def train(rank, a, h):
if h.num_gpus > 1:
init_process_group(backend=h.dist_config['dist_backend'],
init_method=h.dist_config['dist_url'],
world_size=h.dist_config['world_size'] * h.num_gpus,
rank=rank)
torch.cuda.manual_seed(h.seed)
torch.cuda.set_device(rank)
device = torch.device('cuda:{:d}'.format(rank))
generator = Generator(h).to(device)
mpd = MultiPeriodDiscriminator().to(device)
msd = MultiScaleDiscriminator().to(device)
if rank == 0:
print(generator)
os.makedirs(a.checkpoint_path, exist_ok=True)
print("checkpoints directory : ", a.checkpoint_path)
if os.path.isdir(a.checkpoint_path):
cp_g = scan_checkpoint(a.checkpoint_path, 'g_')
cp_do = scan_checkpoint(a.checkpoint_path, 'do_')
steps = 0
if cp_g is None or cp_do is None:
state_dict_do = None
last_epoch = -1
else:
state_dict_g = load_checkpoint(cp_g, device)
state_dict_do = load_checkpoint(cp_do, device)
generator.load_state_dict(state_dict_g['generator'])
mpd.load_state_dict(state_dict_do['mpd'])
msd.load_state_dict(state_dict_do['msd'])
steps = state_dict_do['steps'] + 1
last_epoch = state_dict_do['epoch']
if h.num_gpus > 1:
generator = DistributedDataParallel(generator,
device_ids=[rank]).to(device)
mpd = DistributedDataParallel(mpd, device_ids=[rank]).to(device)
msd = DistributedDataParallel(msd, device_ids=[rank]).to(device)
optim_g = torch.optim.AdamW(generator.parameters(),
h.learning_rate,
betas=[h.adam_b1, h.adam_b2])
optim_d = torch.optim.AdamW(itertools.chain(msd.parameters(),
mpd.parameters()),
h.learning_rate,
betas=[h.adam_b1, h.adam_b2])
if state_dict_do is not None:
optim_g.load_state_dict(state_dict_do['optim_g'])
optim_d.load_state_dict(state_dict_do['optim_d'])
scheduler_g = torch.optim.lr_scheduler.ExponentialLR(optim_g,
gamma=h.lr_decay,
last_epoch=last_epoch)
scheduler_d = torch.optim.lr_scheduler.ExponentialLR(optim_d,
gamma=h.lr_decay,
last_epoch=last_epoch)
training_filelist, validation_filelist = get_dataset_filelist(a)
trainset = MelDataset(training_filelist,
h.segment_size,
h.n_fft,
h.num_mels,
h.hop_size,
h.win_size,
h.sampling_rate,
h.fmin,
h.fmax,
n_cache_reuse=0,
shuffle=False if h.num_gpus > 1 else True,
fmax_loss=h.fmax_for_loss,
device=device,
fine_tuning=a.fine_tuning,
base_mels_path=a.input_mels_dir)
train_sampler = DistributedSampler(trainset) if h.num_gpus > 1 else None
train_loader = DataLoader(trainset,
num_workers=h.num_workers,
shuffle=False,
sampler=train_sampler,
batch_size=h.batch_size,
pin_memory=True,
drop_last=True)
if rank == 0:
validset = MelDataset(validation_filelist,
h.segment_size,
h.n_fft,
h.num_mels,
h.hop_size,
h.win_size,
h.sampling_rate,
h.fmin,
h.fmax,
False,
False,
n_cache_reuse=0,
fmax_loss=h.fmax_for_loss,
device=device,
fine_tuning=a.fine_tuning,
base_mels_path=a.input_mels_dir)
validation_loader = DataLoader(validset,
num_workers=1,
shuffle=False,
sampler=None,
batch_size=1,
pin_memory=True,
drop_last=True)
sw = SummaryWriter(os.path.join(a.checkpoint_path, 'logs'))
generator.train()
mpd.train()
msd.train()
for epoch in range(max(0, last_epoch), a.training_epochs):
if rank == 0:
start = time.time()
print("Epoch: {}".format(epoch + 1))
if h.num_gpus > 1:
train_sampler.set_epoch(epoch)
for i, batch in enumerate(train_loader):
if rank == 0:
start_b = time.time()
x, y, _, y_mel = batch
x = torch.autograd.Variable(x.to(device, non_blocking=True))
