Open-Sora/scripts/train-vae-v2.py

import os
from glob import glob

import colossalai
import torch
import torch.distributed as dist
import wandb
from colossalai.booster import Booster
from colossalai.booster.plugin import LowLevelZeroPlugin
from colossalai.cluster import DistCoordinator
from colossalai.nn.optimizer import HybridAdam
from colossalai.utils import get_current_device
from einops import rearrange
from tqdm import tqdm

from opensora.acceleration.checkpoint import set_grad_checkpoint
from opensora.acceleration.parallel_states import (
    get_data_parallel_group,
    set_data_parallel_group,
    set_sequence_parallel_group,
)
from opensora.acceleration.plugin import ZeroSeqParallelPlugin
from opensora.datasets import prepare_dataloader, prepare_variable_dataloader
from opensora.models.vae.vae_3d_v2 import AdversarialLoss, DiscriminatorLoss, LeCamEMA, VEALoss, pad_at_dim
from opensora.registry import DATASETS, MODELS, build_module
from opensora.utils.ckpt_utils import create_logger, load_json, save_json
from opensora.utils.config_utils import (
    create_experiment_workspace,
    create_tensorboard_writer,
    parse_configs,
    save_training_config,
)
from opensora.utils.misc import all_reduce_mean, format_numel_str, get_model_numel, to_torch_dtype


def main():
    # ======================================================
    # 1. args & cfg
    # ======================================================
    cfg = parse_configs(training=True)

    # ======================================================
    # 2. runtime variables & colossalai launch
    # ======================================================
    assert torch.cuda.is_available(), "Training currently requires at least one GPU."

    # 2.1. colossalai init distributed training
    colossalai.launch_from_torch({})
    coordinator = DistCoordinator()

    exp_dir = None
    if coordinator.is_master():  # only create directory for master
        exp_name, exp_dir = create_experiment_workspace(cfg)
        save_training_config(cfg._cfg_dict, exp_dir)
    dist.barrier()

    # get exp dir for non-master process
    if exp_dir is None:
        # experiment_index = len(glob(f"{cfg.outputs}/*")) - 1
        # model_name = cfg.model["type"].replace("/", "-")
        # exp_name = f"{experiment_index:03d}-{model_name}"
        # exp_dir = f"{cfg.outputs}/{exp_name}"
        exp_name, exp_dir = create_experiment_workspace(cfg, get_last_workspace=True)
        assert os.path.exists(exp_dir)

    device = get_current_device()
    assert cfg.dtype in ["fp16", "bf16"], f"Unknown mixed precision {cfg.dtype}"
    dtype = to_torch_dtype(cfg.dtype)

    # 2.2. init logger, tensorboard & wandb
    if not coordinator.is_master():
        logger = create_logger(None)
    else:
        print(cfg)
        logger = create_logger(exp_dir)
        logger.info(f"Experiment directory created at {exp_dir}")

        writer = create_tensorboard_writer(exp_dir)
        if cfg.wandb:
            # wandb.init(project="opensora-vae", name=exp_name, config=cfg._cfg_dict)
            # NOTE: here we use the outputs folder name to store running records of different experiments (since frequent interruption)
            name = os.path.basename(os.path.normpath(cfg.outputs))
            wandb.init(project="opensora-vae", name=name, config=cfg._cfg_dict)

    # 2.3. initialize ColossalAI booster
    if cfg.plugin == "zero2":
        plugin = LowLevelZeroPlugin(
            stage=2,
            precision=cfg.dtype,
            initial_scale=2**16,
            max_norm=cfg.grad_clip,
        )
        set_data_parallel_group(dist.group.WORLD)
    elif cfg.plugin == "zero2-seq":
        plugin = ZeroSeqParallelPlugin(
            sp_size=cfg.sp_size,
            stage=2,
            precision=cfg.dtype,
            initial_scale=2**16,
            max_norm=cfg.grad_clip,
        )
        set_sequence_parallel_group(plugin.sp_group)
        set_data_parallel_group(plugin.dp_group)
    else:
        raise ValueError(f"Unknown plugin {cfg.plugin}")
    booster = Booster(plugin=plugin)

