train_dreamcreature_sdxl.py

#!/usr/bin/env python
# coding=utf-8
# Copyright 2023 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""Fine-tuning script for Stable Diffusion XL for text2image with support for LoRA."""

import argparse
import copy
import itertools
import logging
import math
import os
import random
import shutil
from pathlib import Path
from typing import Dict

import datasets
import diffusers
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 diffusers import (
    AutoencoderKL,
    DDPMScheduler,
    StableDiffusionXLPipeline,
    UNet2DConditionModel,
)
from diffusers.loaders import LoraLoaderMixin
from diffusers.models.lora import LoRALinearLayer
from diffusers.optimization import get_scheduler
from diffusers.training_utils import compute_snr
from diffusers.utils import check_min_version, is_wandb_available
from diffusers.utils.import_utils import is_xformers_available
from huggingface_hub import create_repo, upload_folder
from packaging import version
from torchvision import transforms
from torchvision.transforms.functional import crop
from tqdm.auto import tqdm
from transformers import PretrainedConfig

from dreamcreature.attn_processor import AttnProcessorCustom
from dreamcreature.dataset import DreamCreatureDataset
from dreamcreature.dino import DINO
from dreamcreature.kmeans_segmentation import KMeansSegmentation
from dreamcreature.loss import dreamcreature_loss
from dreamcreature.mapper import TokenMapper
from dreamcreature.pipeline_xl import DreamCreatureSDXLPipeline
from dreamcreature.text_encoder import CustomCLIPTextModel, CustomCLIPTextModelWithProjection
from dreamcreature.tokenizer import MultiTokenCLIPTokenizer
from utils import add_tokens, tokenize_prompt, get_attn_processors

IMAGENET_TEMPLATES = [
    "a photo of a {}",
    "a rendering of a {}",
    "a cropped photo of the {}",
    "the photo of a {}",
    "a photo of a clean {}",
    "a photo of a dirty {}",
    "a dark photo of the {}",
    "a photo of my {}",
    "a photo of the cool {}",
    "a close-up photo of a {}",
    "a bright photo of the {}",
    "a cropped photo of a {}",
    "a photo of the {}",
    "a good photo of the {}",
    "a photo of one {}",
    "a close-up photo of the {}",
    "a rendition of the {}",
    "a photo of the clean {}",
    "a rendition of a {}",
    "a photo of a nice {}",
    "a good photo of a {}",
    "a photo of the nice {}",
    "a photo of the small {}",
    "a photo of the weird {}",
    "a photo of the large {}",
    "a photo of a cool {}",
    "a photo of a small {}",
]

# Will error if the minimal version of diffusers is not installed. Remove at your own risks.
check_min_version("0.25.0.dev0")

logger = get_logger(__name__)


# TODO: This function should be removed once training scripts are rewritten in PEFT
def text_encoder_lora_state_dict(text_encoder):
    state_dict = {}

    def text_encoder_attn_modules(text_encoder):
        from transformers import CLIPTextModel, CLIPTextModelWithProjection

        attn_modules = []

        if isinstance(text_encoder, (CLIPTextModel, CLIPTextModelWithProjection)):
            for i, layer in enumerate(text_encoder.text_model.encoder.layers):
                name = f"text_model.encoder.layers.{i}.self_attn"
                mod = layer.self_attn
                attn_modules.append((name, mod))

        return attn_modules

    for name, module in text_encoder_attn_modules(text_encoder):
        for k, v in module.q_proj.lora_linear_layer.state_dict().items():
            state_dict[f"{name}.q_proj.lora_linear_layer.{k}"] = v

        for k, v in module.k_proj.lora_linear_layer.state_dict().items():
            state_dict[f"{name}.k_proj.lora_linear_layer.{k}"] = v

        for k, v in module.v_proj.lora_linear_layer.state_dict().items():
            state_dict[f"{name}.v_proj.lora_linear_layer.{k}"] = v

        for k, v in module.out_proj.lora_linear_layer.state_dict().items():
            state_dict[f"{name}.out_proj.lora_linear_layer.{k}"] = v

    return state_dict


def save_model_card(
        repo_id: str,
        images=None,
        base_model=str,
        dataset_name=str,
        train_text_encoder=False,
        repo_folder=None,
        vae_path=None,
):
    img_str = ""
    for i, image in enumerate(images):
        image.save(os.path.join(repo_folder, f"image_{i}.png"))
        img_str += f"![img_{i}](./image_{i}.png)\n"

    yaml = f"""
---
license: creativeml-openrail-m
base_model: {base_model}
dataset: {dataset_name}
tags:
- stable-diffusion-xl
- stable-diffusion-xl-diffusers
- text-to-image
- diffusers
- lora
inference: true
---
    """
    model_card = f"""
# LoRA text2image fine-tuning - {repo_id}

These are LoRA adaption weights for {base_model}. The weights were fine-tuned on the {dataset_name} dataset. You can find some example images in the following. \n
{img_str}

LoRA for the text encoder was enabled: {train_text_encoder}.

Special VAE used for training: {vae_path}.
"""
    with open(os.path.join(repo_folder, "README.md"), "w") as f:
        f.write(yaml + model_card)


def import_model_class_from_model_name_or_path(
        pretrained_model_name_or_path: str, revision: str, subfolder: str = "text_encoder"
):
    text_encoder_config = PretrainedConfig.from_pretrained(
        pretrained_model_name_or_path, subfolder=subfolder, revision=revision
    )
    model_class = text_encoder_config.architectures[0]

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

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

        return CLIPTextModelWithProjection
    else:
        raise ValueError(f"{model_class} is not supported.")


