train_unconditional_ort.py

import argparse
import inspect
import math
import os
from pathlib import Path
from typing import Optional

import torch
import torch.nn.functional as F

from accelerate import Accelerator
from accelerate.logging import get_logger
from datasets import load_dataset
from diffusers import DDPMPipeline, DDPMScheduler, UNet2DModel
from diffusers.optimization import get_scheduler
from diffusers.training_utils import EMAModel
from diffusers.utils import check_min_version
from huggingface_hub import HfFolder, Repository, create_repo, whoami
from onnxruntime.training.ortmodule import ORTModule
from torchvision.transforms import (
    CenterCrop,
    Compose,
    InterpolationMode,
    Normalize,
    RandomHorizontalFlip,
    Resize,
    ToTensor,
)
from tqdm.auto import tqdm


# Will error if the minimal version of diffusers is not installed. Remove at your own risks.
check_min_version("0.10.0.dev0")


logger = get_logger(__name__)


def _extract_into_tensor(arr, timesteps, broadcast_shape):
    """
    Extract values from a 1-D numpy array for a batch of indices.

    :param arr: the 1-D numpy array.
    :param timesteps: a tensor of indices into the array to extract.
    :param broadcast_shape: a larger shape of K dimensions with the batch
                            dimension equal to the length of timesteps.
    :return: a tensor of shape [batch_size, 1, ...] where the shape has K dims.
    """
    if not isinstance(arr, torch.Tensor):
        arr = torch.from_numpy(arr)
    res = arr[timesteps].float().to(timesteps.device)
    while len(res.shape) < len(broadcast_shape):
        res = res[..., None]
    return res.expand(broadcast_shape)


def parse_args():
    parser = argparse.ArgumentParser(description="Simple example of a training script.")
    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 HF 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(
        "--output_dir",
        type=str,
        default="ddpm-model-64",
        help="The output directory where the model predictions and checkpoints will be written.",
    )
    parser.add_argument("--overwrite_output_dir", action="store_true")
    parser.add_argument(
        "--cache_dir",
        type=str,
        default=None,
        help="The directory where the downloaded models and datasets will be stored.",
    )
    parser.add_argument(
        "--resolution",
        type=int,
        default=64,
        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=16, help="Batch size (per device) for the training dataloader."
    )
    parser.add_argument(
        "--eval_batch_size", type=int, default=16, help="The number of images to generate for evaluation."
    )
    parser.add_argument(
        "--dataloader_num_workers",
        type=int,
        default=0,
        help=(
            "The number of subprocesses to use for data loading. 0 means that the data will be loaded in the main"
            " process."
        ),
    )
    parser.add_argument("--num_epochs", type=int, default=100)
    parser.add_argument("--save_images_epochs", type=int, default=10, help="How often to save images during training.")
    parser.add_argument(
        "--save_model_epochs", type=int, default=10, help="How often to save the model during training."
    )
    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(
        "--learning_rate",
        type=float,
        default=1e-4,
        help="Initial learning rate (after the potential warmup period) to use.",
    )
    parser.add_argument(
        "--lr_scheduler",
        type=str,
        default="cosine",
        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("--adam_beta1", type=float, default=0.95, 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-6, help="Weight decay magnitude for the Adam optimizer."
    )
    parser.add_argument("--adam_epsilon", type=float, default=1e-08, help="Epsilon value for the Adam optimizer.")
    parser.add_argument(
        "--use_ema",
        action="store_true",
        help="Whether to use Exponential Moving Average for the final model weights.",
    )
    parser.add_argument("--ema_inv_gamma", type=float, default=1.0, help="The inverse gamma value for the EMA decay.")
    parser.add_argument("--ema_power", type=float, default=3 / 4, help="The power value for the EMA decay.")
    parser.add_argument("--ema_max_decay", type=float, default=0.9999, help="The maximum decay magnitude for EMA.")
    parser.add_argument("--push_to_hub", action="store_true", help="Whether or not to push the model to the Hub.")
    parser.add_argument("--hub_token", type=str, default=None, help="The token to use to push to the Model Hub.")
    parser.add_argument(
        "--hub_model_id",
        type=str,
        default=None,
        help="The name of the repository to keep in sync with the local `output_dir`.",
    )
    parser.add_argument(
        "--hub_private_repo", action="store_true", help="Whether or not to create a private repository."
    )
    parser.add_argument(
        "--logger",
        type=str,
        default="tensorboard",
        choices=["tensorboard", "wandb"],
        help=(
            "Whether to use [tensorboard](https://www.tensorflow.org/tensorboard) or [wandb](https://www.wandb.ai)"
            " for experiment tracking and logging of model metrics and model checkpoints"
        ),
    )
    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("--local_rank", type=int, default=-1, help="For distributed training: local_rank")
    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."
        ),
    )
    parser.add_argument(
        "--prediction_type",
        type=str,
        default="epsilon",
        choices=["epsilon", "sample"],
        help="Whether the model should predict the 'epsilon'/noise error or directly the reconstructed image 'x0'.",
    )

