train_ms.py

import argparse
import datetime
import gc
import os
import platform

import torch
import torch.distributed as dist
from huggingface_hub import HfApi
from torch.cuda.amp import GradScaler, autocast
from torch.nn import functional as F
from torch.nn.parallel import DistributedDataParallel as DDP
from torch.utils.data import DataLoader
from torch.utils.tensorboard import SummaryWriter
from tqdm import tqdm

# logging.getLogger("numba").setLevel(logging.WARNING)
import default_style
from config import get_config
from data_utils import (
    DistributedBucketSampler,
    TextAudioSpeakerCollate,
    TextAudioSpeakerLoader,
)
from losses import discriminator_loss, feature_loss, generator_loss, kl_loss
from mel_processing import mel_spectrogram_torch, spec_to_mel_torch
from style_bert_vits2.logging import logger
from style_bert_vits2.models import commons, utils
from style_bert_vits2.models.hyper_parameters import HyperParameters
from style_bert_vits2.models.models import (
    DurationDiscriminator,
    MultiPeriodDiscriminator,
    SynthesizerTrn,
)
from style_bert_vits2.nlp.symbols import SYMBOLS
from style_bert_vits2.utils.stdout_wrapper import SAFE_STDOUT


torch.backends.cuda.matmul.allow_tf32 = True
torch.backends.cudnn.allow_tf32 = (
    True  # If encontered training problem,please try to disable TF32.
)
torch.set_float32_matmul_precision("medium")
torch.backends.cuda.sdp_kernel("flash")
torch.backends.cuda.enable_flash_sdp(True)
torch.backends.cuda.enable_mem_efficient_sdp(
    True
)  # Not available if torch version is lower than 2.0
torch.backends.cuda.enable_math_sdp(True)

config = get_config()
global_step = 0

api = HfApi()


def run():
    # Command line configuration is not recommended unless necessary, use config.yml
    parser = argparse.ArgumentParser()
    parser.add_argument(
        "-c",
        "--config",
        type=str,
        default=config.train_ms_config.config_path,
        help="JSON file for configuration",
    )
    parser.add_argument(
        "-m",
        "--model",
        type=str,
        help="数据集文件夹路径，请注意，数据不再默认放在/logs文件夹下。如果需要用命令行配置，请声明相对于根目录的路径",
        default=config.dataset_path,
    )
    parser.add_argument(
        "--assets_root",
        type=str,
        help="Root directory of model assets needed for inference.",
        default=config.assets_root,
    )
    parser.add_argument(
        "--skip_default_style",
        action="store_true",
        help="Skip saving default style config and mean vector.",
    )
    parser.add_argument(
        "--no_progress_bar",
        action="store_true",
        help="Do not show the progress bar while training.",
    )
    parser.add_argument(
        "--speedup",
        action="store_true",
        help="Speed up training by disabling logging and evaluation.",
    )
    parser.add_argument(
        "--repo_id",
        help="Huggingface model repo id to backup the model.",
        default=None,
    )
    parser.add_argument(
        "--not_use_custom_batch_sampler",
        help="Don't use custom batch sampler for training, which was used in the version < 2.5",
        action="store_true",
    )
    args = parser.parse_args()

    # Set log file
    model_dir = os.path.join(args.model, config.train_ms_config.model_dir)
    timestamp = datetime.datetime.now().strftime("%Y%m%d_%H%M%S")
    logger.add(os.path.join(args.model, f"train_{timestamp}.log"))

    # Parsing environment variables
    envs = config.train_ms_config.env
    for env_name, env_value in envs.items():
        if env_name not in os.environ.keys():
            logger.info(f"Loading configuration from config {env_value!s}")
            os.environ[env_name] = str(env_value)
    logger.info(
        "Loading environment variables \nMASTER_ADDR: {},\nMASTER_PORT: {},\nWORLD_SIZE: {},\nRANK: {},\nLOCAL_RANK: {}".format(
            os.environ["MASTER_ADDR"],
            os.environ["MASTER_PORT"],
            os.environ["WORLD_SIZE"],
            os.environ["RANK"],
            os.environ["LOCAL_RANK"],
        )
    )

    backend = "nccl"
    if platform.system() == "Windows":
        backend = "gloo"  # If Windows,switch to gloo backend.
    dist.init_process_group(
        backend=backend,
        init_method="env://",
        timeout=datetime.timedelta(seconds=300),
    )  # Use torchrun instead of mp.spawn
    rank = dist.get_rank()
    local_rank = int(os.environ["LOCAL_RANK"])
    n_gpus = dist.get_world_size()

    hps = HyperParameters.load_from_json(args.config)
    # This is needed because we have to pass values to `train_and_evaluate()`
    hps.model_dir = model_dir
    hps.speedup = args.speedup
    hps.repo_id = args.repo_id

