class Synthesizer(object):
def load_model(self, model_path, model_name, model_config, use_cuda):
model_config = os.path.join(model_path, model_config)
self.model_file = os.path.join(model_path, model_name)
print(" > Loading model ...")
print(" | > model config: ", model_config)
print(" | > model file: ", self.model_file)
config = load_config(model_config)
self.config = config
self.use_cuda = use_cuda
self.ap = AudioProcessor(**config.audio)
self.model = Tacotron(61, config.embedding_size, self.ap.num_freq,
self.ap.num_mels, config.r)
# load model state
if use_cuda:
cp = torch.load(self.model_file)
else:
cp = torch.load(self.model_file,
map_location=lambda storage, loc: storage)
# load the model
self.model.load_state_dict(cp['model'])
if use_cuda:
self.model.cuda()
self.model.eval()
def save_wav(self, wav, path):
# wav *= 32767 / max(1e-8, np.max(np.abs(wav)))
wav = np.array(wav)
self.ap.save_wav(wav, path)
def tts(self, text):
text_cleaner = [self.config.text_cleaner]
wavs = []
for sen in text.split('.'):
if len(sen) < 3:
continue
sen = sen.strip()
sen += '.'
print(sen)
sen = sen.strip()
seq = np.array(
phoneme_to_sequence(sen, text_cleaner,
self.config.phoneme_language))
chars_var = torch.from_numpy(seq).unsqueeze(0).long()
if self.use_cuda:
chars_var = chars_var.cuda()
mel_out, linear_out, alignments, stop_tokens = self.model.forward(
chars_var)
linear_out = linear_out[0].data.cpu().numpy()
wav = self.ap.inv_spectrogram(linear_out.T)
wavs += list(wav)
wavs += [0] * 10000
out = io.BytesIO()
self.save_wav(wavs, out)
return out
def tacotron(pretrained=True, **kwargs):
model = Tacotron(embed_dims=hp.tts_embed_dims,
num_chars=len(symbols),
encoder_dims=hp.tts_encoder_dims,
decoder_dims=hp.tts_decoder_dims,
n_mels=hp.num_mels,
fft_bins=hp.num_mels,
postnet_dims=hp.tts_postnet_dims,
encoder_K=hp.tts_encoder_K,
lstm_dims=hp.tts_lstm_dims,
postnet_K=hp.tts_postnet_K,
num_highways=hp.tts_num_highways,
dropout=hp.tts_dropout,
stop_threshold=hp.tts_stop_threshold)
if pretrained:
state_dict = fetch_and_load_state_dict("tacotron")
state_dict["decoder.r"] = state_dict.pop("r")
state_dict["stop_threshold"] = torch.tensor(hp.tts_stop_threshold,
dtype=torch.float32)
model.load_state_dict(state_dict)
return model
def main(args):
# setup output paths and read configs
c = load_config(args.config_path)
_ = os.path.dirname(os.path.realpath(__file__))
OUT_PATH = os.path.join(_, c.output_path)
OUT_PATH = create_experiment_folder(OUT_PATH)
CHECKPOINT_PATH = os.path.join(OUT_PATH, 'checkpoints')
shutil.copyfile(args.config_path, os.path.join(OUT_PATH, 'config.json'))
# save config to tmp place to be loaded by subsequent modules.
file_name = str(os.getpid())
tmp_path = os.path.join("/tmp/", file_name+'_tts')
pickle.dump(c, open(tmp_path, "wb"))
# setup tensorboard
LOG_DIR = OUT_PATH
tb = SummaryWriter(LOG_DIR)
# Ctrl+C handler to remove empty experiment folder
def signal_handler(signal, frame):
print(" !! Pressed Ctrl+C !!")
