def get_collate_function(model_name, n_frames_per_step):
if model_name == 'Tacotron2':
collate_fn = TextMelCollate(n_frames_per_step)
elif model_name == 'WaveGlow':
collate_fn = torch.utils.data.dataloader.default_collate
else:
raise NotImplementedError(
"unknown collate function requested: {}".format(model_name))
return collate_fn
def main():
parser = argparse.ArgumentParser(
description='PyTorch TTS Data Pre-processing')
parser = parse_args(parser)
args, unk_args = parser.parse_known_args()
if len(unk_args) > 0:
raise ValueError(f'Invalid options {unk_args}')
if args.extract_pitch_char:
assert args.extract_durations, "Durations required for pitch extraction"
DLLogger.init(backends=[
JSONStreamBackend(Verbosity.DEFAULT, args.log_file),
StdOutBackend(Verbosity.VERBOSE)
])
for k, v in vars(args).items():
DLLogger.log(step="PARAMETER", data={k: v})
model = load_and_setup_model(
'Tacotron2',
parser,
args.tacotron2_checkpoint,
amp=False,
device=torch.device('cuda' if args.cuda else 'cpu'),
forward_is_infer=False,
ema=False)
if args.train_mode:
model.train()
# n_mel_channels arg has been consumed by model's arg parser
args.n_mel_channels = model.n_mel_channels
for datum in ('mels', 'mels_teacher', 'attentions', 'durations',
'pitch_mel', 'pitch_char', 'pitch_trichar'):
if getattr(args, f'extract_{datum}'):
Path(args.dataset_path, datum).mkdir(parents=False, exist_ok=True)
filenames = [
Path(l.split('|')[0]).stem for l in open(args.wav_text_filelist, 'r')
]
# Compatibility with Tacotron2 Data loader
args.n_speakers = 1
dataset = FilenamedLoader(filenames,
args.dataset_path,
args.wav_text_filelist,
args,
load_mel_from_disk=False)
# TextMelCollate supports only n_frames_per_step=1
data_loader = DataLoader(dataset,
batch_size=args.batch_size,
shuffle=False,
sampler=None,
num_workers=0,
collate_fn=TextMelCollate(1),
pin_memory=False,
drop_last=False)
pitch_vecs = {'mel': {}, 'char': {}, 'trichar': {}}
for i, batch in enumerate(data_loader):
tik = time.time()
fnames = batch[-1]
x, _, _ = batch_to_gpu(batch[:-1])
_, text_lens, mels_padded, _, mel_lens = x
for j, mel in enumerate(mels_padded):
fpath = Path(args.dataset_path, 'mels', fnames[j] + '.pt')
torch.save(mel[:, :mel_lens[j]].cpu(), fpath)
with torch.no_grad():
out_mels, out_mels_postnet, _, alignments = model.forward(x)
if args.extract_mels_teacher:
for j, mel in enumerate(out_mels_postnet):
fpath = Path(args.dataset_path, 'mels_teacher',
fnames[j] + '.pt')
torch.save(mel[:, :mel_lens[j]].cpu(), fpath)
if args.extract_attentions:
for j, ali in enumerate(alignments):
ali = ali[:mel_lens[j], :text_lens[j]]
fpath = Path(args.dataset_path, 'attentions',
fnames[j] + '.pt')
torch.save(ali.cpu(), fpath)
durations = []
if args.extract_durations:
for j, ali in enumerate(alignments):
text_len = text_lens[j]
ali = ali[:mel_lens[j], :text_len]
dur = torch.histc(torch.argmax(ali, dim=1),
min=0,
max=text_len - 1,
bins=text_len)
durations.append(dur)
fpath = Path(args.dataset_path, 'durations', fnames[j] + '.pt')
torch.save(dur.cpu().int(), fpath)
if args.extract_pitch_mel or args.extract_pitch_char or args.extract_pitch_trichar:
for j, dur in enumerate(durations):
fpath = Path(args.dataset_path, 'pitch_char',
fnames[j] + '.pt')
wav = Path(args.dataset_path, 'wavs', fnames[j] + '.wav')
p_mel, p_char, p_trichar = calculate_pitch(
str(wav),
dur.cpu().numpy())
pitch_vecs['mel'][fnames[j]] = p_mel
pitch_vecs['char'][fnames[j]] = p_char
pitch_vecs['trichar'][fnames[j]] = p_trichar
nseconds = time.time() - tik
DLLogger.log(step=f'{i+1}/{len(data_loader)} ({nseconds:.2f}s)',
data={})
if args.extract_pitch_mel:
normalize_pitch_vectors(pitch_vecs['mel'])
for fname, pitch in pitch_vecs['mel'].items():