y = torch.autograd.Variable(y.to(device, non_blocking=True))
y_mel = torch.autograd.Variable(y_mel.to(device,
non_blocking=True))
y = y.unsqueeze(1)
y_g_hat = generator(x)
y_g_hat_mel = mel_spectrogram(y_g_hat.squeeze(1), h.n_fft,
h.num_mels, h.sampling_rate,
h.hop_size, h.win_size, h.fmin,
h.fmax_for_loss)
optim_d.zero_grad()
# MPD
y_df_hat_r, y_df_hat_g, _, _ = mpd(y, y_g_hat.detach())
loss_disc_f, losses_disc_f_r, losses_disc_f_g = discriminator_loss(
y_df_hat_r, y_df_hat_g)
# MSD
y_ds_hat_r, y_ds_hat_g, _, _ = msd(y, y_g_hat.detach())
loss_disc_s, losses_disc_s_r, losses_disc_s_g = discriminator_loss(
y_ds_hat_r, y_ds_hat_g)
loss_disc_all = loss_disc_s + loss_disc_f
loss_disc_all.backward()
optim_d.step()
# Generator
optim_g.zero_grad()
# L1 Mel-Spectrogram Loss
loss_mel = F.l1_loss(y_mel, y_g_hat_mel) * 45
y_df_hat_r, y_df_hat_g, fmap_f_r, fmap_f_g = mpd(y, y_g_hat)
y_ds_hat_r, y_ds_hat_g, fmap_s_r, fmap_s_g = msd(y, y_g_hat)
loss_fm_f = feature_loss(fmap_f_r, fmap_f_g)
loss_fm_s = feature_loss(fmap_s_r, fmap_s_g)
loss_gen_f, losses_gen_f = generator_loss(y_df_hat_g)
loss_gen_s, losses_gen_s = generator_loss(y_ds_hat_g)
loss_gen_all = loss_gen_s + loss_gen_f + loss_fm_s + loss_fm_f + loss_mel
loss_gen_all.backward()
optim_g.step()
if rank == 0:
# STDOUT logging
if steps % a.stdout_interval == 0:
with torch.no_grad():
mel_error = F.l1_loss(y_mel, y_g_hat_mel).item()
print(
'Steps : {:d}, Gen Loss Total : {:4.3f}, Mel-Spec. Error : {:4.3f}, s/b : {:4.3f}'
.format(steps, loss_gen_all, mel_error,
time.time() - start_b))
# checkpointing
if steps % a.checkpoint_interval == 0 and steps != 0:
checkpoint_path = "{}/g_{:08d}".format(
a.checkpoint_path, steps)
save_checkpoint(
checkpoint_path, {
'generator': (generator.module if h.num_gpus > 1
else generator).state_dict()
})
checkpoint_path = "{}/do_{:08d}".format(
a.checkpoint_path, steps)
save_checkpoint(
checkpoint_path, {
'mpd': (mpd.module
if h.num_gpus > 1 else mpd).state_dict(),
'msd': (msd.module
if h.num_gpus > 1 else msd).state_dict(),
'optim_g':
optim_g.state_dict(),
'optim_d':
optim_d.state_dict(),
'steps':
steps,
'epoch':
epoch
})
# Tensorboard summary logging
if steps % a.summary_interval == 0:
sw.add_scalar("training/gen_loss_total", loss_gen_all,
steps)
sw.add_scalar("training/mel_spec_error", mel_error, steps)
# Validation
if steps % a.validation_interval == 0: # and steps != 0:
generator.eval()
torch.cuda.empty_cache()
val_err_tot = 0
with torch.no_grad():
for j, batch in enumerate(validation_loader):
x, y, _, y_mel = batch
y_g_hat = generator(x.to(device))
y_mel = torch.autograd.Variable(
y_mel.to(device, non_blocking=True))
y_g_hat_mel = mel_spectrogram(
y_g_hat.squeeze(1), h.n_fft, h.num_mels,
h.sampling_rate, h.hop_size, h.win_size,
h.fmin, h.fmax_for_loss)
val_err_tot += F.l1_loss(y_mel, y_g_hat_mel).item()
if j <= 4:
if steps == 0:
sw.add_audio('gt/y_{}'.format(j), y[0],
steps, h.sampling_rate)
sw.add_figure('gt/y_spec_{}'.format(j),
plot_spectrogram(x[0]),
steps)
sw.add_audio('generated/y_hat_{}'.format(j),
y_g_hat[0], steps,
h.sampling_rate)
y_hat_spec = mel_spectrogram(
y_g_hat.squeeze(1), h.n_fft, h.num_mels,
h.sampling_rate, h.hop_size, h.win_size,
h.fmin, h.fmax)
sw.add_figure(
'generated/y_hat_spec_{}'.format(j),
plot_spectrogram(
y_hat_spec.squeeze(0).cpu().numpy()),
steps)
val_err = val_err_tot / (j + 1)
