train_text_to_image_decoder.py

#!/usr/bin/env python
# coding=utf-8
# Copyright 2023 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and

import argparse
import logging
import math
import os
import shutil
from pathlib import Path

import accelerate
import datasets
import numpy as np
import torch
import torch.nn.functional as F
import torch.utils.checkpoint
import transformers
from accelerate import Accelerator
from accelerate.logging import get_logger
from accelerate.state import AcceleratorState
from accelerate.utils import ProjectConfiguration, set_seed
from datasets import load_dataset
from huggingface_hub import create_repo, upload_folder
from packaging import version
from PIL import Image
from tqdm import tqdm
from transformers import CLIPImageProcessor, CLIPVisionModelWithProjection
from transformers.utils import ContextManagers
from torchvision import transforms

import diffusers
from diffusers import AutoPipelineForText2Image, DDPMScheduler, UNet2DConditionModel, VQModel
from diffusers.optimization import get_scheduler
from diffusers.training_utils import EMAModel, compute_snr
from diffusers.utils import check_min_version, is_wandb_available, make_image_grid
from diffusers.utils.import_utils import is_xformers_available


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("0.26.0")

logger = get_logger(__name__, log_level="INFO")


def save_model_card(
    args,
    repo_id: str,
    images=None,
    repo_folder=None,
):
    img_str = ""
    if len(images) > 0:
        image_grid = make_image_grid(images, 1, len(args.validation_prompts))
        image_grid.save(os.path.join(repo_folder, "val_imgs_grid.png"))
        img_str += "![val_imgs_grid](./val_imgs_grid.png)\n"

    yaml = f"""
---
license: creativeml-openrail-m
base_model: {args.pretrained_decoder_model_name_or_path}
prior:
- ECLIPSE-Community/Lambda-ECLIPSE-Prior-v1.0
tags:
- λ-ECLIPSE
- ECLIPSE
- kandinsky
- text-to-image
- diffusers
inference: true
---
    """
    model_card = f"""
# Finetuning - {repo_id}

This pipeline was finetuned from **{args.pretrained_decoder_model_name_or_path}**. Below are some example images generated with the finetuned pipeline using the following prompts: {args.validation_prompts}: \n
{img_str}

Note: these example images should not follow the original target concept!! This is to ensure that there is not overfitting.
## Pipeline usage

## Training info

These are the key hyperparameters used during training:

* Epochs: {args.num_train_epochs}
* Learning rate: {args.learning_rate}
* Batch size: {args.train_batch_size}
* Gradient accumulation steps: {args.gradient_accumulation_steps}
* Image resolution: {args.resolution}
* Mixed-precision: {args.mixed_precision}

"""
    wandb_info = ""
    if is_wandb_available():
        wandb_run_url = None
        if wandb.run is not None:
            wandb_run_url = wandb.run.url

    if wandb_run_url is not None:
        wandb_info = f"""
More information on all the CLI arguments and the environment are available on your [`wandb` run page]({wandb_run_url}).
"""

    model_card += wandb_info

    with open(os.path.join(repo_folder, "README.md"), "w") as f:
        f.write(yaml + model_card)


def log_validation(vae, image_encoder, image_processor, unet, args, accelerator, weight_dtype, epoch):
    logger.info("Running validation... ")

    pipeline = AutoPipelineForText2Image.from_pretrained(
        args.pretrained_decoder_model_name_or_path,
        vae=accelerator.unwrap_model(vae),
        prior_image_encoder=accelerator.unwrap_model(image_encoder),
        prior_image_processor=image_processor,
        unet=accelerator.unwrap_model(unet),
        torch_dtype=weight_dtype,
    )
    pipeline = pipeline.to(accelerator.device)
    pipeline.set_progress_bar_config(disable=True)

    if args.enable_xformers_memory_efficient_attention:
        pipeline.enable_xformers_memory_efficient_attention()

    if args.seed is None:
        generator = None
    else:
        generator = torch.Generator(device=accelerator.device).manual_seed(args.seed)

    images = []
    for i in range(len(args.validation_prompts)):
        with torch.autocast("cuda"):
            image = pipeline(args.validation_prompts[i], num_inference_steps=20, generator=generator).images[0]

        images.append(image)

