train_ms_cluster.py

# flake8: noqa: E402

import os

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

from transformers import get_linear_schedule_with_warmup

logging.getLogger("numba").setLevel(logging.WARNING)
import commons
import utils
from data_utils import (
    TextAudioSpeakerLoader,
    TextAudioSpeakerCollate,
    DistributedBucketSampler,
)
from models import (
    SynthesizerTrn,
    MultiPeriodDiscriminator,
    DurationDiscriminator,
)
from losses import generator_loss, discriminator_loss, feature_loss, kl_loss
from mel_processing import mel_spectrogram_torch, spec_to_mel_torch
from text.symbols import symbols

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.cudnn.benchmark = True
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)
global_step = 0


def run():
    dist.init_process_group(
        backend="nccl",
        init_method="env://",  # Due to some training problem,we proposed to use gloo instead of nccl.
    )  # Use torchrun instead of mp.spawn
    rank = dist.get_rank()
    n_gpus = dist.get_world_size()
    hps = utils.get_hparams()
    torch.manual_seed(hps.train.seed)
    torch.cuda.set_device(rank)
    global global_step
    if rank == 0:
        logger = utils.get_logger(hps.model_dir)
        logger.info(hps)
        utils.check_git_hash(hps.model_dir)
        writer = SummaryWriter(log_dir=hps.model_dir)
        writer_eval = SummaryWriter(log_dir=os.path.join(hps.model_dir, "eval"))
    train_dataset = TextAudioSpeakerLoader(hps.data.training_files, hps.data)
    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,
    )
    collate_fn = TextAudioSpeakerCollate()
    train_loader = DataLoader(
        train_dataset,
        num_workers=8,
        shuffle=False,
        pin_memory=True,
        collate_fn=collate_fn,
        batch_sampler=train_sampler,
        persistent_workers=True,
        prefetch_factor=4,
    )  # DataLoader config could be adjusted.
    if rank == 0:
        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 (
        "use_noise_scaled_mas" in hps.model.keys()
        and hps.model.use_noise_scaled_mas is True
    ):
        print("Using noise scaled MAS for VITS2")
        mas_noise_scale_initial = 0.01
        noise_scale_delta = 2e-6
    else:
        print("Using normal MAS for VITS1")
        mas_noise_scale_initial = 0.0
        noise_scale_delta = 0.0
    if (
        "use_duration_discriminator" in hps.model.keys()
        and hps.model.use_duration_discriminator is True
    ):
        print("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(rank)
    if (
        "use_spk_conditioned_encoder" in hps.model.keys()
        and 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:
        print("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,
    ).cuda(rank)

    net_d = MultiPeriodDiscriminator(hps.model.use_spectral_norm).cuda(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=[rank], find_unused_parameters=True)
    net_d = DDP(net_d, device_ids=[rank], find_unused_parameters=True)
    if net_dur_disc is not None:
        net_dur_disc = DDP(net_dur_disc, device_ids=[rank], find_unused_parameters=True)
    try:
        if net_dur_disc is not None:
            _, _, dur_resume_lr, epoch_str = utils.load_checkpoint(
                utils.latest_checkpoint_path(hps.model_dir, "DUR_*.pth"),
                net_dur_disc,
                None,
                skip_optimizer=hps.train.skip_optimizer
                if "skip_optimizer" in hps.train
                else True,
            )
            _, _, g_resume_lr, epoch_str = utils.load_checkpoint(
                utils.latest_checkpoint_path(hps.model_dir, "G_*.pth"),
                net_g,
                None,
                skip_optimizer=hps.train.skip_optimizer
                if "skip_optimizer" in hps.train
                else True,
            )
            _, _, d_resume_lr, epoch_str = utils.load_checkpoint(
                utils.latest_checkpoint_path(hps.model_dir, "D_*.pth"),
                net_d,
                None,
                skip_optimizer=hps.train.skip_optimizer
                if "skip_optimizer" in hps.train
                else True,
            )
            # 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
            # if not optim_dur_disc.param_groups[0].get("initial_lr"):
            #     optim_dur_disc.param_groups[0]["initial_lr"] = dur_resume_lr

        epoch_str = 1
        global_step = (epoch_str - 1) * len(train_loader)
    except Exception as e:
        print(e)
        epoch_str = 1
        global_step = 0

    training_steps = len(train_loader) * hps.train.epochs
    warmup_steps = training_steps * hps.train.warmup_ratio
    if rank == 0:
        print(f"Total training steps {len(train_loader)} * {hps.train.epochs} = {training_steps}")
        print(f"Warmup steps {warmup_steps}")
    scheduler_g = get_linear_schedule_with_warmup(optim_g, warmup_steps, training_steps)
    # scheduler_g = torch.optim.lr_scheduler.ExponentialLR(
    #     optim_g, gamma=hps.train.lr_decay, last_epoch=epoch_str - 2
    # )
    scheduler_d = get_linear_schedule_with_warmup(optim_d, warmup_steps, training_steps)
    # scheduler_d = torch.optim.lr_scheduler.ExponentialLR(
    #     optim_d, gamma=hps.train.lr_decay, last_epoch=epoch_str - 2
    # )
    if net_dur_disc is not None:
        if not optim_dur_disc.param_groups[0].get("initial_lr"):
            optim_dur_disc.param_groups[0]["initial_lr"] = dur_resume_lr
        scheduler_dur_disc = get_linear_schedule_with_warmup(optim_dur_disc, warmup_steps, training_steps)
        # scheduler_dur_disc = torch.optim.lr_scheduler.ExponentialLR(
        #     optim_dur_disc, gamma=hps.train.lr_decay, last_epoch=epoch_str - 2
        # )
    else:
        scheduler_dur_disc = None
    scaler = GradScaler(enabled=hps.train.fp16_run)