sw.add_scalar("validation/mel_spec_error", val_err,
steps)
generator.train()
steps += 1
scheduler_g.step()
scheduler_d.step()
if rank == 0:
print('Time taken for epoch {} is {} sec\n'.format(
epoch + 1, int(time.time() - start)))
def train(rank, a, h):
if h.num_gpus > 1:
init_process_group(backend=h.dist_config['dist_backend'],
init_method=h.dist_config['dist_url'],
world_size=h.dist_config['world_size'] * h.num_gpus,
rank=rank)
torch.cuda.manual_seed(h.seed)
device = torch.device('cuda:{:d}'.format(rank))
generator = Generator(h).to(device)
mpd = MultiPeriodDiscriminator(
h["discriminator_periods"] if "discriminator_periods" in
h.keys() else None).to(device)
msd = MultiScaleDiscriminator().to(device)
if rank == 0:
print(generator)
os.makedirs(a.checkpoint_path, exist_ok=True)
print("checkpoints directory : ", a.checkpoint_path)
if os.path.isdir(a.checkpoint_path):
cp_g = scan_checkpoint(a.checkpoint_path, 'g_')
cp_do = scan_checkpoint(a.checkpoint_path, 'do_')
steps = 0
if cp_g is not None:
state_dict_g = load_checkpoint(cp_g, device)
gsd = generator.state_dict()
gsd.update({
k: v
for k, v in state_dict_g['generator'].items()
if k in gsd and state_dict_g['generator'][k].shape == gsd[k].shape
})
missing_keys = {
k: v
for k, v in state_dict_g['generator'].items()
if not (k in gsd
and state_dict_g['generator'][k].shape == gsd[k].shape)
}.keys()
generator.load_state_dict(gsd)
del gsd, state_dict_g
if cp_do is None or len(missing_keys) or a.from_zero:
state_dict_do = None
last_epoch = -1
else:
state_dict_do = load_checkpoint(cp_do, device)
mpd.load_state_dict(state_dict_do['mpd'])
del state_dict_do['mpd']
msd.load_state_dict(state_dict_do['msd'])
del state_dict_do['msd']
steps = state_dict_do['steps'] + 1
last_epoch = state_dict_do['epoch']
if h.num_gpus > 1:
generator = DistributedDataParallel(generator,
device_ids=[rank]).to(device)
mpd = DistributedDataParallel(mpd, device_ids=[rank]).to(device)
msd = DistributedDataParallel(msd, device_ids=[rank]).to(device)
optim_g = torch.optim.AdamW(generator.parameters(),
h.learning_rate,
betas=[h.adam_b1, h.adam_b2])
optim_d = torch.optim.AdamW(itertools.chain(msd.parameters(),
mpd.parameters()),
h.learning_rate,
betas=[h.adam_b1, h.adam_b2])
if state_dict_do is not None:
optim_g.load_state_dict(state_dict_do['optim_g'])
optim_d.load_state_dict(state_dict_do['optim_d'])
del state_dict_do
scheduler_g = torch.optim.lr_scheduler.ExponentialLR(optim_g,
gamma=h.lr_decay,
last_epoch=last_epoch)
scheduler_d = torch.optim.lr_scheduler.ExponentialLR(optim_d,
gamma=h.lr_decay,
last_epoch=last_epoch)
training_filelist, validation_filelist = get_dataset_filelist(
a, h.segment_size, h.sampling_rate)
trainset = MelDataset(training_filelist,
h.segment_size,
h.n_fft,
h.num_mels,
h.hop_size,
h.win_size,
h.sampling_rate,
h.fmin,
h.fmax,
n_cache_reuse=0,
shuffle=False if h.num_gpus > 1 else True,
fmax_loss=h.fmax_for_loss,
device=device,
fine_tuning=a.fine_tuning,
trim_non_voiced=a.trim_non_voiced)
STFT = STFT_Class(h.sampling_rate, h.num_mels, h.n_fft, h.win_size,
h.hop_size, h.fmin, h.fmax)
train_sampler = DistributedSampler(trainset) if h.num_gpus > 1 else None
train_loader = DataLoader(trainset,
num_workers=h.num_workers,
shuffle=False,
sampler=train_sampler,
batch_size=h.batch_size,
pin_memory=True,
drop_last=True)
assert len(train_loader), 'No audio files in dataset!'