def parse_args(input_args=None):
    parser = argparse.ArgumentParser(description="Simple example of a training script.")
    parser.add_argument(
        "--pretrained_model_name_or_path",
        type=str,
        default=None,
        required=True,
        help="Path to pretrained model or model identifier from huggingface.co/models.",
    )
    parser.add_argument(
        "--pretrained_vae_model_name_or_path",
        type=str,
        default=None,
        help="Path to pretrained VAE model with better numerical stability. More details: https://github.com/huggingface/diffusers/pull/4038.",
    )
    parser.add_argument(
        "--revision",
        type=str,
        default=None,
        required=False,
        help="Revision of pretrained model identifier from huggingface.co/models.",
    )
    parser.add_argument(
        "--variant",
        type=str,
        default=None,
        help="Variant of the model files of the pretrained model identifier from huggingface.co/models, 'e.g.' fp16",
    )
    parser.add_argument(
        "--dataset_name",
        type=str,
        default=None,
        help=(
            "The name of the Dataset (from the HuggingFace hub) to train on (could be your own, possibly private,"
            " dataset). It can also be a path pointing to a local copy of a dataset in your filesystem,"
            " or to a folder containing files that 🤗 Datasets can understand."
        ),
    )
    parser.add_argument(
        "--dataset_config_name",
        type=str,
        default=None,
        help="The config of the Dataset, leave as None if there's only one config.",
    )
    parser.add_argument(
        "--train_data_dir",
        type=str,
        default=None,
        help=(
            "A folder containing the training data. Folder contents must follow the structure described in"
            " https://huggingface.co/docs/datasets/image_dataset#imagefolder. In particular, a `metadata.jsonl` file"
            " must exist to provide the captions for the images. Ignored if `dataset_name` is specified."
        ),
    )
    parser.add_argument(
        "--image_column", type=str, default="image", help="The column of the dataset containing an image."
    )
    parser.add_argument(
        "--caption_column",
        type=str,
        default="text",
        help="The column of the dataset containing a caption or a list of captions.",
    )
    parser.add_argument(
        "--validation_prompt",
        type=str,
        default=None,
        help="A prompt that is used during validation to verify that the model is learning.",
    )
    parser.add_argument(
        "--num_validation_images",
        type=int,
        default=4,
        help="Number of images that should be generated during validation with `validation_prompt`.",
    )
    parser.add_argument(
        "--validation_epochs",
        type=int,
        default=1,
        help=(
            "Run fine-tuning validation every X epochs. The validation process consists of running the prompt"
            " `args.validation_prompt` multiple times: `args.num_validation_images`."
        ),
    )
    parser.add_argument(
        "--max_train_samples",
        type=int,
        default=None,
        help=(
            "For debugging purposes or quicker training, truncate the number of training examples to this "
            "value if set."
        ),
    )
    parser.add_argument(
        "--output_dir",
        type=str,
        default="sd-model-finetuned-lora",
        help="The output directory where the model predictions and checkpoints will be written.",
    )
    parser.add_argument(
        "--cache_dir",
        type=str,
        default=None,
        help="The directory where the downloaded models and datasets will be stored.",
    )
    parser.add_argument("--seed", type=int, default=None, help="A seed for reproducible training.")
    parser.add_argument(
        "--resolution",
        type=int,
        default=1024,
        help=(
            "The resolution for input images, all the images in the train/validation dataset will be resized to this"
            " resolution"
        ),
    )
    parser.add_argument(
        "--center_crop",
        default=False,
        action="store_true",
        help=(
            "Whether to center crop the input images to the resolution. If not set, the images will be randomly"
            " cropped. The images will be resized to the resolution first before cropping."
        ),
    )
    parser.add_argument(
        "--random_flip",
        action="store_true",
        help="whether to randomly flip images horizontally",
    )
    parser.add_argument(
        "--train_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=16, help="Batch size (per device) for the training dataloader."
    )
    parser.add_argument("--num_train_epochs", type=int, default=100)
    parser.add_argument(
        "--max_train_steps",
        type=int,
        default=None,
        help="Total number of training steps to perform.  If provided, overrides num_train_epochs.",
    )
    parser.add_argument(
        "--checkpointing_steps",
        type=int,
        default=500,
        help=(
            "Save a checkpoint of the training state every X updates. These checkpoints can be used both as final"
            " checkpoints in case they are better than the last checkpoint, and are also suitable for resuming"
            " training using `--resume_from_checkpoint`."
        ),
    )
    parser.add_argument(
        "--checkpoints_total_limit",
        type=int,
        default=None,
        help=("Max number of checkpoints to store."),
    )
    parser.add_argument(
        "--resume_from_checkpoint",
        type=str,
        default=None,
        help=(
            "Whether training should be resumed from a previous checkpoint. Use a path saved by"
            ' `--checkpointing_steps`, or `"latest"` to automatically select the last available checkpoint.'
        ),
    )
    parser.add_argument(
        "--gradient_accumulation_steps",
        type=int,
        default=1,
        help="Number of updates steps to accumulate before performing a backward/update pass.",
    )
    parser.add_argument(
        "--gradient_checkpointing",
        action="store_true",
        help="Whether or not to use gradient checkpointing to save memory at the expense of slower backward pass.",
    )
    parser.add_argument(
        "--learning_rate",
        type=float,
        default=1e-4,
        help="Initial learning rate (after the potential warmup period) to use.",
    )
    parser.add_argument(
        "--scale_lr",
        action="store_true",
        default=False,
        help="Scale the learning rate by the number of GPUs, gradient accumulation steps, and batch size.",
    )
    parser.add_argument(
        "--lr_scheduler",
        type=str,
        default="constant",
        help=(
            'The scheduler type to use. Choose between ["linear", "cosine", "cosine_with_restarts", "polynomial",'
            ' "constant", "constant_with_warmup"]'
        ),
    )
    parser.add_argument(
        "--lr_warmup_steps", type=int, default=500, help="Number of steps for the warmup in the lr scheduler."
    )
    parser.add_argument(
        "--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(
        "--allow_tf32",
        action="store_true",
        help=(
            "Whether or not to allow TF32 on Ampere GPUs. Can be used to speed up training. For more information, see"
            " https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices"
        ),
    )
    parser.add_argument(
        "--dataloader_num_workers",
        type=int,
        default=0,
        help=(
            "Number of subprocesses to use for data loading. 0 means that the data will be loaded in the main process."
        ),
    )
    parser.add_argument(
        "--use_8bit_adam", action="store_true", help="Whether or not to use 8-bit Adam from bitsandbytes."
    )
    parser.add_argument("--adam_beta1", type=float, default=0.9, help="The beta1 parameter for the Adam optimizer.")
    parser.add_argument("--adam_beta2", type=float, default=0.999, help="The beta2 parameter for the Adam optimizer.")
    parser.add_argument("--adam_weight_decay", type=float, default=1e-2, help="Weight decay to use.")
    parser.add_argument("--adam_epsilon", type=float, default=1e-08, help="Epsilon value for the Adam optimizer")
    parser.add_argument("--max_grad_norm", default=1.0, type=float, help="Max gradient norm.")
    parser.add_argument("--push_to_hub", action="store_true", help="Whether or not to push the model to the Hub.")
    parser.add_argument("--hub_token", type=str, default=None, help="The token to use to push to the Model Hub.")
    parser.add_argument(
        "--prediction_type",
        type=str,
        default=None,
        help="The prediction_type that shall be used for training. Choose between 'epsilon' or 'v_prediction' or leave `None`. If left to `None` the default prediction type of the scheduler: `noise_scheduler.config.prediciton_type` is chosen.",
    )
    parser.add_argument(
        "--hub_model_id",
        type=str,
        default=None,
        help="The name of the repository to keep in sync with the local `output_dir`.",
    )
    parser.add_argument(
        "--logging_dir",
        type=str,
        default="logs",
        help=(
            "[TensorBoard](https://www.tensorflow.org/tensorboard) log directory. Will default to"
            " *output_dir/runs/**CURRENT_DATETIME_HOSTNAME***."
        ),
    )
    parser.add_argument(
        "--report_to",
        type=str,
        default="tensorboard",
        help=(
            'The integration to report the results and logs to. Supported platforms are `"tensorboard"`'
            ' (default), `"wandb"` and `"comet_ml"`. Use `"all"` to report to all integrations.'
        ),
    )
    parser.add_argument(
        "--mixed_precision",
        type=str,
        default=None,
        choices=["no", "fp16", "bf16"],
        help=(
            "Whether to use mixed precision. Choose between fp16 and bf16 (bfloat16). Bf16 requires PyTorch >="
            " 1.10.and an Nvidia Ampere GPU.  Default to the value of accelerate config of the current system or the"
            " flag passed with the `accelerate.launch` command. Use this argument to override the accelerate config."
        ),
    )
    parser.add_argument("--local_rank", type=int, default=-1, help="For distributed training: local_rank")
    parser.add_argument(
        "--enable_xformers_memory_efficient_attention", action="store_true", help="Whether or not to use xformers."
    )
    parser.add_argument("--noise_offset", type=float, default=0, help="The scale of noise offset.")
    parser.add_argument(
        "--rank",
        type=int,
        default=4,
        help=("The dimension of the LoRA update matrices."),
    )