    # 比较路径是否相同
    if os.path.realpath(args.config) != os.path.realpath(
        config.train_ms_config.config_path
    ):
        with open(args.config, encoding="utf-8") as f:
            data = f.read()
        os.makedirs(os.path.dirname(config.train_ms_config.config_path), exist_ok=True)
        with open(config.train_ms_config.config_path, "w", encoding="utf-8") as f:
            f.write(data)

    """
    Path constants are a bit complicated...
    TODO: Refactor or rename these?
    (Both `config.yml` and `config.json` are used, which is confusing I think.)

    args.model: For saving all info needed for training.
        default: `Data/{model_name}`.
    hps.model_dir := model_dir: For saving checkpoints (for resuming training).
        default: `Data/{model_name}/models`.
        (Use `hps` since we have to pass `model_dir` to `train_and_evaluate()`.

    args.assets_root: The root directory of model assets needed for inference.
        default: config.assets_root == `model_assets`.

    config.out_dir: The directory for model assets of this model (for inference).
        default: `model_assets/{model_name}`.
    """

    if args.repo_id is not None:
        # First try to upload config.json to check if the repo exists
        try:
            api.upload_file(
                path_or_fileobj=args.config,
                path_in_repo=f"Data/{config.model_name}/config.json",
                repo_id=hps.repo_id,
            )
        except Exception as e:
            logger.error(e)
            logger.error(
                f"Failed to upload files to the repo {hps.repo_id}. Please check if the repo exists and you have logged in using `huggingface-cli login`."
            )
            raise e
        # Upload Data dir for resuming training
        api.upload_folder(
            repo_id=hps.repo_id,
            folder_path=config.dataset_path,
            path_in_repo=f"Data/{config.model_name}",
            delete_patterns="*.pth",  # Only keep the latest checkpoint
            run_as_future=True,
        )

    os.makedirs(config.out_dir, exist_ok=True)

    if not args.skip_default_style:
        default_style.save_styles_by_dirs(
            os.path.join(args.model, "wavs"),
            config.out_dir,
            config_path=args.config,
            config_output_path=os.path.join(config.out_dir, "config.json"),
        )

    torch.manual_seed(hps.train.seed)
    torch.cuda.set_device(local_rank)

    global global_step
    writer = None
    writer_eval = None
    if rank == 0 and not args.speedup:
        # logger = utils.get_logger(hps.model_dir)
        # logger.info(hps)
        utils.check_git_hash(model_dir)
        writer = SummaryWriter(log_dir=model_dir)
        writer_eval = SummaryWriter(log_dir=os.path.join(model_dir, "eval"))
    train_dataset = TextAudioSpeakerLoader(hps.data.training_files, hps.data)
    collate_fn = TextAudioSpeakerCollate()
    if not args.not_use_custom_batch_sampler:
        train_sampler = DistributedBucketSampler(
            train_dataset,
            hps.train.batch_size,
            [32, 300, 400, 500, 600, 700, 800, 900, 1000],
            num_replicas=n_gpus,
            rank=rank,
            shuffle=True,
        )
        train_loader = DataLoader(
            train_dataset,
            # メモリ消費量を減らそうとnum_workersを1にしてみる
            # num_workers=min(config.train_ms_config.num_workers, os.cpu_count() // 2),
            num_workers=1,
            shuffle=False,
            pin_memory=True,
            collate_fn=collate_fn,
            batch_sampler=train_sampler,
            # batch_size=hps.train.batch_size,
            persistent_workers=True,
            # これもメモリ消費量を減らそうとしてコメントアウト
            # prefetch_factor=6,
        )
    else:
        train_loader = DataLoader(
            train_dataset,
            # メモリ消費量を減らそうとnum_workersを1にしてみる
            # num_workers=min(config.train_ms_config.num_workers, os.cpu_count() // 2),
            num_workers=1,
            shuffle=True,
            pin_memory=True,
            collate_fn=collate_fn,
            # batch_sampler=train_sampler,
            batch_size=hps.train.batch_size,
            persistent_workers=True,
            # これもメモリ消費量を減らそうとしてコメントアウト
            # prefetch_factor=6,
        )
    eval_dataset = None
    eval_loader = None
    if rank == 0 and not args.speedup:
        eval_dataset = TextAudioSpeakerLoader(hps.data.validation_files, hps.data)
        eval_loader = DataLoader(
            eval_dataset,
            num_workers=0,
            shuffle=False,
            batch_size=1,
            pin_memory=True,
            drop_last=False,
            collate_fn=collate_fn,
        )
    if hps.model.use_noise_scaled_mas is True:
        logger.info("Using noise scaled MAS for VITS2")
        mas_noise_scale_initial = 0.01
        noise_scale_delta = 2e-6
    else:
        logger.info("Using normal MAS for VITS1")
        mas_noise_scale_initial = 0.0
        noise_scale_delta = 0.0
    if hps.model.use_duration_discriminator is True:
        logger.info("Using duration discriminator for VITS2")
        net_dur_disc = DurationDiscriminator(
            hps.model.hidden_channels,
            hps.model.hidden_channels,
            3,
            0.1,
            gin_channels=hps.model.gin_channels if hps.data.n_speakers != 0 else 0,
        ).cuda(local_rank)
    if hps.model.use_spk_conditioned_encoder is True:
        if hps.data.n_speakers == 0:
            raise ValueError(
                "n_speakers must be > 0 when using spk conditioned encoder to train multi-speaker model"
            )
    else:
        logger.info("Using normal encoder for VITS1")