remove_experiment_folder(OUT_PATH)
sys.exit(1)
signal.signal(signal.SIGINT, signal_handler)
# Setup the dataset
dataset = LJSpeechDataset(os.path.join(c.data_path, 'metadata.csv'),
os.path.join(c.data_path, 'wavs'),
c.r,
c.sample_rate,
c.text_cleaner,
c.num_mels,
c.min_level_db,
c.frame_shift_ms,
c.frame_length_ms,
c.preemphasis,
c.ref_level_db,
c.num_freq,
c.power
)
dataloader = DataLoader(dataset, batch_size=c.batch_size,
shuffle=True, collate_fn=dataset.collate_fn,
drop_last=True, num_workers=c.num_loader_workers)
# setup the model
model = Tacotron(c.embedding_size,
c.hidden_size,
c.num_mels,
c.num_freq,
c.r)
# plot model on tensorboard
dummy_input = dataset.get_dummy_data()
## TODO: onnx does not support RNN fully yet
# model_proto_path = os.path.join(OUT_PATH, "model.proto")
# onnx.export(model, dummy_input, model_proto_path, verbose=True)
# tb.add_graph_onnx(model_proto_path)
if use_cuda:
model = nn.DataParallel(model.cuda())
optimizer = optim.Adam(model.parameters(), lr=c.lr)
if args.restore_step:
checkpoint = torch.load(os.path.join(
args.restore_path, 'checkpoint_%d.pth.tar' % args.restore_step))
model.load_state_dict(checkpoint['model'])
optimizer.load_state_dict(checkpoint['optimizer'])
print("\n > Model restored from step %d\n" % args.restore_step)
start_epoch = checkpoint['step'] // len(dataloader)
best_loss = checkpoint['linear_loss']
else:
start_epoch = 0
print("\n > Starting a new training")
num_params = count_parameters(model)
print(" | > Model has {} parameters".format(num_params))
model = model.train()
if not os.path.exists(CHECKPOINT_PATH):
os.mkdir(CHECKPOINT_PATH)
if use_cuda:
criterion = nn.L1Loss().cuda()
else:
criterion = nn.L1Loss()
n_priority_freq = int(3000 / (c.sample_rate * 0.5) * c.num_freq)
#lr_scheduler = ReduceLROnPlateau(optimizer, factor=c.lr_decay,
# patience=c.lr_patience, verbose=True)
epoch_time = 0
best_loss = float('inf')
for epoch in range(0, c.epochs):
print("\n | > Epoch {}/{}".format(epoch, c.epochs))
progbar = Progbar(len(dataset) / c.batch_size)
for num_iter, data in enumerate(dataloader):
start_time = time.time()
text_input = data[0]
text_lengths = data[1]
linear_input = data[2]
mel_input = data[3]
current_step = num_iter + args.restore_step + epoch * len(dataloader) + 1
# setup lr
current_lr = lr_decay(c.lr, current_step)
for params_group in optimizer.param_groups:
params_group['lr'] = current_lr
optimizer.zero_grad()
# Add a single frame of zeros to Mel Specs for better end detection
#try:
# mel_input = np.concatenate((np.zeros(
# [c.batch_size, 1, c.num_mels], dtype=np.float32),
# mel_input[:, 1:, :]), axis=1)
#except:
# raise TypeError("not same dimension")
# convert inputs to variables
text_input_var = Variable(text_input)
mel_spec_var = Variable(mel_input)
linear_spec_var = Variable(linear_input, volatile=True)
# sort sequence by length.
# TODO: might be unnecessary
sorted_lengths, indices = torch.sort(
text_lengths.view(-1), dim=0, descending=True)
sorted_lengths = sorted_lengths.long().numpy()
text_input_var = text_input_var[indices]
mel_spec_var = mel_spec_var[indices]
linear_spec_var = linear_spec_var[indices]
if use_cuda:
text_input_var = text_input_var.cuda()
mel_spec_var = mel_spec_var.cuda()
linear_spec_var = linear_spec_var.cuda()
mel_output, linear_output, alignments =\
model.forward(text_input_var, mel_spec_var,
input_lengths= torch.autograd.Variable(torch.cuda.LongTensor(sorted_lengths)))
mel_loss = criterion(mel_output, mel_spec_var)
#linear_loss = torch.abs(linear_output - linear_spec_var)
#linear_loss = 0.5 * \
#torch.mean(linear_loss) + 0.5 * \
#torch.mean(linear_loss[:, :n_priority_freq, :])
linear_loss = 0.5 * criterion(linear_output, linear_spec_var) \
+ 0.5 * criterion(linear_output[:, :, :n_priority_freq],
linear_spec_var[: ,: ,:n_priority_freq])
loss = mel_loss + linear_loss
# loss = loss.cuda()
loss.backward()
grad_norm = nn.utils.clip_grad_norm(model.parameters(), 1.) ## TODO: maybe no need
optimizer.step()
step_time = time.time() - start_time
epoch_time += step_time
progbar.update(num_iter+1, values=[('total_loss', loss.data[0]),
('linear_loss', linear_loss.data[0]),
('mel_loss', mel_loss.data[0]),
('grad_norm', grad_norm)])
# Plot Learning Stats
tb.add_scalar('Loss/TotalLoss', loss.data[0], current_step)
tb.add_scalar('Loss/LinearLoss', linear_loss.data[0],
current_step)
tb.add_scalar('Loss/MelLoss', mel_loss.data[0], current_step)
tb.add_scalar('Params/LearningRate', optimizer.param_groups[0]['lr'],
current_step)
tb.add_scalar('Params/GradNorm', grad_norm, current_step)
tb.add_scalar('Time/StepTime', step_time, current_step)
align_img = alignments[0].data.cpu().numpy()
align_img = plot_alignment(align_img)
tb.add_image('Attn/Alignment', align_img, current_step)
if current_step % c.save_step == 0:
if c.checkpoint:
# save model
save_checkpoint(model, optimizer, linear_loss.data[0],
OUT_PATH, current_step, epoch)
# Diagnostic visualizations
const_spec = linear_output[0].data.cpu().numpy()
gt_spec = linear_spec_var[0].data.cpu().numpy()
const_spec = plot_spectrogram(const_spec, dataset.ap)
gt_spec = plot_spectrogram(gt_spec, dataset.ap)
tb.add_image('Spec/Reconstruction', const_spec, current_step)
tb.add_image('Spec/GroundTruth', gt_spec, current_step)
align_img = alignments[0].data.cpu().numpy()
align_img = plot_alignment(align_img)
tb.add_image('Attn/Alignment', align_img, current_step)
# Sample audio
audio_signal = linear_output[0].data.cpu().numpy()
dataset.ap.griffin_lim_iters = 60
audio_signal = dataset.ap.inv_spectrogram(audio_signal.T)
try:
tb.add_audio('SampleAudio', audio_signal, current_step,
sample_rate=c.sample_rate)
except:
print("\n > Error at audio signal on TB!!")
