def main(libri_dir=DATASET_DIR):
libri = read_librispeech_structure(libri_dir)
batch = stochastic_mini_batch(libri, batch_size=BATCH_NUM_TRIPLETS)
batch_size = BATCH_NUM_TRIPLETS * 3 # A triplet has 3 parts.
x, y = batch.to_inputs()
b = x[0]
num_frames = b.shape[0]
print('num_frames = ', num_frames)
model = convolutional_model(batch_input_shape=[batch_size * num_frames] + list(b.shape[1:]),
batch_size=batch_size, num_frames=num_frames)
model.compile(optimizer='adam', loss=deep_speaker_loss)
print(model.summary())
grad_steps = 0
orig_time = time()
while True:
batch = stochastic_mini_batch(libri, batch_size=BATCH_NUM_TRIPLETS)
x, _ = batch.to_inputs()
# output.shape = (3, 383, 32, 32, 3) something like this
# explanation = (batch_size, num_frames, width, height, channels)
x = np.reshape(x, (batch_size * num_frames, b.shape[2], b.shape[2], b.shape[3]))
# we don't need to use the targets y, because we know by the convention that:
# we have [anchors, positive examples, negative examples]. The loss only uses x and
# can determine if a sample is an anchor, positive or negative sample.
stub_targets = np.random.uniform(size=(x.shape[0], 1))
# result = model.predict(x, batch_size=x.shape[0])
# print(result.shape)
# np.set_printoptions(precision=2)
# print(result[0:20, 0:5])
loss = model.train_on_batch(x, stub_targets)
print('batch #{0} processed in {1:.2f}s, training loss = {2}.'.format(grad_steps, time() - orig_time, loss))
grad_steps += 1
orig_time = time()
def main():
from librispeech_wav_reader import read_librispeech_structure
libri = read_librispeech_structure(c.DATASET_DIR)
stochastic_mini_batch(libri, 3)
def main(libri_dir=c.DATASET_DIR):
logging.info('Looking for audio [wav] files in {}.'.format(libri_dir))
libri = read_librispeech_structure(libri_dir)
if len(libri) == 0:
logging.warning(
'Have you converted flac files to wav? If not, run audio/convert_flac_2_wav.sh'
)
exit(1)
batch = stochastic_mini_batch(libri, batch_size=c.BATCH_NUM_TRIPLETS)
batch_size = c.BATCH_NUM_TRIPLETS * 3 # A triplet has 3 parts.
x, y = batch.to_inputs()
b = x[0]
num_frames = b.shape[0]
logging.info('num_frames = {}'.format(num_frames))
batch_shape = [batch_size * num_frames] + list(b.shape[1:])
logging.info('batch shape: {}'.format(batch_shape))
logging.info('batch size: {}'.format(batch_size))
model = convolutional_model(batch_input_shape=batch_shape,
batch_size=batch_size,
num_frames=num_frames)
logging.info(model.summary())
logging.info('Compiling the model...')
model.compile(optimizer='adam', loss=deep_speaker_loss)
logging.info('[DONE]')
grad_steps = 0
last_checkpoint = get_last_checkpoint_if_any(c.CHECKPOINT_FOLDER)
if last_checkpoint is not None:
logging.info('Found checkpoint [{}]. Resume from here...'.format(
last_checkpoint))
model.load_weights(last_checkpoint)
grad_steps = int(last_checkpoint.split('_')[-2])
logging.info('[DONE]')
logging.info('Starting training...')
orig_time = time()
while True:
grad_steps += 1
batch = stochastic_mini_batch(libri, batch_size=c.BATCH_NUM_TRIPLETS)
x, _ = batch.to_inputs()
# output.shape = (3, 383, 32, 32, 3) something like this
# explanation = (batch_size, num_frames, width, height, channels)
logging.info('x.shape before reshape: {}'.format(x.shape))
x = np.reshape(
x, (batch_size * num_frames, b.shape[2], b.shape[2], b.shape[3]))
logging.info('x.shape after reshape: {}'.format(x.shape))
# we don't need to use the targets y, because we know by the convention that:
# we have [anchors, positive examples, negative examples]. The loss only uses x and
# can determine if a sample is an anchor, positive or negative sample.
stub_targets = np.random.uniform(size=(x.shape[0], 1))
# result = model.predict(x, batch_size=x.shape[0])
# logging.info(result.shape)
# np.set_printoptions(precision=2)
# logging.info(result[0:20, 0:5])
logging.info('-' * 80)
logging.info('== Presenting batch #{0}'.format(grad_steps))
logging.info(batch.libri_batch)
loss = model.train_on_batch(x, stub_targets)
logging.info(
'== Processed in {0:.2f}s by the network, training loss = {1}.'.
format(time() - orig_time, loss))
orig_time = time()
# record training loss
with open(c.LOSS_FILE, "a") as f:
f.write("{0},{1}\n".format(grad_steps, loss))
# checkpoints are really heavy so let's just keep the last one.
create_dir_and_delete_content(c.CHECKPOINT_FOLDER)
model.save_weights('{0}/model_{1}_{2:.5f}.h5'.format(
c.CHECKPOINT_FOLDER, grad_steps, loss))