train_stage1.py

import argparse
import logging
import math
import os
from pathlib import Path
import itertools
import numpy as np
import torch.utils.checkpoint
import transformers
from accelerate import Accelerator
from accelerate.logging import get_logger
from accelerate.utils import ProjectConfiguration, set_seed
from datasets import load_dataset
from PIL import Image
from torchvision import transforms
from tqdm.auto import tqdm
from transformers import AutoTokenizer, PretrainedConfig

import diffusers
from diffusers import (
    AutoencoderKL,
    UNet2DConditionModel,
    DDPMScheduler,
    UniPCMultistepScheduler,
)
from diffusers.optimization import get_scheduler
from diffusers.utils import check_min_version, is_wandb_available
import torch.nn.functional as F
import albumentations as A
import cv2
from ref_encoder.latent_controlnet import ControlNetModel
from utils.pipeline_cn import StableDiffusionControlNetPipeline

# Will error if the minimal version of diffusers is not installed. Remove at your own risks.
check_min_version("0.23.0")

logger = get_logger(__name__)

def concatenate_images(image_files, output_file, type="pil"):
    if type == "np":
        image_files = [Image.fromarray(img) for img in image_files]
    images = image_files  # list
    max_height = max(img.height for img in images)
    images = [img.resize((img.width, max_height)) for img in images]
    total_width = sum(img.width for img in images)
    combined = Image.new('RGB', (total_width, max_height))
    x_offset = 0
    for img in images:
        combined.paste(img, (x_offset, 0))
        x_offset += img.width
    combined.save(output_file)

def image_grid(imgs, rows, cols):
    assert len(imgs) == rows * cols
    w, h = imgs[0].size
    grid = Image.new("RGB", size=(cols * w, rows * h))
    for i, img in enumerate(imgs):
        grid.paste(img, box=(i % cols * w, i // cols * h))
    return grid

def log_validation(vae, text_encoder, tokenizer, unet, controlnet, args, accelerator, weight_dtype, step):
    logger.info("Running validation... ")
    controlnet = accelerator.unwrap_model(controlnet)
    pipeline = StableDiffusionControlNetPipeline.from_pretrained(
        args.pretrained_model_name_or_path,
        vae=vae,
        text_encoder=text_encoder,
        tokenizer=tokenizer,
        unet=unet,
        controlnet=controlnet,
        safety_checker=None,
        revision=args.revision,
        torch_dtype=weight_dtype,
    )
    pipeline.scheduler = UniPCMultistepScheduler.from_config(pipeline.scheduler.config)
    pipeline = pipeline.to(accelerator.device)
    pipeline.set_progress_bar_config(disable=True)

    validation_ids = args.validation_ids
    validation_path = os.path.join(args.output_dir, "validation", f"step-{step}")
    os.makedirs(validation_path, exist_ok=True)
    _num = 0
    for validation_id in validation_ids:
        _num += 1
        validation_id = Image.open(validation_id).convert("RGB").resize((512, 512))
        for num in range(args.num_validation_images):
            with torch.autocast("cuda"):
                sample = pipeline(
                    prompt="",
                    negative_prompt="",
                    num_inference_steps=30,
                    guidance_scale=1.000001,
                    width=512,
                    height=512,
                    image=validation_id,
                    controlnet_conditioning_scale=1.,
                    generator=None,
                ).images[0]
                concatenate_images([validation_id, sample],
                                   output_file=os.path.join(validation_path, str(num)+str(_num)+".jpg"), type="pil")


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

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

        return CLIPTextModel
    elif model_class == "RobertaSeriesModelWithTransformation":
        from diffusers.pipelines.alt_diffusion.modeling_roberta_series import RobertaSeriesModelWithTransformation

