train_lsd.py

import argparse
import copy
import gc
import hashlib
import importlib
import itertools
import logging
import math
import os
import shutil
import warnings
from pathlib import Path

import numpy as np
import torch
import torch.nn.functional as F
import torch.utils.checkpoint
import transformers
from accelerate import Accelerator
from accelerate.logging import get_logger
from accelerate.utils import ProjectConfiguration, set_seed
from huggingface_hub import create_repo
from packaging import version
from PIL import Image
from tqdm.auto import tqdm

import diffusers
from diffusers import (
    AutoencoderKL,
    DDPMScheduler,
    StableDiffusionPipeline,
    UNet2DConditionModel,
)
# from diffusers.training_utils import compute_snr # diffusers is still working on this, uncomment in future versions
from diffusers.utils import check_min_version, is_wandb_available
from diffusers.utils.import_utils import is_xformers_available

from einops import rearrange, reduce, repeat

from src.models.backbone import UNetEncoder
from src.models.slot_attn import MultiHeadSTEVESA
from src.models.unet_with_pos import UNet2DConditionModelWithPos
from src.data.dataset import GlobDataset

from src.parser import parse_args

from src.models.utils import ColorMask

if is_wandb_available():
    import wandb

logger = get_logger(__name__)


@torch.no_grad()
def log_validation(
    val_dataset,
    backbone,
    slot_attn,
    unet,
    vae,
    scheduler,
    args,
    accelerator,
    weight_dtype,
    global_step,
):
    logger.info(
        f"Running validation... \n."
    )
    unet = accelerator.unwrap_model(unet)
    backbone = accelerator.unwrap_model(backbone)
    slot_attn = accelerator.unwrap_model(slot_attn)

    colorizer = ColorMask(
        num_slots=slot_attn.config.num_slots,
        log_img_size=256,
        norm_mean=0,
        norm_std=1,
    )

    val_dataloader = torch.utils.data.DataLoader(
        val_dataset,
        batch_size=args.val_batch_size,
        shuffle=False,
        num_workers=args.dataloader_num_workers,
    )

    # We train on the simplified learning objective. If we were previously predicting a variance, we need the scheduler to ignore it
    scheduler_args = {}

    if "variance_type" in scheduler.config:
        variance_type = scheduler.config.variance_type

        if variance_type in ["learned", "learned_range"]:
            variance_type = "fixed_small"

        scheduler_args["variance_type"] = variance_type

    # use a more efficient scheduler at test time
    module = importlib.import_module("diffusers")
    scheduler_class = getattr(module, args.validation_scheduler)
    scheduler = scheduler_class.from_config(
        scheduler.config, **scheduler_args)

    pipeline = StableDiffusionPipeline(
        vae=vae,
        text_encoder=None,
        tokenizer=None,
        unet=unet,
        scheduler=scheduler,
        safety_checker=None,
        feature_extractor=None,
        requires_safety_checker=None,
    )

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

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

    num_digits = len(str(args.max_train_steps))
    folder_name = f"image_logging_{global_step:0{num_digits}}"
    image_log_dir = os.path.join(accelerator.logging_dir, folder_name, )
    os.makedirs(image_log_dir, exist_ok=True)

    images = []
    image_count = 0

    for batch_idx, batch in enumerate(val_dataloader):

        pixel_values = batch["pixel_values"].to(
            device=accelerator.device, dtype=weight_dtype)

        with torch.autocast("cuda"):
            model_input = vae.encode(pixel_values).latent_dist.sample()
            pixel_values_recon = vae.decode(model_input).sample

            if args.backbone_config == "pretrain_dino":
                pixel_values_vit = batch["pixel_values_vit"].to(device=accelerator.device, 
                                                                dtype=weight_dtype)
                feat = backbone(pixel_values_vit)
            else:
                feat = backbone(pixel_values)
            slots, attn = slot_attn(feat[:, None])  # for the time dimension
            slots = slots[:, 0]
            images_gen = pipeline(
                prompt_embeds=slots,
                height=args.resolution,
                width=args.resolution,
                num_inference_steps=25,
                generator=generator,
                guidance_scale=1.,
                output_type="pt",
            ).images

