padding_idx=hparams.padding_idx,
use_memory_mask=hparams.use_memory_mask,
)
model = model.cuda()
#model = DataParallelFix(model)
optimizer = optim.Adam(model.parameters(),
lr=hparams.initial_learning_rate, betas=(
hparams.adam_beta1, hparams.adam_beta2),
weight_decay=hparams.weight_decay)
# Load checkpoint
if checkpoint_path:
print("Load checkpoint from: {}".format(checkpoint_path))
checkpoint = torch.load(checkpoint_path)
model.load_state_dict(checkpoint["state_dict"])
optimizer.load_state_dict(checkpoint["optimizer"])
try:
global_step = int(checkpoint["global_step"])
global_epoch = int(checkpoint["global_epoch"])
except:
# TODO
pass
# Setup tensorboard logger
tensorboard_logger.configure(log_path)
#print(hparams_debug_string())
# Train!
try:
)
vocoder_model = WaveLSTM(
n_vocab=257,
embedding_dim=256,
mel_dim=hparams.num_mels,
linear_dim=hparams.num_freq,
r=hparams.outputs_per_step,
padding_idx=hparams.padding_idx,
use_memory_mask=hparams.use_memory_mask,
)
#checkpoint = torch.load(checkpoint_path)
#checkpoints_dir = os.path.dirname(checkpoint_path)
acousticmodel.load_state_dict(checkpoint_acousticmodel["state_dict"])
acousticmodel.decoder.max_decoder_steps = max_decoder_steps
os.makedirs(dst_dir, exist_ok=True)
vocoder_checkpoint_path = 'exp/exp_vocoding_bsz4seqlen8_cloneofwavernn/checkpoints/checkpoint_step1400000.pth'
vocoder_model.load_state_dict(checkpoint_vocoder["state_dict"])
with open(text_list_file_path, "rb") as f:
lines = f.readlines()
for idx, line in enumerate(lines):
fname = line.decode("utf-8").split()[0].zfill(8)
cmd = 'cp vox/wav/' + fname + '.wav ' + dst_dir + '/' + fname + '_original.wav'
print(cmd)
os.system(cmd)
text = ' '.join(k for k in line.decode("utf-8").split()[1:])
