train_SDFT.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.

import argparse
import functools
import gc
import logging
import math
import os
import random
from pathlib import Path

import datasets
import numpy as np
import torch
import torch.nn.functional as F
import torch.utils.checkpoint
import transformers
from accelerate import Accelerator
from accelerate.logging import get_logger
from accelerate.utils import DistributedDataParallelKwargs, ProjectConfiguration, set_seed
from datasets import Dataset, DatasetDict, Image
from packaging import version
from peft import LoraConfig
from peft.utils import get_peft_model_state_dict
from torchvision import transforms
from tqdm.auto import tqdm
from transformers import CLIPTextModel, CLIPTokenizer

import diffusers
from diffusers import AutoencoderKL, DDPMScheduler, StableDiffusionPipeline, UNet2DConditionModel

from diffusers.optimization import get_scheduler
from diffusers.training_utils import compute_snr
from diffusers.utils import check_min_version, convert_state_dict_to_diffusers
from diffusers.utils.import_utils import is_xformers_available
from diffusers.utils.torch_utils import is_compiled_module

from glob import glob
from os.path import join as ospj
from torchvision.utils import save_image

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

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


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="stabilityai/stable-diffusion-2-inpainting",
        help="Path to pretrained model or model identifier from huggingface.co/models.",
    )
    parser.add_argument(
        "--revision",
        type=str,
        default=None,
        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(
        "--image_column", 
        type=str, 
        default="image", 
        help="The column of the dataset containing an image."
    )
    parser.add_argument(
        "--mask_column", 
        type=str, 
        default="mask_image", 
        help="The column of the dataset containing an mask."
    )
    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(
        "--data_split",
        type=float,
        default=1.0,
        help="The split of Train/Eval dataset.",
    )
    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(
        "--data_dir",
        type=str,
        default=None,
        help="The data directory.",
    )
    parser.add_argument(
        "--output_dir",
        type=str,
        default="output/SDFT_weights",
        help="The output directory where the model predictions and checkpoints will be written.",
    )
    parser.add_argument(
        "--cache_dir",
        type=str,
        default="./cached",
        help="The directory where the downloaded models and datasets will be stored.",
    )
    parser.add_argument("--seed", type=int, default=0, 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(
        "--center_crop",
        type=bool,
        default=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",
        type=bool,
        default=False,
        help="whether to randomly flip images horizontally",
    )
    parser.add_argument(
        "--train_batch_size",
        type=int, 
        default=8, 
        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=100,
        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(
        "--gradient_accumulation_steps",
        type=int,
        default=8,
        help="Number of updates steps to accumulate before performing a backward/update 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",
        type=bool,
        default=True,
        help="Scale the learning rate by the number of GPUs, gradient accumulation steps, and batch size.",
    )
    parser.add_argument(
        "--lr_scheduler",
        type=str,
        default="linear",
        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=0, 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",
        type=bool,
        default=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("--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(
        "--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(
        "--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", 
        type=bool,
        default=True,
        help="Whether or not to use xformers."
    )
    parser.add_argument(
        "--enable_lora",
        type=bool,
        default=True,
        help=("whether to use lora fine-tune"),
    )
    parser.add_argument(
        "--rank",
        type=int,
        default=32,
        help=("The dimension of the LoRA update matrices."),
    )
    parser.add_argument(
        "--data_index",
        type=float,
        default=1.1,
        help=("Data case index"),
    )

    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

    return args


def compute_vae_encodings(batch, vae):

    batch.pop(args.image_column)
    batch.pop(args.mask_column)
    batch.pop(args.caption_column)

    images = batch.pop("pixel_values")
    mask_images = batch.pop("mask_images")

    pixel_values = torch.stack(list(images))
    pixel_values = pixel_values.to(memory_format=torch.contiguous_format).float()
    pixel_values = pixel_values.to(vae.device, dtype=vae.dtype)

    mask_images = torch.stack(list(mask_images))
    mask_images = mask_images.to(memory_format=torch.contiguous_format).float()
    mask_images = mask_images.to(vae.device, dtype=vae.dtype)

    masked_images = pixel_values * (1 - mask_images)

    with torch.no_grad():
        model_input = vae.encode(pixel_values).latent_dist.sample()
        masked = vae.encode(masked_images).latent_dist.sample()

    model_input = model_input * vae.config.scaling_factor
    masked = masked * vae.config.scaling_factor
    mask = F.interpolate(mask_images, size=(model_input.shape[2:]))

    return {
        "model_input": model_input.cpu(),
        "mask": mask.cpu(), 
        "masked": masked.cpu(),
        "input_ids": batch["input_ids"].cpu()
    }


def main(args):

    args.data_dir = f"output/SDFT_pseudo_pairs"

    logging_dir = Path(args.output_dir, args.logging_dir)
    os.system(f"rm -rf {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]
    )