        grid_image = colorizer.get_heatmap(img=(pixel_values * 0.5 + 0.5),
                                           attn=reduce(
                                               attn[:, 0], 'b num_h (h w) s -> b s h w', h=int(np.sqrt(attn.shape[-2])), 
                                               reduction='mean'
                                           ),
                                           recon=[pixel_values_recon * 0.5 + 0.5, images_gen])
        ndarr = grid_image.mul(255).add_(0.5).clamp_(0, 255).permute(1, 2, 0).to('cpu', torch.uint8).numpy()
        im = Image.fromarray(ndarr)
        images.append(im)
        img_path = os.path.join(image_log_dir, f"image_{batch_idx:02}.jpg")
        im.save(img_path, optimize=True, quality=95)
        image_count += pixel_values.shape[0]
        if image_count >= args.num_validation_images:
            break

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

    del pipeline
    torch.cuda.empty_cache()

    return images


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, # for manually handled case, should not pass it here
        mixed_precision=args.mixed_precision,
        log_with=args.report_to,
        project_config=accelerator_project_config,
    )

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

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

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

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

        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 scheduler and models
    if args.unet_config == "pretrain_sd":
        noise_scheduler = DDPMScheduler.from_pretrained(
        args.pretrained_model_name, subfolder="scheduler")
    else:
        noise_scheduler_config = DDPMScheduler.load_config(args.scheduler_config)
        noise_scheduler = DDPMScheduler.from_config(noise_scheduler_config)

    vae = AutoencoderKL.from_pretrained(
        args.pretrained_model_name, subfolder="vae")
    
    if os.path.exists(args.backbone_config):
        train_backbone = True
        backbone_config = UNetEncoder.load_config(args.backbone_config)
        backbone = UNetEncoder.from_config(backbone_config)
    elif args.backbone_config == "pretrain_dino":
        train_backbone = False
        dinov2 = torch.hub.load("facebookresearch/dinov2", "dinov2_vitb14")
        class DINOBackbone(torch.nn.Module):
            def __init__(self, dinov2):
                super().__init__()
                self.dinov2 = dinov2

            def forward(self, x):
                enc_out = self.dinov2.forward_features(x)
                return rearrange(
                    enc_out["x_norm_patchtokens"], 
                    "b (h w ) c -> b c h w",
                    h=int(np.sqrt(enc_out["x_norm_patchtokens"].shape[-2]))
                )
        backbone = DINOBackbone(dinov2)
    else:
        raise ValueError(
            f"Unknown unet config {args.unet_config}")

    slot_attn_config = MultiHeadSTEVESA.load_config(args.slot_attn_config)
    slot_attn = MultiHeadSTEVESA.from_config(slot_attn_config)

    if os.path.exists(args.unet_config):
        train_unet = True
        unet_config = UNet2DConditionModelWithPos.load_config(args.unet_config)
        unet = UNet2DConditionModelWithPos.from_config(unet_config)
    elif args.unet_config == "pretrain_sd":
        train_unet = False
        unet = UNet2DConditionModel.from_pretrained(
            args.pretrained_model_name, subfolder="unet", revision=args.revision
        )
    else:
        raise ValueError(
            f"Unknown unet config {args.unet_config}")

    # 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:
            for model in models:

                # continue if not one of [UNetEncoder, MultiHeadSTEVESA, UNet2DConditionModelWithPos]
                if not isinstance(model, (UNetEncoder, MultiHeadSTEVESA, UNet2DConditionModelWithPos)):
                    continue

                sub_dir = model._get_name().lower()

                model.save_pretrained(os.path.join(output_dir, sub_dir))

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

    def load_model_hook(models, input_dir):
        while len(models) > 0:
            # pop models so that they are not loaded again
            model = models.pop()

            sub_dir = model._get_name().lower()

            if isinstance(model, UNetEncoder):
                # load diffusers style into model
                load_model = UNetEncoder.from_pretrained(
                    input_dir, subfolder=sub_dir)
                model.register_to_config(**load_model.config)
            elif isinstance(model, MultiHeadSTEVESA):
                load_model = MultiHeadSTEVESA.from_pretrained(
                    input_dir, subfolder=sub_dir)
                model.register_to_config(**load_model.config)
            elif isinstance(model, UNet2DConditionModelWithPos):
                load_model = UNet2DConditionModelWithPos.from_pretrained(
                    input_dir, subfolder=sub_dir)
                model.register_to_config(**load_model.config)
            else:
                raise ValueError(
                    f"Unknown model type {type(model)}")