print(audio_signal.max())
print(audio_signal.min())
# average loss after the epoch
avg_epoch_loss = np.mean(
progbar.sum_values['linear_loss'][0] / max(1, progbar.sum_values['linear_loss'][1]))
best_loss = save_best_model(model, optimizer, avg_epoch_loss,
best_loss, OUT_PATH,
current_step, epoch)
#lr_scheduler.step(loss.data[0])
tb.add_scalar('Time/EpochTime', epoch_time, epoch)
epoch_time = 0
def main(args):
# Setup the dataset
train_dataset = LJSpeechDataset(os.path.join(c.data_path, 'train_metadata.csv'),
os.path.join(c.data_path, 'wavs'),
c.r,
c.sample_rate,
c.text_cleaner,
c.num_mels,
c.min_level_db,
c.frame_shift_ms,
c.frame_length_ms,
c.preemphasis,
c.ref_level_db,
c.num_freq,
c.power,
min_seq_len=c.min_seq_len
)
train_loader = DataLoader(train_dataset, batch_size=c.batch_size,
shuffle=False, collate_fn=train_dataset.collate_fn,
drop_last=False, num_workers=c.num_loader_workers,
pin_memory=True)
'''
val_dataset = LJSpeechDataset(os.path.join(c.data_path, 'valid_metadata.csv'),
os.path.join(c.data_path, 'wavs'),
c.r,
c.sample_rate,
c.text_cleaner,
c.num_mels,
c.min_level_db,
c.frame_shift_ms,
c.frame_length_ms,
c.preemphasis,
c.ref_level_db,
c.num_freq,
c.power
)
val_loader = DataLoader(val_dataset, batch_size=c.eval_batch_size,
shuffle=False, collate_fn=val_dataset.collate_fn,
drop_last=False, num_workers=4,
pin_memory=True)
'''
model = Tacotron(c.embedding_size,
c.num_freq,
c.num_mels,
c.r)
optimizer = optim.Adam(model.parameters(), lr=c.lr)
optimizer_st = optim.Adam(model.decoder.stopnet.parameters(), lr=c.lr)
criterion = L1LossMasked()
criterion_st = nn.BCELoss()
if args.restore_path:
checkpoint = torch.load(args.restore_path)
model.load_state_dict(checkpoint['model'])
optimizer = optim.Adam(model.parameters(), lr=c.lr)
optimizer.load_state_dict(checkpoint['optimizer'])
for state in optimizer.state.values():
for k, v in state.items():
if torch.is_tensor(v):
state[k] = v.cuda()
print(" > Model restored from step %d" % checkpoint['step'])
start_epoch = checkpoint['step'] // len(train_loader)
best_loss = checkpoint['linear_loss']
start_epoch = 0
args.restore_step = checkpoint['step']
optimizer_st = optim.Adam(model.decoder.stopnet.parameters(), lr=c.lr)
else:
args.restore_step = 0
print("\n > Starting a new training")
if use_cuda:
model = nn.DataParallel(model.cuda())
criterion.cuda()
criterion_st.cuda()
num_params = count_parameters(model)
print(" | > Model has {} parameters".format(num_params))
if not os.path.exists(CHECKPOINT_PATH):
os.mkdir(CHECKPOINT_PATH)
if 'best_loss' not in locals():
best_loss = float('inf')
for epoch in range(0, c.epochs):
train_loss, current_step = train(
model, criterion, criterion_st, train_loader, optimizer, optimizer_st, epoch)
#val_loss = evaluate(model, criterion, criterion_st, val_loader, current_step)
best_loss = save_best_model(model, optimizer, train_loss, best_loss, OUT_PATH, current_step, epoch)
class Synthesizer(object):
def load_model(self, model_path, model_config, wavernn_path, use_cuda):
self.model_file = model_path
print(" > Loading model ...")