    # 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)

    # Load scheduler, tokenizer and models.
    noise_scheduler = DDPMScheduler.from_pretrained(args.pretrained_model_name_or_path, subfolder="scheduler")
    tokenizer = CLIPTokenizer.from_pretrained(
        args.pretrained_model_name_or_path, subfolder="tokenizer", revision=args.revision
    )
    text_encoder = CLIPTextModel.from_pretrained(
        args.pretrained_model_name_or_path, subfolder="text_encoder", revision=args.revision
    )
    vae = AutoencoderKL.from_pretrained(
        args.pretrained_model_name_or_path, subfolder="vae", revision=args.revision, variant=args.variant
    )
    unet = UNet2DConditionModel.from_pretrained(
        args.pretrained_model_name_or_path, subfolder="unet", revision=args.revision, variant=args.variant
    )
    # freeze parameters of models to save more memory
    unet.requires_grad_(False)
    vae.requires_grad_(False)
    text_encoder.requires_grad_(False)

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

    # Freeze the unet parameters before adding adapters
    # for param in unet.parameters():
    #     param.requires_grad_(False)

    unet_lora_config = LoraConfig(
        r=args.rank,
        lora_alpha=args.rank,
        init_lora_weights="gaussian",
        target_modules=["attn1.to_k", "attn1.to_q", "attn1.to_v", "attn1.to_out.0"],
    )

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

    # Add adapter and make sure the trainable params are in float32.
    if args.enable_lora: 
        unet.add_adapter(unet_lora_config)
    else:
        for name, param in unet.named_parameters():
            if "attn1" in name:
                param.requires_grad_(True)
    if args.mixed_precision == "fp16":
        for param in unet.parameters():
            # only upcast trainable parameters (LoRA) into fp32
            if param.requires_grad:
                param.data = param.to(torch.float32)

    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")

    # 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
        )

    optimizer_class = torch.optim.AdamW

    # Optimizer creation
    params_to_optimize = []
    print("Parameter to optimize:")
    print("="*60)
    pnum = 0.0
    pnum_tr = 0.0
    for name, param in unet.named_parameters():
        pnum += param.numel()
        if param.requires_grad:
            print(name)
            pnum_tr += param.numel()
            params_to_optimize.append(param)
    print("-"*60)
    print(f"Total trainable parameters: {pnum_tr/1e6}M of {pnum/1e6}M.")
    print("="*60)

    optimizer = optimizer_class(
        params_to_optimize,
        lr=args.learning_rate,
        betas=(args.adam_beta1, args.adam_beta2),
        weight_decay=args.adam_weight_decay,
        eps=args.adam_epsilon,
    )

    # Get the datasets
    dataset = DatasetDict()

    image_list = sorted(glob(ospj(args.data_dir, "ref_rgb*.png")))
    mask_list = sorted(glob(ospj(args.data_dir, "warp_back_mask*.png")))
    text_list = ["" for _ in range(len(image_list))]
    data_split = int(len(image_list)*args.data_split)

    train_data_dict = {
        "image": image_list[:data_split],
        "mask_image": mask_list[:data_split],
        "text": text_list[:data_split],
    }
    eval_data_dict = {
        "image": image_list[data_split:],
        "mask_image": mask_list[data_split:],
        "text": text_list[data_split:],
    }

    dataset["train"] = Dataset.from_dict(train_data_dict).cast_column("image", Image()).cast_column("mask_image", Image())
    dataset["eval"] = Dataset.from_dict(eval_data_dict).cast_column("image", Image()).cast_column("mask_image", Image())

    # Preprocessing the datasets.
    # We need to tokenize input captions and transform the images.
    def tokenize_captions(examples, is_train=True):
        captions = []
        for caption in examples[args.caption_column]:
            if isinstance(caption, str):
                captions.append(caption)
            elif isinstance(caption, (list, np.ndarray)):
                # take a random caption if there are multiple
                captions.append(random.choice(caption) if is_train else caption[0])
            else:
                raise ValueError(
                    f"Caption column `{args.caption_column}` should contain either strings or lists of strings."
                )
        inputs = tokenizer(
            captions, max_length=tokenizer.model_max_length, padding="max_length", truncation=True, return_tensors="pt"
        )
        return inputs.input_ids