    net_g = SynthesizerTrn(
        len(SYMBOLS),
        hps.data.filter_length // 2 + 1,
        hps.train.segment_size // hps.data.hop_length,
        n_speakers=hps.data.n_speakers,
        mas_noise_scale_initial=mas_noise_scale_initial,
        noise_scale_delta=noise_scale_delta,
        # hps.model 以下のすべての値を引数に渡す
        use_spk_conditioned_encoder=hps.model.use_spk_conditioned_encoder,
        use_noise_scaled_mas=hps.model.use_noise_scaled_mas,
        use_mel_posterior_encoder=hps.model.use_mel_posterior_encoder,
        use_duration_discriminator=hps.model.use_duration_discriminator,
        use_wavlm_discriminator=hps.model.use_wavlm_discriminator,
        inter_channels=hps.model.inter_channels,
        hidden_channels=hps.model.hidden_channels,
        filter_channels=hps.model.filter_channels,
        n_heads=hps.model.n_heads,
        n_layers=hps.model.n_layers,
        kernel_size=hps.model.kernel_size,
        p_dropout=hps.model.p_dropout,
        resblock=hps.model.resblock,
        resblock_kernel_sizes=hps.model.resblock_kernel_sizes,
        resblock_dilation_sizes=hps.model.resblock_dilation_sizes,
        upsample_rates=hps.model.upsample_rates,
        upsample_initial_channel=hps.model.upsample_initial_channel,
        upsample_kernel_sizes=hps.model.upsample_kernel_sizes,
        n_layers_q=hps.model.n_layers_q,
        use_spectral_norm=hps.model.use_spectral_norm,
        gin_channels=hps.model.gin_channels,
        slm=hps.model.slm,
    ).cuda(local_rank)

    if getattr(hps.train, "freeze_ZH_bert", False):
        logger.info("Freezing ZH bert encoder !!!")
        for param in net_g.enc_p.bert_proj.parameters():
            param.requires_grad = False

    if getattr(hps.train, "freeze_EN_bert", False):
        logger.info("Freezing EN bert encoder !!!")
        for param in net_g.enc_p.en_bert_proj.parameters():
            param.requires_grad = False

    if getattr(hps.train, "freeze_JP_bert", False):
        logger.info("Freezing JP bert encoder !!!")
        for param in net_g.enc_p.ja_bert_proj.parameters():
            param.requires_grad = False
    if getattr(hps.train, "freeze_style", False):
        logger.info("Freezing style encoder !!!")
        for param in net_g.enc_p.style_proj.parameters():
            param.requires_grad = False

    if getattr(hps.train, "freeze_decoder", False):
        logger.info("Freezing decoder !!!")
        for param in net_g.dec.parameters():
            param.requires_grad = False