    parser.add_argument('--filename', default='train.txt')
    parser.add_argument('--code_filename', default='train_caps_better_m8_k256.txt')
    parser.add_argument('--repeat', default=1, type=int)

    parser.add_argument('--scheduler_steps', default=1000, type=int, help='scheduler step, if turbo, set to 4')
    parser.add_argument('--num_parts', type=int, default=4, help="Number of parts")
    parser.add_argument('--num_k_per_part', type=int, default=256, help='Number of k')

    parser.add_argument('--mapper_lr_scale', default=1, type=float)
    parser.add_argument('--mapper_lr', default=0.0001, type=float)
    parser.add_argument('--attn_loss', default=0, type=float)
    parser.add_argument('--projection_nlayers', default=3, type=int)

    parser.add_argument('--masked_training', action='store_true')
    parser.add_argument('--drop_tokens', action='store_true')
    parser.add_argument('--drop_rate', type=float, default=0.5)
    parser.add_argument('--drop_counts', default='half')

    parser.add_argument('--class_name', default='')
    parser.add_argument('--no_pe', action='store_true')
    parser.add_argument('--vector_shuffle', action='store_true')

    parser.add_argument('--use_gt_label', action='store_true')
    parser.add_argument('--bg_code', default=7, type=int)  # for gt_label
    parser.add_argument('--fg_idx', default=0, type=int)
    parser.add_argument('--use_templates', action='store_true')

    parser.add_argument('--filter_class', default=None, type=int, help='debugging purpose')

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

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

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

    return args


def unet_attn_processors_state_dict(unet) -> Dict[str, torch.tensor]:
    """
    Returns:
        a state dict containing just the attention processor parameters.
    """
    attn_processors = get_attn_processors(unet)

    attn_processors_state_dict = {}

    for attn_processor_key, attn_processor in attn_processors.items():
        for parameter_key, parameter in attn_processor.state_dict().items():
            attn_processors_state_dict[f"{attn_processor_key}.{parameter_key}"] = parameter

    return attn_processors_state_dict


def encode_prompt(text_encoders, text_input_ids_list, placeholder_token_ids, mapper_outputs):
    prompt_embeds_list = []

    for i, text_encoder in enumerate(text_encoders):
        text_input_ids = text_input_ids_list[i]

        modified_hs = text_encoder.text_model.forward_embeddings_with_mapper(text_input_ids,
                                                                             None,
                                                                             mapper_outputs[i],
                                                                             placeholder_token_ids)

        prompt_embeds = text_encoder(text_input_ids,
                                     hidden_states=modified_hs,
                                     output_hidden_states=True)
        # prompt_embeds = text_encoder(
        #     text_input_ids.to(text_encoder.device),
        #     output_hidden_states=True,
        # )

