train_broken.py

import logging
import multiprocessing
import os
import time
import signal
import sys
import os
import threading

import torch

# 启用 SDPA 优化，使用 Flash kernel
if hasattr(torch.backends.cuda, 'sdp_kernel'):
    torch.backends.cuda.sdp_enabled = True
    torch.backends.cuda.sdp_kernel = "flash"

import torch.distributed as dist
import torch.multiprocessing as mp
from torch.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
"""
# === xFormers patch start ===
import xformers.ops as xops
from torch.nn import MultiheadAttention

def _patched_mha_forward(self, query, key, value, **kwargs):
    # memory_efficient_attention 返回 (attn_output, None)
    return xops.memory_efficient_attention(query, key, value), None

MultiheadAttention.forward = _patched_mha_forward
# === xFormers patch end ===
"""
import modules.commons as commons
import utils
from data_utils import TextAudioCollate, TextAudioSpeakerLoader
from models import (
    MultiPeriodDiscriminator,
    SynthesizerTrn,
)
from modules.losses import discriminator_loss, feature_loss, generator_loss, kl_loss
from modules.mel_processing import mel_spectrogram_torch, spec_to_mel_torch

logging.getLogger('matplotlib').setLevel(logging.WARNING)
logging.getLogger('numba').setLevel(logging.WARNING)

terminate_flag = threading.Event()

global_step = 0
start_time = time.time()

# os.environ['TORCH_DISTRIBUTED_DEBUG'] = 'INFO'

torch.set_float32_matmul_precision('high')  # 加速矩阵乘法
torch.backends.cuda.matmul.allow_tf32 = True  # 允许TF32矩阵乘法
torch.backends.cudnn.allow_tf32 = True  # 允许TF32卷积
torch.backends.cudnn.benchmark = True  # 加速卷积运算

def main():
    """Assume Single Node Multi GPUs Training Only"""
    assert torch.cuda.is_available(), "CPU training is not allowed."
    hps = utils.get_hparams()

    n_gpus = torch.cuda.device_count()
    os.environ['MASTER_ADDR'] = 'localhost'
    os.environ['MASTER_PORT'] = hps.train.port

    mp.spawn(run, nprocs=n_gpus, args=(n_gpus, hps,))


def run(rank, n_gpus, hps):
    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"))
    
    # for pytorch on win, backend use gloo    
    dist.init_process_group(backend=  'gloo' if os.name == 'nt' else 'nccl', init_method='env://', world_size=n_gpus, rank=rank)
    torch.manual_seed(hps.train.seed)
    torch.cuda.set_device(rank)
    collate_fn = TextAudioCollate()
    all_in_mem = hps.train.all_in_mem   # If you have enough memory, turn on this option to avoid disk IO and speed up training.
    train_dataset = TextAudioSpeakerLoader(hps.data.training_files, hps, all_in_mem=all_in_mem)
    num_workers = 5 if multiprocessing.cpu_count() > 4 else multiprocessing.cpu_count()
    if all_in_mem:
        num_workers = 0
    train_loader = DataLoader(train_dataset, num_workers=num_workers, shuffle=False, pin_memory=True, persistent_workers=True, prefetch_factor=hps.train.batch_size*4,
                              batch_size=hps.train.batch_size, collate_fn=collate_fn)
    if rank == 0:
        eval_dataset = TextAudioSpeakerLoader(hps.data.validation_files, hps, all_in_mem=all_in_mem,vol_aug = False)
        eval_loader = DataLoader(eval_dataset, num_workers=1, shuffle=False,
                                 batch_size=1, pin_memory=False,
                                 drop_last=False, collate_fn=collate_fn)

    net_g = SynthesizerTrn(
        hps.data.filter_length // 2 + 1,
        hps.train.segment_size // hps.data.hop_length,
        **hps.model).cuda(rank)
    net_d = MultiPeriodDiscriminator(hps.model.use_spectral_norm).cuda(rank)
    optim_g = torch.optim.AdamW(
        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)
    net_g = DDP(net_g, device_ids=[rank])  # , find_unused_parameters=True)
    net_d = DDP(net_d, device_ids=[rank])


    skip_optimizer = False
    try:
        _, _, _, epoch_str = utils.load_checkpoint(utils.latest_checkpoint_path(hps.model_dir, "G_*.pth"), net_g,
                                                   optim_g, skip_optimizer)
        _, _, _, epoch_str = utils.load_checkpoint(utils.latest_checkpoint_path(hps.model_dir, "D_*.pth"), net_d,
                                                   optim_d, skip_optimizer)
        epoch_str = max(epoch_str, 1)
        name=utils.latest_checkpoint_path(hps.model_dir, "D_*.pth")
        global_step=int(name[name.rfind("_")+1:name.rfind(".")])+1
        #global_step = (epoch_str - 1) * len(train_loader)
    except Exception:
        print("load old checkpoint failed...")
        epoch_str = 1
        global_step = 0
    if skip_optimizer:
        epoch_str = 1
        global_step = 0