    # ======================================================
    # 3. build dataset and dataloader
    # ======================================================
    dataset = build_module(cfg.dataset, DATASETS)
    logger.info(f"Dataset contains {len(dataset)} samples.")
    dataloader_args = dict(
        dataset=dataset,
        batch_size=cfg.batch_size,
        num_workers=cfg.num_workers,
        seed=cfg.seed,
        shuffle=True,
        drop_last=True,
        pin_memory=True,
        process_group=get_data_parallel_group(),
    )
    # TODO: use plugin's prepare dataloader
    if cfg.bucket_config is None:
        dataloader = prepare_dataloader(**dataloader_args)
    else:
        dataloader = prepare_variable_dataloader(
            bucket_config=cfg.bucket_config,
            num_bucket_build_workers=cfg.num_bucket_build_workers,
            **dataloader_args,
        )
    if cfg.dataset.type == "VideoTextDataset":
        total_batch_size = cfg.batch_size * dist.get_world_size() // cfg.sp_size
        logger.info(f"Total batch size: {total_batch_size}")

    # ======================================================
    # 4. build model
    # ======================================================
    # 4.1. build model

    if cfg.get("use_pipeline") == True:
        # use 2D VAE, then temporal VAE
        vae_2d = build_module(cfg.vae_2d, MODELS)

    vae = build_module(cfg.model, MODELS, device=device)
    vae_numel, vae_numel_trainable = get_model_numel(vae)
    logger.info(
        f"Trainable vae params: {format_numel_str(vae_numel_trainable)}, Total model params: {format_numel_str(vae_numel)}"
    )

    discriminator = build_module(cfg.discriminator, MODELS, device=device)
    discriminator_numel, discriminator_numel_trainable = get_model_numel(discriminator)
    logger.info(
        f"Trainable discriminator params: {format_numel_str(discriminator_numel_trainable)}, Total model params: {format_numel_str(discriminator_numel)}"
    )

    # LeCam Initialization
    lecam_ema = LeCamEMA(decay=cfg.ema_decay, dtype=dtype, device=device)

    # 4.3. move to device
    if cfg.get("use_pipeline") == True:
        vae_2d.to(device, dtype).eval()  # eval mode, not training!

    vae = vae.to(device, dtype)
    discriminator = discriminator.to(device, dtype)

    # 4.5. setup optimizer
    # vae optimizer
    optimizer = HybridAdam(
        filter(lambda p: p.requires_grad, vae.parameters()), lr=cfg.lr, weight_decay=0, adamw_mode=True
    )
    lr_scheduler = None
    # disc optimizer
    disc_optimizer = HybridAdam(
        filter(lambda p: p.requires_grad, discriminator.parameters()), lr=cfg.lr, weight_decay=0, adamw_mode=True
    )
    disc_lr_scheduler = None

    # 4.6. prepare for training
    if cfg.grad_checkpoint:
        set_grad_checkpoint(vae)
        set_grad_checkpoint(discriminator)
    vae.train()
    discriminator.train()

    # =======================================================
    # 5. boost model for distributed training with colossalai
    # =======================================================
    torch.set_default_dtype(dtype)
    vae, optimizer, _, dataloader, lr_scheduler = booster.boost(
        model=vae, optimizer=optimizer, lr_scheduler=lr_scheduler, dataloader=dataloader
    )
    torch.set_default_dtype(torch.float)
    num_steps_per_epoch = len(dataloader)
    logger.info("Boost vae for distributed training")

    discriminator, disc_optimizer, _, _, disc_lr_scheduler = booster.boost(
        model=discriminator, optimizer=disc_optimizer, lr_scheduler=disc_lr_scheduler
    )
    logger.info("Boost discriminator for distributed training")

    # =======================================================
    # 6. training loop
    # =======================================================
    start_epoch = start_step = log_step = sampler_start_idx = 0
    running_loss = 0.0
    running_disc_loss = 0.0