        # We are only ALWAYS interested in the pooled output of the final text encoder
        pooled_prompt_embeds = prompt_embeds[0]
        prompt_embeds = prompt_embeds.hidden_states[-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 collate_fn(args, tokenizer_one, tokenizer_two, placeholder_token):
    # Preprocessing the datasets.
    train_resize = transforms.Resize(args.resolution, interpolation=transforms.InterpolationMode.BILINEAR)
    train_crop = transforms.CenterCrop(args.resolution) if args.center_crop else transforms.RandomCrop(args.resolution)
    train_flip = transforms.RandomHorizontalFlip(p=1.0)
    train_transforms = transforms.Compose(
        [
            transforms.ToTensor(),
            transforms.Normalize([0.5], [0.5]),
        ]
    )

    def f(examples):
        # image aug
        original_sizes = []
        all_images = []
        crop_top_lefts = []
        captions = []
        raw_images = []
        appeared_tokens = []
        codes = []
        for i in range(len(examples)):
            ##### original sdxl process #####
            image = examples[i]['pixel_values'].convert('RGB')
            original_sizes.append((image.height, image.width))
            image = train_resize(image)
            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)
            else:
                y1, x1, h, w = train_crop.get_params(image, (args.resolution, args.resolution))
                image = crop(image, y1, x1, h, w)
            if args.random_flip and random.random() < 0.5:
                # flip
                x1 = image.width - x1
                image = train_flip(image)
            crop_top_left = (y1, x1)
            crop_top_lefts.append(crop_top_left)
            raw_images.append(image)
            image = train_transforms(image)
            all_images.append(image)

            ##### dreamcreature caption #####
            if args.use_templates and random.random() <= 0.5:  # 50% using templates
                if args.class_name != '':
                    caption = random.choice(IMAGENET_TEMPLATES).format(f'{placeholder_token} {args.class_name}')
                else:
                    caption = random.choice(IMAGENET_TEMPLATES).format(placeholder_token)
            else:
                if args.class_name != '':
                    caption = f'{placeholder_token} {args.class_name}'
                else:
                    caption = placeholder_token

            tokens = tokenizer_one.token_map[placeholder_token][:args.num_parts]
            tokens = [tokens[a] for a in examples[i]['appeared']]

            if args.vector_shuffle or args.drop_tokens:
                tokens = copy.copy(tokens)
                random.shuffle(tokens)

            if args.drop_tokens and random.random() < args.drop_rate and len(tokens) >= 2:
                # randomly drop half of the tokens
                if args.drop_counts == 'half':
                    tokens = tokens[:len(tokens) // 2]
                else:
                    tokens = tokens[:int(args.drop_counts)]

            caption = caption.replace(placeholder_token, ' '.join(tokens))
            captions.append(caption)

            appeared = [int(t.split('_')[1]) for t in tokens]  # <part>_i
            # examples[i]['appeared'] = appeared

            appeared_tokens.append(appeared)

            code = examples[i]['codes']
            codes.append(code)

        tokens_one = tokenize_prompt(tokenizer_one, captions)
        tokens_two = tokenize_prompt(tokenizer_two, captions)

        ##### start stacking #####
        pixel_values = torch.stack([image for image in all_images])
        pixel_values = pixel_values.to(memory_format=torch.contiguous_format).float()
        original_sizes = [s for s in original_sizes]
        crop_top_lefts = [c for c in crop_top_lefts]
        input_ids_one = torch.stack([t for t in tokens_one])
        input_ids_two = torch.stack([t for t in tokens_two])

        codes = torch.stack(codes, dim=0)

        collate_output = {
            "original_sizes": original_sizes,
            "crop_top_lefts": crop_top_lefts,
            "pixel_values": pixel_values,
            "input_ids_one": input_ids_one,
            "input_ids_two": input_ids_two,
            "raw_images": raw_images,
            "appeared_tokens": appeared_tokens,
            "codes": codes
        }

        return collate_output

    return f


def setup_attn_processors(unet, args):
    attn_size = args.resolution // 32
    attn_procs = {}
    for name in unet.attn_processors.keys():
        attn_procs[name] = AttnProcessorCustom(attn_size)
    unet.set_attn_processor(attn_procs)


def init_for_pipeline(args):
    tokenizer_one = MultiTokenCLIPTokenizer.from_pretrained(
        args.pretrained_model_name_or_path,
        subfolder="tokenizer",
        revision=args.revision,
        use_fast=False,
    )
    tokenizer_two = MultiTokenCLIPTokenizer.from_pretrained(
        args.pretrained_model_name_or_path,
        subfolder="tokenizer_2",
        revision=args.revision,
        use_fast=False,
    )
    text_encoder_cls_one = CustomCLIPTextModel
    text_encoder_cls_two = CustomCLIPTextModelWithProjection
    text_encoder_one = text_encoder_cls_one.from_pretrained(
        args.pretrained_model_name_or_path, subfolder="text_encoder", revision=args.revision, variant=args.variant
    )
    text_encoder_two = text_encoder_cls_two.from_pretrained(
        args.pretrained_model_name_or_path, subfolder="text_encoder_2", revision=args.revision, variant=args.variant
    )

    OUT_DIMS = 768 + 1280  # 2048
    simple_mapper = TokenMapper(args.num_parts,
                                args.num_k_per_part,
                                OUT_DIMS,
                                args.projection_nlayers)
    return text_encoder_one, text_encoder_two, tokenizer_one, tokenizer_two, simple_mapper


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

    accelerator_project_config = ProjectConfiguration(project_dir=args.output_dir, logging_dir=logging_dir)
    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],
    )

    if args.report_to == "wandb":
        if not is_wandb_available():
            raise ImportError("Make sure to install wandb if you want to use it for logging during training.")
        import wandb

    # Make one log on every process with the configuration for debugging.
    logging.basicConfig(
        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
        datefmt="%m/%d/%Y %H:%M:%S",
        level=logging.INFO,
    )
    logger.info(accelerator.state, main_process_only=False)
    if accelerator.is_local_main_process:
        datasets.utils.logging.set_verbosity_warning()
        transformers.utils.logging.set_verbosity_warning()
        diffusers.utils.logging.set_verbosity_info()
    else:
        datasets.utils.logging.set_verbosity_error()
        transformers.utils.logging.set_verbosity_error()
        diffusers.utils.logging.set_verbosity_error()

