def validation(generator,
validation_loader,
sw,
h,
steps,
device,
first=False):
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.VOC_sampling_rate,
h.hop_size, h.win_size, h.fmin,
h.fmax_for_loss)
# print(y_mel.shape, y_g_hat_mel.shape)
if y_g_hat_mel.size(2) > y_mel.size(2):
y_g_hat_mel = y_g_hat_mel[:, :, :y_mel.size(2)]
else:
y_mel = y_mel[:, :, :y_g_hat_mel.size(2)]
val_err_tot += F.l1_loss(y_mel, y_g_hat_mel).item()
if j <= 4:
if steps == 0 or first:
sw.add_audio('gt/y_{}'.format(j), y[0], steps,
h.VOC_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.VOC_sampling_rate)
y_hat_spec = mel_spectrogram(y_g_hat.squeeze(1), h.n_fft,
h.num_mels, h.VOC_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()
def get_mel(x, sr=None):
if h.VOC_sampling_rate != h.TTS_sampling_rate:
if sr != h.TTS_sampling_rate:
audio_resample = torchaudio.transforms.Resample(orig_freq=sr, new_freq=h.TTS_sampling_rate, resampling_method='sinc_interpolation')
x = audio_resample(x)
mel = mel_spectrogram(x, h.n_fft, h.num_mels, h.TTS_sampling_rate, h.hop_size, h.win_size, h.fmin, h.fmax)
mel = torch.nn.functional.interpolate(mel.unsqueeze(0), scale_factor=(1, int(h.VOC_sampling_rate/h.TTS_sampling_rate)), mode='bilinear', align_corners=True).squeeze(0)
else:
if sr != h.VOC_sampling_rate:
audio_resample = torchaudio.transforms.Resample(orig_freq=sr, new_freq=h.VOC_sampling_rate, resampling_method='sinc_interpolation')
x = audio_resample(x)
mel = mel_spectrogram(x, h.n_fft, h.num_mels, h.VOC_sampling_rate, h.hop_size, h.win_size, h.fmin, h.fmax)
return mel
def Pattern_Generate(path,
n_fft: int,
num_mels: int,
sample_rate: int,
hop_size: int,
win_size: int,
fmin: int,
fmax: int,
center: bool = False,
top_db=60):
try:
audio, _ = librosa.load(path, sr=sample_rate)
except Exception as e:
return None, None
audio = librosa.effects.trim(audio,
top_db=top_db,
frame_length=512,
hop_length=256)[0]
audio = librosa.util.normalize(audio) * 0.95
mel = mel_spectrogram(y=torch.from_numpy(audio).float().unsqueeze(0),
n_fft=n_fft,
num_mels=num_mels,
sampling_rate=sample_rate,
hop_size=hop_size,
win_size=win_size,
fmin=fmin,
fmax=fmax,
center=center).squeeze(0).T.numpy()
return audio, mel
def Pattern_Generate(path,
n_fft: int,
num_mels: int,
sample_rate: int,
hop_size: int,
win_size: int,
fmin: int,
fmax: int,
center: bool = False,
top_db=60):
audio, _ = librosa.load(path, sr=sample_rate)
audio = librosa.effects.trim(audio,
top_db=top_db,
frame_length=512,
hop_length=256)[0]
audio = librosa.util.normalize(audio) * 0.95
audio = audio[:audio.shape[0] - (audio.shape[0] % hop_size)]
spect = spectrogram(y=torch.from_numpy(audio).float().unsqueeze(0),
n_fft=n_fft,
hop_size=hop_size,
win_size=win_size,
center=center).squeeze(0).T.numpy()
mel = mel_spectrogram(y=torch.from_numpy(audio).float().unsqueeze(0),
n_fft=n_fft,
num_mels=num_mels,
sampling_rate=sample_rate,
hop_size=hop_size,
win_size=win_size,
fmin=fmin,
fmax=fmax,
center=center).squeeze(0).T.numpy()
log_f0 = np.log(rapt(
x=audio * 32768,
fs=sample_rate,
hopsize=hop_size,
))
if log_f0.shape[0] != mel.shape[0]:
print(path, audio.shape[0], log_f0.shape[0], mel.shape[0])