    warmup_epoch = hps.train.warmup_epochs
    scheduler_g = torch.optim.lr_scheduler.ExponentialLR(optim_g, gamma=hps.train.lr_decay, last_epoch=epoch_str - 2)
    scheduler_d = torch.optim.lr_scheduler.ExponentialLR(optim_d, gamma=hps.train.lr_decay, last_epoch=epoch_str - 2)

    scaler = GradScaler(enabled=hps.train.fp16_run)

    for epoch in range(epoch_str, hps.train.epochs + 1):
        # set up warm-up learning rate
        if epoch <= warmup_epoch:
            for param_group in optim_g.param_groups:
                param_group['lr'] = hps.train.learning_rate / warmup_epoch * epoch
            for param_group in optim_d.param_groups:
                param_group['lr'] = hps.train.learning_rate / warmup_epoch * epoch
        # training
        if rank == 0:
            train_and_evaluate(rank, epoch, hps, [net_g, net_d], [optim_g, optim_d], [scheduler_g, scheduler_d], scaler,
                               [train_loader, eval_loader], logger, [writer, writer_eval])
        else:
            train_and_evaluate(rank, epoch, hps, [net_g, net_d], [optim_g, optim_d], [scheduler_g, scheduler_d], scaler,
                               [train_loader, None], None, None)
        # update learning rate
        scheduler_g.step()
        scheduler_d.step()


def train_and_evaluate(rank, epoch, hps, nets, optims, schedulers, scaler, loaders, logger, writers):
    net_g, net_d = nets
    optim_g, optim_d = optims
    scheduler_g, scheduler_d = schedulers
    train_loader, eval_loader = loaders
    if writers is not None:
        writer, writer_eval = writers
    
    accum_steps = getattr(hps.train, "grad_accumulate", 1)
    
    half_type = torch.bfloat16 if hps.train.half_type=="bf16" else torch.float16

    # train_loader.batch_sampler.set_epoch(epoch)
    global global_step

    net_g.train()
    net_d.train()
    #optim_d.zero_grad()
    #optim_g.zero_grad()
    
    for batch_idx, items in enumerate(train_loader):
        c, f0, spec, y, spk, lengths, uv,volume = items
        g = spk.cuda(rank, non_blocking=True)
        spec, y = spec.cuda(rank, non_blocking=True), y.cuda(rank, non_blocking=True)
        c = c.cuda(rank, non_blocking=True)
        f0 = f0.cuda(rank, non_blocking=True)
        uv = uv.cuda(rank, non_blocking=True)
        lengths = lengths.cuda(rank, non_blocking=True)
        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)
        
        up_optim=(batch_idx + 1) % accum_steps == 0
        with autocast(enabled=hps.train.fp16_run, dtype=half_type,device_type='cuda'):
            y_hat, ids_slice, z_mask, \
            (z, z_p, m_p, logs_p, m_q, logs_q), pred_lf0, norm_lf0, lf0 = net_g(c, f0, uv, spec, g=g, c_lengths=lengths,
                                                                                spec_lengths=lengths,vol = volume)

            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
            # —— 1）Disc forward / backward —— 
            y_d_hat_r, y_d_hat_g, _, _ = net_d(y, y_hat.detach())

            with autocast(enabled=False, dtype=half_type,device_type='cuda'):
                loss_disc, losses_disc_r, losses_disc_g = discriminator_loss(y_d_hat_r, y_d_hat_g)
                loss_disc_all = loss_disc
        
        loss_disc_all = loss_disc_all #/ accum_steps
        if up_optim:
            scaler.scale(loss_disc_all).backward()
        else:
            with net_d.no_sync():
                scaler.scale(loss_disc_all).backward()

        # 只有当累积步数到达时，才实际更新 D
        if up_optim:
            scaler.unscale_(optim_d)
            grad_norm_d = commons.clip_grad_value_(net_d.parameters(), None)
            scaler.step(optim_d)
            #scaler.update() # 不在 discriminator 更新里调用，交给 generator 完成
            optim_d.zero_grad()

        # 冻结d
        """
        for p in net_d.parameters():
            p.requires_grad = False
        """
        with autocast(enabled=hps.train.fp16_run, dtype=half_type,device_type='cuda'):
            # Generator
            y_d_hat_r, y_d_hat_g, fmap_r, fmap_g = net_d(y, y_hat)
            with autocast(enabled=False, dtype=half_type,device_type='cuda'):
                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_lf0 = F.mse_loss(pred_lf0, lf0) if net_g.module.use_automatic_f0_prediction else 0
                loss_gen_all = loss_gen + loss_fm + loss_mel + loss_kl + loss_lf0
        # —— 2）Gen forward / backward —— 
        loss_gen_all = loss_gen_all #/ accum_steps
        if up_optim:
            scaler.scale(loss_gen_all).backward()
        else:
            with net_g.no_sync():
                scaler.scale(loss_gen_all).backward()
        # 解冻d
        """
        for p in net_d.parameters():
            p.requires_grad = True
        """

        if up_optim:
            scaler.unscale_(optim_g)
            grad_norm_g = commons.clip_grad_value_(net_g.parameters(), None)
            scaler.step(optim_g)
            scaler.update()
            optim_g.zero_grad()