fpath = Path(args.dataset_path, 'pitch_mel', fname + '.pt')
torch.save(torch.from_numpy(pitch), fpath)
if args.extract_pitch_char:
mean, std = normalize_pitch_vectors(pitch_vecs['char'])
for fname, pitch in pitch_vecs['char'].items():
fpath = Path(args.dataset_path, 'pitch_char', fname + '.pt')
torch.save(torch.from_numpy(pitch), fpath)
save_stats(args.dataset_path, args.wav_text_filelist, 'pitch_char',
mean, std)
if args.extract_pitch_trichar:
normalize_pitch_vectors(pitch_vecs['trichar'])
for fname, pitch in pitch_vecs['trichar'].items():
fpath = Path(args.dataset_path, 'pitch_trichar', fname + '.pt')
torch.save(torch.from_numpy(pitch), fpath)
DLLogger.flush()
def main():
parser = argparse.ArgumentParser(description='PyTorch Tacotron 2 Training')
parser = parse_args(parser)
args, _ = parser.parse_known_args()
LOGGER.set_model_name("Tacotron2_PyT")
LOGGER.set_backends([
dllg.StdOutBackend(log_file=None,
logging_scope=dllg.TRAIN_ITER_SCOPE,
iteration_interval=1),
dllg.JsonBackend(log_file=os.path.join(
args.output_directory, args.log_file) if args.rank == 0 else None,
logging_scope=dllg.TRAIN_ITER_SCOPE,
iteration_interval=1)
])
LOGGER.timed_block_start("run")
LOGGER.register_metric(tags.TRAIN_ITERATION_LOSS,
metric_scope=dllg.TRAIN_ITER_SCOPE)
LOGGER.register_metric("iter_time", metric_scope=dllg.TRAIN_ITER_SCOPE)
LOGGER.register_metric("epoch_time", metric_scope=dllg.EPOCH_SCOPE)
LOGGER.register_metric("run_time", metric_scope=dllg.RUN_SCOPE)
LOGGER.register_metric("val_iter_loss", metric_scope=dllg.EPOCH_SCOPE)
LOGGER.register_metric("train_epoch_frames/sec",
metric_scope=dllg.EPOCH_SCOPE)
LOGGER.register_metric("train_epoch_avg_frames/sec",
metric_scope=dllg.EPOCH_SCOPE)
LOGGER.register_metric("train_epoch_avg_loss",
metric_scope=dllg.EPOCH_SCOPE)
log_hardware()
parser = parse_tacotron2_args(parser)
args = parser.parse_args()
log_args(args)
torch.backends.cudnn.enabled = args.cudnn_enabled
torch.backends.cudnn.benchmark = args.cudnn_benchmark
distributed_run = args.world_size > 1
if distributed_run:
init_distributed(args, args.world_size, args.rank, args.group_name)
os.makedirs(args.output_directory, exist_ok=True)
LOGGER.log(key=tags.RUN_START)
run_start_time = time.time()
model = get_tacotron2_model(args,
len(args.training_anchor_dirs),
is_training=True)
if not args.amp_run and distributed_run:
model = DDP(model)
model.restore_checkpoint(
os.path.join(args.output_directory, args.latest_checkpoint_file))
optimizer = torch.optim.Adam(model.parameters(),
lr=args.init_lr,
weight_decay=args.weight_decay)
if args.amp_run:
model, optimizer = amp.initialize(model, optimizer, opt_level="O1")
if distributed_run:
model = DDP(model)
criterion = Tacotron2Loss()
collate_fn = TextMelCollate(args)
train_dataset = TextMelDataset(args, args.training_anchor_dirs)
train_loader = DataLoader(train_dataset,
num_workers=2,
shuffle=False,
batch_size=args.batch_size //
len(args.training_anchor_dirs),
pin_memory=False,
drop_last=True,
collate_fn=collate_fn)
# valate_dataset = TextMelDataset(args, args.validation_anchor_dirs)
model.train()
elapsed_epochs = model.get_elapsed_epochs()
epochs = args.epochs - elapsed_epochs
iteration = elapsed_epochs * len(train_loader)
LOGGER.log(key=tags.TRAIN_LOOP)
for epoch in range(1, epochs + 1):
LOGGER.epoch_start()
epoch_start_time = time.time()
epoch += elapsed_epochs
LOGGER.log(key=tags.TRAIN_EPOCH_START, value=epoch)
# used to calculate avg frames/sec over epoch
reduced_num_frames_epoch = 0
# used to calculate avg loss over epoch
train_epoch_avg_loss = 0.0
train_epoch_avg_frames_per_sec = 0.0
num_iters = 0
adjust_learning_rate(optimizer, epoch, args)
for i, batch in enumerate(train_loader):
print(f"Batch: {i}/{len(train_loader)} epoch {epoch}")