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

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

    vae.requires_grad_(False)
    if not train_backbone:
        try:
            backbone.requires_grad_(False)
        except:
            pass
    if not train_unet:
        unet.requires_grad_(False)
        
    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()
            try:
                backbone.enable_xformers_memory_efficient_attention()
            except AttributeError:
                pass
        else:
            raise ValueError(
                "xformers is not available. Make sure it is installed correctly")

    if args.gradient_checkpointing:
        unet.enable_gradient_checkpointing()
        try:
            backbone.enable_gradient_checkpointing()
        except AttributeError:
                pass

    # Check that all trainable models are in full precision
    low_precision_error_string = (
        "Please make sure to always have all model weights in full float32 precision when starting training - even if"
        " doing mixed precision training. copy of the weights should still be float32."
    )

    if train_unet and accelerator.unwrap_model(unet).dtype != torch.float32:
        raise ValueError(
            f"Unet loaded as datatype {accelerator.unwrap_model(unet).dtype}. {low_precision_error_string}"
        )

    if train_backbone and accelerator.unwrap_model(backbone).dtype != torch.float32:
        raise ValueError(
            f"Backbone loaded as datatype {accelerator.unwrap_model(backbone).dtype}. {low_precision_error_string}"
        )

    if accelerator.unwrap_model(slot_attn).dtype != torch.float32:
        raise ValueError(
            f"Slot Attn loaded as datatype {accelerator.unwrap_model(slot_attn).dtype}. {low_precision_error_string}"
        )

    # 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


    params_to_optimize = list(slot_attn.parameters()) + \
        (list(backbone.parameters()) if train_backbone else []) + \
        (list(unet.parameters()) if train_unet else [])
    params_group = [
        {'params': list(slot_attn.parameters()) + \
         (list(backbone.parameters()) if train_backbone else []),
         'lr': args.learning_rate * args.encoder_lr_scale}
    ]
    if train_unet:
        params_group.append(
            {'params': unet.parameters(), "lr": args.learning_rate}
        )

    optimizer = optimizer_class(
        params_group,
        lr=args.learning_rate,
        betas=(args.adam_beta1, args.adam_beta2),
        weight_decay=args.adam_weight_decay,
        eps=args.adam_epsilon,
    )
    
    # implement your lr_sceduler here, here I use constant functions as 
    # the template for your reference
    lr_scheduler = torch.optim.lr_scheduler.LambdaLR(
        optimizer, lr_lambda=[lambda _: 1, lambda _: 1] if train_unet else [lambda _: 1]
        )

    train_dataset = GlobDataset(
        root=args.dataset_root,
        img_size=args.resolution,
        img_glob=args.dataset_glob,
        data_portion=(0.0, args.train_split_portion),
        vit_norm=args.backbone_config == "pretrain_dino",
        random_flip=args.flip_images,
        vit_input_resolution=args.vit_input_resolution
    )

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

    # validation set is only for visualization
    val_dataset = GlobDataset(
        root=args.dataset_root,
        img_size=args.resolution,
        img_glob=args.dataset_glob,
        data_portion=(args.train_split_portion if args.train_split_portion < 1. else 0.9, 1.0),
        vit_norm=args.backbone_config == "pretrain_dino",
        vit_input_resolution=args.vit_input_resolution
    )

    # 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

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

    if train_backbone:
        backbone = accelerator.prepare(backbone)
    if train_unet:
        unet = accelerator.prepare(unet)

    # 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 vae device and cast to weight_dtype
    vae.to(accelerator.device, dtype=weight_dtype)
    if not train_backbone:
        try:
            backbone.to(accelerator.device, dtype=weight_dtype)
        except:
            pass
    if not train_unet:
        unet.to(accelerator.device, dtype=weight_dtype)