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


def parse_args(input_args=None):
    parser = argparse.ArgumentParser(description="Simple example of training script.")
    parser.add_argument("--noise_offset", type=float, default=0.1, help="The scale of noise offset.")
    parser.add_argument(
        "--pretrained_model_name_or_path",
        type=str,
        default="/share/zhangyuxuan/project/workspace/sd_model_v1-5",
        help="Path to pretrained model or model identifier from huggingface.co/models."
    )
    parser.add_argument(
        "--controlnet_model_name_or_path",
        type=str,
        default=None,
        help="Path to pretrained controlnet model or model identifier from huggingface.co/models."
        " If not specified controlnet weights are initialized from unet.",
    )
    parser.add_argument(
        "--train_data_dir",
        type=str,
        default="",
        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("--source_column", type=str, default="image")
    parser.add_argument("--target_column", type=str, default="image")
    parser.add_argument(
        "--revision",
        type=str,
        default=None,
        required=False,
        help=(
            "Revision of pretrained model identifier from huggingface.co/models. Trainable model components should be"
            " float32 precision."
        ),
    )
    parser.add_argument(
        "--output_dir",
        type=str,
        default="train_lr1e-5_refunet",
        help="The output directory where the model predictions and checkpoints will be written.",
    )
    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=1000)
    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=1000,
        help=(
            "Save a checkpoint of the training state every X updates. Checkpoints can be used for resuming training via `--resume_from_checkpoint`. "
            "In the case that the checkpoint is better than the final trained model, the checkpoint can also be used for inference."
            "Using a checkpoint for inference requires separate loading of the original pipeline and the individual checkpointed model components."
            "See https://huggingface.co/docs/diffusers/main/en/training/dreambooth#performing-inference-using-a-saved-checkpoint for step by step"
            "instructions."
        ),
    )
    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-5,
        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(
        "--lr_num_cycles",
        type=int,
        default=1,
        help="Number of hard resets of the lr in cosine_with_restarts scheduler.",
    )
    parser.add_argument("--lr_power", type=float, default=1.0, help="Power factor of the polynomial scheduler.")
    parser.add_argument(
        "--use_8bit_adam", action="store_true", help="Whether or not to use 8-bit Adam from bitsandbytes."
    )
    parser.add_argument(
        "--dataloader_num_workers",
        type=int,
        default=8,
        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=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(
        "--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="no",
        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(
        "--enable_xformers_memory_efficient_attention", action="store_true", help="Whether or not to use xformers."
    )

    parser.add_argument(
        "--max_train_samples",
        type=int,
        default=None,
        help=(
            "For debugging purposes or quicker training, truncate the number of training examples to this "
            "value if set."
        ),
    )
    parser.add_argument(
        "--proportion_empty_prompts",
        type=float,
        default=0,
        help="Proportion of image prompts to be replaced with empty strings. Defaults to 0 (no prompt replacement).",
    )
    parser.add_argument(
        "--validation_ids",
        type=str,
        default=["", ""],
        nargs="+",
        help=(
            "A set of prompts evaluated every `--validation_steps` and logged to `--report_to`."
            " Provide either a matching number of `--validation_image`s, a single `--validation_image`"
            " to be used with all prompts, or a single prompt that will be used with all `--validation_image`s."
        ),
    )
    parser.add_argument(
        "--validation_hairs",
        type=str,
        default=["", ""],
        nargs="+",
        help=(
            "A set of paths to the controlnet conditioning image be evaluated every `--validation_steps`"
            " and logged to `--report_to`. Provide either a matching number of `--validation_prompt`s, a"
            " a single `--validation_prompt` to be used with all `--validation_image`s, or a single"
            " `--validation_image` that will be used with all `--validation_prompt`s."
        ),
    )
    parser.add_argument(
        "--num_validation_images",
        type=int,
        default=3,
        help="Number of images to be generated for each `--validation_image`, `--validation_prompt` pair",
    )
    parser.add_argument(
        "--validation_steps",
        type=int,
        default=10,
        help=(
            "Run validation every X steps. Validation consists of running the prompt"
            " `args.validation_prompt` multiple times: `args.num_validation_images`"
            " and logging the images."
        ),
    )
    parser.add_argument(
        "--tracker_project_name",
        type=str,
        default="train",
        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"
        ),
    )

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

    if args.resolution % 8 != 0:
        raise ValueError(
            "`--resolution` must be divisible by 8 for consistently sized encoded images between the VAE and the controlnet encoder."
        )

    return args


def make_train_dataset(args, tokenizer, accelerator):

    if args.train_data_dir is not None:
        dataset = load_dataset('json', data_files=args.train_data_dir)
    column_names = dataset["train"].column_names