print(" | > model config: ", model_config)
print(" | > model file: ", self.model_file)
config = load_config(model_config)
self.config = config
self.use_cuda = use_cuda
self.use_phonemes = config.use_phonemes
self.ap = AudioProcessor(**config.audio)
if self.use_phonemes:
self.input_size = len(phonemes)
self.input_adapter = lambda sen: phoneme_to_sequence(sen, [self.config.text_cleaner], self.config.phoneme_language)
else:
self.input_size = len(symbols)
self.input_adapter = lambda sen: text_to_sequence(sen, [self.config.text_cleaner])
self.model = Tacotron(self.input_size, config.embedding_size, self.ap.num_freq, self.ap.num_mels, config.r, attn_windowing=True)
self.model.decoder.max_decoder_steps = 8000
# load model state
if use_cuda:
cp = torch.load(self.model_file)
else:
cp = torch.load(self.model_file, map_location=lambda storage, loc: storage)
# load the model
self.model.load_state_dict(cp['model'])
if use_cuda:
self.model.cuda()
self.model.eval()
self.vocoder=WaveRNNVocoder.Vocoder()
self.vocoder.loadWeights(wavernn_path)
self.firwin = signal.firwin(1025, [65, 7600], pass_zero=False, fs=16000)
def save_wav(self, wav, path):
# wav *= 32767 / max(1e-8, np.max(np.abs(wav)))
wav = np.array(wav)
self.ap.save_wav(wav, path)
#split text into chunks that are smaller than maxlen. Preferably, split on punctuation.
def ttmel(self, text):
mel_ret = []
text_list = split_text(text, maxlen)
for t in text_list:
if len(t) < 3:
continue
seq = np.array(self.input_adapter(t))
chars_var = torch.from_numpy(seq).unsqueeze(0).long()
if self.use_cuda:
chars_var = chars_var.cuda()
mel_out, _, alignments, stop_tokens = self.model.forward(chars_var)
mel_out = mel_out[0].data.cpu().numpy().T
mel_ret.append(mel_out)
return np.hstack(mel_ret)
def tts(self, mel):
wav = self.vocoder.melToWav(mel)
return wav
def main(args):
model = Tacotron(c.embedding_size, ap.num_freq, ap.num_mels, c.r)
print(" | > Num output units : {}".format(ap.num_freq), flush=True)
optimizer = optim.Adam(model.parameters(), lr=c.lr, weight_decay=0)
optimizer_st = optim.Adam(
model.decoder.stopnet.parameters(), lr=c.lr, weight_decay=0)
criterion = L1LossMasked()
criterion_st = nn.BCELoss()
if args.restore_path:
checkpoint = torch.load(args.restore_path)
try:
model.load_state_dict(checkpoint['model'])
except:
model_dict = model.state_dict()
# Partial initialization: if there is a mismatch with new and old layer, it is skipped.
# 1. filter out unnecessary keys
pretrained_dict = {
k: v
for k, v in checkpoint['model'].items() if k in model_dict
}
# 2. overwrite entries in the existing state dict
model_dict.update(pretrained_dict)
# 3. load the new state dict
model.load_state_dict(model_dict)
if use_cuda:
model = model.cuda()
criterion.cuda()
criterion_st.cuda()
optimizer.load_state_dict(checkpoint['optimizer'])
print(
" > Model restored from step %d" % checkpoint['step'], flush=True)
start_epoch = checkpoint['epoch']
best_loss = checkpoint['linear_loss']
args.restore_step = checkpoint['step']
else:
args.restore_step = 0
print("\n > Starting a new training", flush=True)
if use_cuda:
model = model.cuda()
criterion.cuda()
criterion_st.cuda()
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(" | > Model has {} parameters".format(num_params), flush=True)
if not os.path.exists(CHECKPOINT_PATH):
os.mkdir(CHECKPOINT_PATH)
if 'best_loss' not in locals():
best_loss = float('inf')
for epoch in range(0, c.epochs):
train_loss, current_step = train(model, criterion, criterion_st,
optimizer, optimizer_st,
scheduler, ap, epoch)
val_loss = evaluate(model, criterion, criterion_st, ap,
current_step)
print(
" | > Train Loss: {:.5f} Validation Loss: {:.5f}".format(
train_loss, val_loss),
flush=True)
best_loss = save_best_model(model, optimizer, train_loss, best_loss,
OUT_PATH, current_step, epoch)
def main(args):
# 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)
model = Tacotron(num_chars=num_chars,
embedding_dim=c.embedding_size,
linear_dim=ap.num_freq,
mel_dim=ap.num_mels,
r=c.r,
memory_size=c.memory_size)
optimizer = optim.Adam(model.parameters(), lr=c.lr, weight_decay=0)
optimizer_st = optim.Adam(model.decoder.stopnet.parameters(),
lr=c.lr,
weight_decay=0)
criterion = L1LossMasked()
criterion_st = nn.BCELoss()
if args.restore_path:
checkpoint = torch.load(args.restore_path)
try:
model.load_state_dict(checkpoint['model'])
optimizer.load_state_dict(checkpoint['optimizer'])
except:
print(" > Partial model initialization.")
