def test_librispeech():
"""Summary
"""
batch_size = 24
filter_length = 2
n_stages = 7
n_layers_per_stage = 9
n_hidden = 48
n_skip = 384
total_length = 16000
sequence_length = get_sequence_length(n_stages, n_layers_per_stage)
prime_length = sequence_length
ckpt_path = 'wavenet/wavenet_filterlen{}_batchsize{}_sequencelen{}_stages{}_layers{}_hidden{}_skips{}/'.format(
filter_length, batch_size, sequence_length, n_stages,
n_layers_per_stage, n_hidden, n_skip)
dataset = librispeech.get_dataset()
batch = next(librispeech.batch_generator(dataset, batch_size,
prime_length))[0]
sess = tf.Session()
net = create_wavenet(batch_size=batch_size,
filter_length=filter_length,
n_hidden=n_hidden,
n_skip=n_skip,
n_layers_per_stage=n_layers_per_stage,
n_stages=n_stages,
shift=False)
init_op = tf.group(tf.global_variables_initializer(),
tf.local_variables_initializer())
sess.run(init_op)
saver = tf.train.Saver()
if tf.train.latest_checkpoint(ckpt_path) is not None:
saver.restore(sess, tf.train.latest_checkpoint(ckpt_path))
else:
print('Could not find checkpoint')
synth = np.zeros([batch_size, total_length], dtype=np.float32)
synth[:, :prime_length] = batch
print('Synthesize...')
for sample_i in range(0, total_length - prime_length):
print('{}/{}/{}'.format(sample_i, prime_length, total_length),
end='\r')
probs = sess.run(net["probs"],
feed_dict={
net["X"]:
synth[:, sample_i:sample_i + sequence_length]
})
idxs = sample_categorical(probs)
idxs = idxs.reshape((batch_size, sequence_length))
if sample_i == 0:
audio = wnu.inv_mu_law_numpy(idxs - 128)
synth[:, :prime_length] = audio
else:
audio = wnu.inv_mu_law_numpy(idxs[:, -1] - 128)
synth[:, prime_length + sample_i] = audio
for i in range(batch_size):
wavfile.write('synthesis-{}.wav'.format(i), 16000, synth[i])
def test_librispeech():
"""Summary
"""
batch_size = 24
filter_length = 2
n_stages = 7
n_layers_per_stage = 9
n_hidden = 48
n_skip = 384
total_length = 16000
sequence_length = get_sequence_length(n_stages, n_layers_per_stage)
prime_length = sequence_length
ckpt_path = 'wavenet/wavenet_filterlen{}_batchsize{}_sequencelen{}_stages{}_layers{}_hidden{}_skips{}/'.format(
filter_length, batch_size, sequence_length, n_stages,
n_layers_per_stage, n_hidden, n_skip)
dataset = librispeech.get_dataset()
batch = next(
librispeech.batch_generator(dataset, batch_size, prime_length))[0]
sess = tf.Session()
net = create_wavenet(
batch_size=batch_size,
filter_length=filter_length,
n_hidden=n_hidden,
n_skip=n_skip,
n_layers_per_stage=n_layers_per_stage,
n_stages=n_stages,
shift=False)
init_op = tf.group(tf.global_variables_initializer(),
tf.local_variables_initializer())
sess.run(init_op)
saver = tf.train.Saver()
if tf.train.latest_checkpoint(ckpt_path) is not None:
saver.restore(sess, tf.train.latest_checkpoint(ckpt_path))
else:
print('Could not find checkpoint')
synth = np.zeros([batch_size, total_length], dtype=np.float32)
synth[:, :prime_length] = batch
print('Synthesize...')
for sample_i in range(0, total_length - prime_length):
print('{}/{}/{}'.format(sample_i, prime_length, total_length), end='\r')
probs = sess.run(
net["probs"],
feed_dict={net["X"]: synth[:, sample_i:sample_i + sequence_length]})
idxs = sample_categorical(probs)
idxs = idxs.reshape((batch_size, sequence_length))
if sample_i == 0:
audio = wnu.inv_mu_law_numpy(idxs - 128)
synth[:, :prime_length] = audio
else:
audio = wnu.inv_mu_law_numpy(idxs[:, -1] - 128)
synth[:, prime_length + sample_i] = audio
for i in range(batch_size):
wavfile.write('synthesis-{}.wav'.format(i), 16000, synth[i])
def test_librispeech():
"""Summary
"""
prime_length = 6144
total_length = 16000 * 3
batch_size = 32
n_stages = 6
n_layers_per_stage = 9
n_hidden = 32
filter_length = 2
n_skip = 256
onehot = False
sequence_length = get_sequence_length(n_stages, n_layers_per_stage)
ckpt_path = 'vctk-wavenet/wavenet_filterlen{}_batchsize{}_sequencelen{}_stages{}_layers{}_hidden{}_skips{}/'.format(
filter_length, batch_size, sequence_length,
n_stages, n_layers_per_stage, n_hidden, n_skip)
dataset = librispeech.get_dataset()
batch = next(librispeech.batch_generator(dataset,
batch_size, prime_length))[0]
with tf.Graph().as_default(), tf.Session() as sess:
net = create_generation_model(batch_size=batch_size,
filter_length=filter_length,
n_hidden=n_hidden,
n_skip=n_skip,
n_layers_per_stage=n_layers_per_stage,
n_stages=n_stages,
onehot=onehot)
saver = tf.train.Saver()
if tf.train.latest_checkpoint(ckpt_path) is not None:
saver.restore(sess, tf.train.latest_checkpoint(ckpt_path))
else:
print('Could not find checkpoint')
sess.run(net['init_ops'])
synth = np.zeros([batch_size, total_length], dtype=np.float32)
synth[:, :prime_length] = batch
print('Synthesize...')