return audio, spect, mel, log_f0
def get_mel(x):
return mel_spectrogram(x, h.n_fft, h.num_mels, h.sampling_rate, h.hop_size,
h.win_size, h.fmin, h.fmax)
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 fit(a, epochs):
if h.num_gpus > 1:
init_distributed(a.rank, h.num_gpus, a.group_name, h.dist_config['dist_backend'], h.dist_config['dist_url'])
generator = Generator().to(device)
discriminator = MultiScaleDiscriminator().to(device)
if h.num_gpus > 1:
generator = apply_gradient_allreduce(generator)
discriminator = apply_gradient_allreduce(discriminator)
g_optim = torch.optim.Adam(generator.parameters(), h.learning_rate, betas=[h.adam_b1, h.adam_b2])
d_optim = torch.optim.Adam(discriminator.parameters(), h.learning_rate, betas=[h.adam_b1, h.adam_b2])
steps = 0
if a.cp_g != "" and a.cp_d != "":
generator, g_optim, steps = load_checkpoint(a.cp_g, generator, g_optim)
discriminator, d_optim, steps = load_checkpoint(a.cp_d, discriminator, d_optim)
steps += 1
with open(a.input_train_metafile, 'r', encoding='utf-8') as fi:
training_files = [os.path.join(a.input_wavs_dir, x.split('|')[0] + '.wav')
for x in fi.read().split('\n') if len(x) > 0]
with open(a.input_valid_metafile, 'r', encoding='utf-8') as fi:
validation_files = [os.path.join(a.input_wavs_dir, x.split('|')[0] + '.wav')
for x in fi.read().split('\n') if len(x) > 0]
trainset = MelDataset(training_files, h.segment_size, h.n_fft, h.num_mels,
h.hop_size, h.win_size, h.sampling_rate, 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=False,
drop_last=True)
if a.rank == 0:
validset = MelDataset(validation_files, 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)
valid_loader = DataLoader(validset, num_workers=1, shuffle=False,
sampler=None,
batch_size=1,
pin_memory=False,
drop_last=True)
if a.rank == 0:
os.makedirs(a.cps, exist_ok=True)
print("checkpoints directory : ", a.cps)
sw = SummaryWriter(os.path.join(a.cps, 'logs'))
epoch_offset = max(0, int(steps / len(train_loader)))
generator.train()
discriminator.train()
for epoch in range(epoch_offset, epochs):
start = time.time()
if a.rank == 0:
print("Epoch: {}".format(epoch))
for i, batch in enumerate(train_loader):
start_b = time.time()
x, y, _ = batch
x = torch.autograd.Variable(x.to(device))
y = torch.autograd.Variable(y.to(device))
y = y.unsqueeze(1)
g_optim.zero_grad()
y_ghat = generator(x)
y_dhat_r, y_dhat_g, fmap_r, fmap_g = discriminator(y, y_ghat)
loss_fm = feature_loss(fmap_r, fmap_g)
loss_gen = generator_loss(y_dhat_g) + loss_fm
if h.num_gpus > 1:
reduced_loss_gen = reduce_tensor(loss_gen.data, h.num_gpus).item()
else:
reduced_loss_gen = loss_gen.item()
loss_gen.backward()
g_optim.step()
d_optim.zero_grad()
y_ghat = y_ghat.detach()
y_dhat_r, y_dhat_g, _, _ = discriminator(y, y_ghat)
loss_disc, losses_disc_r, losses_disc_g = discriminator_loss(y_dhat_r, y_dhat_g)
if h.num_gpus > 1:
reduced_loss_disc = reduce_tensor(loss_disc.data, h.num_gpus).item()
else:
reduced_loss_disc = loss_disc.item()
loss_disc.backward()
d_optim.step()
if a.rank == 0 and steps % a.stdout_interval == 0:
print('Steps : {:d}, Gen Loss : {:4.3f}, Disc Loss : {:4.3f}, s/b : {:4.3f}'.