partial_init_flag = True
model_dict = model.state_dict()
# Partial initialization: if there is a mismatch with new and old layer, it is skipped.
# 1. filter out unnecessary keys
pretrained_dict = {
k: v
for k, v in checkpoint['model'].items() if k in model_dict
}
# 2. filter out different size layers
pretrained_dict = {
k: v
for k, v in pretrained_dict.items()
if v.numel() == model_dict[k].numel()
}
# 3. overwrite entries in the existing state dict
model_dict.update(pretrained_dict)
# 4. load the new state dict
model.load_state_dict(model_dict)
print(" | > {} / {} layers are initialized".format(
len(pretrained_dict), len(model_dict)))
if use_cuda:
model = model.cuda()
criterion.cuda()
criterion_st.cuda()
for group in optimizer.param_groups:
group['lr'] = c.lr
print(" > Model restored from step %d" % checkpoint['step'],
flush=True)
start_epoch = checkpoint['epoch']
best_loss = checkpoint['linear_loss']
args.restore_step = checkpoint['step']
else:
args.restore_step = 0
if use_cuda:
model = model.cuda()
criterion.cuda()
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')
for epoch in range(0, c.epochs):
train_loss, current_step = train(model, criterion, criterion_st,
optimizer, optimizer_st, scheduler,
ap, epoch)
val_loss = evaluate(model, criterion, criterion_st, ap, current_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, current_step, epoch)
def main(args):
dataset = importlib.import_module('datasets.' + c.dataset)
Dataset = getattr(dataset, 'MyDataset')
audio = importlib.import_module('utils.' + c.audio_processor)
AudioProcessor = getattr(audio, 'AudioProcessor')
ap = AudioProcessor(sample_rate=c.sample_rate,
num_mels=c.num_mels,
min_level_db=c.min_level_db,
frame_shift_ms=c.frame_shift_ms,
frame_length_ms=c.frame_length_ms,
ref_level_db=c.ref_level_db,
num_freq=c.num_freq,
power=c.power,
preemphasis=c.preemphasis)
# Setup the dataset
train_dataset = Dataset(c.data_path,
c.meta_file_train,
c.r,
c.text_cleaner,
ap=ap,
min_seq_len=c.min_seq_len)
train_loader = DataLoader(train_dataset,
batch_size=c.batch_size,
shuffle=False,
collate_fn=train_dataset.collate_fn,
drop_last=False,
num_workers=c.num_loader_workers,
pin_memory=True)
if c.run_eval:
val_dataset = Dataset(c.data_path,
c.meta_file_val,
c.r,
c.text_cleaner,
ap=ap)
val_loader = DataLoader(val_dataset,
batch_size=c.eval_batch_size,
shuffle=False,
collate_fn=val_dataset.collate_fn,
drop_last=False,
num_workers=4,
pin_memory=True)
else:
val_loader = None
model = Tacotron(c.embedding_size, ap.num_freq, c.num_mels, c.r)
print(" | > Num output units : {}".format(ap.num_freq), flush=True)
optimizer = optim.Adam(model.parameters(), lr=c.lr)
optimizer_st = optim.Adam(model.decoder.stopnet.parameters(), lr=c.lr)
criterion = L1LossMasked()
criterion_st = nn.BCELoss()
if args.restore_path:
checkpoint = torch.load(args.restore_path)
model.load_state_dict(checkpoint['model'])
if use_cuda:
model = model.cuda()
criterion.cuda()
criterion_st.cuda()
optimizer.load_state_dict(checkpoint['optimizer'])
# optimizer_st.load_state_dict(checkpoint['optimizer_st'])
for state in optimizer.state.values():
for k, v in state.items():
if torch.is_tensor(v):
state[k] = v.cuda()
print(" > Model restored from step %d" % checkpoint['step'],
flush=True)
start_epoch = checkpoint['step'] // len(train_loader)
best_loss = checkpoint['linear_loss']
args.restore_step = checkpoint['step']
else:
args.restore_step = 0
print("\n > Starting a new training", flush=True)
if use_cuda:
model = model.cuda()
criterion.cuda()
criterion_st.cuda()
scheduler = AnnealLR(optimizer, warmup_steps=c.warmup_steps)
num_params = count_parameters(model)
print(" | > Model has {} parameters".format(num_params), flush=True)
if not os.path.exists(CHECKPOINT_PATH):
os.mkdir(CHECKPOINT_PATH)
if 'best_loss' not in locals():