if rank == 0:
validset = MelDataset(validation_filelist,
h.segment_size,
h.n_fft,
h.num_mels,
h.hop_size,
h.win_size,
h.sampling_rate,
h.fmin,
h.fmax,
False,
False,
n_cache_reuse=0,
fmax_loss=h.fmax_for_loss,
device=device,
fine_tuning=a.fine_tuning,
trim_non_voiced=a.trim_non_voiced)
validation_loader = DataLoader(validset,
num_workers=h.num_workers,
shuffle=False,
sampler=None,
batch_size=1,
pin_memory=True,
drop_last=True)
sw = SummaryWriter(os.path.join(a.checkpoint_path, 'logs'),
max_queue=10000)
sw.logged_gt_plots = False
if h.num_gpus > 1:
import gc
gc.collect()
torch.cuda.empty_cache()
generator.train()
mpd.train()
msd.train()
for epoch in range(max(0, last_epoch), a.training_epochs):
if rank == 0:
start = time.time()
print("Epoch: {}".format(epoch + 1))
if h.num_gpus > 1:
train_sampler.set_epoch(epoch)
for i, batch in enumerate(train_loader):
if rank == 0:
start_b = time.time()
x, y, _, y_mel = batch
x = torch.autograd.Variable(x.to(device, non_blocking=True))
y = torch.autograd.Variable(y.to(device, non_blocking=True))
y_mel = torch.autograd.Variable(y_mel.to(device,
non_blocking=True))
y = y.unsqueeze(1)
y_g_hat = generator(x)
y_g_hat_mel = STFT.get_mel(y_g_hat.squeeze(1))
optim_d.zero_grad()
# MPD
y_df_hat_r, y_df_hat_g, _, _ = mpd(y, y_g_hat.detach())
loss_disc_f, losses_disc_f_r, losses_disc_f_g = discriminator_loss(
y_df_hat_r, y_df_hat_g)
# MSD
y_ds_hat_r, y_ds_hat_g, _, _ = msd(y, y_g_hat.detach())
loss_disc_s, losses_disc_s_r, losses_disc_s_g = discriminator_loss(
y_ds_hat_r, y_ds_hat_g)
loss_disc_all = loss_disc_s + loss_disc_f
loss_disc_all.backward()
optim_d.step()
# Generator
optim_g.zero_grad()
# L1 Mel-Spectrogram Loss
loss_mel = F.l1_loss(y_mel, y_g_hat_mel)
y_df_hat_r, y_df_hat_g, fmap_f_r, fmap_f_g = mpd(y, y_g_hat)
y_ds_hat_r, y_ds_hat_g, fmap_s_r, fmap_s_g = msd(y, y_g_hat)
loss_fm_f = feature_loss(fmap_f_r, fmap_f_g)
loss_fm_s = feature_loss(fmap_s_r, fmap_s_g)
loss_gen_f, losses_gen_f = generator_loss(y_df_hat_g)
loss_gen_s, losses_gen_s = generator_loss(y_ds_hat_g)
loss_gen_all = loss_gen_s + loss_gen_f + loss_fm_s + loss_fm_f + loss_mel * 45
loss_gen_all.backward()
optim_g.step()
if rank == 0:
torch.set_grad_enabled(False)
# STDOUT logging
if steps % a.stdout_interval == 0:
print(
'Steps : {:d}, Gen Loss Total : {:4.3f}, Mel-Spec. Error : {:4.3f}, s/b : {:4.3f}'
.format(steps, loss_gen_all, loss_mel.item(),
time.time() - start_b))
# checkpointing
if steps % a.checkpoint_interval == 0 and steps != 0:
checkpoint_path = "{}/g_{:08d}".format(
a.checkpoint_path, steps)
save_checkpoint(
checkpoint_path, {
'generator': (generator.module if h.num_gpus > 1
else generator).state_dict()
})
checkpoint_path = "{}/do_{:08d}".format(
a.checkpoint_path, steps)
save_checkpoint(
checkpoint_path, {
'mpd': (mpd.module
if h.num_gpus > 1 else mpd).state_dict(),
'msd': (msd.module
if h.num_gpus > 1 else msd).state_dict(),
'optim_g':
optim_g.state_dict(),
'optim_d':
optim_d.state_dict(),
'steps':
steps,
'epoch':
epoch
})
del_old_checkpoints(a.checkpoint_path, 'g_',
a.n_models_to_keep)
del_old_checkpoints(a.checkpoint_path, 'do_',
a.n_models_to_keep)
# Tensorboard summary logging
if steps % a.summary_interval == 0:
sw.add_scalar("training/gen_loss_total", loss_gen_all,
steps)
sw.add_scalar("training/mel_spec_error", loss_mel.item(),
steps)
# Validation
if steps % a.validation_interval == 0: # and steps != 0:
print("Validating...")
n_audios_to_plot = 6
generator.eval()
torch.cuda.empty_cache()
val_err_tot = 0
for j, batch in tqdm(enumerate(validation_loader),
total=len(validation_loader)):
x, y, _, y_mel = batch
y_g_hat = generator(x.to(device))
y_hat_spec = STFT.get_mel(y_g_hat.squeeze(1))
val_err_tot += F.l1_loss(y_mel,
y_hat_spec.to(y_mel)).item()
if j < n_audios_to_plot and not sw.logged_gt_plots:
sw.add_audio(f'gt/y_{j}', y[0], steps,
h.sampling_rate)
sw.add_figure(f'spec_{j:02}/gt_spec',
plot_spectrogram(y_mel[0]), steps)
if j < n_audios_to_plot:
sw.add_audio(f'generated/y_hat_{j}', y_g_hat[0],
steps, h.sampling_rate)
sw.add_figure(
f'spec_{j:02}/pred_spec',
plot_spectrogram(
y_hat_spec.squeeze(0).cpu().numpy()),
steps)
if j > 64: # I am NOT patient enough to complete an entire validation cycle with over 1536 files.