    # If passed along, set the training seed now.
    if args.seed is not None:
        set_seed(args.seed)

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

        if args.push_to_hub:
            repo_id = create_repo(
                repo_id=args.hub_model_id or Path(args.output_dir).name, exist_ok=True, token=args.hub_token
            ).repo_id

    # Load the tokenizers (replace AutoTokenizer with the custom MultiTokenCLIPTokenizer)
    tokenizer_one = MultiTokenCLIPTokenizer.from_pretrained(
        args.pretrained_model_name_or_path,
        subfolder="tokenizer",
        revision=args.revision,
        use_fast=False,
    )
    tokenizer_two = MultiTokenCLIPTokenizer.from_pretrained(
        args.pretrained_model_name_or_path,
        subfolder="tokenizer_2",
        revision=args.revision,
        use_fast=False,
    )
    # import correct text encoder classes
    # text_encoder_cls_one = import_model_class_from_model_name_or_path(
    #     args.pretrained_model_name_or_path, args.revision
    # )
    # text_encoder_cls_two = import_model_class_from_model_name_or_path(
    #     args.pretrained_model_name_or_path, args.revision, subfolder="text_encoder_2"
    # )
    text_encoder_cls_one = CustomCLIPTextModel
    text_encoder_cls_two = CustomCLIPTextModelWithProjection

    # Load scheduler and models
    noise_scheduler = DDPMScheduler.from_pretrained(args.pretrained_model_name_or_path,
                                                    subfolder="scheduler",
                                                    num_train_steps=args.scheduler_steps)
    text_encoder_one = text_encoder_cls_one.from_pretrained(
        args.pretrained_model_name_or_path, subfolder="text_encoder", revision=args.revision, variant=args.variant
    )
    text_encoder_two = text_encoder_cls_two.from_pretrained(
        args.pretrained_model_name_or_path, subfolder="text_encoder_2", revision=args.revision, variant=args.variant
    )
    vae_path = (
        args.pretrained_model_name_or_path
        if args.pretrained_vae_model_name_or_path is None
        else args.pretrained_vae_model_name_or_path
    )
    vae = AutoencoderKL.from_pretrained(
        vae_path,
        subfolder="vae" if args.pretrained_vae_model_name_or_path is None else None,
        revision=args.revision,
        variant=args.variant,
    )
    unet = UNet2DConditionModel.from_pretrained(
        args.pretrained_model_name_or_path, subfolder="unet", revision=args.revision, variant=args.variant
    )

    ##### dreamcreature init #####
    OUT_DIMS = 768 + 1280  # 2048

    dino = DINO()
    seg = KMeansSegmentation(args.train_data_dir + '/pretrained_kmeans.pth',
                             args.fg_idx,
                             args.bg_code,
                             args.num_parts,
                             args.num_k_per_part)

    simple_mapper = TokenMapper(args.num_parts,
                                args.num_k_per_part,
                                OUT_DIMS,
                                args.projection_nlayers)

    # 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)
    dino.requires_grad_(False)

    ##### dreamcreature, add sub-concepts token ids ####
    placeholder_token = "<part>"
    initializer_token = None
    placeholder_token_ids_one = add_tokens(tokenizer_one,
                                           text_encoder_one,
                                           placeholder_token,
                                           args.num_parts,
                                           initializer_token)
    placeholder_token_ids_two = add_tokens(tokenizer_two,
                                           text_encoder_two,
                                           placeholder_token,
                                           args.num_parts,
                                           initializer_token)

    # For mixed precision training we cast all non-trainable weigths (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

    # Move unet, vae and text_encoder to device and cast to weight_dtype
    # The VAE is in float32 to avoid NaN losses.
    unet.to(accelerator.device, dtype=weight_dtype)
    if args.pretrained_vae_model_name_or_path is None:
        vae.to(accelerator.device, dtype=torch.float32)
    else:
        vae.to(accelerator.device, dtype=weight_dtype)
    text_encoder_one.to(accelerator.device, dtype=weight_dtype)
    text_encoder_two.to(accelerator.device, dtype=weight_dtype)
    simple_mapper.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")

    # now we will add new LoRA weights to the attention layers
    # Set correct lora layers
    unet_lora_parameters = []
    for attn_processor_name, attn_processor in unet.attn_processors.items():
        # Parse the attention module.
        attn_module = unet
        for n in attn_processor_name.split(".")[:-1]:
            attn_module = getattr(attn_module, n)

        # Set the `lora_layer` attribute of the attention-related matrices.
        attn_module.to_q.set_lora_layer(
            LoRALinearLayer(
                in_features=attn_module.to_q.in_features, out_features=attn_module.to_q.out_features, rank=args.rank
            )
        )
        attn_module.to_k.set_lora_layer(
            LoRALinearLayer(
                in_features=attn_module.to_k.in_features, out_features=attn_module.to_k.out_features, rank=args.rank
            )
        )
        attn_module.to_v.set_lora_layer(
            LoRALinearLayer(
                in_features=attn_module.to_v.in_features, out_features=attn_module.to_v.out_features, rank=args.rank
            )
        )
        attn_module.to_out[0].set_lora_layer(
            LoRALinearLayer(
                in_features=attn_module.to_out[0].in_features,
                out_features=attn_module.to_out[0].out_features,
                rank=args.rank,
            )
        )

        # Accumulate the LoRA params to optimize.
        unet_lora_parameters.extend(attn_module.to_q.lora_layer.parameters())
        unet_lora_parameters.extend(attn_module.to_k.lora_layer.parameters())
        unet_lora_parameters.extend(attn_module.to_v.lora_layer.parameters())
        unet_lora_parameters.extend(attn_module.to_out[0].lora_layer.parameters())

    setup_attn_processors(unet, args)

