def test_train_step():
input_dummy = torch.randint(0, 24, (8, 128)).long().to(device)
input_lengths = torch.randint(100, 129, (8, )).long().to(device)
input_lengths[-1] = 128
mel_spec = torch.rand(8, 120, c.audio['num_mels']).to(device)
linear_spec = torch.rand(8, 120, c.audio['num_freq']).to(device)
mel_lengths = torch.randint(20, 120, (8, )).long().to(device)
mel_lengths[-1] = 120
stop_targets = torch.zeros(8, 120, 1).float().to(device)
speaker_ids = torch.randint(0, 5, (8, )).long().to(device)
for idx in mel_lengths:
stop_targets[:, int(idx.item()):, 0] = 1.0
stop_targets = stop_targets.view(input_dummy.shape[0],
stop_targets.size(1) // c.r, -1)
stop_targets = (stop_targets.sum(2) >
0.0).unsqueeze(2).float().squeeze()
criterion = L1LossMasked(seq_len_norm=False).to(device)
criterion_st = nn.BCEWithLogitsLoss().to(device)
model = Tacotron(
num_chars=32,
num_speakers=5,
gst=True,
postnet_output_dim=c.audio['num_freq'],
decoder_output_dim=c.audio['num_mels'],
r=c.r,
memory_size=c.memory_size
).to(device) #FIXME: missing num_speakers parameter to Tacotron ctor
model.train()
print(model)
print(" > Num parameters for Tacotron GST model:%s" %
(count_parameters(model)))
model_ref = copy.deepcopy(model)
count = 0
for param, param_ref in zip(model.parameters(),
model_ref.parameters()):
assert (param - param_ref).sum() == 0, param
count += 1
optimizer = optim.Adam(model.parameters(), lr=c.lr)
for _ in range(10):
mel_out, linear_out, align, stop_tokens = model.forward(
input_dummy, input_lengths, mel_spec, mel_lengths, speaker_ids)
optimizer.zero_grad()
loss = criterion(mel_out, mel_spec, mel_lengths)
stop_loss = criterion_st(stop_tokens, stop_targets)
loss = loss + criterion(linear_out, linear_spec,
mel_lengths) + stop_loss
loss.backward()
optimizer.step()
# check parameter changes
count = 0
for param, param_ref in zip(model.parameters(),
model_ref.parameters()):
# ignore pre-higway layer since it works conditional
assert (param != param_ref).any(
), "param {} with shape {} not updated!! \n{}\n{}".format(
count, param.shape, param, param_ref)
count += 1
def test_in_out(self):
# test input == target
layer = L1LossMasked()
dummy_input = T.ones(4, 8, 128).float()
dummy_target = T.ones(4, 8, 128).float()
dummy_length = (T.ones(4) * 8).long()
output = layer(dummy_input, dummy_target, dummy_length)
assert output.item() == 0.0
# test input != target
dummy_input = T.ones(4, 8, 128).float()
dummy_target = T.zeros(4, 8, 128).float()
dummy_length = (T.ones(4) * 8).long()
output = layer(dummy_input, dummy_target, dummy_length)
assert output.item() == 1.0, "1.0 vs {}".format(output.data[0])
# test if padded values of input makes any difference
dummy_input = T.ones(4, 8, 128).float()
dummy_target = T.zeros(4, 8, 128).float()
dummy_length = (T.arange(5, 9)).long()
mask = ((sequence_mask(dummy_length).float() - 1.0) *
100.0).unsqueeze(2)
output = layer(dummy_input + mask, dummy_target, dummy_length)
assert output.item() == 1.0, "1.0 vs {}".format(output.data[0])
dummy_input = T.rand(4, 8, 128).float()
dummy_target = dummy_input.detach()
dummy_length = (T.arange(5, 9)).long()
mask = ((sequence_mask(dummy_length).float() - 1.0) *
100.0).unsqueeze(2)
output = layer(dummy_input + mask, dummy_target, dummy_length)
assert output.item() == 0, "0 vs {}".format(output.data[0])
def test_train_step(self):
input = torch.randint(0, 24, (8, 128)).long().to(device)
mel_spec = torch.rand(8, 30, c.num_mels).to(device)
linear_spec = torch.rand(8, 30, c.num_freq).to(device)
mel_lengths = torch.randint(20, 30, (8, )).long().to(device)
criterion = L1LossMasked().to(device)
model = Tacotron(c.embedding_size, c.num_freq, c.num_mels,
c.r).to(device)
model.train()
model_ref = copy.deepcopy(model)
count = 0
for param, param_ref in zip(model.parameters(),
model_ref.parameters()):
assert (param - param_ref).sum() == 0, param
count += 1
optimizer = optim.Adam(model.parameters(), lr=c.lr)
for i in range(5):
mel_out, linear_out, align = model.forward(input, mel_spec)
optimizer.zero_grad()
loss = criterion(mel_out, mel_spec, mel_lengths)
loss = 0.5 * loss + 0.5 * criterion(linear_out, linear_spec,
mel_lengths)
loss.backward()
optimizer.step()
# check parameter changes
count = 0
for param, param_ref in zip(model.parameters(),
model_ref.parameters()):
# ignore pre-higway layer since it works conditional
if count not in [139, 59]:
assert (param != param_ref).any(
), "param {} with shape {} not updated!! \n{}\n{}".format(
count, param.shape, param, param_ref)
count += 1
def test_train_step(self):
input = torch.randint(0, 24, (8, 128)).long().to(device)
mel_spec = torch.rand(8, 30, c.num_mels).to(device)
linear_spec = torch.rand(8, 30, c.num_freq).to(device)
mel_lengths = torch.randint(20, 30, (8, )).long().to(device)
stop_targets = torch.zeros(8, 30, 1).float().to(device)
for idx in mel_lengths:
stop_targets[:, int(idx.item()):, 0] = 1.0
stop_targets = stop_targets.view(input.shape[0],
stop_targets.size(1) // c.r, -1)
stop_targets = (stop_targets.sum(2) > 0.0).unsqueeze(2).float()
criterion = L1LossMasked().to(device)
criterion_st = nn.BCELoss().to(device)
model = Tacotron(c.embedding_size, c.num_freq, c.num_mels,
c.r).to(device)
model.train()
model_ref = copy.deepcopy(model)
count = 0
for param, param_ref in zip(model.parameters(),
model_ref.parameters()):
assert (param - param_ref).sum() == 0, param
count += 1
optimizer = optim.Adam(model.parameters(), lr=c.lr)
for i in range(5):
mel_out, linear_out, align, stop_tokens = model.forward(
input, mel_spec)
assert stop_tokens.data.max() <= 1.0
assert stop_tokens.data.min() >= 0.0
optimizer.zero_grad()
loss = criterion(mel_out, mel_spec, mel_lengths)
stop_loss = criterion_st(stop_tokens, stop_targets)
loss = loss + criterion(linear_out, linear_spec,
mel_lengths) + stop_loss
loss.backward()
optimizer.step()
# check parameter changes
count = 0
for param, param_ref in zip(model.parameters(),
model_ref.parameters()):
# ignore pre-higway layer since it works conditional
if count not in [145, 59]:
assert (param != param_ref).any(
), "param {} with shape {} not updated!! \n{}\n{}".format(
count, param.shape, param, param_ref)
count += 1
def test_in_out(self):
layer = L1LossMasked()
dummy_input = T.ones(4, 8, 128).float()
dummy_target = T.ones(4, 8, 128).float()
dummy_length = (T.ones(4) * 8).long()
output = layer(dummy_input, dummy_target, dummy_length)
assert output.item() == 0.0
dummy_input = T.ones(4, 8, 128).float()
dummy_target = T.zeros(4, 8, 128).float()
dummy_length = (T.ones(4) * 8).long()
output = layer(dummy_input, dummy_target, dummy_length)
assert output.item() == 1.0, "1.0 vs {}".format(output.data[0])
dummy_input = T.ones(4, 8, 128).float()
dummy_target = T.zeros(4, 8, 128).float()
dummy_length = (T.arange(5, 9)).long()
mask = ((sequence_mask(dummy_length).float() - 1.0) *
100.0).unsqueeze(2)
output = layer(dummy_input + mask, dummy_target, dummy_length)
assert output.item() == 1.0, "1.0 vs {}".format(output.data[0])
def test_in_out(self):
layer = L1LossMasked()
dummy_input = T.autograd.Variable(T.ones(4, 8, 128).float())
dummy_target = T.autograd.Variable(T.ones(4, 8, 128).float())
dummy_length = T.autograd.Variable((T.ones(4) * 8).long())
output = layer(dummy_input, dummy_target, dummy_length)
assert output.shape[0] == 0
assert len(output.shape) == 1
assert output.data[0] == 0.0
dummy_input = T.autograd.Variable(T.ones(4, 8, 128).float())
dummy_target = T.autograd.Variable(T.zeros(4, 8, 128).float())
dummy_length = T.autograd.Variable((T.ones(4) * 8).long())
output = layer(dummy_input, dummy_target, dummy_length)
assert output.data[0] == 1.0, "1.0 vs {}".format(output.data[0])
dummy_input = T.autograd.Variable(T.ones(4, 8, 128).float())
dummy_target = T.autograd.Variable(T.zeros(4, 8, 128).float())
dummy_length = T.autograd.Variable((T.arange(5, 9)).long())
mask = ((_sequence_mask(dummy_length).float() - 1.0) *
100.0).unsqueeze(2)
output = layer(dummy_input + mask, dummy_target, dummy_length)
assert output.data[0] == 1.0, "1.0 vs {}".format(output.data[0])
def main(args): #pylint: disable=redefined-outer-name
# Audio processor
ap = AudioProcessor(**c.audio)
# DISTRUBUTED
if num_gpus > 1:
init_distributed(args.rank, num_gpus, args.group_id,
c.distributed["backend"], c.distributed["url"])
num_chars = len(phonemes) if c.use_phonemes else len(symbols)
if c.use_speaker_embedding:
speakers = get_speakers(c.data_path, c.meta_file_train, c.dataset)
if args.restore_path:
prev_out_path = os.path.dirname(args.restore_path)
speaker_mapping = load_speaker_mapping(prev_out_path)
assert all([speaker in speaker_mapping
for speaker in speakers]), "As of now you, you cannot " \
"introduce new speakers to " \
"a previously trained model."