break
sw.logged_gt_plots = True
val_err = val_err_tot / (j + 1)
sw.add_scalar("validation/mel_spec_error", val_err, steps)
generator.train()
print(f"Done. Val_loss = {val_err}")
torch.set_grad_enabled(True)
steps += 1
scheduler_g.step()
scheduler_d.step()
if rank == 0:
print('Time taken for epoch {} is {} sec\n'.format(
epoch + 1, int(time.time() - start)))
def train(rank, args, hp, hp_str):
# if hp.train.num_gpus > 1:
# init_process_group(backend=hp.dist.dist_backend, init_method=hp.dist.dist_url,
# world_size=hp.dist.world_size * hp.train.num_gpus, rank=rank)
torch.cuda.manual_seed(hp.train.seed)
device = torch.device('cuda:{:d}'.format(rank))
generator = Generator(hp.model.in_channels,
hp.model.out_channels).to(device)
specd = SpecDiscriminator().to(device)
msd = MultiScaleDiscriminator().to(device)
stft_loss = MultiResolutionSTFTLoss()
if rank == 0:
print(generator)
os.makedirs(hp.logs.chkpt_dir, exist_ok=True)
print("checkpoints directory : ", hp.logs.chkpt_dir)
if os.path.isdir(hp.logs.chkpt_dir):
cp_g = scan_checkpoint(hp.logs.chkpt_dir, 'g_')
cp_do = scan_checkpoint(hp.logs.chkpt_dir, 'do_')
steps = 0
if cp_g is None or cp_do is None:
state_dict_do = None
last_epoch = -1
else:
state_dict_g = load_checkpoint(cp_g, device)
state_dict_do = load_checkpoint(cp_do, device)
generator.load_state_dict(state_dict_g['generator'])
specd.load_state_dict(state_dict_do['specd'])
msd.load_state_dict(state_dict_do['msd'])
steps = state_dict_do['steps'] + 1
last_epoch = state_dict_do['epoch']
if hp.train.num_gpus > 1:
generator = DistributedDataParallel(generator,
device_ids=[rank]).to(device)
specd = DistributedDataParallel(specd, device_ids=[rank]).to(device)
msd = DistributedDataParallel(msd, device_ids=[rank]).to(device)
optim_g = torch.optim.AdamW(
generator.parameters(),
hp.train.adamG.lr,
betas=[hp.train.adamG.beta1, hp.train.adamG.beta2])
optim_d = torch.optim.AdamW(
itertools.chain(msd.parameters(), specd.parameters()),
hp.train.adamD.lr,
betas=[hp.train.adamD.beta1, hp.train.adamD.beta2])
if state_dict_do is not None:
optim_g.load_state_dict(state_dict_do['optim_g'])
optim_d.load_state_dict(state_dict_do['optim_d'])
# scheduler_g = torch.optim.lr_scheduler.ExponentialLR(optim_g, gamma=hp.train.adam.lr_decay, last_epoch=last_epoch)
# scheduler_d = torch.optim.lr_scheduler.ExponentialLR(optim_d, gamma=hp.train.adam.lr_decay, last_epoch=last_epoch)
training_filelist, validation_filelist = get_dataset_filelist(args)
trainset = MelDataset(training_filelist,
hp.data.input_wavs,
hp.data.output_wavs,
hp.audio.segment_length,
hp.audio.filter_length,
hp.audio.n_mel_channels,
hp.audio.hop_length,
hp.audio.win_length,
hp.audio.sampling_rate,
hp.audio.mel_fmin,
hp.audio.mel_fmax,
n_cache_reuse=0,
shuffle=False if hp.train.num_gpus > 1 else True,
fmax_loss=None,
device=device)
train_sampler = DistributedSampler(
trainset) if hp.train.num_gpus > 1 else None
train_loader = DataLoader(trainset,
num_workers=hp.train.num_workers,
shuffle=False,
sampler=train_sampler,
batch_size=hp.train.batch_size,
pin_memory=True,
drop_last=True)
if rank == 0:
validset = MelDataset(validation_filelist,
hp.data.input_wavs,
hp.data.output_wavs,
hp.audio.segment_length,
hp.audio.filter_length,
hp.audio.n_mel_channels,
hp.audio.hop_length,
hp.audio.win_length,
hp.audio.sampling_rate,
hp.audio.mel_fmin,
hp.audio.mel_fmax,
split=False,
shuffle=False,
n_cache_reuse=0,
fmax_loss=None,
device=device)
validation_loader = DataLoader(validset,
num_workers=1,
shuffle=False,
sampler=None,
batch_size=1,
pin_memory=True,
drop_last=True)
sw = SummaryWriter(os.path.join(hp.logs.chkpt_dir, 'logs'))
generator.train()
specd.train()
msd.train()
with_postnet = False
for epoch in range(max(0, last_epoch), args.training_epochs):
if rank == 0:
start = time.time()
print("Epoch: {}".format(epoch + 1))