    # The text encoder comes from 🤗 transformers, so we cannot directly modify it.
    # So, instead, we monkey-patch the forward calls of its attention-blocks.
    if args.train_text_encoder:
        # ensure that dtype is float32, even if rest of the model that isn't trained is loaded in fp16
        text_lora_parameters_one = LoraLoaderMixin._modify_text_encoder(
            text_encoder_one, dtype=torch.float32, rank=args.rank
        )
        text_lora_parameters_two = LoraLoaderMixin._modify_text_encoder(
            text_encoder_two, dtype=torch.float32, rank=args.rank
        )

    # 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
            mapper_to_save = None

            for model in models:
                if isinstance(model, type(accelerator.unwrap_model(unet))):
                    unet_lora_layers_to_save = unet_attn_processors_state_dict(model)
                elif isinstance(model, type(accelerator.unwrap_model(text_encoder_one))):
                    text_encoder_one_lora_layers_to_save = text_encoder_lora_state_dict(model)
                elif isinstance(model, type(accelerator.unwrap_model(text_encoder_two))):
                    text_encoder_two_lora_layers_to_save = text_encoder_lora_state_dict(model)
                elif isinstance(model, TokenMapper):
                    mapper_to_save = model.state_dict()
                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,
            )
            torch.save(mapper_to_save, output_dir + '/hash_mapper.pth')

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

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

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

        lora_state_dict, network_alphas = LoraLoaderMixin.lora_state_dict(input_dir)
        LoraLoaderMixin.load_lora_into_unet(lora_state_dict, network_alphas=network_alphas, unet=unet_)

        text_encoder_state_dict = {k: v for k, v in lora_state_dict.items() if "text_encoder." in k}
        LoraLoaderMixin.load_lora_into_text_encoder(
            text_encoder_state_dict, network_alphas=network_alphas, text_encoder=text_encoder_one_
        )

        text_encoder_2_state_dict = {k: v for k, v in lora_state_dict.items() if "text_encoder_2." in k}
        LoraLoaderMixin.load_lora_into_text_encoder(
            text_encoder_2_state_dict, network_alphas=network_alphas, text_encoder=text_encoder_two_
        )
        mapper_.load_state_dict(torch.load(input_dir + '/hash_mapper.pth'))

    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:
        torch.backends.cuda.matmul.allow_tf32 = True

    if args.scale_lr:
        args.learning_rate = (
                args.learning_rate * args.gradient_accumulation_steps * args.train_batch_size * accelerator.num_processes
        )

    # Use 8-bit Adam for lower memory usage or to fine-tune the model in 16GB GPUs
    if args.use_8bit_adam:
        try:
            import bitsandbytes as bnb
        except ImportError:
            raise ImportError(
                "To use 8-bit Adam, please install the bitsandbytes library: `pip install bitsandbytes`."
            )

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

    extra_params = list(simple_mapper.parameters())
    mapper_lr = args.learning_rate * args.mapper_lr_scale if args.learning_rate != 0 else args.mapper_lr

    # Optimizer creation
    params_to_optimize = (
        itertools.chain(unet_lora_parameters, text_lora_parameters_one, text_lora_parameters_two)
        if args.train_text_encoder
        else unet_lora_parameters
    )
    optimizer = optimizer_class(
        [{'params': params_to_optimize},
         {'params': extra_params, 'lr': mapper_lr}],
        lr=args.learning_rate,
        betas=(args.adam_beta1, args.adam_beta2),
        weight_decay=args.adam_weight_decay,
        eps=args.adam_epsilon,
    )

    # create
    train_dataset = DreamCreatureDataset(args.train_data_dir,
                                         args.filename,
                                         code_filename=args.code_filename,
                                         num_parts=args.num_parts,
                                         num_k_per_part=args.num_k_per_part,
                                         repeat=args.repeat,
                                         use_gt_label=args.use_gt_label,
                                         bg_code=args.bg_code)

    with accelerator.main_process_first():
        if args.filter_class is not None:
            train_dataset.filter_by_class(args.filter_class)
            print('selected', len(train_dataset))
        if args.max_train_samples is not None:
            train_dataset.set_max_samples(args.max_train_samples, args.seed)

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

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

    lr_scheduler = get_scheduler(
        args.lr_scheduler,
        optimizer=optimizer,
        num_warmup_steps=args.lr_warmup_steps * args.gradient_accumulation_steps,
        num_training_steps=args.max_train_steps * args.gradient_accumulation_steps,
    )

    # Prepare everything with our `accelerator`.
    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
        )
    simple_mapper = accelerator.prepare(simple_mapper)

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

    # We need to initialize the trackers we use, and also store our configuration.
    # The trackers initializes automatically on the main process.
    if accelerator.is_main_process:
        accelerator.init_trackers("text2image-fine-tune", config=vars(args))

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

    logger.info("***** Running training *****")
    logger.info(f"  Num examples = {len(train_dataset)}")
    logger.info(f"  Num Epochs = {args.num_train_epochs}")
    logger.info(f"  Instantaneous batch size per device = {args.train_batch_size}")
    logger.info(f"  Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}")
    logger.info(f"  Gradient Accumulation steps = {args.gradient_accumulation_steps}")
    logger.info(f"  Total optimization steps = {args.max_train_steps}")
    global_step = 0
    first_epoch = 0

    # Potentially load in the weights and states from a previous save
    if args.resume_from_checkpoint:
        if args.resume_from_checkpoint != "latest":
            path = os.path.basename(args.resume_from_checkpoint)
        else:
            # Get the most recent checkpoint
            dirs = os.listdir(args.output_dir)
            dirs = [d for d in dirs if d.startswith("checkpoint")]
            dirs = sorted(dirs, key=lambda x: int(x.split("-")[1]))
            path = dirs[-1] if len(dirs) > 0 else None