    # 6.1. resume training
    if cfg.load is not None:
        logger.info("Loading checkpoint")
        booster.load_model(vae, os.path.join(cfg.load, "model"))
        booster.load_model(discriminator, os.path.join(cfg.load, "discriminator"))
        booster.load_optimizer(optimizer, os.path.join(cfg.load, "optimizer"))
        booster.load_optimizer(disc_optimizer, os.path.join(cfg.load, "disc_optimizer"))
        if lr_scheduler is not None:
            booster.load_lr_scheduler(lr_scheduler, os.path.join(cfg.load, "lr_scheduler"))
        if disc_lr_scheduler is not None:
            booster.load_lr_scheduler(disc_lr_scheduler, os.path.join(cfg.load, "disc_lr_scheduler"))

        # LeCam EMA for discriminator
        lecam_path = os.path.join(cfg.load, "lecam_states.json")
        if cfg.lecam_loss_weight is not None and os.path.exists(lecam_path):
            lecam_state = load_json(lecam_path)
            lecam_ema_real, lecam_ema_fake = lecam_state["lecam_ema_real"], lecam_state["lecam_ema_fake"]
            lecam_ema = LeCamEMA(
                decay=cfg.ema_decay, ema_real=lecam_ema_real, ema_fake=lecam_ema_fake, dtype=dtype, device=device
            )

        running_states = load_json(os.path.join(cfg.load, "running_states.json"))
        dist.barrier()
        start_epoch, start_step, sampler_start_idx = (
            running_states["epoch"],
            running_states["step"],
            running_states["sample_start_index"],
        )
        logger.info(f"Loaded checkpoint {cfg.load} at epoch {start_epoch} step {start_step}")
    logger.info(f"Training for {cfg.epochs} epochs with {num_steps_per_epoch} steps per epoch")

    dataloader.sampler.set_start_index(sampler_start_idx)

    # 6.2 Define loss functions
    vae_loss_fn = VEALoss(
        logvar_init=cfg.logvar_init,
        perceptual_loss_weight=cfg.perceptual_loss_weight,
        kl_loss_weight=cfg.kl_loss_weight,
        device=device,
        dtype=dtype,
    )

    adversarial_loss_fn = AdversarialLoss(
        discriminator_factor=cfg.discriminator_factor,
        discriminator_start=cfg.discriminator_start,
        generator_factor=cfg.generator_factor,
        generator_loss_type=cfg.generator_loss_type,
    )

    disc_loss_fn = DiscriminatorLoss(
        discriminator_factor=cfg.discriminator_factor,
        discriminator_start=cfg.discriminator_start,
        discriminator_loss_type=cfg.discriminator_loss_type,
        lecam_loss_weight=cfg.lecam_loss_weight,
        gradient_penalty_loss_weight=cfg.gradient_penalty_loss_weight,
    )

    # 6.3. training loop

    # calculate discriminator_time_padding
    disc_time_downsample_factor = 2 ** len(cfg.discriminator.channel_multipliers)
    if cfg.dataset.num_frames % disc_time_downsample_factor != 0:
        disc_time_padding = disc_time_downsample_factor - cfg.dataset.num_frames % disc_time_downsample_factor
    else:
        disc_time_padding = 0
    video_contains_first_frame = cfg.video_contains_first_frame

    for epoch in range(start_epoch, cfg.epochs):
        dataloader.sampler.set_epoch(epoch)
        dataloader_iter = iter(dataloader)
        logger.info(f"Beginning epoch {epoch}...")

        with tqdm(
            range(start_step, num_steps_per_epoch),
            desc=f"Epoch {epoch}",
            disable=not coordinator.is_master(),
            total=num_steps_per_epoch,
            initial=start_step,
        ) as pbar:
            # with profile(
            #     activities=[ProfilerActivity.CPU, ProfilerActivity.CUDA],
            #     schedule=torch.profiler.schedule(
            #             wait=1,
            #             warmup=1,
            #             active=2,
            #             repeat=2,
            #         ),
            #     on_trace_ready=torch.profiler.tensorboard_trace_handler('/home/shenchenhui/log'),
            #     with_stack=True,
            #     record_shapes=True,
            #     profile_memory=True,
            # ) as p: # trace efficiency

            for step in pbar:
                # with profile(
                #     activities=[ProfilerActivity.CPU, ProfilerActivity.CUDA],
                #     with_stack=True,
                # ) as p: # trace efficiency

