizers, text_encoders
        gc.collect()
        if torch.cuda.is_available():
            torch.cuda.empty_cache()

    # If custom instance prompts are NOT provided (i.e. the instance prompt is used for all images),
    # pack the statically computed variables appropriately here. This is so that we don't
    # have to pass them to the dataloader.

    # if not train_dataset.custom_instance_prompts:
    #     if not args.train_text_encoder:
    #         prompt_embeds = instance_prompt_hidden_states
    #         unet_add_text_embeds = instance_pooled_prompt_embeds
    #         if args.with_prior_preservation:
    #             prompt_embeds = torch.cat([prompt_embeds, class_prompt_hidden_states], dim=0)
    #             unet_add_text_embeds = torch.cat([unet_add_text_embeds, class_pooled_prompt_embeds], dim=0)
    #     # if we're optmizing the text encoder (both if instance prompt is used for all images or custom prompts) we need to tokenize and encode the
    #     # batch prompts on all training steps
    #     else:
    #         tokens_one = tokenize_prompt(tokenizer_one, args.instance_prompt)
    #         tokens_two = tokenize_prompt(tokenizer_two, args.instance_prompt)
    #         if args.with_prior_preservation:
    #             class_tokens_one = tokenize_prompt(tokenizer_one, args.class_prompt)
    #             class_tokens_two = tokenize_prompt(tokenizer_two, args.class_prompt)
    #             tokens_one = torch.cat([tokens_one, class_tokens_one], dim=0)
    #             tokens_two = torch.cat([tokens_two, class_tokens_two], dim=0)

    # 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 * accelerator.num_processes,
        num_training_steps=args.max_train_steps * accelerator.num_processes,
        num_cycles=args.lr_num_cycles,
        power=args.lr_power,
    )

    # Prepare everything with our `accelerator`.
    if args.train_text_encoder:
        (
            unet,
            text_encoder_one,
            text_encoder_two,
            optimizer,
            train_dataloader,
            lr_scheduler,
        ) = accelerator.prepare(
            unet,
            text_encoder_one,
            text_encoder_two,
            optimizer,
            train_dataloader,
            lr_scheduler,
        )
    else:
        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:
        tracker_name = (
            "dreambooth-lora-sd-xl"
            if "playground" not in args.pretrained_model_name_or_path
            else "dreambooth-lora-playground"
        )
        accelerator.init_trackers(tracker_name, 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}")
    global_step = 0
    first_epoch = 0

    # Potentially load in the weights and states from a previous save
    if args.resume_from_checkpoint:
        if args.resume_from_checkpoint != "latest":
            path = os.path.basename(args.resume_from_checkpoint)
        else:
            # Get the mos recent checkpoint
            dirs = os.listdir(args.output_dir)
            dirs = [d for d in dirs if d.startswith("checkpoint")]
            dirs = sorted(dirs, key=lambda x: int(x.split("-")[1]))
            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,
    )

    def get_sigmas(timesteps, n_dim=4, dtype=torch.float32):
        sigmas = noise_scheduler.sigmas.to(device=accelerator.device, dtype=dtype)
        schedule_timesteps = noise_scheduler.timesteps.to(accelerator.device)
        timesteps = timesteps.to(accelerator.device)

        step_indices = [(schedule_timesteps == t).nonzero().item() for t in timesteps]

        sigma = sigmas[step_indices].flatten()
        while len(sigma.shape) < n_dim:
            sigma = sigma.unsqueeze(-1)
        return sigma

    del text_encoder_one, text_encoder_two
    del pipe.text_encoder
    del pipe.text_encoder_2
    gc.collect()
    if torch.cuda.is_available():
        torch.cuda.empty_cache()

    from utils import inject_fooocus_inpaint_head

    for epoch in range(first_epoch, args.num_train_epochs):
        unet.train()
        if args.train_text_encoder:
            text_encoder_one.train()
            text_encoder_two.train()

            # set top parameter requires_grad = True for gradient checkpointing works
            accelerator.unwrap_model(
                text_encoder_one
            ).text_model.embeddings.requires_grad_(True)
            accelerator.unwrap_model(
                text_encoder_two
            ).text_model.embeddings.requires_grad_(True)

        for step, batch in enumerate(train_dataloader):
            with accelerator.accumulate(unet):
                # pixel_values = batch["pixel_values"].to(dtype=vae.dtype)
                assert len(batch["pixel_values"]) == 1, "only support batch size 1"
                # (inpaint_head_feature, pixel_values) = batch["pixel_values"][0]

                (inpaint_head_feature, model_input, latent_mask, prompt) = batch[
                    "pixel_values"
                ][0]

                inject_fooocus_inpaint_head(
                    unet, inpaint_head_feature, parallel_train=True
                )

                prompt_embeds = prompt_embeds_dict[prompt].to(accelerator.device)
                # print(f"training on {prompt}")
                unet_add_text_embeds = unet_add_text_embeds_dict[prompt].to(
                    accelerator.device
                )