LOGGER.iteration_start()
iter_start_time = time.time()
LOGGER.log(key=tags.TRAIN_ITER_START, value=i)
# start = time.perf_counter()
optimizer.zero_grad()
x, y, num_frames = batch_to_gpu(batch)
y_pred = model(x)
loss = criterion(y_pred, y)
if distributed_run:
reduced_loss = reduce_tensor(loss.data, args.world_size).item()
reduced_num_frames = reduce_tensor(num_frames.data, 1).item()
else:
reduced_loss = loss.item()
reduced_num_frames = num_frames.item()
if np.isnan(reduced_loss):
raise Exception("loss is NaN")
LOGGER.log(key=tags.TRAIN_ITERATION_LOSS, value=reduced_loss)
train_epoch_avg_loss += reduced_loss
num_iters += 1
# accumulate number of frames processed in this epoch
reduced_num_frames_epoch += reduced_num_frames
if args.amp_run:
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
grad_norm = torch.nn.utils.clip_grad_norm_(
amp.master_params(optimizer), args.grad_clip_thresh)
else:
loss.backward()
grad_norm = torch.nn.utils.clip_grad_norm_(
model.parameters(), args.grad_clip_thresh)
optimizer.step()
iteration += 1
LOGGER.log(key=tags.TRAIN_ITER_STOP, value=i)
iter_stop_time = time.time()
iter_time = iter_stop_time - iter_start_time
frames_per_sec = reduced_num_frames / iter_time
train_epoch_avg_frames_per_sec += frames_per_sec
LOGGER.log(key="train_iter_frames/sec", value=frames_per_sec)
LOGGER.log(key="iter_time", value=iter_time)
LOGGER.iteration_stop()
LOGGER.log(key=tags.TRAIN_EPOCH_STOP, value=epoch)
epoch_stop_time = time.time()
epoch_time = epoch_stop_time - epoch_start_time
LOGGER.log(key="train_epoch_frames/sec",
value=(reduced_num_frames_epoch / epoch_time))
LOGGER.log(key="train_epoch_avg_frames/sec",
value=(train_epoch_avg_frames_per_sec /
num_iters if num_iters > 0 else 0.0))
LOGGER.log(key="train_epoch_avg_loss",
value=(train_epoch_avg_loss /
num_iters if num_iters > 0 else 0.0))
LOGGER.log(key="epoch_time", value=epoch_time)
LOGGER.log(key=tags.EVAL_START, value=epoch)
# validate(model, criterion, valate_dataset, iteration, collate_fn, distributed_run, args)
LOGGER.log(key=tags.EVAL_STOP, value=epoch)
# Store latest checkpoint in each epoch
model.elapse_epoch()
checkpoint_path = os.path.join(args.output_directory,
args.latest_checkpoint_file)
torch.save(model.state_dict(), checkpoint_path)
# Plot alignemnt
if epoch % args.epochs_per_alignment == 0 and args.rank == 0:
alignments = y_pred[3].data.numpy()
index = np.random.randint(len(alignments))
plot_alignment(
alignments[index].transpose(0, 1), # [enc_step, dec_step]
os.path.join(args.output_directory,
f"align_{epoch:04d}_{iteration}.png"),
info=
f"{datetime.now().strftime('%Y-%m-%d %H:%M')} Epoch={epoch:04d} Iteration={iteration} Average loss={train_epoch_avg_loss/num_iters:.5f}"
)
# Save checkpoint
if epoch % args.epochs_per_checkpoint == 0 and args.rank == 0:
checkpoint_path = os.path.join(args.output_directory,
f"checkpoint_{epoch:04d}.pt")
print(
f"Saving model and optimizer state at epoch {epoch:04d} to {checkpoint_path}"
)
torch.save(model.state_dict(), checkpoint_path)
# Save evaluation
# save_sample(model, args.tacotron2_checkpoint, args.phrase_path,
# os.path.join(args.output_directory, f"sample_{epoch:04d}_{iteration}.wav"), args.sampling_rate)
LOGGER.epoch_stop()
run_stop_time = time.time()
run_time = run_stop_time - run_start_time
LOGGER.log(key="run_time", value=run_time)
LOGGER.log(key=tags.RUN_FINAL)
print("training time", run_stop_time - run_start_time)
LOGGER.timed_block_stop("run")
if args.rank == 0:
LOGGER.finish()
def get_collate_function(model_name):
return {
'Tacotron2': lambda _: TextMelCollate(n_frames_per_step=1),
'WaveGlow': lambda _: torch.utils.data.dataloader.default_collate,
'FastPitch': TextMelAliCollate
}[model_name]()