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

            initial_global_step = global_step
            first_epoch = global_step // num_update_steps_per_epoch

    else:
        initial_global_step = 0

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

    for epoch in range(first_epoch, args.num_train_epochs):
        unet.train()
        if args.train_text_encoder:
            text_encoder_one.train()
            text_encoder_two.train()
        train_loss = 0.0
        train_diff_loss = 0.0
        train_attn_loss = 0.0
        for step, batch in enumerate(train_dataloader):
            with accelerator.accumulate(unet, simple_mapper):
                # Convert images to latent space
                if args.pretrained_vae_model_name_or_path is not None:
                    pixel_values = batch["pixel_values"].to(dtype=weight_dtype)
                else:
                    pixel_values = batch["pixel_values"]

                model_input = vae.encode(pixel_values).latent_dist.sample()
                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)

                # Sample noise that we'll add to the latents
                noise = torch.randn_like(model_input)
                if args.noise_offset:
                    # https://www.crosslabs.org//blog/diffusion-with-offset-noise
                    noise += args.noise_offset * torch.randn(
                        (model_input.shape[0], model_input.shape[1], 1, 1), device=model_input.device
                    )

                bsz = model_input.shape[0]
                # Sample a random timestep for each image
                timesteps = torch.randint(
                    0, noise_scheduler.config.num_train_timesteps, (bsz,), device=model_input.device
                )
                timesteps = timesteps.long()

                # 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)

                # 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

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

                # Predict the noise residual
                unet_added_conditions = {"time_ids": add_time_ids}
                # prompt_embeds, pooled_prompt_embeds = encode_prompt(
                #     text_encoders=[text_encoder_one, text_encoder_two],
                #     tokenizers=None,
                #     prompt=None,
                #     text_input_ids_list=[batch["input_ids_one"], batch["input_ids_two"]],
                # )
                mapper_outputs = simple_mapper(batch['codes'])
                prompt_embeds, pooled_prompt_embeds = encode_prompt(
                    text_encoders=[text_encoder_one, text_encoder_two],
                    text_input_ids_list=[batch["input_ids_one"], batch["input_ids_two"]],
                    placeholder_token_ids=placeholder_token_ids_one,
                    mapper_outputs=[mapper_outputs[..., :768], mapper_outputs[..., 768:]]
                )

                unet_added_conditions.update({"text_embeds": pooled_prompt_embeds})
                model_pred = unet(
                    noisy_model_input, timesteps, prompt_embeds, added_cond_kwargs=unet_added_conditions
                ).sample

                # Get the target for loss depending on the prediction type
                if args.prediction_type is not None:
                    # set prediction_type of scheduler if defined
                    noise_scheduler.register_to_config(prediction_type=args.prediction_type)

                if noise_scheduler.config.prediction_type == "epsilon":
                    target = noise
                elif noise_scheduler.config.prediction_type == "v_prediction":
                    target = noise_scheduler.get_velocity(model_input, noise, timesteps)
                else:
                    raise ValueError(f"Unknown prediction type {noise_scheduler.config.prediction_type}")

                if args.snr_gamma is None:
                    loss = F.mse_loss(model_pred.float(), target.float(), reduction="none")
                    attn_loss, max_attn = dreamcreature_loss(batch,
                                                             unet,
                                                             dino,
                                                             seg,
                                                             placeholder_token_ids_one,
                                                             accelerator)
                    if args.masked_training:
                        masks = batch['masks'].unsqueeze(1).to(accelerator.device)
                        loss_image_mask = F.interpolate(masks.float(),
                                                        size=target.shape[-2:],
                                                        mode='bilinear') * torch.ones_like(target)
                        loss = loss * loss_image_mask
                        loss = loss.sum() / loss_image_mask.sum()
                    else:
                        loss = loss.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)
                    if noise_scheduler.config.prediction_type == "v_prediction":
                        # Velocity objective requires that we add one to SNR values before we divide by them.
                        snr = snr + 1
                    mse_loss_weights = (
                            torch.stack([snr, args.snr_gamma * torch.ones_like(timesteps)], dim=1).min(dim=1)[0] / snr
                    )

                    loss = F.mse_loss(model_pred.float(), target.float(), reduction="none")
                    attn_loss, max_attn = dreamcreature_loss(batch,
                                                             unet,
                                                             dino,
                                                             seg,
                                                             placeholder_token_ids_one,
                                                             accelerator)
                    if args.masked_training:
                        masks = batch['masks'].unsqueeze(1).to(accelerator.device)
                        loss_image_mask = F.interpolate(masks.float(),
                                                        size=target.shape[-2:],
                                                        mode='bilinear') * torch.ones_like(target)
                        loss = loss * loss_image_mask
                        loss = loss.sum(dim=list(range(1, len(loss.shape)))) * mse_loss_weights
                        loss = loss.sum() / loss_image_mask.sum()
                    else:
                        loss = loss.mean(dim=list(range(1, len(loss.shape)))) * mse_loss_weights
                        loss = loss.mean()

                diff_loss = loss.clone().detach()
                avg_diff_loss = accelerator.gather(diff_loss.repeat(args.train_batch_size)).mean()
                train_diff_loss += avg_diff_loss.item() / args.gradient_accumulation_steps

                avg_attn_loss = accelerator.gather(attn_loss.repeat(args.train_batch_size)).mean()
                train_attn_loss += avg_attn_loss.item() / args.gradient_accumulation_steps

                loss += args.attn_loss * attn_loss

                # Gather the losses across all processes for logging (if we use distributed training).
                avg_loss = accelerator.gather(loss.repeat(args.train_batch_size)).mean()
                train_loss += avg_loss.item() / args.gradient_accumulation_steps

                # Backpropagate
                accelerator.backward(loss)
                if accelerator.sync_gradients:
                    params_to_clip = (
                        itertools.chain(unet_lora_parameters, text_lora_parameters_one, text_lora_parameters_two)
                        if args.train_text_encoder
                        else unet_lora_parameters
                    )
                    params_to_clip = list(params_to_clip) + extra_params
                    accelerator.clip_grad_norm_(params_to_clip, args.max_grad_norm)
                optimizer.step()
                lr_scheduler.step()
                optimizer.zero_grad()