        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_kl]
                """
                losses = [loss_disc * accum_steps,
                    loss_gen  * accum_steps,
                    loss_fm   * accum_steps,
                    loss_mel  * accum_steps,
                    loss_kl   * accum_steps]
                """
                reference_loss=0
                for i in losses:
                    reference_loss += i
                logger.info('Train Epoch: {} [{:.0f}%]'.format(
                    epoch,
                    100. * batch_idx / len(train_loader)))
                logger.info(f"Losses: {[x.item() for x in losses]}, step: {global_step}, lr: {lr}, reference_loss: {reference_loss}")
                if "grad_norm_d" in locals():
                    scalar_dict = {
                        "loss/g/total": loss_gen_all, #* accum_steps,
                        "loss/d/total": loss_disc_all, #* accum_steps,
                        "learning_rate": lr,
                        "grad_norm_d": grad_norm_d,
                        "grad_norm_g": grad_norm_g,
                    }
                    scalar_dict.update({
                        "loss/g/fm": loss_fm, #* accum_steps,
                        "loss/g/mel": loss_mel, #* accum_steps,
                        "loss/g/kl": loss_kl, #* accum_steps,
                        "loss/g/lf0": loss_lf0, #* accum_steps
                    })

                    # 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())
                    }

                    if net_g.module.use_automatic_f0_prediction:
                        image_dict.update({
                            "all/lf0": utils.plot_data_to_numpy(lf0[0, 0, :].cpu().numpy(),
                                                                  pred_lf0[0, 0, :].detach().cpu().numpy()),
                            "all/norm_lf0": utils.plot_data_to_numpy(lf0[0, 0, :].cpu().numpy(),
                                                                       norm_lf0[0, 0, :].detach().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: # 不保存G_0.pth和D_0.pth
                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)),
                                      Async=True)
                utils.save_checkpoint(net_d, optim_d, hps.train.learning_rate, epoch,
                                      os.path.join(hps.model_dir, "D_{}.pth".format(global_step)),
                                      Async=True)
                keep_ckpts = getattr(hps.train, 'keep_ckpts', 0)
                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 terminate_flag.is_set() and up_optim:
            save_model_on_exit(rank, hps, net_g, net_d, optim_g, optim_d, epoch)
            sys.exit(0)


    if rank == 0:
        global start_time
        now = time.time()
        durtaion = format(now - start_time, '.2f')
        logger.info(f'====> Epoch: {epoch}, cost {durtaion} s')
        start_time = now


def evaluate(hps, generator, eval_loader, writer_eval):
    generator.eval()
    image_dict = {}
    audio_dict = {}
    with torch.inference_mode(): # 比torch.no_grad()快一点
        for batch_idx, items in enumerate(eval_loader):
            c, f0, spec, y, spk, _, uv,volume = items
            g = spk[:1].cuda(0)
            spec, y = spec[:1].cuda(0), y[:1].cuda(0)
            c = c[:1].cuda(0)
            f0 = f0[:1].cuda(0)
            uv= uv[:1].cuda(0)
            if volume is not None:
                volume = volume[:1].cuda(0)
            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,_ = generator.module.infer(c, f0, uv, g=g,vol = volume)

            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
            )

            audio_dict.update({
                f"gen/audio_{batch_idx}": y_hat[0],
                f"gt/audio_{batch_idx}": y[0]
            })
        image_dict.update({
            "gen/mel": utils.plot_spectrogram_to_numpy(y_hat_mel[0].cpu().numpy()),
            "gt/mel": utils.plot_spectrogram_to_numpy(mel[0].cpu().numpy())
        })
    utils.summarize(
        writer=writer_eval,
        global_step=global_step,
        images=image_dict,
        audios=audio_dict,
        audio_sampling_rate=hps.data.sampling_rate
    )
    generator.train()

def save_model_on_exit(rank, hps, net_g, net_d, optim_g, optim_d, epoch):
    if rank == 0:
        print("\n[INFO] Saving model before exiting...")
        from utils import save_checkpoint  # 延迟导入防止循环
        save_checkpoint(net_g, optim_g, hps.train.learning_rate, epoch,
                        os.path.join(hps.model_dir, f"G_{global_step}.pth"))
        save_checkpoint(net_d, optim_d, hps.train.learning_rate, epoch,
                        os.path.join(hps.model_dir, f"D_{global_step}.pth"))

def handle_sigint(signum, frame):
    if not terminate_flag.is_set():
        print("\n[INFO] Received Ctrl+C, will save and exit after current iteration. Press again to force exit.")
        terminate_flag.set()
    else:
        print("\n[INFO] Forced termination.")
        sys.exit(1)

signal.signal(signal.SIGINT, handle_sigint)

if __name__ == "__main__":
    main()