    for epoch in range(epoch_str, hps.train.epochs + 1):
        if rank == 0:
            train_and_evaluate(
                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],
            )
        else:
            train_and_evaluate(
                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,
            )
        # scheduler_g.step()
        # scheduler_d.step()
        # if net_dur_disc is not None:
        #     scheduler_dur_disc.step()


def train_and_evaluate(
    rank, epoch, hps, nets, optims, schedulers, scaler, loaders, logger, writers
):
    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,
    ) in enumerate(tqdm(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(rank, non_blocking=True), x_lengths.cuda(
            rank, non_blocking=True
        )
        spec, spec_lengths = spec.cuda(rank, non_blocking=True), spec_lengths.cuda(
            rank, non_blocking=True
        )
        y, y_lengths = y.cuda(rank, non_blocking=True), y_lengths.cuda(
            rank, non_blocking=True
        )
        speakers = speakers.cuda(rank, non_blocking=True)
        tone = tone.cuda(rank, non_blocking=True)
        language = language.cuda(rank, non_blocking=True)
        bert = bert.cuda(rank, non_blocking=True)
        ja_bert = ja_bert.cuda(rank, non_blocking=True)

        with autocast(enabled=hps.train.fp16_run):
            (
                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,
            )
            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),
                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=False):
                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=False):
                    # 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(), hps.train.clipping_grad_norm)
                scaler.step(optim_dur_disc)
                scheduler_dur_disc.step()

        optim_d.zero_grad()
        scaler.scale(loss_disc_all).backward()
        scaler.unscale_(optim_d)
        grad_norm_d = commons.clip_grad_value_(net_d.parameters(), hps.train.clipping_grad_norm)
        scaler.step(optim_d)
        scheduler_d.step()

        with autocast(enabled=hps.train.fp16_run):
            # 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=False):
                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)
        grad_norm_g = commons.clip_grad_value_(net_g.parameters(), hps.train.clipping_grad_norm)
        scaler.step(optim_g)
        scaler.update()
        scheduler_g.step()

        if rank == 0:
            if global_step % hps.train.log_interval == 0:
                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(
                    {"loss/g/{}".format(i): v for i, v in enumerate(losses_gen)}
                )
                scalar_dict.update(
                    {"loss/d_r/{}".format(i): v for i, v in enumerate(losses_disc_r)}
                )
                scalar_dict.update(
                    {"loss/d_g/{}".format(i): v for i, v in enumerate(losses_disc_g)}
                )

                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:
                evaluate(hps, net_g, eval_loader, writer_eval)
                utils.save_checkpoint(
                    net_g,
                    optim_g,
                    hps.train.learning_rate,
                    epoch,
                    os.path.join(hps.model_dir, "G_{}.pth".format(global_step)),
                )
                utils.save_checkpoint(
                    net_d,
                    optim_d,
                    hps.train.learning_rate,
                    epoch,
                    os.path.join(hps.model_dir, "D_{}.pth".format(global_step)),
                )
                if net_dur_disc is not None:
                    utils.save_checkpoint(
                        net_dur_disc,
                        optim_dur_disc,
                        hps.train.learning_rate,
                        epoch,
                        os.path.join(hps.model_dir, "DUR_{}.pth".format(global_step)),
                    )
                keep_ckpts = getattr(hps.train, "keep_ckpts", 5)
                if keep_ckpts > 0:
                    utils.clean_checkpoints(
                        path_to_models=hps.model_dir,
                        n_ckpts_to_keep=keep_ckpts,
                        sort_by_time=True,
                    )

        global_step += 1

    if rank == 0:
        logger.info("====> Epoch: {}".format(epoch))


def evaluate(hps, generator, eval_loader, writer_eval):
    generator.eval()
    image_dict = {}
    audio_dict = {}
    print("Evaluating ...")
    with torch.no_grad():
        for batch_idx, (
            x,
            x_lengths,
            spec,
            spec_lengths,
            y,
            y_lengths,
            speakers,
            tone,
            language,
            bert,
            ja_bert,
        ) in enumerate(tqdm(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()
            tone = tone.cuda()
            language = language.cuda()
            for use_sdp in [True, False]:
                y_hat, attn, mask, *_ = generator.module.infer(
                    x,
                    x_lengths,
                    speakers,
                    tone,
                    language,
                    bert,
                    ja_bert,
                    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

                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()
                        )
                    }
                )
                audio_dict.update(
                    {
                        f"gen/audio_{batch_idx}_{use_sdp}": y_hat[
                            0, :, : y_hat_lengths[0]
                        ]
                    }
                )
                image_dict.update(
                    {
                        f"gt/mel_{batch_idx}": utils.plot_spectrogram_to_numpy(
                            mel[0].cpu().numpy()
                        )
                    }
                )
                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()