best_loss = float('inf')
for epoch in range(0, c.epochs):
train_loss, current_step = train(model, criterion, criterion_st,
train_loader, optimizer, optimizer_st,
scheduler, ap, epoch)
val_loss = evaluate(model, criterion, criterion_st, val_loader, ap,
current_step)
print(" | > Train Loss: {:.5f} Validation Loss: {:.5f}".format(
train_loss, val_loss),
flush=True)
best_loss = save_best_model(model, optimizer, train_loss, best_loss,
OUT_PATH, current_step, epoch)
def main(args):
# setup output paths and read configs
c = load_config(args.config_path)
_ = os.path.dirname(os.path.realpath(__file__))
OUT_PATH = os.path.join(_, c.output_path)
OUT_PATH = create_experiment_folder(OUT_PATH)
CHECKPOINT_PATH = os.path.join(OUT_PATH, 'checkpoints')
shutil.copyfile(args.config_path, os.path.join(OUT_PATH, 'config.json'))
# Ctrl+C handler to remove empty experiment folder
def signal_handler(signal, frame):
print(" !! Pressed Ctrl+C !!")
remove_experiment_folder(OUT_PATH)
sys.exit(0)
signal.signal(signal.SIGINT, signal_handler)
dataset = LJSpeechDataset(os.path.join(c.data_path, 'metadata.csv'),
os.path.join(c.data_path, 'wavs'), c.r,
c.sample_rate, c.text_cleaner)
model = Tacotron(c.embedding_size, c.hidden_size, c.num_mels, c.num_freq,
c.r)
if use_cuda:
model = nn.DataParallel(model.cuda())
optimizer = optim.Adam(model.parameters(), lr=c.lr)
try:
checkpoint = torch.load(
os.path.join(CHECKPOINT_PATH,
'checkpoint_%d.pth.tar' % args.restore_step))
model.load_state_dict(checkpoint['model'])
optimizer.load_state_dict(checkpoint['optimizer'])
print("\n > Model restored from step %d\n" % args.restore_step)
except:
print("\n > Starting a new training\n")
model = model.train()
if not os.path.exists(CHECKPOINT_PATH):
os.mkdir(CHECKPOINT_PATH)
if use_cuda:
criterion = nn.L1Loss().cuda()
else:
criterion = nn.L1Loss()
n_priority_freq = int(3000 / (c.sample_rate * 0.5) * c.num_freq)
for epoch in range(c.epochs):
dataloader = DataLoader(dataset,
batch_size=c.batch_size,
shuffle=True,
collate_fn=dataset.collate_fn,
drop_last=True,
num_workers=32)
progbar = Progbar(len(dataset) / c.batch_size)
for i, data in enumerate(dataloader):
text_input = data[0]
magnitude_input = data[1]
mel_input = data[2]
current_step = i + args.restore_step + epoch * len(dataloader) + 1
optimizer.zero_grad()
try:
mel_input = np.concatenate(
(np.zeros([c.batch_size, 1, c.num_mels],
dtype=np.float32), mel_input[:, 1:, :]),
axis=1)
except:
raise TypeError("not same dimension")
if use_cuda:
text_input_var = Variable(torch.from_numpy(text_input).type(
torch.cuda.LongTensor),
requires_grad=False).cuda()
mel_input_var = Variable(torch.from_numpy(mel_input).type(
torch.cuda.FloatTensor),
requires_grad=False).cuda()
mel_spec_var = Variable(torch.from_numpy(mel_input).type(
torch.cuda.FloatTensor),
requires_grad=False).cuda()
linear_spec_var = Variable(
torch.from_numpy(magnitude_input).type(
torch.cuda.FloatTensor),
requires_grad=False).cuda()
else:
text_input_var = Variable(torch.from_numpy(text_input).type(
torch.LongTensor),
requires_grad=False)
mel_input_var = Variable(torch.from_numpy(mel_input).type(
torch.FloatTensor),
requires_grad=False)
mel_spec_var = Variable(torch.from_numpy(mel_input).type(
torch.FloatTensor),
requires_grad=False)
linear_spec_var = Variable(
torch.from_numpy(magnitude_input).type(torch.FloatTensor),
requires_grad=False)
mel_output, linear_output, alignments =\
model.forward(text_input_var, mel_input_var)
mel_loss = criterion(mel_output, mel_spec_var)
linear_loss = torch.abs(linear_output - linear_spec_var)
linear_loss = 0.5 * \
torch.mean(linear_loss) + 0.5 * \
torch.mean(linear_loss[:, :n_priority_freq, :])
loss = mel_loss + linear_loss
loss = loss.cuda()
start_time = time.time()
loss.backward()
nn.utils.clip_grad_norm(model.parameters(), 1.)