    for tracker in accelerator.trackers:
        if tracker.name == "tensorboard":
            np_images = np.stack([np.asarray(img) for img in images])
            tracker.writer.add_images("validation", np_images, epoch, dataformats="NHWC")
        elif tracker.name == "wandb":
            tracker.log(
                {
                    "validation": [
                        wandb.Image(image, caption=f"{i}: {args.validation_prompts[i]}")
                        for i, image in enumerate(images)
                    ]
                }
            )
        else:
            logger.warn(f"image logging not implemented for {tracker.name}")

    del pipeline
    torch.cuda.empty_cache()

    return images


def parse_args():
    parser = argparse.ArgumentParser(description="Simple example of finetuning Kandinsky 2.2.")
    parser.add_argument(
        "--pretrained_decoder_model_name_or_path",
        type=str,
        default="kandinsky-community/kandinsky-2-2-decoder",
        required=False,
        help="Path to pretrained model or model identifier from huggingface.co/models.",
    )
    parser.add_argument(
        "--pretrained_prior_model_name_or_path",
        type=str,
        default="kandinsky-community/kandinsky-2-2-prior",
        required=False,
        help="Path to pretrained model or model identifier from huggingface.co/models.",
    )
    parser.add_argument(
        "--dataset_config_name",
        type=str,
        default=None,
        help="The config of the Dataset, leave as None if there's only one config.",
    )
    parser.add_argument(
        "--train_data_dir",
        type=str,
        default=None,
        help=(
            "A folder containing the training data. Folder contents must follow the structure described in"
            " https://huggingface.co/docs/datasets/image_dataset#imagefolder. In particular, a `metadata.jsonl` file"
            " must exist to provide the captions for the images. Ignored if `dataset_name` is specified."
        ),
    )
    parser.add_argument(
        "--image_column", type=str, default="image", help="The column of the dataset containing an image."
    )
    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(
        "--validation_prompts",
        type=str,
        default=None,
        nargs="+",
        help=("A set of prompts evaluated every `--validation_epochs` and logged to `--report_to`."),
    )
    parser.add_argument(
        "--output_dir",
        type=str,
        default="kandi_2_2-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=512,
        help=(
            "The resolution for input images, all the images in the train/validation dataset will be resized to this"
            " resolution"
        ),
    )
    parser.add_argument(
        "--train_batch_size", type=int, default=1, 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(
        "--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="learning rate",
    )
    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(
        "--snr_gamma",
        type=float,
        default=None,
        help="SNR weighting gamma to be used if rebalancing the loss. Recommended value is 5.0. "
        "More details here: https://arxiv.org/abs/2303.09556.",
    )
    parser.add_argument(
        "--use_8bit_adam", action="store_true", help="Whether or not to use 8-bit Adam from bitsandbytes."
    )
    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("--use_ema", action="store_true", help="Whether to use EMA model.")
    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("--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=0.0,
        required=False,
        help="weight decay_to_use",
    )
    parser.add_argument("--adam_epsilon", type=float, default=1e-08, help="Epsilon value for the Adam optimizer")
    parser.add_argument("--max_grad_norm", default=1.0, type=float, help="Max gradient norm.")
    parser.add_argument("--push_to_hub", action="store_true", help="Whether or not to push the model to the Hub.")
    parser.add_argument("--hub_token", type=str, default=None, help="The token to use to push to the Model Hub.")
    parser.add_argument(
        "--hub_model_id",
        type=str,
        default=None,
        help="The name of the repository to keep in sync with the local `output_dir`.",
    )
    parser.add_argument(
        "--logging_dir",
        type=str,
        default="logs",
        help=(
            "[TensorBoard](https://www.tensorflow.org/tensorboard) log directory. Will default to"
            " *output_dir/runs/**CURRENT_DATETIME_HOSTNAME***."
        ),
    )
    parser.add_argument(
        "--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(
        "--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("--local_rank", type=int, default=-1, help="For distributed training: local_rank")
    parser.add_argument(
        "--checkpointing_steps",
        type=int,
        default=500,
        help=(
            "Save a checkpoint of the training state every X updates. These checkpoints are only 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(
        "--enable_xformers_memory_efficient_attention", action="store_true", help="Whether or not to use xformers."
    )
    parser.add_argument(
        "--validation_epochs",
        type=int,
        default=5,
        help="Run validation every X epochs.",
    )
    parser.add_argument(
        "--tracker_project_name",
        type=str,
        default="text2image-fine-tune",
        help=(
            "The `project_name` argument passed to Accelerator.init_trackers for"
            " more information see https://huggingface.co/docs/accelerate/v0.17.0/en/package_reference/accelerator#accelerate.Accelerator"
        ),
    )
    parser.add_argument(
        "--instance_data_dir",
        type=str,
        default=None,
        required=True,
        help="A folder containing the training data of instance images.",
    )
    parser.add_argument(
        "--subject_data_dir",
        type=str,
        default=None,
        required=True,
        help="A folder containing the inference data of segmented images for prior.",
    )