    net_d = MultiPeriodDiscriminator(hps.model.use_spectral_norm).cuda(local_rank)
    optim_g = torch.optim.AdamW(
        filter(lambda p: p.requires_grad, net_g.parameters()),
        hps.train.learning_rate,
        betas=hps.train.betas,
        eps=hps.train.eps,
    )
    optim_d = torch.optim.AdamW(
        net_d.parameters(),
        hps.train.learning_rate,
        betas=hps.train.betas,
        eps=hps.train.eps,
    )
    if net_dur_disc is not None:
        optim_dur_disc = torch.optim.AdamW(
            net_dur_disc.parameters(),
            hps.train.learning_rate,
            betas=hps.train.betas,
            eps=hps.train.eps,
        )
    else:
        optim_dur_disc = None
    net_g = DDP(net_g, device_ids=[local_rank])
    net_d = DDP(net_d, device_ids=[local_rank])
    dur_resume_lr = None
    if net_dur_disc is not None:
        net_dur_disc = DDP(
            net_dur_disc, device_ids=[local_rank], find_unused_parameters=True
        )

    if utils.is_resuming(model_dir):
        if net_dur_disc is not None:
            _, _, dur_resume_lr, epoch_str = utils.checkpoints.load_checkpoint(
                utils.checkpoints.get_latest_checkpoint_path(model_dir, "DUR_*.pth"),
                net_dur_disc,
                optim_dur_disc,
                skip_optimizer=hps.train.skip_optimizer,
            )
            if not optim_dur_disc.param_groups[0].get("initial_lr"):
                optim_dur_disc.param_groups[0]["initial_lr"] = dur_resume_lr
        _, optim_g, g_resume_lr, epoch_str = utils.checkpoints.load_checkpoint(
            utils.checkpoints.get_latest_checkpoint_path(model_dir, "G_*.pth"),
            net_g,
            optim_g,
            skip_optimizer=hps.train.skip_optimizer,
        )
        _, optim_d, d_resume_lr, epoch_str = utils.checkpoints.load_checkpoint(
            utils.checkpoints.get_latest_checkpoint_path(model_dir, "D_*.pth"),
            net_d,
            optim_d,
            skip_optimizer=hps.train.skip_optimizer,
        )
        if not optim_g.param_groups[0].get("initial_lr"):
            optim_g.param_groups[0]["initial_lr"] = g_resume_lr
        if not optim_d.param_groups[0].get("initial_lr"):
            optim_d.param_groups[0]["initial_lr"] = d_resume_lr

        epoch_str = max(epoch_str, 1)
        # global_step = (epoch_str - 1) * len(train_loader)
        global_step = int(
            utils.get_steps(
                utils.checkpoints.get_latest_checkpoint_path(model_dir, "G_*.pth")
            )
        )
        logger.info(
            f"******************Found the model. Current epoch is {epoch_str}, gloabl step is {global_step}*********************"
        )
    else:
        try:
            _ = utils.safetensors.load_safetensors(
                os.path.join(model_dir, "G_0.safetensors"), net_g
            )
            _ = utils.safetensors.load_safetensors(
                os.path.join(model_dir, "D_0.safetensors"), net_d
            )
            if net_dur_disc is not None:
                _ = utils.safetensors.load_safetensors(
                    os.path.join(model_dir, "DUR_0.safetensors"), net_dur_disc
                )
            logger.info("Loaded the pretrained models.")
        except Exception as e:
            logger.warning(e)
            logger.warning(
                "It seems that you are not using the pretrained models, so we will train from scratch."
            )
        finally:
            epoch_str = 1
            global_step = 0

    def lr_lambda(epoch):
        """
        Learning rate scheduler for warmup and exponential decay.
        - During the warmup period, the learning rate increases linearly.
        - After the warmup period, the learning rate decreases exponentially.
        """
        if epoch < hps.train.warmup_epochs:
            return float(epoch) / float(max(1, hps.train.warmup_epochs))
        else:
            return hps.train.lr_decay ** (epoch - hps.train.warmup_epochs)

    scheduler_last_epoch = epoch_str - 2
    scheduler_g = torch.optim.lr_scheduler.LambdaLR(
        optim_g, lr_lambda=lr_lambda, last_epoch=scheduler_last_epoch
    )
    scheduler_d = torch.optim.lr_scheduler.LambdaLR(
        optim_d, lr_lambda=lr_lambda, last_epoch=scheduler_last_epoch
    )
    if net_dur_disc is not None:
        scheduler_dur_disc = torch.optim.lr_scheduler.LambdaLR(
            optim_dur_disc, lr_lambda=lr_lambda, last_epoch=scheduler_last_epoch
        )
    else:
        scheduler_dur_disc = None
    scaler = GradScaler(enabled=hps.train.bf16_run)
    logger.info("Start training.")