optimizer.step()
time_per_step = time.time() - start_time
progbar.update(i,
values=[('total_loss', loss.data[0]),
('linear_loss', linear_loss.data[0]),
('mel_loss', mel_loss.data[0])])
if current_step % c.save_step == 0:
checkpoint_path = 'checkpoint_{}.pth.tar'.format(current_step)
checkpoint_path = os.path.join(OUT_PATH, checkpoint_path)
save_checkpoint(
{
'model': model.state_dict(),
'optimizer': optimizer.state_dict(),
'step': current_step,
'total_loss': loss.data[0],
'linear_loss': linear_loss.data[0],
'mel_loss': mel_loss.data[0],
'date': datetime.date.today().strftime("%B %d, %Y")
}, checkpoint_path)
print(" > Checkpoint is saved : {}".format(checkpoint_path))
if current_step in c.decay_step:
optimizer = adjust_learning_rate(optimizer, current_step)
def main(args):
# Setup the dataset
train_dataset = LJSpeechDataset(os.path.join(c.data_path, 'metadata_train.csv'),
os.path.join(c.data_path, 'wavs'),
c.r,
c.sample_rate,
c.text_cleaner,
c.num_mels,
c.min_level_db,
c.frame_shift_ms,
c.frame_length_ms,
c.preemphasis,
c.ref_level_db,
c.num_freq,
c.power
)
train_loader = DataLoader(train_dataset, batch_size=c.batch_size,
shuffle=False, collate_fn=train_dataset.collate_fn,
drop_last=False, num_workers=c.num_loader_workers,
pin_memory=True)
val_dataset = LJSpeechDataset(os.path.join(c.data_path, 'metadata_val.csv'),
os.path.join(c.data_path, 'wavs'),
c.r,
c.sample_rate,
c.text_cleaner,
c.num_mels,
c.min_level_db,
c.frame_shift_ms,
c.frame_length_ms,
c.preemphasis,
c.ref_level_db,
c.num_freq,
c.power
)
val_loader = DataLoader(val_dataset, batch_size=c.batch_size,
shuffle=False, collate_fn=val_dataset.collate_fn,
drop_last=False, num_workers= 4,
pin_memory=True)
model = Tacotron(c.embedding_size,
c.hidden_size,
c.num_mels,
c.num_freq,
c.r,
use_atten_mask=True)
optimizer = optim.Adam(model.parameters(), lr=c.lr)
if use_cuda:
criterion = nn.L1Loss().cuda()
else:
criterion = nn.L1Loss()
if args.restore_path:
checkpoint = torch.load(args.restore_path)
model.load_state_dict(checkpoint['model'])
optimizer.load_state_dict(checkpoint['optimizer'])
print("\n > Model restored from step %d\n" % checkpoint['step'])
start_epoch = checkpoint['step'] // len(train_loader)
best_loss = checkpoint['linear_loss']
start_epoch = 0
args.restore_step = checkpoint['step']
else:
args.restore_step = 0
print("\n > Starting a new training")
if use_cuda:
model = nn.DataParallel(model.cuda())
num_params = count_parameters(model)
print(" | > Model has {} parameters".format(num_params))
if not os.path.exists(CHECKPOINT_PATH):
os.mkdir(CHECKPOINT_PATH)
if 'best_loss' not in locals():
best_loss = float('inf')
for epoch in range(0, c.epochs):
train_loss, current_step = train(model, criterion, train_loader, optimizer, epoch)
val_loss = evaluate(model, criterion, val_loader, current_step)
best_loss = save_best_model(model, optimizer, val_loss,
best_loss, OUT_PATH,
current_step, epoch)
ap = AudioProcessor(**c.audio)
use_cuda = True
data_loader = setup_loader(c, is_val=False)
num_chars = len(phonemes) if c.use_phonemes else len(symbols)
model = Tacotron(
num_chars=num_chars,
embedding_dim=c.embedding_size,
linear_dim=ap.num_freq,
mel_dim=ap.num_mels,
r=c.r,
memory_size=c.memory_size)
checkpoint = torch.load(MODEL_PATH)
model.load_state_dict(checkpoint['model'])
if use_cuda:
model.cuda()
with torch.no_grad():
if data_loader is not None:
for num_iter, data in enumerate(data_loader):
# print(num_iter * batch_size)
start_time = time.time()
# setup input data
text_input = data[0]
text_lengths = data[1]
linear_input = data[2]
mel_input = data[3]