                # encode batch prompts when custom prompts are provided for each image -
                # if train_dataset.custom_instance_prompts:
                #     if not args.train_text_encoder:
                #         prompt_embeds, unet_add_text_embeds = compute_text_embeddings(
                #             prompts, text_encoders, tokenizers
                #         )
                #     else:
                #         tokens_one = tokenize_prompt(tokenizer_one, prompts)
                #         tokens_two = tokenize_prompt(tokenizer_two, prompts)

                # Convert images to latent space
                # model_input = vae.encode(pixel_values).latent_dist.sample()

                # if latents_mean is None and latents_std is None:
                #     model_input = model_input * vae.config.scaling_factor
                #     if args.pretrained_vae_model_name_or_path is None:
                #         model_input = model_input.to(weight_dtype)
                # else:
                #     latents_mean = latents_mean.to(device=model_input.device, dtype=model_input.dtype)
                #     latents_std = latents_std.to(device=model_input.device, dtype=model_input.dtype)
                #     model_input = (model_input - latents_mean) * vae.config.scaling_factor / latents_std
                #     model_input = model_input.to(dtype=weight_dtype)

                # Sample noise that we'll add to the latents
                noise = torch.randn_like(model_input)
                bsz = model_input.shape[0]

                # Sample a random timestep for each image
                if not args.do_edm_style_training:
                    timesteps = torch.randint(
                        0,
                        noise_scheduler.config.num_train_timesteps,
                        (bsz,),
                        device=model_input.device,
                    )
                    timesteps = timesteps.long()
                else:
                    # in EDM formulation, the model is conditioned on the pre-conditioned noise levels
                    # instead of discrete timesteps, so here we sample indices to get the noise levels
                    # from `scheduler.timesteps`
                    indices = torch.randint(
                        0, noise_scheduler.config.num_train_timesteps, (bsz,)
                    )
                    timesteps = noise_scheduler.timesteps[indices].to(
                        device=model_input.device
                    )

                # Add noise to the model input according to the noise magnitude at each timestep
                # (this is the forward diffusion process)
                noisy_model_input = noise_scheduler.add_noise(
                    model_input, noise, timesteps
                )
                # For EDM-style training, we first obtain the sigmas based on the continuous timesteps.
                # We then precondition the final model inputs based on these sigmas instead of the timesteps.
                # Follow: Section 5 of https://arxiv.org/abs/2206.00364.
                if args.do_edm_style_training:
                    sigmas = get_sigmas(
                        timesteps, len(noisy_model_input.shape), noisy_model_input.dtype
                    )
                    if "EDM" in scheduler_type:
                        inp_noisy_latents = noise_scheduler.precondition_inputs(
                            noisy_model_input, sigmas
                        )
                    else:
                        inp_noisy_latents = noisy_model_input / ((sigmas**2 + 1) ** 0.5)

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

                # Calculate the elements to repeat depending on the use of prior-preservation and custom captions.
                if not train_dataset.custom_instance_prompts:  # None
                    elems_to_repeat_text_embeds = (
                        bsz // 2 if args.with_prior_preservation else bsz
                    )
                else:
                    elems_to_repeat_text_embeds = 1

                # Predict the noise residual
                if not args.train_text_encoder:
                    unet_added_conditions = {
                        "time_ids": add_time_ids,
                        "text_embeds": unet_add_text_embeds.repeat(
                            elems_to_repeat_text_embeds, 1
                        ),
                    }
                    prompt_embeds_input = prompt_embeds.repeat(
                        elems_to_repeat_text_embeds, 1, 1
                    )
                    model_pred = unet(
                        (
                            inp_noisy_latents
                            if args.do_edm_style_training
                            else noisy_model_input
                        ),
                        timesteps,
                        prompt_embeds_input,
                        added_cond_kwargs=unet_added_conditions,
                        return_dict=False,
                    )[0]
                else:
                    unet_added_conditions = {"time_ids": add_time_ids}
                    prompt_embeds, pooled_prompt_embeds = encode_prompt(
                        text_encoders=[text_encoder_one, text_encoder_two],
                        tokenizers=None,
                        prompt=None,
                        text_input_ids_list=[tokens_one, tokens_two],
                    )
                    unet_added_conditions.update(
                        {
                            "text_embeds": pooled_prompt_embeds.repeat(
                                elems_to_repeat_text_embeds, 1
                            )
                        }
                    )
                    prompt_embeds_input = prompt_embeds.repeat(
                        elems_to_repeat_text_embeds, 1, 1
                    )
                    model_pred = unet(
                        (
                            inp_noisy_latents
                            if args.do_edm_style_training
                            else noisy_model_input
                        ),
                        timesteps,
                        prompt_embeds_input,
                        added_cond_kwargs=unet_added_conditions,
                        return_dict=False,
                    )[0]