    parser.add_argument("--ddpm_num_steps", type=int, default=1000)
    parser.add_argument("--ddpm_beta_schedule", type=str, default="linear")
    parser.add_argument(
        "--checkpointing_steps",
        type=int,
        default=500,
        help=(
            "Save a checkpoint of the training state every X updates. These checkpoints are only suitable for resuming"
            " training using `--resume_from_checkpoint`."
        ),
    )
    parser.add_argument(
        "--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.'
        ),
    )

    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

    if args.dataset_name is None and args.train_data_dir is None:
        raise ValueError("You must specify either a dataset name from the hub or a train data directory.")

    return args


def get_full_repo_name(model_id: str, organization: Optional[str] = None, token: Optional[str] = None):
    if token is None:
        token = HfFolder.get_token()
    if organization is None:
        username = whoami(token)["name"]
        return f"{username}/{model_id}"
    else:
        return f"{organization}/{model_id}"


def main(args):
    logging_dir = os.path.join(args.output_dir, args.logging_dir)
    accelerator = Accelerator(
        gradient_accumulation_steps=args.gradient_accumulation_steps,
        mixed_precision=args.mixed_precision,
        log_with=args.logger,
        logging_dir=logging_dir,
    )

    model = UNet2DModel(
        sample_size=args.resolution,
        in_channels=3,
        out_channels=3,
        layers_per_block=2,
        block_out_channels=(128, 128, 256, 256, 512, 512),
        down_block_types=(
            "DownBlock2D",
            "DownBlock2D",
            "DownBlock2D",
            "DownBlock2D",
            "AttnDownBlock2D",
            "DownBlock2D",
        ),
        up_block_types=(
            "UpBlock2D",
            "AttnUpBlock2D",
            "UpBlock2D",
            "UpBlock2D",
            "UpBlock2D",
            "UpBlock2D",
        ),
    )

    if args.use_ema:
        ema_model = EMAModel(
            model.parameters(),
            decay=args.ema_max_decay,
            use_ema_warmup=True,
            inv_gamma=args.ema_inv_gamma,
            power=args.ema_power,
        )

    accepts_prediction_type = "prediction_type" in set(inspect.signature(DDPMScheduler.__init__).parameters.keys())
    if accepts_prediction_type:
        noise_scheduler = DDPMScheduler(
            num_train_timesteps=args.ddpm_num_steps,
            beta_schedule=args.ddpm_beta_schedule,
            prediction_type=args.prediction_type,
        )
    else:
        noise_scheduler = DDPMScheduler(num_train_timesteps=args.ddpm_num_steps, beta_schedule=args.ddpm_beta_schedule)

    optimizer = torch.optim.AdamW(
        model.parameters(),
        lr=args.learning_rate,
        betas=(args.adam_beta1, args.adam_beta2),
        weight_decay=args.adam_weight_decay,
        eps=args.adam_epsilon,
    )

    augmentations = Compose(
        [
            Resize(args.resolution, interpolation=InterpolationMode.BILINEAR),
            CenterCrop(args.resolution),
            RandomHorizontalFlip(),
            ToTensor(),
            Normalize([0.5], [0.5]),
        ]
    )

    if args.dataset_name is not None:
        dataset = load_dataset(
            args.dataset_name,
            args.dataset_config_name,
            cache_dir=args.cache_dir,
            split="train",
        )
    else:
        dataset = load_dataset("imagefolder", data_dir=args.train_data_dir, cache_dir=args.cache_dir, split="train")

    def transforms(examples):
        images = [augmentations(image.convert("RGB")) for image in examples["image"]]
        return {"input": images}

    logger.info(f"Dataset size: {len(dataset)}")

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

    lr_scheduler = get_scheduler(
        args.lr_scheduler,
        optimizer=optimizer,
        num_warmup_steps=args.lr_warmup_steps,
        num_training_steps=(len(train_dataloader) * args.num_epochs) // args.gradient_accumulation_steps,
    )

    model, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
        model, optimizer, train_dataloader, lr_scheduler
    )

    if args.use_ema:
        accelerator.register_for_checkpointing(ema_model)
        ema_model.to(accelerator.device)

    num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)

    model = ORTModule(model)