            # Checks if the accelerator has performed an optimization step behind the scenes
            if accelerator.sync_gradients:
                progress_bar.update(1)
                global_step += 1
                accelerator.log({"train_loss": train_loss,
                                 "diff_loss": train_diff_loss,
                                 "attn_loss": train_attn_loss,
                                 "max_attn": max_attn.item()
                                 }, step=global_step)
                train_loss = 0.0
                train_attn_loss = 0.0
                train_diff_loss = 0.0

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

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

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

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

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

            logs = {"step_loss": diff_loss.detach().item(),
                    "attn_loss": attn_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:
            # todo: change pipeline
            if args.validation_prompt is not None and epoch % args.validation_epochs == 0:
                logger.info(
                    f"Running validation... \n Generating {args.num_validation_images} images with prompt:"
                    f" {args.validation_prompt}."
                )
                # create pipeline
                pipeline = DreamCreatureSDXLPipeline.from_pretrained(
                    args.pretrained_model_name_or_path,
                    vae=vae,
                    tokenizer=tokenizer_one,
                    tokenizer_2=tokenizer_two,
                    text_encoder=accelerator.unwrap_model(text_encoder_one),
                    text_encoder_2=accelerator.unwrap_model(text_encoder_two),
                    unet=accelerator.unwrap_model(unet),
                    revision=args.revision,
                    variant=args.variant,
                    torch_dtype=weight_dtype,
                )
                pipeline.placeholder_token_ids = placeholder_token_ids_one
                pipeline.simple_mapper = accelerator.unwrap_model(simple_mapper)
                pipeline.replace_token = False

                pipeline = pipeline.to(accelerator.device)
                pipeline.set_progress_bar_config(disable=True)

                # run inference
                generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed else None
                pipeline_args = {"prompt": args.validation_prompt}

                num_steps = 4 if 'turbo' in args.pretrained_model_name_or_path else 25
                gs = 0 if 'turbo' in args.pretrained_model_name_or_path else 5.0

                images = [
                    pipeline(**pipeline_args, num_inference_steps=num_steps, guidance_scale=gs,
                             generator=generator, height=args.resolution, width=args.resolution).images[0]
                    for _ in range(args.num_validation_images)
                ]

                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")
                    if tracker.name == "wandb":
                        tracker.log(
                            {
                                "validation": [
                                    wandb.Image(image, caption=f"{i}: {args.validation_prompt}")
                                    for i, image in enumerate(images)
                                ]
                            }
                        )

                del pipeline
                torch.cuda.empty_cache()

    # Save the lora layers
    accelerator.wait_for_everyone()
    if accelerator.is_main_process:
        unet = accelerator.unwrap_model(unet)
        unet_lora_layers = unet_attn_processors_state_dict(unet)

        if args.train_text_encoder:
            text_encoder_one = accelerator.unwrap_model(text_encoder_one)
            text_encoder_lora_layers = text_encoder_lora_state_dict(text_encoder_one)
            text_encoder_two = accelerator.unwrap_model(text_encoder_two)
            text_encoder_2_lora_layers = text_encoder_lora_state_dict(text_encoder_two)
        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,
        )
        torch.save(simple_mapper.to(torch.float32).state_dict(), args.output_dir + '/hash_mapper.pth')

        del unet
        del text_encoder_one
        del text_encoder_two
        del text_encoder_lora_layers
        del text_encoder_2_lora_layers
        del simple_mapper
        torch.cuda.empty_cache()

        # Final inference
        # Load previous pipeline
        text_encoder_one, text_encoder_two, tokenizer_one, tokenizer_two, simple_mapper = init_for_pipeline(args)
        pipeline = DreamCreatureSDXLPipeline.from_pretrained(
            args.pretrained_model_name_or_path,
            vae=vae,
            tokenizer=tokenizer_one,
            tokenizer_2=tokenizer_two,
            text_encoder=text_encoder_one,
            text_encoder_2=text_encoder_two,
            revision=args.revision,
            variant=args.variant,
            torch_dtype=weight_dtype,
        )
        pipeline.placeholder_token_ids = placeholder_token_ids_one
        pipeline.replace_token = False
        pipeline.simple_mapper = simple_mapper
        pipeline.simple_mapper.load_state_dict(torch.load(args.output_dir + '/hash_mapper.pth', map_location='cpu'))
        pipeline.simple_mapper.to(accelerator.device)
        setup_attn_processors(pipeline.unet, args)

        pipeline = pipeline.to(accelerator.device)

        # load attention processors
        pipeline.load_lora_weights(args.output_dir)

        # run inference
        images = []
        if args.validation_prompt and args.num_validation_images > 0:
            num_steps = 4 if 'turbo' in args.pretrained_model_name_or_path else 25
            gs = 0 if 'turbo' in args.pretrained_model_name_or_path else 5.0
            generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed else None
            images = [
                pipeline(args.validation_prompt, num_inference_steps=num_steps,
                         guidance_scale=gs, generator=generator, height=args.resolution,
                         width=args.resolution).images[0]
                for _ in range(args.num_validation_images)
            ]

            for tracker in accelerator.trackers:
                if tracker.name == "tensorboard":
                    np_images = np.stack([np.asarray(img) for img in images])
                    tracker.writer.add_images("test", np_images, epoch, dataformats="NHWC")
                if tracker.name == "wandb":
                    tracker.log(
                        {
                            "test": [
                                wandb.Image(image, caption=f"{i}: {args.validation_prompt}")
                                for i, image in enumerate(images)
                            ]
                        }
                    )

        if args.push_to_hub:
            save_model_card(
                repo_id,
                images=images,
                base_model=args.pretrained_model_name_or_path,
                dataset_name=args.dataset_name,
                train_text_encoder=args.train_text_encoder,
                repo_folder=args.output_dir,
                vae_path=args.pretrained_vae_model_name_or_path,
            )
            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__":
    args = parse_args()
    main(args)