else:
speaker_mapping = {name: i for i, name in enumerate(speakers)}
save_speaker_mapping(OUT_PATH, speaker_mapping)
num_speakers = len(speaker_mapping)
print("Training with {} speakers: {}".format(num_speakers,
", ".join(speakers)))
else:
num_speakers = 0
model = setup_model(num_chars, num_speakers, c)
print(" | > Num output units : {}".format(ap.num_freq), flush=True)
optimizer = RAdam(model.parameters(), lr=c.lr, weight_decay=0)
if c.stopnet and c.separate_stopnet:
optimizer_st = RAdam(model.decoder.stopnet.parameters(),
lr=c.lr,
weight_decay=0)
else:
optimizer_st = None
if c.loss_masking:
criterion = L1LossMasked() if c.model in ["Tacotron", "TacotronGST"
] else MSELossMasked()
else:
criterion = nn.L1Loss() if c.model in ["Tacotron", "TacotronGST"
] else nn.MSELoss()
criterion_st = nn.BCEWithLogitsLoss() if c.stopnet else None
if args.restore_path:
checkpoint = torch.load(args.restore_path)
try:
# TODO: fix optimizer init, model.cuda() needs to be called before
# optimizer restore
# optimizer.load_state_dict(checkpoint['optimizer'])
if c.reinit_layers:
raise RuntimeError
model.load_state_dict(checkpoint['model'])
except:
print(" > Partial model initialization.")
model_dict = model.state_dict()
model_dict = set_init_dict(model_dict, checkpoint, c)
model.load_state_dict(model_dict)
del model_dict
for group in optimizer.param_groups:
group['lr'] = c.lr
print(" > Model restored from step %d" % checkpoint['step'],
flush=True)
args.restore_step = checkpoint['step']
else:
args.restore_step = 0
if use_cuda:
model = model.cuda()
criterion.cuda()
if criterion_st:
criterion_st.cuda()
# DISTRUBUTED
if num_gpus > 1:
model = apply_gradient_allreduce(model)
if c.lr_decay:
scheduler = NoamLR(optimizer,
warmup_steps=c.warmup_steps,
last_epoch=args.restore_step - 1)
else:
scheduler = None
num_params = count_parameters(model)
print("\n > Model has {} parameters".format(num_params), flush=True)
if 'best_loss' not in locals():
best_loss = float('inf')
global_step = args.restore_step
for epoch in range(0, c.epochs):
# set gradual training
if c.gradual_training is not None:
r, c.batch_size = gradual_training_scheduler(global_step, c)
c.r = r
model.decoder.set_r(r)
print(" > Number of outputs per iteration:", model.decoder.r)
train_loss, global_step = train(model, criterion, criterion_st,
optimizer, optimizer_st, scheduler, ap,
global_step, epoch)
val_loss = evaluate(model, criterion, criterion_st, ap, global_step,
epoch)
print(" | > Training Loss: {:.5f} Validation Loss: {:.5f}".format(
train_loss, val_loss),
flush=True)
target_loss = train_loss
if c.run_eval:
target_loss = val_loss
best_loss = save_best_model(model, optimizer, target_loss, best_loss,
OUT_PATH, global_step, epoch)
model = setup_model(num_chars, num_speakers, C)
checkpoint = torch.load(MODEL_FILE)
model.load_state_dict(checkpoint['model'])
print(checkpoint['step'])
model.eval()
if use_cuda:
model = model.cuda()
# ### Generate model outputs
import pickle
file_idxs = []
losses = []
postnet_losses = []
criterion = L1LossMasked()
for data in tqdm(loader):
# setup input data
text_input = data[0]
text_lengths = data[1]
speaker_names = data[2]
linear_input = data[3] if C.model in ["Tacotron", "TacotronGST"] else None
mel_input = data[4]
mel_lengths = data[5]
stop_targets = data[6]
item_idx = data[7]
if C.use_speaker_embedding:
speaker_ids = [
speaker_mapping[speaker_name] for speaker_name in speaker_names
]