def decode():
"""Decoding the inputs using current model."""
tf.logging.info("Get TEST sets number.")
num_batch = get_num_batch(FLAGS.test_list_file, infer=True)
with tf.Graph().as_default():
with tf.device('/cpu:0'):
with tf.name_scope('input'):
data_list = read_list(FLAGS.test_list_file)
test_utt_id, test_inputs, _ = get_batch(
data_list,
batch_size=1,
input_size=FLAGS.input_dim,
output_size=FLAGS.output_dim,
left=FLAGS.left_context,
right=FLAGS.right_context,
num_enqueuing_threads=FLAGS.num_threads,
num_epochs=1,
infer=True)
# test_inputs = tf.squeeze(test_inputs, axis=[0])
devices = []
for i in xrange(FLAGS.num_gpu):
device_name = ("/gpu:%d" % i)
print('Using device: ', device_name)
devices.append(device_name)
# Prevent exhausting all the gpu memories.
config = tf.ConfigProto()
config.gpu_options.per_process_gpu_memory_fraction = 0.4
#config.gpu_options.allow_growth = True
config.allow_soft_placement = True
set_session(tf.Session(config=config))
# execute the session
with tf.Session(config=config) as sess:
# Create two models with tr_inputs and cv_inputs individually.
with tf.name_scope('model'):
model = DNNTrainer(sess,
FLAGS,
devices,
test_inputs,
labels=None,
cross_validation=True)
show_all_variables()
init = tf.group(tf.global_variables_initializer(),
tf.local_variables_initializer())
print("Initializing variables ...")
sess.run(init)
if model.load(model.save_dir, moving_average=False):
print("[*] Load Moving Average model SUCCESS")
else:
print("[!] Load failed. Checkpoint not found. Exit now.")
sys.exit(1)
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
cmvn_filename = os.path.join(FLAGS.data_dir, "train_cmvn.npz")
if os.path.isfile(cmvn_filename):
cmvn = np.load(cmvn_filename)
else:
tf.logging.fatal("%s not exist, exit now." % cmvn_filename)
sys.exit(1)
out_dir_name = os.path.join('/Work18/2017/linan/SE/my_enh',
FLAGS.save_dir, FLAGS.savetestdir)
# out_dir_name = os.path.join(FLAGS.save_dir, 'test')
if not os.path.exists(out_dir_name):
os.makedirs(out_dir_name)
write_scp_path = os.path.join(out_dir_name, 'feats.scp')
write_ark_path = os.path.join(out_dir_name, 'feats.ark')
writer = ArkWriter(write_scp_path)
outputs = model.generator(test_inputs, None, reuse=True)
outputs = tf.reshape(outputs, [-1, model.output_dim])
print('shape is', np.shape(outputs))
try:
for batch in range(num_batch):
if coord.should_stop():
break
# outputs = model.generator(test_inputs, None, reuse=True)
# outputs = tf.reshape(outputs, [-1, model.output_dim])
utt_id, activations = sess.run([test_utt_id, outputs])
# sequence = activations * cmvn['stddev_labels'] + \
# cmvn['mean_labels']
sequence = activations
save_result = np.vstack(sequence)
dir_load = FLAGS.savetestdir
dir_load = dir_load.split('/')[-1]
mode = FLAGS.mode
if mode == 'use_org':
inputs_path = os.path.join(
'workspace/features/spectrogram/test', dir_load,
'%s.wav.p' % utt_id[0])
data = cPickle.load(open(inputs_path, 'rb'))
[mixed_complx_x] = data
#tf.logging.info("Write inferred %s to %s" %(utt_id[0], np.shape(save_result)))
save_result = np.exp(save_result)
n_window = cfg.n_window
s = recover_wav(save_result, mixed_complx_x,
cfg.n_overlap, np.hamming)
s *= np.sqrt(
(np.hamming(n_window)**2
).sum()) # Scaler for compensate the amplitude
# change after spectrogram and IFFT.
print("start enhance wav file")
# Write out enhanced wav.
out_path = os.path.join("workspace", "enh_wavs",
"test", dir_load,
"%s.enh.wav" % utt_id[0])
print("have enhanced all the wav")
pp_data.create_folder(os.path.dirname(out_path))
pp_data.write_audio(out_path, s, 16000)
elif mode == 'g_l':
inputs_path = os.path.join(
'workspace/features/spectrogram/test', dir_load,
'%s.wav.p' % utt_id[0])
data = cPickle.load(open(inputs_path, 'rb'))
[mixed_complx_x] = data
save_result = np.exp(save_result)
s = save_result
s = audio_utilities.reconstruct_signal_griffin_lim(
s, mixed_complx_x, 512, 256, 15)
#s = recover_wav(save_result,mixed_complx_x,cfg.n_overlap, np.hamming)
s *= np.sqrt((np.hamming(cfg.n_window)**2).sum())
#s = audio._griffin_lim(s)
out_path = os.path.join("workspace", "enh_wavs",
"test2", dir_load,
"%s.enh.wav" % utt_id[0])
pp_data.create_folder(os.path.dirname(out_path))
pp_data.write_audio(out_path, s, 16000)
tf.logging.info("Write inferred%s" % (np.shape(s)))
#writer.write_next_utt(write_ark_path, utt_id[0], save_result)
tf.logging.info("Write inferred %s to %s" %
(utt_id[0], out_path))
except Exception, e:
# Report exceptions to the coordinator.
coord.request_stop(e)
finally:
def decode():
"""Decoding the inputs using current model."""