    # Handle the repository creation
    if accelerator.is_main_process:
        if args.push_to_hub:
            if args.hub_model_id is None:
                repo_name = get_full_repo_name(Path(args.output_dir).name, token=args.hub_token)
            else:
                repo_name = args.hub_model_id
            create_repo(repo_name, exist_ok=True, token=args.hub_token)
            repo = Repository(args.output_dir, clone_from=repo_name, token=args.hub_token)

            with open(os.path.join(args.output_dir, ".gitignore"), "w+") as gitignore:
                if "step_*" not in gitignore:
                    gitignore.write("step_*\n")
                if "epoch_*" not in gitignore:
                    gitignore.write("epoch_*\n")
        elif args.output_dir is not None:
            os.makedirs(args.output_dir, exist_ok=True)

    if accelerator.is_main_process:
        run = os.path.split(__file__)[-1].split(".")[0]
        accelerator.init_trackers(run)

    global_step = 0
    first_epoch = 0
    if args.resume_from_checkpoint:
        if args.resume_from_checkpoint != "latest":
            path = os.path.basename(args.resume_from_checkpoint)
        else:
            # Get the most recent checkpoint
            dirs = os.listdir(args.output_dir)
            dirs = [d for d in dirs if d.startswith("checkpoint")]
            dirs = sorted(dirs, key=lambda x: int(x.split("-")[1]))
            path = dirs[-1] if len(dirs) > 0 else None

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

            resume_global_step = global_step * args.gradient_accumulation_steps
            first_epoch = global_step // num_update_steps_per_epoch
            resume_step = resume_global_step % (num_update_steps_per_epoch * args.gradient_accumulation_steps)

    for epoch in range(first_epoch, args.num_epochs):
        model.train()
        progress_bar = tqdm(total=num_update_steps_per_epoch, disable=not accelerator.is_local_main_process)
        progress_bar.set_description(f"Epoch {epoch}")
        for step, batch in enumerate(train_dataloader):
            # Skip steps until we reach the resumed step
            if args.resume_from_checkpoint and epoch == first_epoch and step < resume_step:
                if step % args.gradient_accumulation_steps == 0:
                    progress_bar.update(1)
                continue

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

            # Add noise to the clean images according to the noise magnitude at each timestep
            # (this is the forward diffusion process)
            noisy_images = noise_scheduler.add_noise(clean_images, noise, timesteps)

            with accelerator.accumulate(model):
                # Predict the noise residual
                model_output = model(noisy_images, timesteps, return_dict=False)[0]

                if args.prediction_type == "epsilon":
                    loss = F.mse_loss(model_output, noise)  # this could have different weights!
                elif args.prediction_type == "sample":
                    alpha_t = _extract_into_tensor(
                        noise_scheduler.alphas_cumprod, timesteps, (clean_images.shape[0], 1, 1, 1)
                    )
                    snr_weights = alpha_t / (1 - alpha_t)
                    loss = snr_weights * F.mse_loss(
                        model_output, clean_images, reduction="none"
                    )  # use SNR weighting from distillation paper
                    loss = loss.mean()
                else:
                    raise ValueError(f"Unsupported prediction type: {args.prediction_type}")

                accelerator.backward(loss)

                if accelerator.sync_gradients:
                    accelerator.clip_grad_norm_(model.parameters(), 1.0)
                optimizer.step()
                lr_scheduler.step()
                if args.use_ema:
                    ema_model.step(model.parameters())
                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 global_step % args.checkpointing_steps == 0:
                    if accelerator.is_main_process:
                        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 = {"loss": loss.detach().item(), "lr": lr_scheduler.get_last_lr()[0], "step": global_step}
            if args.use_ema:
                logs["ema_decay"] = ema_model.decay
            progress_bar.set_postfix(**logs)
            accelerator.log(logs, step=global_step)
        progress_bar.close()

        accelerator.wait_for_everyone()

        # Generate sample images for visual inspection
        if accelerator.is_main_process:
            if epoch % args.save_images_epochs == 0 or epoch == args.num_epochs - 1:
                pipeline = DDPMPipeline(
                    unet=accelerator.unwrap_model(ema_model.averaged_model if args.use_ema else model),
                    scheduler=noise_scheduler,
                )

                generator = torch.Generator(device=pipeline.device).manual_seed(0)
                # run pipeline in inference (sample random noise and denoise)
                images = pipeline(
                    generator=generator,
                    batch_size=args.eval_batch_size,
                    output_type="numpy",
                ).images

                # denormalize the images and save to tensorboard
                images_processed = (images * 255).round().astype("uint8")

                if args.logger == "tensorboard":
                    accelerator.get_tracker("tensorboard").add_images(
                        "test_samples", images_processed.transpose(0, 3, 1, 2), epoch
                    )

            if epoch % args.save_model_epochs == 0 or epoch == args.num_epochs - 1:
                # save the model
                pipeline.save_pretrained(args.output_dir)
                if args.push_to_hub:
                    repo.push_to_hub(commit_message=f"Epoch {epoch}", blocking=False)
        accelerator.wait_for_everyone()

    accelerator.end_training()


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