    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

    return args

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

    def __init__(
        self,
        instance_data_root,
        subject_data_root,
        instance_prompt,
        tokenizer=None,
        class_data_root=None,
        class_prompt=None,
        class_num=None,
        size=512,
        center_crop=False,
        encoder_hidden_states=None,
        class_prompt_encoder_hidden_states=None,
        tokenizer_max_length=None,
    ):
        self.size = size
        self.center_crop = center_crop
        self.tokenizer = tokenizer
        self.encoder_hidden_states = encoder_hidden_states
        self.class_prompt_encoder_hidden_states = class_prompt_encoder_hidden_states
        self.tokenizer_max_length = tokenizer_max_length

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

        self.subject_data_root = Path(subject_data_root)
        if not self.subject_data_root.exists():
            raise ValueError("Segmented images root doesn't exists.")

        self.instance_images_path = list(Path(instance_data_root).iterdir())
        self.subject_images_path = list(Path(subject_data_root).iterdir())
        self.num_instance_images = len(self.instance_images_path)
        self.num_subject_images = len(self.subject_images_path)
        self.instance_prompt = instance_prompt
        self._length = self.num_instance_images
        self.class_data_root = None

        self.image_transforms = transforms.Compose(
            [
                transforms.Resize(size, interpolation=transforms.InterpolationMode.BILINEAR),
                transforms.CenterCrop(size) if center_crop else transforms.RandomCrop(size),
            ]
        )

    def __len__(self):
        return self._length

    def __getitem__(self, index):
        example = {}
        instance_image = Image.open(self.instance_images_path[index % self.num_instance_images])
        instance_image = self.image_transforms(instance_image)
        instance_image = np.array(instance_image).astype(np.float32) / 127.5 - 1
        instance_image = torch.from_numpy(np.transpose(instance_image, [2, 0, 1]))
    
        subject_image = Image.open(self.subject_images_path[index % self.num_subject_images])
        subject_image = self.image_transforms(subject_image)
        subject_image = np.array(subject_image).astype(np.float32) / 127.5 - 1
        subject_image = torch.from_numpy(np.transpose(subject_image, [2, 0, 1]))

        # if not instance_image.mode == "RGB":
        #     instance_image = instance_image.convert("RGB")
        # example["instance_pil_images"] = instance_image
        example["instance_images"] = instance_image
        example["instance_pil_images"] = [str(self.instance_images_path[index % self.num_instance_images])]
        example["subject_pil_images"] = [str(self.subject_images_path[index % self.num_subject_images])]
        example["instance_prompt"] = self.instance_prompt

        return example



def main():
    args = parse_args()
    logging_dir = os.path.join(args.output_dir, args.logging_dir)
    accelerator_project_config = ProjectConfiguration(
        total_limit=args.checkpoints_total_limit, project_dir=args.output_dir, logging_dir=logging_dir
    )
    accelerator = Accelerator(
        gradient_accumulation_steps=args.gradient_accumulation_steps,
        mixed_precision=args.mixed_precision,
        log_with=args.report_to,
        project_config=accelerator_project_config,
    )