    # Preprocessing the datasets.
    train_image_transforms = transforms.Compose(
        [
            transforms.Resize(args.resolution, interpolation=transforms.InterpolationMode.BILINEAR),
            transforms.CenterCrop(args.resolution) if args.center_crop else transforms.RandomCrop(args.resolution),
            transforms.RandomHorizontalFlip() if args.random_flip else transforms.Lambda(lambda x: x),
            transforms.ToTensor(),
            transforms.Normalize([0.5], [0.5]),
        ]
    )

    train_mask_transforms = transforms.Compose(
        [
            transforms.Resize(args.resolution, interpolation=transforms.InterpolationMode.BILINEAR),
            transforms.CenterCrop(args.resolution) if args.center_crop else transforms.RandomCrop(args.resolution),
            transforms.RandomHorizontalFlip() if args.random_flip else transforms.Lambda(lambda x: x),
            transforms.ToTensor()
        ]
    )

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

    def preprocess_train(examples):
        if "model_input" in examples:
            return examples
        
        images = [image.convert("RGB") for image in examples[args.image_column]]
        mask_images = [mask.convert("L") for mask in examples[args.mask_column]]

        examples["pixel_values"] = [train_image_transforms(image) for image in images]
        examples["mask_images"] = [train_mask_transforms(mask) for mask in mask_images]
        examples["input_ids"] = tokenize_captions(examples) # torch.zeros((len(images), 77), dtype=torch.int64)

        return examples

    compute_vae_encodings_fn = functools.partial(compute_vae_encodings, vae=vae)
    with accelerator.main_process_first():
        from datasets.fingerprint import Hasher

        # fingerprint used by the cache for the other processes to load the result
        # details: https://github.com/huggingface/diffusers/pull/4038#discussion_r1266078401
        new_fingerprint_for_vae = Hasher.hash("vae")
        train_dataset = dataset["train"].with_transform(preprocess_train)
        train_dataset = train_dataset.map(
            compute_vae_encodings_fn,
            batched=True,
            batch_size=args.train_batch_size, #* accelerator.num_processes * args.gradient_accumulation_steps,
            new_fingerprint=new_fingerprint_for_vae
        )

    del vae
    gc.collect()
    torch.cuda.empty_cache()

    def collate_fn(examples):
        model_input = torch.stack([torch.tensor(example["model_input"]) for example in examples])
        mask = torch.stack([torch.tensor(example["mask"]) for example in examples])
        masked = torch.stack([torch.tensor(example["masked"]) for example in examples])
        input_ids = torch.stack([torch.tensor(example["input_ids"]) for example in examples])

        return {
            "model_input": model_input,
            "mask": mask,
            "masked": masked,
            "input_ids": input_ids
        }

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

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

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

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

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

    # We need to initialize the trackers we use, and also store our configuration.
    # The trackers initializes automatically on the main process.
    if accelerator.is_main_process:
        accelerator.init_trackers("fine-tune-sd2_lora", 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
    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()
        train_loss = 0.0
        for step, batch in enumerate(train_dataloader):
            with accelerator.accumulate(unet): 
                # Sample noise that we'll add to the latents
                model_input = batch["model_input"].to(accelerator.device)
                mask = batch["mask"].to(accelerator.device)
                masked = batch["masked"].to(accelerator.device)

                noise = torch.randn_like(model_input)

                bsz = model_input.shape[0]
                # Sample a random timestep for each image without bias.
                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)
                # Concat mask for inpainting
                noisy_model_input = torch.cat([noisy_model_input, mask, masked], dim=1)

                # Get the text embedding for conditioning
                encoder_hidden_states = text_encoder(batch["input_ids"], return_dict=False)[0]

                # Predict the noise residual and compute loss
                model_pred = unet(noisy_model_input, timesteps, encoder_hidden_states, return_dict=False)[0]

                # 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="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")
                    loss = loss.mean(dim=list(range(1, len(loss.shape)))) * mse_loss_weights
                    loss = loss.mean()

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

                # Backpropagate
                accelerator.backward(loss)
                if accelerator.sync_gradients:
                    accelerator.clip_grad_norm_(params_to_optimize, 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}, step=global_step)
                train_loss = 0.0

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

            if global_step >= args.max_train_steps:
                break

    # Save the lora layers
    accelerator.wait_for_everyone()
    if accelerator.is_main_process:
        unet = unet.to(torch.float32)
        unwrapped_unet = unwrap_model(unet)

        if args.enable_lora:
            unet_lora_state_dict = convert_state_dict_to_diffusers(get_peft_model_state_dict(unwrapped_unet))
            StableDiffusionPipeline.save_lora_weights(
                save_directory=args.output_dir,
                unet_lora_layers=unet_lora_state_dict,
                safe_serialization=True,
            )
        else:
            pipeline = StableDiffusionPipeline.from_pretrained(
                args.pretrained_model_name_or_path,
                unet=unet,
                revision=args.revision,
                variant=args.variant,
                torch_dtype=weight_dtype,
            )
            pipeline.save_pretrained(args.output_dir)

        del unet
        torch.cuda.empty_cache()

    accelerator.end_training()

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