                # SCH: calc global step at the start
                global_step = epoch * num_steps_per_epoch + step

                batch = next(dataloader_iter)
                x = batch["video"].to(device, dtype)  # [B, C, T, H, W]

                # supprt for image or video inputs
                assert x.ndim in {4, 5}, f"received input of {x.ndim} dimensions"  # either image or video
                assert (
                    x.shape[-2:] == cfg.dataset.image_size
                ), f"received input size {x.shape[-2:]}, but config image size is {cfg.dataset.image_size}"
                is_image = x.ndim == 4
                if is_image:
                    video = rearrange(x, "b c ... -> b c 1 ...")
                    video_contains_first_frame = True
                else:
                    video = x

                #  ===== Spatial VAE =====
                if cfg.get("use_pipeline") == True:
                    with torch.no_grad():
                        video = vae_2d.encode(video)

                #  ====== VAE ======
                recon_video, posterior = vae(
                    video,
                    video_contains_first_frame=video_contains_first_frame,
                )

                #  ====== Generator Loss ======
                # simple nll loss
                nll_loss, weighted_nll_loss, weighted_kl_loss = vae_loss_fn(
                    video, recon_video, posterior, split="train"
                )

                adversarial_loss = torch.tensor(0.0)
                # adversarial loss
                if global_step > cfg.discriminator_start:
                    # padded videos for GAN
                    fake_video = pad_at_dim(recon_video, (disc_time_padding, 0), value=0.0, dim=2)
                    fake_logits = discriminator(fake_video.contiguous())
                    adversarial_loss = adversarial_loss_fn(
                        fake_logits,
                        nll_loss,
                        vae.module.get_last_layer(),
                        global_step,
                        is_training=vae.training,
                    )

                vae_loss = weighted_nll_loss + weighted_kl_loss + adversarial_loss

                optimizer.zero_grad()
                # Backward & update
                booster.backward(loss=vae_loss, optimizer=optimizer)
                # # NOTE: clip gradients? this is done in Open-Sora-Plan
                # torch.nn.utils.clip_grad_norm_(vae.parameters(), 1) # NOTE: done by grad_clip
                optimizer.step()

                # Log loss values:
                all_reduce_mean(vae_loss)  # NOTE: this is to get average loss for logging
                running_loss += vae_loss.item()

                #  ====== Discriminator Loss ======
                if global_step > cfg.discriminator_start:
                    # if video_contains_first_frame:
                    # Since we don't have enough T frames, pad anyways
                    real_video = pad_at_dim(video, (disc_time_padding, 0), value=0.0, dim=2)
                    fake_video = pad_at_dim(recon_video, (disc_time_padding, 0), value=0.0, dim=2)

                    if cfg.gradient_penalty_loss_weight is not None and cfg.gradient_penalty_loss_weight > 0.0:
                        real_video = real_video.requires_grad_()
                        real_logits = discriminator(
                            real_video.contiguous()
                        )  # SCH: not detached for now for gradient_penalty calculation
                    else:
                        real_logits = discriminator(real_video.contiguous().detach())

                    fake_logits = discriminator(fake_video.contiguous().detach())

                    lecam_ema_real, lecam_ema_fake = lecam_ema.get()

                    weighted_d_adversarial_loss, lecam_loss, gradient_penalty_loss = disc_loss_fn(
                        real_logits,
                        fake_logits,
                        global_step,
                        lecam_ema_real=lecam_ema_real,
                        lecam_ema_fake=lecam_ema_fake,
                        real_video=real_video if cfg.gradient_penalty_loss_weight is not None else None,
                    )
                    disc_loss = weighted_d_adversarial_loss + lecam_loss + gradient_penalty_loss
                    if cfg.lecam_loss_weight is not None:
                        ema_real = torch.mean(real_logits.clone().detach()).to(device, dtype)
                        ema_fake = torch.mean(fake_logits.clone().detach()).to(device, dtype)
                        all_reduce_mean(ema_real)
                        all_reduce_mean(ema_fake)
                        lecam_ema.update(ema_real, ema_fake)

                    disc_optimizer.zero_grad()
                    # Backward & update
                    booster.backward(loss=disc_loss, optimizer=disc_optimizer)
                    # # NOTE: TODO: clip gradients? this is done in Open-Sora-Plan
                    # torch.nn.utils.clip_grad_norm_(discriminator.parameters(), 1) # NOTE: done by grad_clip
                    disc_optimizer.step()