    diff = abs(
        epoch_str * len(train_loader) - (hps.train.epochs + 1) * len(train_loader)
    )
    pbar = None
    if not args.no_progress_bar:
        pbar = tqdm(
            total=global_step + diff,
            initial=global_step,
            smoothing=0.05,
            file=SAFE_STDOUT,
        )
    initial_step = global_step

    for epoch in range(epoch_str, hps.train.epochs + 1):
        if rank == 0:
            train_and_evaluate(
                rank,
                local_rank,
                epoch,
                hps,
                [net_g, net_d, net_dur_disc],
                [optim_g, optim_d, optim_dur_disc],
                [scheduler_g, scheduler_d, scheduler_dur_disc],
                scaler,
                [train_loader, eval_loader],
                logger,
                [writer, writer_eval],
                pbar,
                initial_step,
            )
        else:
            train_and_evaluate(
                rank,
                local_rank,
                epoch,
                hps,
                [net_g, net_d, net_dur_disc],
                [optim_g, optim_d, optim_dur_disc],
                [scheduler_g, scheduler_d, scheduler_dur_disc],
                scaler,
                [train_loader, None],
                None,
                None,
                pbar,
                initial_step,
            )
        scheduler_g.step()
        scheduler_d.step()
        if net_dur_disc is not None:
            scheduler_dur_disc.step()

        if epoch == hps.train.epochs:
            # Save the final models
            assert optim_g is not None
            utils.checkpoints.save_checkpoint(
                net_g,
                optim_g,
                hps.train.learning_rate,
                epoch,
                os.path.join(model_dir, f"G_{global_step}.pth"),
            )
            assert optim_d is not None
            utils.checkpoints.save_checkpoint(
                net_d,
                optim_d,
                hps.train.learning_rate,
                epoch,
                os.path.join(model_dir, f"D_{global_step}.pth"),
            )
            if net_dur_disc is not None:
                assert optim_dur_disc is not None
                utils.checkpoints.save_checkpoint(
                    net_dur_disc,
                    optim_dur_disc,
                    hps.train.learning_rate,
                    epoch,
                    os.path.join(model_dir, f"DUR_{global_step}.pth"),
                )
            utils.safetensors.save_safetensors(
                net_g,
                epoch,
                os.path.join(
                    config.out_dir,
                    f"{config.model_name}_e{epoch}_s{global_step}.safetensors",
                ),
                for_infer=True,
            )
            if hps.repo_id is not None:
                future1 = api.upload_folder(
                    repo_id=hps.repo_id,
                    folder_path=config.dataset_path,
                    path_in_repo=f"Data/{config.model_name}",
                    delete_patterns="*.pth",  # Only keep the latest checkpoint
                    run_as_future=True,
                )
                future2 = api.upload_folder(
                    repo_id=hps.repo_id,
                    folder_path=config.out_dir,
                    path_in_repo=f"model_assets/{config.model_name}",
                    run_as_future=True,
                )
                try:
                    future1.result()
                    future2.result()
                except Exception as e:
                    logger.error(e)

    if pbar is not None:
        pbar.close()


def train_and_evaluate(
    rank,
    local_rank,
    epoch,
    hps: HyperParameters,
    nets,
    optims,
    schedulers,
    scaler,
    loaders,
    logger,
    writers,
    pbar: tqdm,
    initial_step: int,
):
    net_g, net_d, net_dur_disc = nets
    optim_g, optim_d, optim_dur_disc = optims
    scheduler_g, scheduler_d, scheduler_dur_disc = schedulers
    train_loader, eval_loader = loaders
    if writers is not None:
        writer, writer_eval = writers