    # 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
    noise_scheduler = DDPMScheduler.from_pretrained(args.pretrained_decoder_model_name_or_path, subfolder="scheduler")
    image_processor = CLIPImageProcessor.from_pretrained(
        args.pretrained_prior_model_name_or_path, subfolder="image_processor"
    )

    def deepspeed_zero_init_disabled_context_manager():
        """
        returns either a context list that includes one that will disable zero.Init or an empty context list
        """
        deepspeed_plugin = AcceleratorState().deepspeed_plugin if accelerate.state.is_initialized() else None
        if deepspeed_plugin is None:
            return []

        return [deepspeed_plugin.zero3_init_context_manager(enable=False)]

    weight_dtype = torch.float32
    if accelerator.mixed_precision == "fp16":
        weight_dtype = torch.float16
    elif accelerator.mixed_precision == "bf16":
        weight_dtype = torch.bfloat16
    with ContextManagers(deepspeed_zero_init_disabled_context_manager()):
        vae = VQModel.from_pretrained(
            args.pretrained_decoder_model_name_or_path, subfolder="movq", torch_dtype=weight_dtype
        ).eval()
        image_encoder = CLIPVisionModelWithProjection.from_pretrained(
            args.pretrained_prior_model_name_or_path, subfolder="image_encoder", torch_dtype=weight_dtype
        ).eval()
    unet = UNet2DConditionModel.from_pretrained(args.pretrained_decoder_model_name_or_path, subfolder="unet")

    # Freeze vae and image_encoder
    vae.requires_grad_(False)
    image_encoder.requires_grad_(False)

    # Set unet to trainable.
    unet.train()

    # Create EMA for the unet.
    if args.use_ema:
        ema_unet = UNet2DConditionModel.from_pretrained(args.pretrained_decoder_model_name_or_path, subfolder="unet")
        ema_unet = EMAModel(ema_unet.parameters(), model_cls=UNet2DConditionModel, model_config=ema_unet.config)
        ema_unet.to(accelerator.device)
    if args.enable_xformers_memory_efficient_attention:
        if is_xformers_available():
            import xformers

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

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

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

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

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

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

                # load diffusers style into model
                load_model = UNet2DConditionModel.from_pretrained(input_dir, subfolder="unet")
                model.register_to_config(**load_model.config)

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

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

    if args.gradient_checkpointing:
        unet.enable_gradient_checkpointing()

    # Enable TF32 for faster training on Ampere GPUs,
    # cf https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices
    if args.allow_tf32:
        torch.backends.cuda.matmul.allow_tf32 = True

    if args.use_8bit_adam:
        try:
            import bitsandbytes as bnb
        except ImportError:
            raise ImportError(
                "Please install bitsandbytes to use 8-bit Adam. You can do so by running `pip install bitsandbytes`"
            )

        optimizer_cls = bnb.optim.AdamW8bit
    else:
        optimizer_cls = torch.optim.AdamW

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

    def center_crop(image):
        width, height = image.size
        new_size = min(width, height)
        left = (width - new_size) / 2
        top = (height - new_size) / 2
        right = (width + new_size) / 2
        bottom = (height + new_size) / 2
        return image.crop((left, top, right, bottom))

    def train_transforms(img):
        img = center_crop(img)
        img = img.resize((args.resolution, args.resolution), resample=Image.BICUBIC, reducing_gap=1)
        img = np.array(img).astype(np.float32) / 127.5 - 1
        img = torch.from_numpy(np.transpose(img, [2, 0, 1]))
        return img

    # Dataset and DataLoaders creation:
    # TODO: Add agrs
    pre_computed_encoder_hidden_states = None
    validation_prompt_encoder_hidden_states = None
    validation_prompt_negative_prompt_embeds = None
    pre_computed_class_prompt_encoder_hidden_states = None
    train_dataset = DreamBoothDataset(
        instance_data_root=args.instance_data_dir,
        subject_data_root=args.subject_data_dir,
        instance_prompt=args.validation_prompts,
        class_data_root=None,
        class_prompt=None,
        class_num=None,
        # tokenizer=tokenizer,
        size=512,
        center_crop=True,
        encoder_hidden_states=pre_computed_encoder_hidden_states,
        class_prompt_encoder_hidden_states=pre_computed_class_prompt_encoder_hidden_states,
        # tokenizer_max_length=args.tokenizer_max_length,
    )

    train_dataloader = torch.utils.data.DataLoader(
        train_dataset,
        shuffle=True,
        # collate_fn=collate_fn,
        batch_size=args.train_batch_size,
        num_workers=args.dataloader_num_workers,
    )

