def predict(desired_sample_rate, fragment_length, _log, seed, _seed, _config,
predict_seconds, data_dir, batch_size, fragment_stride,
nb_output_bins, learn_all_outputs, run_dir,
predict_use_softmax_as_input, use_ulaw, predict_initial_input,
**kwargs):
checkpoint_dir = os.path.join(run_dir, 'checkpoints')
last_checkpoint = sorted(os.listdir(checkpoint_dir))[-1]
epoch = int(re.match(r'checkpoint\.(\d+?)-.*', last_checkpoint).group(1))
_log.info('Using checkpoint from epoch: %s' % epoch)
sample_dir = os.path.join(run_dir, 'samples')
if not os.path.exists(sample_dir):
os.mkdir(sample_dir)
sample_name = make_sample_name(epoch)
sample_filename = os.path.join(sample_dir, sample_name)
_log.info('Saving to "%s"' % sample_filename)
sample_stream = make_sample_stream(desired_sample_rate, sample_filename)
model = build_model()
model.load_weights(os.path.join(checkpoint_dir, last_checkpoint))
if predict_initial_input is not '':
_log.info('Taking first %d (%.2fs) from \'%s\' as initial input.' %
(fragment_length, fragment_length / desired_sample_rate,
predict_initial_input))
wav = dataset.process_wav(desired_sample_rate, predict_initial_input,
use_ulaw)
outputs = list(dataset.one_hot(wav[0:fragment_length]))
else:
_log.info('Taking sample from test dataset as initial input.' %
(fragment_length, predict_initial_input))
data_generators, _ = dataset.generators(data_dir, desired_sample_rate,
fragment_length, batch_size,
fragment_stride,
nb_output_bins,
learn_all_outputs, use_ulaw)
outputs = list(data_generators['test'].next()[0][-1])
# write_samples(sample_stream, outputs)
for i in tqdm(xrange(int(desired_sample_rate * predict_seconds))):
prediction_seed = np.expand_dims(
np.array(outputs[i:i + fragment_length]), 0)
output = model.predict(prediction_seed)
output_dist = output[0][-1]
output_val = draw_sample(output_dist)
if predict_use_softmax_as_input:
outputs.append(output_dist)
else:
outputs.append(output_val)
write_samples(sample_stream, [output_val])
sample_stream.close()
_log.info("Done!")
def get_generators(batch_size, data_dir, desired_sample_rate, fragment_length, fragment_stride, learn_all_outputs,
nb_output_bins, use_ulaw, test_factor, data_dir_structure, randomize_batch_order, _rnd, random_train_batches):
if data_dir_structure == 'flat':
return dataset.generators(data_dir, desired_sample_rate, fragment_length, batch_size,
fragment_stride, nb_output_bins, learn_all_outputs, use_ulaw, randomize_batch_order,
_rnd, random_train_batches)
elif data_dir_structure == 'vctk':
return dataset.generators_vctk(data_dir, desired_sample_rate, fragment_length, batch_size,
fragment_stride, nb_output_bins, learn_all_outputs, use_ulaw, test_factor,
randomize_batch_order, _rnd, random_train_batches)
else:
raise ValueError('data_dir_structure must be "flat" or "vctk", is %s' % data_dir_structure)
def get_generators(batch_size, data_dir, desired_sample_rate, fragment_length, fragment_stride, learn_all_outputs,
nb_output_bins, use_ulaw, test_factor, data_dir_structure, randomize_batch_order, _rnd,
random_train_batches):
if data_dir_structure == 'flat':
return dataset.generators(data_dir, desired_sample_rate, fragment_length, batch_size,
fragment_stride, nb_output_bins, learn_all_outputs, use_ulaw, randomize_batch_order,
_rnd, random_train_batches)
elif data_dir_structure == 'vctk':