format(steps, reduced_loss_gen, reduced_loss_disc, time.time() - start_b))
if a.rank == 0 and steps % a.checkpoint_interval == 0 and steps != 0:
checkpoint_path = "{}/g_{:08d}".format(a.cps, steps)
save_checkpoint(generator, g_optim, h.learning_rate, steps, checkpoint_path)
checkpoint_path = "{}/d_{:08d}".format(a.cps, steps)
save_checkpoint(discriminator, d_optim, h.learning_rate, steps, checkpoint_path)
if a.rank == 0 and steps % a.summary_interval == 0:
sw.add_scalar("training/gen_loss", reduced_loss_gen, steps)
sw.add_scalar("training/disc_loss", reduced_loss_disc, steps)
for i, (r, g) in enumerate(zip(losses_disc_r, losses_disc_g)):
sw.add_scalar("training/disc{:d}_loss_r".format(i+1), r, steps)
sw.add_scalar("training/disc{:d}_loss_g".format(i+1), g, steps)
for i, (r, g) in enumerate(zip(y_dhat_r, y_dhat_g)):
sw.add_histogram("training/disc{:d}_r_output".format(i+1), r, steps)
sw.add_histogram("training/disc{:d}_g_output".format(i+1), g, steps)
sw.add_histogram("training/gen_output", y_ghat, steps)
sw.add_audio('training_gt/y', y[0], steps, h.sampling_rate)
sw.add_audio('training_predicted/y_hat', y_ghat[0], steps, h.sampling_rate)
if a.rank == 0 and steps % a.validation_interval == 0: # and steps != 0:
for i, batch in enumerate(valid_loader):
x, y, _ = batch
y_ghat = generator(x.to(device))
sw.add_audio('validation_gt/y_{}'.format(i), y[0], steps, h.sampling_rate)
sw.add_audio('validation_predicted/y_hat_{}'.format(i), y_ghat[0], steps, h.sampling_rate)
# print(plot_spectrogram(x[i]))
sw.add_figure('validation_gt/y_spec_{}'.format(i), plot_spectrogram(x[0]), steps)
y_hat_spec = mel_spectrogram(y_ghat.detach().cpu().numpy()[0][0], h.n_fft, h.num_mels, h.sampling_rate, h.hop_size, h.win_size,
h.fmin, h.fmax, center=False)
sw.add_figure('validation_predicted/y_hat_spec_{}'.format(i), plot_spectrogram(y_hat_spec), steps)
if i == 4:
break
steps += 1
if a.rank == 0:
print('Time taken for epoch {} is {} sec\n'.format(epoch + 1, int(time.time()-start)))
def eval(rank, a, h, speaker_mapping=None):
torch.cuda.manual_seed(h.seed)
device = torch.device('cuda:{:d}'.format(rank))
generator = Generator(h).to(device)
mpd = MultiPeriodDiscriminator().to(device)
msd = MultiScaleDiscriminator().to(device)
with open(a.input_validation_file, 'r', encoding='utf-8') as fi:
validation_filelist = [os.path.join(a.input_wavs_dir, x.split('|')[0] + '.wav')
for x in fi.read().split('\n') if len(x) > 0]
'''validset = MelDataset(validation_filelist, h.segment_size, h.n_fft, h.num_mels,
h.hop_size, h.win_size, h.VOC_sampling_rate, h.TTS_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=a.batch_size,
pin_memory=True,
drop_last=True)'''
validset = MelDataset(validation_filelist, h.segment_size, h.n_fft, h.num_mels,
h.hop_size, h.win_size, h.VOC_sampling_rate, h.TTS_sampling_rate, h.fmin, h.fmax, n_cache_reuse=0,
shuffle=False, 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=h.num_workers, shuffle=False,
sampler=None,
batch_size=h.batch_size,
pin_memory=True,
drop_last=True)
min_loss = 999999
best_checkpoint = ''
# list all checkpoints
cp_list = glob.glob(os.path.join(a.checkpoint_path, 'g_*'))
for checkpoint_g in cp_list:
generator = Generator(h).to(device)
state_dict_g = load_checkpoint(checkpoint_g, device)
generator.load_state_dict(state_dict_g['generator'])
generator.eval()
total_checkpoint_loss = 0