if hp.train.num_gpus > 1:
train_sampler.set_epoch(epoch)
for i, batch in enumerate(train_loader):
if rank == 0:
start_b = time.time()
if steps > hp.train.postnet_start_steps:
with_postnet = True
x, y, file, _, y_mel_loss = batch
x = torch.autograd.Variable(x.to(device, non_blocking=True))
y = torch.autograd.Variable(y.to(device, non_blocking=True))
y_mel_loss = torch.autograd.Variable(
y_mel_loss.to(device, non_blocking=True))
# y_mel = torch.autograd.Variable(y_mel.to(device, non_blocking=True))
x = x.unsqueeze(1)
y = y.unsqueeze(1)
before_y_g_hat, y_g_hat = generator(x, with_postnet)
if y_g_hat is not None:
y_g_hat_mel = mel_spectrogram(
y_g_hat.squeeze(1), hp.audio.filter_length,
hp.audio.n_mel_channels, hp.audio.sampling_rate,
hp.audio.hop_length, hp.audio.win_length,
hp.audio.mel_fmin, None)
if steps > hp.train.discriminator_train_start_steps:
for _ in range(hp.train.rep_discriminator):
optim_d.zero_grad()
# SpecD
y_df_hat_r, y_df_hat_g, _, _ = specd(
y_mel_loss, y_g_hat_mel.detach())
loss_disc_f, losses_disc_f_r, losses_disc_f_g = discriminator_loss(
y_df_hat_r, y_df_hat_g)
# MSD
y_ds_hat_r, y_ds_hat_g, _, _ = msd(y, y_g_hat.detach())
loss_disc_s, losses_disc_s_r, losses_disc_s_g = discriminator_loss(
y_ds_hat_r, y_ds_hat_g)
loss_disc_all = loss_disc_s + loss_disc_f
loss_disc_all.backward()
optim_d.step()
before_y_g_hat_mel = mel_spectrogram(
before_y_g_hat.squeeze(1), hp.audio.filter_length,
hp.audio.n_mel_channels, hp.audio.sampling_rate,
hp.audio.hop_length, hp.audio.win_length, hp.audio.mel_fmin,
None)
# Generator
optim_g.zero_grad()
# L1 Mel-Spectrogram Loss
# before_loss_mel = F.l1_loss(y_mel_loss, before_y_g_hat_mel)
sc_loss, mag_loss = stft_loss(
before_y_g_hat[:, :, :y.size(2)].squeeze(1), y.squeeze(1))
before_loss_mel = sc_loss + mag_loss
# L1 Sample Loss
before_loss_sample = F.l1_loss(y, before_y_g_hat)
loss_gen_all = before_loss_mel + before_loss_sample
if y_g_hat is not None:
# L1 Mel-Spectrogram Loss
# loss_mel = F.l1_loss(y_mel_loss, y_g_hat_mel)
sc_loss_, mag_loss_ = stft_loss(
y_g_hat[:, :, :y.size(2)].squeeze(1), y.squeeze(1))
loss_mel = sc_loss_ + mag_loss_
# L1 Sample Loss
loss_sample = F.l1_loss(y, y_g_hat)
loss_gen_all += loss_mel + loss_sample
if steps > hp.train.discriminator_train_start_steps:
y_df_hat_r, y_df_hat_g, fmap_f_r, fmap_f_g = specd(
y_mel_loss, y_g_hat_mel)
y_ds_hat_r, y_ds_hat_g, fmap_s_r, fmap_s_g = msd(y, y_g_hat)
loss_fm_f = feature_loss(fmap_f_r, fmap_f_g)
loss_fm_s = feature_loss(fmap_s_r, fmap_s_g)
loss_gen_f, losses_gen_f = generator_loss(y_df_hat_g)
loss_gen_s, losses_gen_s = generator_loss(y_ds_hat_g)
loss_gen_all += hp.model.lambda_adv * (
loss_gen_s + loss_gen_f + loss_fm_s + loss_fm_f)
loss_gen_all.backward()
optim_g.step()
if rank == 0:
# STDOUT logging
if steps % args.stdout_interval == 0:
with torch.no_grad():
mel_error = F.l1_loss(y_mel_loss,
before_y_g_hat_mel).item()
sample_error = F.l1_loss(y, before_y_g_hat)
print(
'Steps : {:d}, Gen Loss Total : {:4.3f}, Sample Error: {:4.3f}, '
'Mel-Spec. Error : {:4.3f}, s/b : {:4.3f}'.format(
steps, loss_gen_all, sample_error, mel_error,
time.time() - start_b))
# checkpointing
if steps % hp.logs.save_interval == 0 and steps != 0:
checkpoint_path = "{}/g_{:08d}".format(
hp.logs.chkpt_dir, steps)
save_checkpoint(
checkpoint_path, {
'generator':
(generator.module if hp.train.num_gpus > 1 else
generator).state_dict()
})
checkpoint_path = "{}/do_{:08d}".format(
hp.logs.chkpt_dir, steps)
save_checkpoint(
checkpoint_path, {
'specd': (specd.module if hp.train.num_gpus > 1
else specd).state_dict(),
'msd': (msd.module if hp.train.num_gpus > 1 else
msd).state_dict(),
'optim_g':
optim_g.state_dict(),
'optim_d':
optim_d.state_dict(),
'steps':
steps,
'epoch':
epoch,
'hp_str':
hp_str
})
# Tensorboard summary logging
if steps % hp.logs.summary_interval == 0:
sw.add_scalar("training/gen_loss_total", loss_gen_all,
steps)
sw.add_scalar("training/mel_spec_error", mel_error, steps)
# Validation
if steps % hp.logs.validation_interval == 0: # and steps != 0:
generator.eval()
torch.cuda.empty_cache()
val_err_tot = 0
with torch.no_grad():
for j, batch in enumerate(validation_loader):
x, y, file, y_mel, y_mel_loss = batch
x = x.unsqueeze(1)
y = y.unsqueeze(1).to(device)
before_y_g_hat, y_g_hat = generator(x.to(device))
y_mel_loss = torch.autograd.Variable(
y_mel_loss.to(device, non_blocking=True))
y_g_hat_mel = mel_spectrogram(
before_y_g_hat.squeeze(1),
hp.audio.filter_length,
hp.audio.n_mel_channels,
hp.audio.sampling_rate, hp.audio.hop_length,
hp.audio.win_length, hp.audio.mel_fmin, None)
val_err_tot += F.l1_loss(y_mel_loss,
y_g_hat_mel).item()
val_err_tot += F.l1_loss(y, before_y_g_hat).item()
if y_g_hat is not None:
val_err_tot += F.l1_loss(y, y_g_hat).item()
if j <= 4:
if steps == 0:
sw.add_audio('gt_noise/y_{}'.format(j),
x[0], steps,
hp.audio.sampling_rate)
sw.add_audio('gt_clean/y_{}'.format(j),
y[0], steps,
hp.audio.sampling_rate)
sw.add_figure(
'gt/y_spec_clean_{}'.format(j),
plot_spectrogram(y_mel[0]), steps)
sw.add_audio('generated/y_hat_{}'.format(j),
before_y_g_hat[0], steps,
hp.audio.sampling_rate)
if y_g_hat is not None:
sw.add_audio(
'generated/y_hat_after_{}'.format(j),
y_g_hat[0], steps,
hp.audio.sampling_rate)
y_hat_spec = mel_spectrogram(
before_y_g_hat.squeeze(1),
hp.audio.filter_length,
hp.audio.n_mel_channels,
hp.audio.sampling_rate,
hp.audio.hop_length, hp.audio.win_length,
hp.audio.mel_fmin, None)
sw.add_figure(
'generated/y_hat_spec_{}'.format(j),
plot_spectrogram(
y_hat_spec.squeeze(0).cpu().numpy()),
steps)
val_err = val_err_tot / (j + 1)
sw.add_scalar("validation/mel_spec_error", val_err,
steps)
generator.train()
steps += 1
# scheduler_g.step()
# scheduler_d.step()
if rank == 0:
print('Time taken for epoch {} is {} sec\n'.format(
epoch + 1, int(time.time() - start)))
def train(rank: int, cfg: DictConfig):
print(OmegaConf.to_yaml(cfg))
if cfg.train.n_gpu > 1:
init_process_group(backend=cfg.train.dist_config['dist_backend'],
init_method=cfg.train.dist_config['dist_url'],
world_size=cfg.train.dist_config['world_size'] *
cfg.train.n_gpu,
rank=rank)
device = torch.device(
'cuda:{:d}'.format(rank) if torch.cuda.is_available() else 'cpu')
generator = Generator(sum(cfg.model.feature_dims), *cfg.model.cond_dims,
**cfg.model.generator).to(device)
discriminator = Discriminator(**cfg.model.discriminator).to(device)
if rank == 0:
print(generator)
os.makedirs(cfg.train.ckpt_dir, exist_ok=True)
print("checkpoints directory : ", cfg.train.ckpt_dir)
if os.path.isdir(cfg.train.ckpt_dir):
cp_g = scan_checkpoint(cfg.train.ckpt_dir, 'g_')
cp_do = scan_checkpoint(cfg.train.ckpt_dir, 'd_')
steps = 1
if cp_g is None or cp_do is None:
state_dict_do = None
last_epoch = -1
else:
state_dict_g = load_checkpoint(cp_g, device)
state_dict_do = load_checkpoint(cp_do, device)
generator.load_state_dict(state_dict_g['generator'])
discriminator.load_state_dict(state_dict_do['discriminator'])
steps = state_dict_do['steps'] + 1
last_epoch = state_dict_do['epoch']
if cfg.train.n_gpu > 1:
generator = DistributedDataParallel(generator,
device_ids=[rank]).to(device)
discriminator = DistributedDataParallel(discriminator,
device_ids=[rank]).to(device)
optim_g = RAdam(generator.parameters(), cfg.opt.lr, betas=cfg.opt.betas)
optim_d = RAdam(discriminator.parameters(),
cfg.opt.lr,
betas=cfg.opt.betas)
if state_dict_do is not None:
optim_g.load_state_dict(state_dict_do['optim_g'])
optim_d.load_state_dict(state_dict_do['optim_d'])
scheduler_g = torch.optim.lr_scheduler.ExponentialLR(
optim_g, gamma=cfg.opt.lr_decay, last_epoch=last_epoch)
scheduler_d = torch.optim.lr_scheduler.ExponentialLR(