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

    lr_scheduler = get_scheduler(
        args.lr_scheduler,
        optimizer=optimizer,
        num_warmup_steps=args.lr_warmup_steps * args.gradient_accumulation_steps,
        num_training_steps=args.max_train_steps * args.gradient_accumulation_steps,
    )
    unet, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
        unet, optimizer, train_dataloader, lr_scheduler
    )
    # Move image_encode and vae to gpu and cast to weight_dtype
    image_encoder.to(accelerator.device, dtype=weight_dtype)
    vae.to(accelerator.device, dtype=weight_dtype)
    # We need to recalculate our total training steps as the size of the training dataloader may have changed.
    num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
    if overrode_max_train_steps:
        args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
    # Afterwards we recalculate our number of training epochs
    args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)

    # 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_config = dict(vars(args))
        tracker_config.pop("validation_prompts")
        accelerator.init_trackers(args.tracker_project_name, tracker_config)

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

    logger.info("***** Running training *****")
    logger.info(f"  Num examples = {len(train_dataset)}")
    logger.info(f"  Num Epochs = {args.num_train_epochs}")
    logger.info(f"  Instantaneous batch size per device = {args.train_batch_size}")
    logger.info(f"  Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}")
    logger.info(f"  Gradient Accumulation steps = {args.gradient_accumulation_steps}")
    logger.info(f"  Total optimization steps = {args.max_train_steps}")
    global_step = 0
    first_epoch = 0
    if args.resume_from_checkpoint:
        if args.resume_from_checkpoint != "latest":
            path = os.path.basename(args.resume_from_checkpoint)
        else:
            # Get the most recent checkpoint
            dirs = os.listdir(args.output_dir)
            dirs = [d for d in dirs if d.startswith("checkpoint")]
            dirs = sorted(dirs, key=lambda x: int(x.split("-")[1]))
            path = dirs[-1] if len(dirs) > 0 else None

        if path is None:
            accelerator.print(
                f"Checkpoint '{args.resume_from_checkpoint}' does not exist. Starting a new training run."
            )
            args.resume_from_checkpoint = None
            initial_global_step = 0
        else:
            accelerator.print(f"Resuming from checkpoint {path}")
            accelerator.load_state(os.path.join(args.output_dir, path))
            global_step = int(path.split("-")[1])

            initial_global_step = global_step
            first_epoch = global_step // num_update_steps_per_epoch
    else:
        initial_global_step = 0

    progress_bar = tqdm(
        range(0, args.max_train_steps),
        initial=initial_global_step,
        desc="Steps",
        # Only show the progress bar once on each machine.
        disable=not accelerator.is_local_main_process,
    )

    for epoch in range(first_epoch, args.num_train_epochs):
        train_loss = 0.0
        for step, batch in enumerate(train_dataloader):
            with accelerator.accumulate(unet):
                # Convert images to latent space
                images = batch["instance_images"].to(weight_dtype)
                latents = vae.encode(images).latents

                images_embeds = []
                for en_i in range(len(batch["instance_pil_images"])):
                    clip_images = image_processor(Image.open(batch["instance_pil_images"][en_i][0]), return_tensors="pt").to(latents.device, weight_dtype)
                    images_embeds.append(image_encoder(**clip_images).image_embeds[0])
                image_embeds = torch.stack(images_embeds)

                # Sample noise that we'll add to the latents
                noise = torch.randn_like(latents)
                bsz = latents.shape[0]
                # Sample a random timestep for each image
                timesteps = torch.randint(0, noise_scheduler.config.num_train_timesteps, (bsz,), device=latents.device)
                timesteps = timesteps.long()

                noisy_latents = noise_scheduler.add_noise(latents, noise, timesteps)

                target = noise

                # Predict the noise residual and compute loss
                added_cond_kwargs = {"image_embeds": image_embeds}
                
                model_pred = unet(noisy_latents, timesteps, None, added_cond_kwargs=added_cond_kwargs).sample[:, :4]

                if args.snr_gamma is None:
                    loss = F.mse_loss(model_pred.float(), target.float(), 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)
                    mse_loss_weights = torch.stack([snr, args.snr_gamma * torch.ones_like(timesteps)], dim=1).min(
                        dim=1
                    )[0]
                    if noise_scheduler.config.prediction_type == "epsilon":
                        mse_loss_weights = mse_loss_weights / snr
                    elif noise_scheduler.config.prediction_type == "v_prediction":
                        mse_loss_weights = mse_loss_weights / (snr + 1)