                    # Log loss values:
                    all_reduce_mean(disc_loss)
                    running_disc_loss += disc_loss.item()
                else:
                    disc_loss = torch.tensor(0.0)
                    weighted_d_adversarial_loss = torch.tensor(0.0)
                    lecam_loss = torch.tensor(0.0)
                    gradient_penalty_loss = torch.tensor(0.0)

                log_step += 1

                # Log to tensorboard
                if coordinator.is_master() and (global_step + 1) % cfg.log_every == 0:
                    avg_loss = running_loss / log_step
                    avg_disc_loss = running_disc_loss / log_step
                    pbar.set_postfix(
                        {"loss": avg_loss, "disc_loss": avg_disc_loss, "step": step, "global_step": global_step}
                    )
                    running_loss = 0
                    log_step = 0
                    running_disc_loss = 0
                    writer.add_scalar("loss", vae_loss.item(), global_step)
                    if cfg.wandb:
                        wandb.log(
                            {
                                "iter": global_step,
                                "num_samples": global_step * total_batch_size,
                                "epoch": epoch,
                                "loss": vae_loss.item(),
                                "kl_loss": weighted_kl_loss.item(),
                                "gen_adv_loss": adversarial_loss.item(),
                                "disc_loss": disc_loss.item(),
                                "lecam_loss": lecam_loss.item(),
                                "r1_grad_penalty": gradient_penalty_loss.item(),
                                "nll_loss": weighted_nll_loss.item(),
                                "avg_loss": avg_loss,
                            },
                            step=global_step,
                        )

                # Save checkpoint
                if cfg.ckpt_every > 0 and (global_step + 1) % cfg.ckpt_every == 0:
                    save_dir = os.path.join(exp_dir, f"epoch{epoch}-global_step{global_step+1}")
                    os.makedirs(os.path.join(save_dir, "model"), exist_ok=True)  # already handled in booster save_model
                    booster.save_model(vae, os.path.join(save_dir, "model"), shard=True)
                    booster.save_model(discriminator, os.path.join(save_dir, "discriminator"), shard=True)
                    booster.save_optimizer(
                        optimizer, os.path.join(save_dir, "optimizer"), shard=True, size_per_shard=4096
                    )
                    booster.save_optimizer(
                        disc_optimizer, os.path.join(save_dir, "disc_optimizer"), shard=True, size_per_shard=4096
                    )

                    if lr_scheduler is not None:
                        booster.save_lr_scheduler(lr_scheduler, os.path.join(save_dir, "lr_scheduler"))
                    if disc_lr_scheduler is not None:
                        booster.save_lr_scheduler(disc_lr_scheduler, os.path.join(save_dir, "disc_lr_scheduler"))

                    running_states = {
                        "epoch": epoch,
                        "step": step + 1,
                        "global_step": global_step + 1,
                        "sample_start_index": (step + 1) * cfg.batch_size,
                    }

                    lecam_ema_real, lecam_ema_fake = lecam_ema.get()
                    lecam_state = {
                        "lecam_ema_real": lecam_ema_real.item(),
                        "lecam_ema_fake": lecam_ema_fake.item(),
                    }
                    if coordinator.is_master():
                        save_json(running_states, os.path.join(save_dir, "running_states.json"))
                        if cfg.lecam_loss_weight is not None:
                            save_json(lecam_state, os.path.join(save_dir, "lecam_states.json"))
                    dist.barrier()

                    logger.info(
                        f"Saved checkpoint at epoch {epoch} step {step + 1} global_step {global_step + 1} to {exp_dir}"
                    )

            # print(p.key_averages().table(sort_by="self_cuda_time_total", row_limit=-1))

        # the continue epochs are not resumed, so we need to reset the sampler start index and start step
        dataloader.sampler.set_start_index(0)
        start_step = 0


if __name__ == "__main__":
    main()