torch.cuda.empty_cache()
with torch.no_grad():
for i, 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.VOC_sampling_rate,
h.hop_size, h.win_size,
h.fmin, h.fmax_for_loss)
# print(y_mel.shape, y_g_hat_mel.shape)
'''if y_g_hat_mel.size(2) > y_mel.size(2):
y_g_hat_mel = y_g_hat_mel[:, :, :y_mel.size(2)]
else:
y_mel = y_mel[:, :, :y_g_hat_mel.size(2)]'''
total_checkpoint_loss += F.l1_loss(y_mel, y_g_hat_mel).item()
total_checkpoint_loss = total_checkpoint_loss/len(validset)
print("Checkpoint", checkpoint_g, "L1 Loss:", total_checkpoint_loss)
if total_checkpoint_loss < min_loss:
print("Partial BEST Checkpoint: ", best_checkpoint, "Checkpoint Loss:", total_checkpoint_loss, "Old loss:", min_loss)
min_loss = total_checkpoint_loss
best_checkpoint = checkpoint_g
print("BEST Final Checkpoint", best_checkpoint, "L1 loss:", min_loss)
# cp best checkpoint
copyfile(best_checkpoint, best_checkpoint.replace("/g_", "/g_best_cv_"))
do_best_checkpoint = best_checkpoint.replace('/g_', '/do_')
copyfile(do_best_checkpoint, do_best_checkpoint.replace("/do_", "/do_best_cv_"))
def train(rank, a, h, resume_run_id=None):
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().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)
cp_g = None
cp_do = None
if resume_run_id:
restored_g = wandb.restore("g_latest")
cp_g = restored_g.name
restored_do = wandb.restore("do_latest")
cp_do = restored_do.name
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,
)
print(f"train dataset size:{len(trainset)}")
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,
)
print(f"valid dataset size:{len(validset)}")
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))
wandb.log(
{
"loss/Gen Loss Total": loss_gen_all,
"loss/Mel-Spec. Error": mel_error,
},
step=steps,
)
# checkpointing
if steps % a.checkpoint_interval == 0 and steps != 0:
# generator
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_name = "g_{:08d}".format(steps)
wandb.save(checkpoint_name)
# also save as latest
checkpoint_path = "{}/g_latest".format(a.checkpoint_path)
save_checkpoint(
checkpoint_path,
{
"generator": (generator.module if h.num_gpus > 1
else generator).state_dict()
},
)
wandb.save("g_latest")
# discriminator
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,
},
)
checkpoint_name = "do_{:08d}".format(steps)
wandb.save(checkpoint_name)
# also save as latest
checkpoint_path = "{}/do_latest".format(a.checkpoint_path)
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,
},
)
wandb.save("do_latest")
# 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():
samples_orig = []
samples_pred = []
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,
)
# log orig audio to wandb
orig_audio = y.squeeze().cpu()
samples_orig.append(
wandb.Audio(
orig_audio,
caption=f"sample {i}",
sample_rate=h.sampling_rate,
))
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,
)
# log pred audio to wandb
pred_audio = y_g_hat.squeeze().cpu()
samples_pred.append(
wandb.Audio(
pred_audio,
caption=f"sample {i}",
sample_rate=h.sampling_rate,
))
val_err = val_err_tot / (j + 1)
sw.add_scalar("validation/mel_spec_error", val_err,
steps)
# log audios to wandb
wandb.log(
{
"audio/generated": samples_pred,
},
step=steps,
)
if steps == 0:
wandb.log(
{
"audio/original": samples_orig,
},
step=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)))