    # 6. Get the column names for input/target.
    if args.source_column is None:
        source_column = column_names[1]
        logger.info(f"source column defaulting to {source_column}")
    else:
        source_column = args.source_column
        if source_column not in column_names:
            raise ValueError(
                f"`--source_column` value '{args.source_column}' not found in dataset columns. Dataset columns are: {', '.join(column_names)}"
            )

    if args.target_column is None:
        target_column = column_names[1]
        logger.info(f"target column defaulting to {target_column}")
    else:
        target_column = args.target_column
        if target_column not in column_names:
            raise ValueError(
                f"`--target_column` value '{args.target_column}' not found in dataset columns. Dataset columns are: {', '.join(column_names)}"
            )

    norm = transforms.Normalize([0.5], [0.5])
    to_tensor = transforms.ToTensor()

    pixel_transform = A.Compose([
        A.SmallestMaxSize(max_size=512),
        A.CenterCrop(512, 512),
        A.Affine(scale=(0.5, 1), translate_percent={"x": (-0.1, 0.1), "y": (-0.1, 0.1)}, rotate=(-10, 10), p=0.8),
    ], additional_targets={'image0': 'image', 'image1': 'image'})

    def imgaug(source_image, target_image):
        source_image = cv2.resize(cv2.cvtColor(source_image, cv2.COLOR_BGR2RGB), [512, 512])
        target_image = cv2.resize(cv2.cvtColor(target_image, cv2.COLOR_BGR2RGB), [512, 512])
        results = pixel_transform(image=source_image, image0=target_image)
        source_image, target_image = norm(to_tensor(results["image"]/255.)), norm(to_tensor(results["image0"]/255.))
        return source_image, target_image

    def preprocess_train(examples):
        source_images = [cv2.imread(image) for image in examples[source_column]]
        target_images = [cv2.imread(image) for image in examples[target_column]]

        pair = [imgaug(image1, image2) for image1, image2 in zip(source_images, target_images)]
        source_images, target_images = zip(*pair)
        source_images_ls = list(source_images)
        target_images_ls = list(target_images)

        examples["source_pixel_values"] = source_images_ls
        examples["target_pixel_values"] = target_images_ls
        return examples

    with accelerator.main_process_first():
        train_dataset = dataset["train"].with_transform(preprocess_train)

    return train_dataset


def collate_fn(examples):
    source_pixel_values = torch.stack([example["source_pixel_values"] for example in examples])
    source_pixel_values = source_pixel_values.to(memory_format=torch.contiguous_format).float()
    target_pixel_values = torch.stack([example["target_pixel_values"] for example in examples])
    target_pixel_values = target_pixel_values.to(memory_format=torch.contiguous_format).float()

    return {
        "source_pixel_values": source_pixel_values,
        "target_pixel_values": target_pixel_values,
    }

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

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

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

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

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

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

    # Load the tokenizer
    if args.pretrained_model_name_or_path:
        tokenizer = AutoTokenizer.from_pretrained(
            args.pretrained_model_name_or_path,
            subfolder="tokenizer",
            revision=args.revision,
            use_fast=False,
        )

    # import correct text encoder class
    text_encoder_cls = import_model_class_from_model_name_or_path(args.pretrained_model_name_or_path, args.revision)

    # Load scheduler and models
    noise_scheduler = DDPMScheduler.from_pretrained(args.pretrained_model_name_or_path, subfolder="scheduler")
    text_encoder = text_encoder_cls.from_pretrained(
        args.pretrained_model_name_or_path, subfolder="text_encoder", revision=args.revision
    ).to(accelerator.device)
    vae = AutoencoderKL.from_pretrained(args.pretrained_model_name_or_path, subfolder="vae", revision=args.revision).to(accelerator.device)
    unet = UNet2DConditionModel.from_pretrained(
        args.pretrained_model_name_or_path, subfolder="unet", revision=args.revision
    ).to(accelerator.device)
    if args.controlnet_model_name_or_path:
        logger.info("Loading existing controlnet weights")
        controlnet = ControlNetModel.from_pretrained(args.controlnet_model_name_or_path).to(accelerator.device)
    else:
        logger.info("Initializing controlnet weights from unet")
        controlnet = ControlNetModel.from_unet(unet).to(accelerator.device)

    vae.requires_grad_(False)
    text_encoder.requires_grad_(False)
    unet.requires_grad_(False)
    controlnet.requires_grad_(True)

    optimizer_class = torch.optim.AdamW
    # Optimizer creation
    params_to_optimize = itertools.chain(controlnet.parameters())
    optimizer = optimizer_class(
        params_to_optimize,
        lr=args.learning_rate,
        betas=(args.adam_beta1, args.adam_beta2),
        weight_decay=args.adam_weight_decay,
        eps=args.adam_epsilon,
    )

    train_dataset = make_train_dataset(args, tokenizer, accelerator)
    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 * accelerator.num_processes,
        num_training_steps=args.max_train_steps * accelerator.num_processes,
        num_cycles=args.lr_num_cycles,
        power=args.lr_power,
    )

