def synthesize(self, args, text, idx):
seq = text_to_sequence(text)
dec_input = np.zeros((1, 200, mel_dim), dtype='float32')
pred = []
for i in range(1, 201):
mel_out, alignment = self.session.run(
[self.mel_output, self.alignment],
feed_dict={
self.model.enc_input: [np.asarray(seq, dtype=np.int32)],
self.model.sequence_length:
np.asarray([len(seq)], dtype=np.int32),
self.model.dec_input:
dec_input
})
if i < 200:
dec_input[:, i, :] = mel_out[5 * i - 1, :]
pred.extend(mel_out[5 * (i - 1):5 * i, :])
np.save(os.path.join(args.save_dir, 'mel-{}'.format(idx)),
pred,
allow_pickle=False)
input_seq = sequence_to_text(seq)
alignment_dir = os.path.join(args.save_dir, 'align-{}.png'.format(idx))
plot_alignment(alignment, alignment_dir, input_seq)
def test_step(text, idx):
seq = text_to_sequence(text)
enc_input = np.asarray([seq], dtype=np.int32)
sequence_length = np.asarray([len(seq)], dtype=np.int32)
dec_input = np.zeros((1, max_iter, mel_dim), dtype=np.float32)
pred = []
for i in range(1, max_iter + 1):
mel_out, alignment = model(enc_input,
sequence_length,
dec_input,
is_training=False)
if i < max_iter:
dec_input[:, i, :] = mel_out[:, reduction * i - 1, :]
pred.extend(mel_out[:, reduction * (i - 1):reduction * i, :])
pred = np.reshape(np.asarray(pred), [-1, mel_dim])
alignment = np.squeeze(alignment, axis=0)
np.save(os.path.join(save_dir, 'mel-{}'.format(idx)),
pred,
allow_pickle=False)
input_seq = sequence_to_text(seq)
alignment_dir = os.path.join(save_dir, 'align-{}.png'.format(idx))
plot_alignment(alignment, alignment_dir, input_seq)
def train(args):
save_dir = './checkpoint/1'
checkpoint_path = os.path.join(save_dir, 'model.ckpt')
coord = tf.train.Coordinator()
feeder = DataFeeder(coord, mode=1)
model = Tacotron()
model.initialize(feeder.enc_input, feeder.sequence_length,
feeder.dec_input, feeder.mel_target)
saver = tf.train.Saver(max_to_keep=5, keep_checkpoint_every_n_hours=2)
with tf.Session() as sess:
try:
sess.run(tf.global_variables_initializer())
if args.step:
load_dir = '{}-{}'.format(checkpoint_path, args.step)
saver.restore(sess, load_dir)
feeder.start_in_session(sess)
while not coord.should_stop():
step, loss, optimize = sess.run(
[model.global_step, model.loss, model.optimize])
print('Step: {}, Loss: {:.5f}'.format(step, loss))
if step % checkpoint_step == 0:
saver.save(sess, checkpoint_path, global_step=step)
input_seq, alignment, pred, target = \
sess.run([model.enc_input[0], model.alignment[0], model.mel_output[0], model.mel_target[0]])
input_seq = sequence_to_text(input_seq)
alignment_dir = os.path.join(
save_dir, 'step-{}-align.png'.format(step))
plot_alignment(alignment, alignment_dir, input_seq)
except Exception as e:
traceback.print_exc()
coord.request_stop(e)
step = tf.Variable(0)
checkpoint_dir = './checkpoint/1'
os.makedirs(checkpoint_dir, exist_ok=True)
checkpoint = tf.train.Checkpoint(optimizer=optimizer, model=model, step=step)
manager = tf.train.CheckpointManager(checkpoint, checkpoint_dir, max_to_keep=5)
checkpoint.restore(manager.latest_checkpoint)
if manager.latest_checkpoint:
print('Restore checkpoint from {}'.format(manager.latest_checkpoint))
try:
for text, dec, mel, text_length in dataset:
loss, pred, alignment = train_step(text, dec, mel, text_length)
checkpoint.step.assign_add(1)
print("Step: {}, Loss: {:.5f}".format(int(checkpoint.step), loss))
if int(checkpoint.step) % checkpoint_step == 0:
checkpoint.save(file_prefix=os.path.join(
checkpoint_dir, 'step-{}'.format(int(checkpoint.step))))
input_seq = sequence_to_text(text[0].numpy())
input_seq = input_seq[:text_length[0].numpy()]
alignment_dir = os.path.join(
checkpoint_dir,
'step-{}-align.png'.format(int(checkpoint.step)))
plot_alignment(alignment, alignment_dir, input_seq)
except Exception:
traceback.print_exc()