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 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):
fragment_length = compute_receptive_field()[0]
_config['fragment_length'] = fragment_length
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.makedirs(sample_dir, exist_ok=True)
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))
model.summary()
if predict_initial_input is None:
outputs = list(dataset.one_hot(np.zeros(fragment_length) + nb_output_bins / 2))
elif predict_initial_input != '':
_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.')
data_generators, _ = get_generators()
outputs = list(data_generators['test'].next()[0][-1])
# write_samples(sample_stream, outputs)
warned_repetition = False
for i in tqdm(range(int(desired_sample_rate * predict_seconds))):
if not warned_repetition:
if np.argmax(outputs[-1]) == np.argmax(outputs[-2]) and np.argmax(outputs[-2]) == np.argmax(outputs[-3]):
warned_repetition = True
_log.warning('Last three predicted outputs where %d' % np.argmax(outputs[-1]))
else:
warned_repetition = False
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 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):
fragment_length = compute_receptive_field()[0]
_config['fragment_length'] = fragment_length
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))
model.summary()
if predict_initial_input is None:
outputs = list(dataset.one_hot(np.zeros(fragment_length) + nb_output_bins / 2))
elif predict_initial_input != '':
_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.')
data_generators, _ = get_generators()
outputs = list(data_generators['test'].next()[0][-1])
# write_samples(sample_stream, outputs)
warned_repetition = False
for i in tqdm(range(int(desired_sample_rate * predict_seconds))):
if not warned_repetition:
if np.argmax(outputs[-1]) == np.argmax(outputs[-2]) and np.argmax(outputs[-2]) == np.argmax(outputs[-3]):
warned_repetition = True
_log.warning('Last three predicted outputs where %d' % np.argmax(outputs[-1]))
else:
warned_repetition = False
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!")