    train_loader.batch_sampler.set_epoch(epoch)
    global global_step

    net_g.train()
    net_d.train()
    if net_dur_disc is not None:
        net_dur_disc.train()
    for batch_idx, (
        x,
        x_lengths,
        spec,
        spec_lengths,
        y,
        y_lengths,
        speakers,
        tone,
        language,
        bert,
        ja_bert,
        en_bert,
        style_vec,
    ) in enumerate(train_loader):
        if net_g.module.use_noise_scaled_mas:
            current_mas_noise_scale = (
                net_g.module.mas_noise_scale_initial
                - net_g.module.noise_scale_delta * global_step
            )
            net_g.module.current_mas_noise_scale = max(current_mas_noise_scale, 0.0)
        x, x_lengths = x.cuda(local_rank, non_blocking=True), x_lengths.cuda(
            local_rank, non_blocking=True
        )
        spec, spec_lengths = spec.cuda(
            local_rank, non_blocking=True
        ), spec_lengths.cuda(local_rank, non_blocking=True)
        y, y_lengths = y.cuda(local_rank, non_blocking=True), y_lengths.cuda(
            local_rank, non_blocking=True
        )
        speakers = speakers.cuda(local_rank, non_blocking=True)
        tone = tone.cuda(local_rank, non_blocking=True)
        language = language.cuda(local_rank, non_blocking=True)
        bert = bert.cuda(local_rank, non_blocking=True)
        ja_bert = ja_bert.cuda(local_rank, non_blocking=True)
        en_bert = en_bert.cuda(local_rank, non_blocking=True)
        style_vec = style_vec.cuda(local_rank, non_blocking=True)

        with autocast(enabled=hps.train.bf16_run, dtype=torch.bfloat16):
            (
                y_hat,
                l_length,
                attn,
                ids_slice,
                x_mask,
                z_mask,
                (z, z_p, m_p, logs_p, m_q, logs_q),
                (hidden_x, logw, logw_),
            ) = net_g(
                x,
                x_lengths,
                spec,
                spec_lengths,
                speakers,
                tone,
                language,
                bert,
                ja_bert,
                en_bert,
                style_vec,
            )
            mel = spec_to_mel_torch(
                spec,
                hps.data.filter_length,
                hps.data.n_mel_channels,
                hps.data.sampling_rate,
                hps.data.mel_fmin,
                hps.data.mel_fmax,
            )
            y_mel = commons.slice_segments(
                mel, ids_slice, hps.train.segment_size // hps.data.hop_length
            )
            y_hat_mel = mel_spectrogram_torch(
                y_hat.squeeze(1).float(),
                hps.data.filter_length,
                hps.data.n_mel_channels,
                hps.data.sampling_rate,
                hps.data.hop_length,
                hps.data.win_length,
                hps.data.mel_fmin,
                hps.data.mel_fmax,
            )

            y = commons.slice_segments(
                y, ids_slice * hps.data.hop_length, hps.train.segment_size
            )  # slice

            # Discriminator
            y_d_hat_r, y_d_hat_g, _, _ = net_d(y, y_hat.detach())
            with autocast(enabled=hps.train.bf16_run, dtype=torch.bfloat16):
                loss_disc, losses_disc_r, losses_disc_g = discriminator_loss(
                    y_d_hat_r, y_d_hat_g
                )
                loss_disc_all = loss_disc
            if net_dur_disc is not None:
                y_dur_hat_r, y_dur_hat_g = net_dur_disc(
                    hidden_x.detach(), x_mask.detach(), logw.detach(), logw_.detach()
                )
                with autocast(enabled=hps.train.bf16_run, dtype=torch.bfloat16):
                    # TODO: I think need to mean using the mask, but for now, just mean all
                    (
                        loss_dur_disc,
                        losses_dur_disc_r,
                        losses_dur_disc_g,
                    ) = discriminator_loss(y_dur_hat_r, y_dur_hat_g)
                    loss_dur_disc_all = loss_dur_disc
                optim_dur_disc.zero_grad()
                scaler.scale(loss_dur_disc_all).backward()
                scaler.unscale_(optim_dur_disc)
                commons.clip_grad_value_(net_dur_disc.parameters(), None)
                scaler.step(optim_dur_disc)

        optim_d.zero_grad()
        scaler.scale(loss_disc_all).backward()
        scaler.unscale_(optim_d)
        if getattr(hps.train, "bf16_run", False):
            torch.nn.utils.clip_grad_norm_(parameters=net_d.parameters(), max_norm=200)
        grad_norm_d = commons.clip_grad_value_(net_d.parameters(), None)
        scaler.step(optim_d)

        with autocast(enabled=hps.train.bf16_run, dtype=torch.bfloat16):
            # Generator
            y_d_hat_r, y_d_hat_g, fmap_r, fmap_g = net_d(y, y_hat)
            if net_dur_disc is not None:
                y_dur_hat_r, y_dur_hat_g = net_dur_disc(hidden_x, x_mask, logw, logw_)
            with autocast(enabled=hps.train.bf16_run, dtype=torch.bfloat16):
                loss_dur = torch.sum(l_length.float())
                loss_mel = F.l1_loss(y_mel, y_hat_mel) * hps.train.c_mel
                loss_kl = kl_loss(z_p, logs_q, m_p, logs_p, z_mask) * hps.train.c_kl