for sample_i in range(total_length - 1):
print('{}/{}/{}'.format(sample_i, prime_length, total_length),
end='\r')
probs = sess.run(
[net["probs"], net["push_ops"]],
feed_dict={net["X"]: synth[:, [sample_i]]})[0]
idxs = sample_categorical(probs)
audio = wnu.inv_mu_law_numpy(idxs - 128)
if sample_i >= prime_length:
synth[:, sample_i + 1] = audio
for i in range(batch_size):
wavfile.write('synthesis-{}.wav'.format(i),
16000, synth[i])
def test_librispeech():
"""Summary
"""
prime_length = 6144
total_length = 16000 * 3
batch_size = 32
n_stages = 6
n_layers_per_stage = 9
n_hidden = 32
filter_length = 2
n_skip = 256
onehot = False
sequence_length = get_sequence_length(n_stages, n_layers_per_stage)
ckpt_path = 'vctk-wavenet/wavenet_filterlen{}_batchsize{}_sequencelen{}_stages{}_layers{}_hidden{}_skips{}/'.format(
filter_length, batch_size, sequence_length,
n_stages, n_layers_per_stage, n_hidden, n_skip)
dataset = librispeech.get_dataset()
batch = next(librispeech.batch_generator(dataset,
batch_size, prime_length))[0]
with tf.Graph().as_default(), tf.Session() as sess:
net = create_generation_model(batch_size=batch_size,
filter_length=filter_length,
n_hidden=n_hidden,
n_skip=n_skip,
n_layers_per_stage=n_layers_per_stage,
n_stages=n_stages,
onehot=onehot)
saver = tf.train.Saver()
if tf.train.latest_checkpoint(ckpt_path) is not None:
saver.restore(sess, tf.train.latest_checkpoint(ckpt_path))
else:
print('Could not find checkpoint')
sess.run(net['init_ops'])
synth = np.zeros([batch_size, total_length], dtype=np.float32)
synth[:, :prime_length] = batch
print('Synthesize...')
for sample_i in range(total_length - 1):
print('{}/{}/{}'.format(sample_i, prime_length, total_length),
end='\r')
probs = sess.run(
[net["probs"], net["push_ops"]],
feed_dict={net["X"]: synth[:, [sample_i]]})[0]
idxs = sample_categorical(probs)
audio = wnu.inv_mu_law_numpy(idxs - 128)
if sample_i >= prime_length:
synth[:, sample_i + 1] = audio
for i in range(batch_size):
wavfile.write('synthesis-{}.wav'.format(i),
16000, synth[i])
def train_librispeech():
dataset = librispeech.get_dataset(convert_to_wav=False)
it_i = 0
n_epochs = 10000
batch_size = 32
n_stages = 6
n_layers_per_stage = 9
n_hidden = 32
filter_length = 2
n_skip = 256
sequence_length = get_sequence_length(n_stages, n_layers_per_stage)
ckpt_path = 'vctk-wavenet/wavenet_filterlen{}_batchsize{}_sequencelen{}_stages{}_layers{}_hidden{}_skips{}/'.format(
filter_length, batch_size, sequence_length,
n_stages, n_layers_per_stage, n_hidden, n_skip)
with tf.Graph().as_default() as g, tf.Session(graph=g) as sess:
net = create_generation_model(n_stages=n_stages,
n_layers_per_stage=n_layers_per_stage,
n_hidden=n_hidden,
batch_size=batch_size,
n_skip=n_skip,
filter_length=filter_length)
batch = librispeech.batch_generator
opt = tf.train.AdamOptimizer(learning_rate=0.00001).minimize(
net['loss'])
sess.run(tf.global_variables_initializer())
saver = tf.train.Saver()
writer = tf.summary.FileWriter(ckpt_path)
if tf.train.latest_checkpoint(ckpt_path) is not None:
saver.restore(sess, tf.train.latest_checkpoint(ckpt_path))
for epoch_i in range(n_epochs):
for batch_xs, batch_hs in batch(dataset, batch_size, sequence_length):
loss, _ = sess.run([net['loss'], opt], feed_dict={
net['X']: batch_xs})
print(loss)
if it_i % 100 == 0:
summary = sess.run(net['summaries'], feed_dict={
net['X']: batch_xs})
writer.add_summary(summary, it_i)
saver.save(sess,
os.path.join(ckpt_path, 'model.ckpt'),
global_step=it_i)
it_i += 1
def test_librispeech_batch_generator():
ds = librispeech.get_dataset()
batch = next(librispeech.batch_generator(ds, batch_size=32))
assert (batch[0].shape == (32, 6144))
def test_librispeech_dataset():
ds = librispeech.get_dataset(convert_to_wav=True)
assert (len(ds) == 106717)
def test_librispeech_batch():
ds = librispeech.get_dataset()
batch = next(librispeech.batch_generator(ds, batch_size=32))
assert(batch[0].shape == (32, 6144))
assert(batch[1].shape == (32,))
def test_librispeech_dataset():
ds = librispeech.get_dataset()
assert(len(ds) == 106717)