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

    # We need to initialize the trackers we use, and also store our configuration.
    # The trackers initializes automatically on the main process.
    if accelerator.is_main_process:
        tracker_config = vars(copy.deepcopy(args))
        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
    accumulate_steps = 0 # necessary for args.gradient_accumulation_steps > 1

    # 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.rstrip('/')) # only the checkpoint folder name is needed, not the full path
        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
            accumulate_steps = global_step * args.gradient_accumulation_steps
            first_epoch = global_step // num_update_steps_per_epoch
    else:
        initial_global_step = 0

    # Only show the progress bar once on each machine.
    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,
        position=0, leave=True
    )

    for epoch in range(first_epoch, args.num_train_epochs):
        if train_unet:
            unet.train()
        if train_backbone:
            backbone.train()
        slot_attn.train()
        for step, batch in enumerate(train_dataloader):
            pixel_values = batch["pixel_values"].to(dtype=weight_dtype)

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

            # Sample noise that we'll add to the model input
            if args.offset_noise:
                noise = torch.randn_like(model_input) + 0.1 * torch.randn(
                    model_input.shape[0], model_input.shape[1], 1, 1, device=model_input.device
                )
            else:
                noise = torch.randn_like(model_input)
            bsz, channels, height, width = model_input.shape
            # 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)

            # timestep is not used, but should we?
            if args.backbone_config == "pretrain_dino":
                pixel_values_vit = batch["pixel_values_vit"].to(dtype=weight_dtype)
                feat = backbone(pixel_values_vit)
            else:
                feat = backbone(pixel_values)
            slots, attn = slot_attn(feat[:, None])  # for the time dimension
            slots = slots[:, 0]

            if not train_unet:
                slots = slots.to(dtype=weight_dtype)

            # Predict the noise residual
            model_pred = unet(
                noisy_model_input, timesteps, slots,
            ).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(
                    model_input, noise, timesteps)
            else:
                raise ValueError(
                    f"Unknown prediction type {noise_scheduler.config.prediction_type}")

            # Compute instance loss
            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)
                base_weight = (
                    torch.stack(
                        [snr, args.snr_gamma * torch.ones_like(timesteps)], dim=1).min(dim=1)[0] / snr
                )

                if noise_scheduler.config.prediction_type == "v_prediction":
                    # Velocity objective needs to be floored to an SNR weight of one.
                    mse_loss_weights = base_weight + 1
                else:
                    # Epsilon and sample both use the same loss weights.
                    mse_loss_weights = base_weight
                loss = F.mse_loss(model_pred.float(),
                                  target.float(), reduction="none")
                loss = loss.mean(
                    dim=list(range(1, len(loss.shape)))) * mse_loss_weights
                loss = loss.mean()

            loss = loss / args.gradient_accumulation_steps

            accelerator.backward(loss)
            accumulate_steps += 1
            # if accelerator.sync_gradients:
            if (accumulate_steps+1) % args.gradient_accumulation_steps == 0:
                params_to_clip = params_to_optimize
                accelerator.clip_grad_norm_(
                    params_to_clip, args.max_grad_norm)
                optimizer.step()
                lr_scheduler.step()
                optimizer.zero_grad(set_to_none=args.set_grads_to_none)

            # Checks if the accelerator has performed an optimization step behind the scenes
            if (accumulate_steps+1) % args.gradient_accumulation_steps == 0:
                progress_bar.update(1)
                global_step += 1

                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}")

                    images = []

                    if global_step % args.validation_steps == 0:
                        images = log_validation(
                            val_dataset=val_dataset,
                            backbone=backbone,
                            slot_attn=slot_attn,
                            unet=unet,
                            vae=vae,
                            scheduler=noise_scheduler,
                            args=args,
                            accelerator=accelerator,
                            weight_dtype=weight_dtype,
                            global_step=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()
    if accelerator.is_main_process:
        save_path = os.path.join(
            args.output_dir, f"checkpoint-{global_step}-last")
        accelerator.save_state(save_path)
        logger.info(f"Saved state to {save_path}")

    accelerator.end_training()


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