def main():
parser = argparse.ArgumentParser(description='PyTorch Tacotron 2 Training')
parser = parse_training_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_dir, 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_dir, 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_dir, args.latest_checkpoint_file))
optimizer = torch.optim.Adam(model.parameters(), lr=args.init_lr, weight_decay=args.weight_decay)
writer = SummaryWriter(args.output_dir)
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)
outputs = model(x)
y_pred = [output.cpu() for output in outputs]
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
writer.add_scalar('Training/Loss', reduced_loss, iteration)
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_dir, args.latest_checkpoint_file)
model.save_checkpoint(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], # [enc_step, dec_step]
os.path.join(args.output_dir, 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_dir, f"checkpoint_{epoch:04d}.pt")
print(f"Saving model and optimizer state at epoch {epoch:04d} to {checkpoint_path}")
model.save_checkpoint(checkpoint_path)
# Save evaluation
# save_sample(model, args.tacotron2_checkpoint, args.phrase_path,
# os.path.join(args.output_dir, 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)
writer.close()
LOGGER.timed_block_stop("run")
if args.rank == 0:
LOGGER.finish()