                loss_fm = feature_loss(fmap_r, fmap_g)
                loss_gen, losses_gen = generator_loss(y_d_hat_g)
                loss_gen_all = loss_gen + loss_fm + loss_mel + loss_dur + loss_kl
                if net_dur_disc is not None:
                    loss_dur_gen, losses_dur_gen = generator_loss(y_dur_hat_g)
                    loss_gen_all += loss_dur_gen
        optim_g.zero_grad()
        scaler.scale(loss_gen_all).backward()
        scaler.unscale_(optim_g)
        if getattr(hps.train, "bf16_run", False):
            torch.nn.utils.clip_grad_norm_(parameters=net_g.parameters(), max_norm=500)
        grad_norm_g = commons.clip_grad_value_(net_g.parameters(), None)
        scaler.step(optim_g)
        scaler.update()

        if rank == 0:
            if global_step % hps.train.log_interval == 0 and not hps.speedup:
                lr = optim_g.param_groups[0]["lr"]
                losses = [loss_disc, loss_gen, loss_fm, loss_mel, loss_dur, loss_kl]
                # logger.info(
                #     "Train Epoch: {} [{:.0f}%]".format(
                #         epoch, 100.0 * batch_idx / len(train_loader)
                #     )
                # )
                # logger.info([x.item() for x in losses] + [global_step, lr])

                scalar_dict = {
                    "loss/g/total": loss_gen_all,
                    "loss/d/total": loss_disc_all,
                    "learning_rate": lr,
                    "grad_norm_d": grad_norm_d,
                    "grad_norm_g": grad_norm_g,
                }
                scalar_dict.update(
                    {
                        "loss/g/fm": loss_fm,
                        "loss/g/mel": loss_mel,
                        "loss/g/dur": loss_dur,
                        "loss/g/kl": loss_kl,
                    }
                )
                scalar_dict.update({f"loss/g/{i}": v for i, v in enumerate(losses_gen)})
                scalar_dict.update(
                    {f"loss/d_r/{i}": v for i, v in enumerate(losses_disc_r)}
                )
                scalar_dict.update(
                    {f"loss/d_g/{i}": v for i, v in enumerate(losses_disc_g)}
                )
                # 以降のログは計算が重い気がするし誰も見てない気がするのでコメントアウト
                # image_dict = {
                #     "slice/mel_org": utils.plot_spectrogram_to_numpy(
                #         y_mel[0].data.cpu().numpy()
                #     ),
                #     "slice/mel_gen": utils.plot_spectrogram_to_numpy(
                #         y_hat_mel[0].data.cpu().numpy()
                #     ),
                #     "all/mel": utils.plot_spectrogram_to_numpy(
                #         mel[0].data.cpu().numpy()
                #     ),
                #     "all/attn": utils.plot_alignment_to_numpy(
                #         attn[0, 0].data.cpu().numpy()
                #     ),
                # }
                utils.summarize(
                    writer=writer,
                    global_step=global_step,
                    # images=image_dict,
                    scalars=scalar_dict,
                )

            if (
                global_step % hps.train.eval_interval == 0
                and global_step != 0
                and initial_step != global_step
            ):
                if not hps.speedup:
                    evaluate(hps, net_g, eval_loader, writer_eval)
                assert hps.model_dir is not None
                utils.checkpoints.save_checkpoint(
                    net_g,
                    optim_g,
                    hps.train.learning_rate,
                    epoch,
                    os.path.join(hps.model_dir, f"G_{global_step}.pth"),
                )
                utils.checkpoints.save_checkpoint(
                    net_d,
                    optim_d,
                    hps.train.learning_rate,
                    epoch,
                    os.path.join(hps.model_dir, f"D_{global_step}.pth"),
                )
                if net_dur_disc is not None:
                    utils.checkpoints.save_checkpoint(
                        net_dur_disc,
                        optim_dur_disc,
                        hps.train.learning_rate,
                        epoch,
                        os.path.join(hps.model_dir, f"DUR_{global_step}.pth"),
                    )
                keep_ckpts = config.train_ms_config.keep_ckpts
                if keep_ckpts > 0:
                    utils.checkpoints.clean_checkpoints(
                        model_dir_path=hps.model_dir,
                        n_ckpts_to_keep=keep_ckpts,
                        sort_by_time=True,
                    )
                # Save safetensors (for inference) to `model_assets/{model_name}`
                utils.safetensors.save_safetensors(
                    net_g,
                    epoch,
                    os.path.join(
                        config.out_dir,
                        f"{config.model_name}_e{epoch}_s{global_step}.safetensors",
                    ),
                    for_infer=True,
                )
                if hps.repo_id is not None:
                    api.upload_folder(
                        repo_id=hps.repo_id,
                        folder_path=config.dataset_path,
                        path_in_repo=f"Data/{config.model_name}",
                        delete_patterns="*.pth",  # Only keep the latest checkpoint
                        run_as_future=True,
                    )
                    api.upload_folder(
                        repo_id=hps.repo_id,
                        folder_path=config.out_dir,
                        path_in_repo=f"model_assets/{config.model_name}",
                        run_as_future=True,
                    )