return dataset.generators_vctk(data_dir, desired_sample_rate, fragment_length, batch_size,
fragment_stride, nb_output_bins, learn_all_outputs, use_ulaw, test_factor,
randomize_batch_order, _rnd, random_train_batches)
else:
raise ValueError('data_dir_structure must be "flat" or "vctk", is %s' % data_dir_structure)
def test_preprocess(desired_sample_rate, fragment_length, _log, seed, _seed,
_config, predict_seconds, data_dir, batch_size,
fragment_stride, nb_output_bins, learn_all_outputs,
run_dir, predict_use_softmax_as_input, use_ulaw, **kwargs):
sample_dir = os.path.join('preprocess_test')
if not os.path.exists(sample_dir):
os.mkdir(sample_dir)
ulaw_str = '_ulaw' if use_ulaw else ''
sample_filename = os.path.join(sample_dir, 'test1%s.wav' % ulaw_str)
sample_stream = make_sample_stream(desired_sample_rate, sample_filename)
data_generators, _ = dataset.generators(data_dir, desired_sample_rate,
fragment_length, batch_size,
fragment_stride, nb_output_bins,
learn_all_outputs, use_ulaw)
outputs = data_generators['test'].next()[0][batch_size - 1].astype('uint8')
write_samples(sample_stream, outputs)
scipy.io.wavfile.write(os.path.join(sample_dir, 'test2%s.wav' % ulaw_str),
desired_sample_rate,
np.argmax(outputs, axis=-1).astype('uint8'))
def main(run_dir, data_dir, nb_epoch, early_stopping_patience,
desired_sample_rate, fragment_length, batch_size, fragment_stride,
nb_output_bins, keras_verbose, _log, seed, _config, debug,
learn_all_outputs, train_only_in_receptive_field, _run, use_ulaw):
if run_dir is None:
run_dir = os.path.join(
'models',
datetime.datetime.now().strftime('run_%Y-%m-%d_%H:%M:%S'))
_config['run_dir'] = run_dir
print_config(_run)
_log.info('Running with seed %d' % seed)
if not debug:
if os.path.exists(run_dir):
raise EnvironmentError('Run with seed %d already exists' % seed)
os.mkdir(run_dir)
checkpoint_dir = os.path.join(run_dir, 'checkpoints')
json.dump(_config, open(os.path.join(run_dir, 'config.json'), 'w'))
_log.info('Loading data...')
data_generators, nb_examples = dataset.generators(
data_dir, desired_sample_rate, fragment_length, batch_size,
fragment_stride, nb_output_bins, learn_all_outputs, use_ulaw)
_log.info('Building model...')
model = build_model(fragment_length)
_log.info(model.summary())
optim = make_optimizer()
_log.info('Compiling Model...')
loss = objectives.categorical_crossentropy
all_metrics = [
metrics.categorical_accuracy, metrics.categorical_mean_squared_error
]
if train_only_in_receptive_field:
loss = skip_out_of_receptive_field(loss)
all_metrics = [skip_out_of_receptive_field(m) for m in all_metrics]
model.compile(optimizer=optim, loss=loss, metrics=all_metrics)
# TODO: Consider gradient weighting making last outputs more important.
callbacks = [
ReduceLROnPlateau(patience=early_stopping_patience / 2,
cooldown=early_stopping_patience / 4,
verbose=1),
EarlyStopping(patience=early_stopping_patience, verbose=1),
]
if not debug:
callbacks.extend([
ModelCheckpoint(os.path.join(
checkpoint_dir, 'checkpoint.{epoch:05d}-{val_loss:.3f}.hdf5'),
save_best_only=True),
CSVLogger(os.path.join(run_dir, 'history.csv')),
])
if not debug:
os.mkdir(checkpoint_dir)
_log.info('Starting Training...')
model.fit_generator(data_generators['train'],
nb_examples['train'],
nb_epoch=nb_epoch,
validation_data=data_generators['test'],
nb_val_samples=nb_examples['test'],
callbacks=callbacks,
verbose=keras_verbose)