    # Prepare everything with our `accelerator`.
    controlnet, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
        controlnet, optimizer, train_dataloader, lr_scheduler
    )

    # For mixed precision training we cast the text_encoder and vae weights to half-precision
    # as these models 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 vae, unet and text_encoder to device and cast to weight_dtype
    vae.to(accelerator.device, dtype=weight_dtype)
    unet.to(accelerator.device, dtype=weight_dtype)
    text_encoder.to(accelerator.device, dtype=weight_dtype)
    controlnet.to(accelerator.device, dtype=torch.float32)

    # 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))
        # tensorboard cannot handle list types for config
        tracker_config.pop("validation_hairs")
        tracker_config.pop("validation_ids")
        accelerator.init_trackers(args.tracker_project_name, config=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 batches each epoch = {len(train_dataloader)}")
    logger.info(f"  Num Epochs = {args.num_train_epochs}")
    logger.info(f"  Instantaneous batch size per device = {args.train_batch_size}")
    logger.info(f"  Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}")
    logger.info(f"  Gradient Accumulation steps = {args.gradient_accumulation_steps}")
    logger.info(f"  Total optimization steps = {args.max_train_steps}")
    global_step = 0
    first_epoch = 0
    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,
    )

    null_text_inputs = tokenizer(
        "", max_length=tokenizer.model_max_length, padding="max_length", truncation=True, return_tensors="pt"
    ).input_ids
    encoder_hidden_states = text_encoder(null_text_inputs.to(device=accelerator.device))[0]

    for epoch in range(first_epoch, args.num_train_epochs):
        for step, batch in enumerate(train_dataloader):
            with accelerator.accumulate(controlnet):

                # Convert images to latent space
                latents = vae.encode(batch["target_pixel_values"].to(dtype=weight_dtype)).latent_dist.sample()
                latents = latents * vae.config.scaling_factor

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

                # Add noise to the latents according to the noise magnitude at each timestep
                # (this is the forward diffusion process)
                noisy_latents = noise_scheduler.add_noise(latents, noise, timesteps)
                # ref_noisy_latents = noise_scheduler.add_noise(ref_latents, noise, timesteps)

                content_latents = vae.encode(batch["source_pixel_values"].to(dtype=weight_dtype)).latent_dist.sample()
                content_latents = content_latents * vae.config.scaling_factor
                down_block_res_samples, mid_block_res_sample = controlnet(
                    noisy_latents,

                    timesteps,
                    encoder_hidden_states=encoder_hidden_states.repeat(bsz, 1, 1),
                    controlnet_cond=content_latents,
                    return_dict=False,
                )

                # Predict the noise residual
                model_pred = unet(
                    noisy_latents,
                    timesteps,
                    encoder_hidden_states=encoder_hidden_states.repeat(bsz, 1, 1).to(dtype=weight_dtype),
                    down_block_additional_residuals=[
                        sample.to(dtype=weight_dtype) for sample in down_block_res_samples
                    ],
                    mid_block_additional_residual=mid_block_res_sample.to(dtype=weight_dtype),
                ).sample

                # Get the target for loss depending on the prediction type
                if noise_scheduler.config.prediction_type == "epsilon":
                    target = noise
                elif noise_scheduler.config.prediction_type == "v_prediction":
                    target = noise_scheduler.get_velocity(latents, noise, timesteps)
                else:
                    raise ValueError(f"Unknown prediction type {noise_scheduler.config.prediction_type}")
                loss = F.mse_loss(model_pred.float(), target.float(), reduction="mean")

                accelerator.backward(loss)
                optimizer.step()
                lr_scheduler.step()
                optimizer.zero_grad()

            # Checks if the accelerator has performed an optimization step behind the scenes
            if accelerator.sync_gradients:
                progress_bar.update(1)
                global_step += 1

                if accelerator.is_main_process:
                    if global_step % args.checkpointing_steps == 0:
                        save_path = os.path.join(args.output_dir, f"checkpoint-{global_step}")
                        accelerator.save_state(save_path, safe_serialization=False)
                        logger.info(f"Saved state to {save_path}")

                    if args.validation_ids is not None and global_step % args.validation_steps == 0:
                        log_validation(
                            vae,
                            text_encoder,
                            tokenizer,
                            unet,
                            controlnet,
                            args,
                            accelerator,
                            weight_dtype,
                            global_step,
                        )

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

            if global_step >= args.max_train_steps:
                break

    # Create the pipeline using using the trained modules and save it.
    accelerator.wait_for_everyone()
    accelerator.end_training()


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