mel_lengths = data[4]
num_chars = len(phonemes) if CONFIG.use_phonemes else len(symbols)
model = Tacotron(num_chars, CONFIG.embedding_size, ap.num_freq, ap.num_mels, CONFIG.r)
texts = []
with open(args.text) as f:
for line in f:
texts.append(line.strip())
if use_cuda:
cp = torch.load(MODEL_PATH)
else:
cp = torch.load(MODEL_PATH, map_location=lambda storage, loc: storage)
model.load_state_dict(cp['model'])
if use_cuda:
model.cuda()
model.eval()
model.decoder.max_decoder_steps = 800
batch_size = 32
for n in range(math.ceil(len(texts) / batch_size)):
batch_texts = texts[n: max(n + batch_size, len(texts))]
wavs, alignments = text2audio(texts, model, CONFIG, use_cuda, ap)
for i, wav in enumerate(wavs):
ap.save_wav(wav, os.path.join(OUT_FOLDER, 'CommonVoice_{}_{}.wav'.format(args.step, n * batch_size + i)))
if save_alignment:
# alignments can be used to train FastSpeech
alignment = alignments[i]
class Synthesizer(object):
"""
Summary:
Config is loaded and the model from the given path is loaded and prepared for inference.
Parameters:
@model_path = model's file directory path
@model_name = model's file name
@model_config = config's file name
@use_cuda = GPU flag
"""
def load_model(self, model_path, model_name, model_config, use_cuda):
#build the config's path
model_config = os.path.join(model_path, model_config)
#build the model's path
model_file = os.path.join(model_path, model_name)
print(" > Loading model ...")
print(" | > Model config path: ", model_config)
print(" | > Model file path: ", model_file)
config = load_config(model_config)
self.use_cuda = use_cuda
self.use_phonemes = config.use_phonemes
self.ap = AudioProcessor(**config.audio)
if self.use_phonemes:
self.input_size = len(phonemes)
self.input_adapter = lambda sen: phoneme_to_sequence(
sen, [config.text_cleaner], config.phoneme_language)
else:
self.input_size = len(symbols)
self.input_adapter = lambda sen: text_to_sequence(
sen, [config.text_cleaner])
self.model = Tacotron(num_chars=config['num_chars'],
embedding_dim=config['embedding_size'],
linear_dim=self.ap.num_freq,
mel_dim=self.ap.num_mels,
r=config['r'])
#load model state
if use_cuda:
cp = torch.load(model_file)
else:
cp = torch.load(model_file,
map_location=lambda storage, loc: storage)
#load the model
self.model.load_state_dict(cp['model'])
#if cuda is enabled & available move tensors to GPU
if use_cuda:
self.model.cuda()
#disables normalization techniques present in code
self.model.eval()
"""
Summary:
Saves the wav at the given path
Parameters:
@wav = wav array
@path = destination path
"""
def save_wav(self, wav, path):
# wav *= 32767 / max(1e-8, np.max(np.abs(wav)))
wav = np.array(wav)
self.ap.save_wav(wav, path)
"""
Summary:
Gets an input, prepares it for the model and returns the predicted output.
Parameters:
@text = input sentence
"""
def tts(self, text, gl_mode=None):
wavs = []
#split the input in sentences
for sen in text.split('.'):
if len(sen) < 3:
continue
sen = sen.strip()
sen += '.'
#print('Input : {}'.format(sen))
#character => phonem => index
seq = np.array(self.input_adapter(sen))
#numpy to pytorch array
chars_var = torch.from_numpy(seq).unsqueeze(0).long()
if self.use_cuda:
chars_var = chars_var.cuda()
#begin the inference
mel_out, linear_out, alignments, stop_tokens = self.model.forward(
chars_var)
#move output tensor to cpu
linear_out = linear_out[0].data.cpu().numpy()
t = time.time()
wav = self.ap.inv_spectrogram(linear_out.T, gl_mode)
t = time.time() - t
wavs += list(wav)
wavs += [0] * 10000
out = io.BytesIO()
self.save_wav(wavs, out)
self.save_wav(wavs, 'gla.wav')
return out