                    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()

                # 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:
                    accelerator.clip_grad_norm_(unet.parameters(), 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:
                if args.use_ema:
                    ema_unet.step(unet.parameters())
                progress_bar.update(1)
                global_step += 1
                accelerator.log({"train_loss": train_loss}, step=global_step)
                train_loss = 0.0

                if global_step % args.checkpointing_steps == 0:
                    if accelerator.is_main_process:
                        # _before_ saving state, check if this save would set us over the `checkpoints_total_limit`
                        if args.checkpoints_total_limit is not None:
                            checkpoints = os.listdir(args.output_dir)
                            checkpoints = [d for d in checkpoints if d.startswith("checkpoint")]
                            checkpoints = sorted(checkpoints, key=lambda x: int(x.split("-")[1]))

                            # before we save the new checkpoint, we need to have at _most_ `checkpoints_total_limit - 1` checkpoints
                            if len(checkpoints) >= args.checkpoints_total_limit:
                                num_to_remove = len(checkpoints) - args.checkpoints_total_limit + 1
                                removing_checkpoints = checkpoints[0:num_to_remove]

                                logger.info(
                                    f"{len(checkpoints)} checkpoints already exist, removing {len(removing_checkpoints)} checkpoints"
                                )
                                logger.info(f"removing checkpoints: {', '.join(removing_checkpoints)}")

                                for removing_checkpoint in removing_checkpoints:
                                    removing_checkpoint = os.path.join(args.output_dir, removing_checkpoint)
                                    shutil.rmtree(removing_checkpoint)

                        save_path = os.path.join(args.output_dir, f"checkpoint-{global_step}")
                        accelerator.save_state(save_path)
                        logger.info(f"Saved state to {save_path}")

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

            if global_step >= args.max_train_steps:
                break

        if accelerator.is_main_process:
            if args.validation_prompts is not None and epoch % args.validation_epochs == 0:
                if args.use_ema:
                    # Store the UNet parameters temporarily and load the EMA parameters to perform inference.
                    ema_unet.store(unet.parameters())
                    ema_unet.copy_to(unet.parameters())
                log_validation(
                    vae,
                    image_encoder,
                    image_processor,
                    unet,
                    args,
                    accelerator,
                    weight_dtype,
                    global_step,
                )
                if args.use_ema:
                    # Switch back to the original UNet parameters.
                    ema_unet.restore(unet.parameters())

    # Create the pipeline using the trained modules and save it.
    accelerator.wait_for_everyone()
    if accelerator.is_main_process:
        unet = accelerator.unwrap_model(unet)
        if args.use_ema:
            ema_unet.copy_to(unet.parameters())

        pipeline = AutoPipelineForText2Image.from_pretrained(
            args.pretrained_decoder_model_name_or_path,
            vae=vae,
            unet=unet,
        )
        pipeline.decoder_pipe.save_pretrained(args.output_dir)

        # Run a final round of inference.
        images = []
        if args.validation_prompts is not None:
            logger.info("Running inference for collecting generated images...")
            pipeline = pipeline.to(accelerator.device)
            pipeline.torch_dtype = weight_dtype
            pipeline.set_progress_bar_config(disable=True)
            pipeline.enable_model_cpu_offload()

            if args.enable_xformers_memory_efficient_attention:
                pipeline.enable_xformers_memory_efficient_attention()

            if args.seed is None:
                generator = None
            else:
                generator = torch.Generator(device=accelerator.device).manual_seed(args.seed)

            for i in range(len(args.validation_prompts)):
                with torch.autocast("cuda"):
                    image = pipeline(args.validation_prompts[i], num_inference_steps=20, generator=generator).images[0]
                images.append(image)

        if args.push_to_hub:
            save_model_card(args, repo_id, images, repo_folder=args.output_dir)
            upload_folder(
                repo_id=repo_id,
                folder_path=args.output_dir,
                commit_message="End of training",
                ignore_patterns=["step_*", "epoch_*"],
            )

    accelerator.end_training()


if __name__ == "__main__":
    main()