        global_step += 1
        if pbar is not None:
            pbar.set_description(
                f"Epoch {epoch}({100.0 * batch_idx / len(train_loader):.0f}%)/{hps.train.epochs}"
            )
            pbar.update()
    # 本家ではこれをスピードアップのために消すと書かれていたので、一応消してみる
    # と思ったけどメモリ使用量が減るかもしれないのでつけてみる
    gc.collect()
    torch.cuda.empty_cache()
    if pbar is None and rank == 0:
        logger.info(f"====> Epoch: {epoch}, step: {global_step}")


def evaluate(hps, generator, eval_loader, writer_eval):
    generator.eval()
    image_dict = {}
    audio_dict = {}
    print()
    logger.info("Evaluating ...")
    with torch.no_grad():
        for batch_idx, (
            x,
            x_lengths,
            spec,
            spec_lengths,
            y,
            y_lengths,
            speakers,
            tone,
            language,
            bert,
            ja_bert,
            en_bert,
            style_vec,
        ) in enumerate(eval_loader):
            x, x_lengths = x.cuda(), x_lengths.cuda()
            spec, spec_lengths = spec.cuda(), spec_lengths.cuda()
            y, y_lengths = y.cuda(), y_lengths.cuda()
            speakers = speakers.cuda()
            bert = bert.cuda()
            ja_bert = ja_bert.cuda()
            en_bert = en_bert.cuda()
            tone = tone.cuda()
            language = language.cuda()
            style_vec = style_vec.cuda()
            for use_sdp in [True, False]:
                y_hat, attn, mask, *_ = generator.module.infer(
                    x,
                    x_lengths,
                    speakers,
                    tone,
                    language,
                    bert,
                    ja_bert,
                    en_bert,
                    style_vec,
                    y=spec,
                    max_len=1000,
                    sdp_ratio=0.0 if not use_sdp else 1.0,
                )
                y_hat_lengths = mask.sum([1, 2]).long() * hps.data.hop_length
                # 以降のログは計算が重い気がするし誰も見てない気がするのでコメントアウト
                # mel = spec_to_mel_torch(
                #     spec,
                #     hps.data.filter_length,
                #     hps.data.n_mel_channels,
                #     hps.data.sampling_rate,
                #     hps.data.mel_fmin,
                #     hps.data.mel_fmax,
                # )
                # y_hat_mel = mel_spectrogram_torch(
                #     y_hat.squeeze(1).float(),
                #     hps.data.filter_length,
                #     hps.data.n_mel_channels,
                #     hps.data.sampling_rate,
                #     hps.data.hop_length,
                #     hps.data.win_length,
                #     hps.data.mel_fmin,
                #     hps.data.mel_fmax,
                # )
                # image_dict.update(
                #     {
                #         f"gen/mel_{batch_idx}": utils.plot_spectrogram_to_numpy(
                #             y_hat_mel[0].cpu().numpy()
                #         )
                #     }
                # )
                # image_dict.update(
                #     {
                #         f"gt/mel_{batch_idx}": utils.plot_spectrogram_to_numpy(
                #             mel[0].cpu().numpy()
                #         )
                #     }
                # )
                audio_dict.update(
                    {
                        f"gen/audio_{batch_idx}_{use_sdp}": y_hat[
                            0, :, : y_hat_lengths[0]
                        ]
                    }
                )
                audio_dict.update({f"gt/audio_{batch_idx}": y[0, :, : y_lengths[0]]})

    utils.summarize(
        writer=writer_eval,
        global_step=global_step,
        images=image_dict,
        audios=audio_dict,
        audio_sampling_rate=hps.data.sampling_rate,
    )
    generator.train()


if __name__ == "__main__":
    run()