                weighting = None
                if args.do_edm_style_training:
                    # Similar to the input preconditioning, the model predictions are also preconditioned
                    # on noised model inputs (before preconditioning) and the sigmas.
                    # Follow: Section 5 of https://arxiv.org/abs/2206.00364.
                    if "EDM" in scheduler_type:
                        model_pred = noise_scheduler.precondition_outputs(
                            noisy_model_input, model_pred, sigmas
                        )
                    else:
                        if noise_scheduler.config.prediction_type == "epsilon":
                            model_pred = model_pred * (-sigmas) + noisy_model_input
                        elif noise_scheduler.config.prediction_type == "v_prediction":
                            model_pred = model_pred * (
                                -sigmas / (sigmas**2 + 1) ** 0.5
                            ) + (noisy_model_input / (sigmas**2 + 1))
                    # We are not doing weighting here because it tends result in numerical problems.
                    # See: https://github.com/huggingface/diffusers/pull/7126#issuecomment-1968523051
                    # There might be other alternatives for weighting as well:
                    # https://github.com/huggingface/diffusers/pull/7126#discussion_r1505404686
                    if "EDM" not in scheduler_type:
                        weighting = (sigmas**-2.0).float()

                # Get the target for loss depending on the prediction type
                if noise_scheduler.config.prediction_type == "epsilon":
                    target = model_input if args.do_edm_style_training else noise
                elif noise_scheduler.config.prediction_type == "v_prediction":
                    target = (
                        model_input
                        if args.do_edm_style_training
                        else noise_scheduler.get_velocity(model_input, 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 prior loss
                    if weighting is not None:
                        prior_loss = torch.mean(
                            (
                                weighting.float()
                                * (model_pred_prior.float() - target_prior.float()) ** 2
                            ).reshape(target_prior.shape[0], -1),
                            1,
                        )
                        prior_loss = prior_loss.mean()
                    else:
                        prior_loss = F.mse_loss(
                            model_pred_prior.float(),
                            target_prior.float(),
                            reduction="mean",
                        )

                if args.snr_gamma is None:
                    if weighting is not None:
                        loss = torch.mean(
                            (
                                weighting.float()
                                * (model_pred.float() - target.float()) ** 2
                            ).reshape(target.shape[0], -1),
                            1,
                        )
                        loss = loss.mean()
                    else:
                        obj_loss_weight = args.loss_reweight_object
                        bg_loss_weight = args.loss_reweight_background
                        mask_reweight = (
                            latent_mask * (obj_loss_weight - bg_loss_weight)
                            + bg_loss_weight
                        )
                        model_pred = model_pred.float() * mask_reweight
                        target = target.float() * mask_reweight
                        loss = F.mse_loss(model_pred, target, reduction="mean")
                else:
                    # Compute loss-weights as per Section 3.4 of https://arxiv.org/abs/2303.09556.
                    # 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)
                    base_weight = (
                        torch.stack(
                            [snr, args.snr_gamma * torch.ones_like(timesteps)], dim=1
                        ).min(dim=1)[0]
                        / snr
                    )

                    if noise_scheduler.config.prediction_type == "v_prediction":
                        # Velocity objective needs to be floored to an SNR weight of one.
                        mse_loss_weights = base_weight + 1
                    else:
                        # Epsilon and sample both use the same loss weights.
                        mse_loss_weights = base_weight

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

                if args.with_prior_preservation:
                    # Add the prior loss to the instance loss.
                    loss = loss + args.prior_loss_weight * prior_loss

                accelerator.backward(loss)
                if accelerator.sync_gradients:
                    params_to_clip = (
                        itertools.chain(
                            unet_lora_parameters,
                            text_lora_parameters_one,
                            text_lora_parameters_two,
                        )
                        if args.train_text_encoder
                        else unet_lora_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

                if accelerator.is_main_process:
                    if global_step % args.checkpointing_steps == 0:
                        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 lora layers
    accelerator.wait_for_everyone()
    if accelerator.is_main_process:
        unet = unwrap_model(unet)
        unet = unet.to(torch.float32)
        unet_lora_layers = convert_state_dict_to_diffusers(
            get_peft_model_state_dict(unet)
        )

        if args.train_text_encoder:
            text_encoder_one = unwrap_model(text_encoder_one)
            text_encoder_lora_layers = convert_state_dict_to_diffusers(
                get_peft_model_state_dict(text_encoder_one.to(torch.float32))
            )
            text_encoder_two = unwrap_model(text_encoder_two)
            text_encoder_2_lora_layers = convert_state_dict_to_diffusers(
                get_peft_model_state_dict(text_encoder_two.to(torch.float32))
            )
        else:
            text_encoder_lora_layers = None
            text_encoder_2_lora_layers = None

        StableDiffusionXLPipeline.save_lora_weights(
            save_directory=args.output_dir,
            unet_lora_layers=unet_lora_layers,
            text_encoder_lora_layers=text_encoder_lora_layers,
            text_encoder_2_lora_layers=text_encoder_2_lora_layers,
        )
        if args.output_kohya_format:
            lora_state_dict = load_file(
                f"{args.output_dir}/pytorch_lora_weights.safetensors"
            )
            peft_state_dict = convert_all_state_dict_to_peft(lora_state_dict)
            kohya_state_dict = convert_state_dict_to_kohya(peft_state_dict)
            save_file(
                kohya_state_dict,
                f"{args.output_dir}/pytorch_lora_weights_kohya.safetensors",
            )

    accelerator.end_training()


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