optim_d, gamma=cfg.opt.lr_decay, last_epoch=last_epoch)
train_filelist = load_dataset_filelist(cfg.dataset.train_list)
trainset = FeatureDataset(cfg.dataset, train_filelist, cfg.data)
train_sampler = DistributedSampler(
trainset) if cfg.train.n_gpu > 1 else None
train_loader = DataLoader(trainset,
batch_size=cfg.train.batch_size,
num_workers=cfg.train.num_workers,
shuffle=True,
sampler=train_sampler,
pin_memory=True,
drop_last=True)
if rank == 0:
val_filelist = load_dataset_filelist(cfg.dataset.test_list)
valset = FeatureDataset(cfg.dataset,
val_filelist,
cfg.data,
segmented=False)
val_loader = DataLoader(valset,
batch_size=1,
num_workers=cfg.train.num_workers,
shuffle=False,
sampler=train_sampler,
pin_memory=True)
sw = SummaryWriter(os.path.join(cfg.train.ckpt_dir, 'logs'))
generator.train()
discriminator.train()
for epoch in range(max(0, last_epoch), cfg.train.epochs):
if rank == 0:
start = time.time()
print("Epoch: {}".format(epoch + 1))
if cfg.train.n_gpu > 1:
train_sampler.set_epoch(epoch)
for y, x_noised_features, x_noised_cond in train_loader:
if rank == 0:
start_b = time.time()
y = y.to(device, non_blocking=True)
x_noised_features = x_noised_features.transpose(1, 2).to(
device, non_blocking=True)
x_noised_cond = x_noised_cond.to(device, non_blocking=True)
z1 = torch.randn(cfg.train.batch_size,
cfg.model.cond_dims[1],
device=device)
z2 = torch.randn(cfg.train.batch_size,
cfg.model.cond_dims[1],
device=device)
y_hat1 = generator(x_noised_features, x_noised_cond, z=z1)
y_hat2 = generator(x_noised_features, x_noised_cond, z=z2)
# Discriminator
real_scores, fake_scores = discriminator(y), discriminator(
y_hat1.detach())
d_loss = discriminator_loss(real_scores, fake_scores)
optim_d.zero_grad()
d_loss.backward(retain_graph=True)
optim_d.step()
# Generator
g_stft_loss = criterion(y, y_hat1) + criterion(
y, y_hat2) - criterion(y_hat1, y_hat2)
g_adv_loss = adversarial_loss(fake_scores)
g_loss = g_adv_loss + g_stft_loss
optim_g.zero_grad()
g_loss.backward()
optim_g.step()
if rank == 0:
# STDOUT logging
if steps % cfg.train.stdout_interval == 0:
with torch.no_grad():
print(
'Steps : {:d}, Gen Loss Total : {:4.3f}, STFT Error : {:4.3f}, s/b : {:4.3f}'
.format(steps, g_loss, g_stft_loss,
time.time() - start_b))
# checkpointing
if steps % cfg.train.checkpoint_interval == 0:
ckpt_dir = "{}/g_{:08d}".format(cfg.train.ckpt_dir, steps)
save_checkpoint(
ckpt_dir, {
'generator':
(generator.module if cfg.train.n_gpu > 1 else
generator).state_dict()
})
ckpt_dir = "{}/do_{:08d}".format(cfg.train.ckpt_dir, steps)
save_checkpoint(
ckpt_dir, {
'discriminator':
(discriminator.module if cfg.train.n_gpu > 1 else
discriminator).state_dict(),
'optim_g':
optim_g.state_dict(),
'optim_d':
optim_d.state_dict(),
'steps':
steps,
'epoch':
epoch
})
# Tensorboard summary logging
if steps % cfg.train.summary_interval == 0:
sw.add_scalar("training/gen_loss_total", g_loss, steps)
sw.add_scalar("training/gen_stft_error", g_stft_loss,
steps)
# Validation
if steps % cfg.train.validation_interval == 0:
generator.eval()
torch.cuda.empty_cache()
val_err_tot = 0
with torch.no_grad():
for j, (y, x_noised_features,
x_noised_cond) in enumerate(val_loader):
y_hat = generator(
x_noised_features.transpose(1, 2).to(device),
x_noised_cond.to(device))
val_err_tot += criterion(y, y_hat).item()
if j <= 4:
# sw.add_audio('noised/y_noised_{}'.format(j), y_noised[0], steps, cfg.data.target_sample_rate)
sw.add_audio('generated/y_hat_{}'.format(j),
y_hat[0], steps,
cfg.data.sample_rate)
sw.add_audio('gt/y_{}'.format(j), y[0], steps,
cfg.data.sample_rate)
val_err = val_err_tot / (j + 1)
sw.add_scalar("validation/stft_error", val_err, steps)
generator.train()
steps += 1
scheduler_g.step()
scheduler_d.step()
if rank == 0:
print('Time taken for epoch {} is {} sec\n'.format(
epoch + 1, int(time.time() - start)))