tf.logging.info("Get TEST sets number.")
num_batch = get_num_batch(FLAGS.test_list_file, infer=True)
with tf.Graph().as_default():
with tf.device('/cpu:0'):
with tf.name_scope('input'):
data_list = read_list(FLAGS.test_list_file)
test_utt_id, test_inputs, _ = get_batch(
data_list,
batch_size=1,
input_size=FLAGS.input_dim,
output_size=FLAGS.output_dim,
left=FLAGS.left_context,
right=FLAGS.right_context,
num_enqueuing_threads=FLAGS.num_threads,
num_epochs=1,
infer=True)
devices = []
for i in xrange(FLAGS.num_gpu):
device_name = ("/gpu:%d" % i)
print('Using device: ', device_name)
devices.append(device_name)
# Prevent exhausting all the gpu memories.
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
config.allow_soft_placement = True
# execute the session
with tf.Session(config=config) as sess:
# Create two models with tr_inputs and cv_inputs individually.
with tf.name_scope('model'):
model = GAN(sess,
FLAGS,
devices,
test_inputs,
labels=None,
cross_validation=True)
show_all_variables()
init = tf.group(tf.global_variables_initializer(),
tf.local_variables_initializer())
print("Initializing variables ...")
sess.run(init)
if model.load(model.save_dir, moving_average=True):
print("[*] Load SUCCESS")
else:
print("[!] Load failed. Checkpoint not found. Exit now.")
sys.exit(1)
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
cmvn_filename = os.path.join(FLAGS.data_dir, "train_cmvn.npz")
if os.path.isfile(cmvn_filename):
cmvn = np.load(cmvn_filename)
else:
tf.logging.fatal("%s not exist, exit now." % cmvn_filename)
sys.exit(1)
out_dir_name = os.path.join(FLAGS.save_dir, 'test')
if not os.path.exists(out_dir_name):
os.makedirs(out_dir_name)
write_scp_path = os.path.join(out_dir_name, 'feats.scp')
write_ark_path = os.path.join(out_dir_name, 'feats.ark')
writer = ArkWriter(write_scp_path)
try:
for batch in range(num_batch):
if coord.should_stop():
break
outputs = model.generator(test_inputs, None, reuse=True)
outputs = tf.reshape(outputs, [-1, model.output_dim])
utt_id, activations = sess.run([test_utt_id, outputs])
sequence = activations * cmvn['stddev_labels'] + \
cmvn['mean_labels']
save_result = np.vstack(sequence)
writer.write_next_utt(write_ark_path, utt_id[0],
save_result)
tf.logging.info("Write inferred %s to %s" %
(utt_id[0], write_ark_path))
except Exception, e:
# Report exceptions to the coordinator.
coord.request_stop(e)
finally:
def decode():
"""Decoding the inputs using current model."""
tf.logging.info("Get TEST sets number.")
num_batch = get_num_batch(FLAGS.test_list_file, infer=True)
with tf.Graph().as_default():
with tf.device('/cpu:0'):
with tf.name_scope('input'):
data_list = read_list(FLAGS.test_list_file)
test_utt_id, test_inputs, \
lengths = get_batch(data_list,
batch_size=1,
input_size=FLAGS.input_dim,
output_size=FLAGS.output_dim,
left_context=FLAGS.left_context,
right_context=FLAGS.right_context,
num_threads=FLAGS.num_threads,
num_epochs=1,
infer=True)
devices = []
for i in range(FLAGS.num_gpu):
device_name = ("/gpu:%d" % i)
print('Using device: ', device_name)
devices.append(device_name)
# Prevent exhausting all the gpu memories.
config = tf.ConfigProto()
#config.gpu_options.allow_growth = True
config.gpu_options.per_process_gpu_memory_fraction = 0.45
config.allow_soft_placement = True
# execute the session
with tf.Session(config=config) as sess:
# Create two models with tr_inputs and cv_inputs individually.
with tf.name_scope('model'):
model = RNNTrainer(sess,
FLAGS,
devices,
test_inputs,
labels=None,
lengths=lengths,
cross_validation=True)
show_all_variables()
init = tf.group(tf.global_variables_initializer(),
tf.local_variables_initializer())
print("Initializing variables ...")
sess.run(init)
if model.load(model.save_dir, moving_average=False):
print("[*] Load model SUCCESS")
else:
print("[!] Load failed. Checkpoint not found. Exit now.")
sys.exit(1)
cmvn_filename = os.path.join(FLAGS.data_dir, "train_cmvn.npz")
if os.path.isfile(cmvn_filename):
cmvn = np.load(cmvn_filename)
else:
tf.logging.fatal("%s not exist, exit now." % cmvn_filename)
sys.exit(1)
out_dir_name = os.path.join(FLAGS.save_dir, 'test')
if not os.path.exists(out_dir_name):
os.makedirs(out_dir_name)
write_scp_path = os.path.join(out_dir_name, 'feats.scp')
write_ark_path = os.path.join(out_dir_name, 'feats.ark')
writer = ArkWriter(write_scp_path)
start = datetime.datetime.now()
outputs = model.generator(test_inputs, None, lengths, reuse=True)
outputs = tf.reshape(outputs, [-1, model.output_dim])
for batch in range(num_batch):
try:
# outputs = model.generator(test_inputs, None, lengths, reuse=True)
# outputs = tf.reshape(outputs, [-1, model.output_dim])
utt_id, activations = sess.run([test_utt_id, outputs])
sequence = activations * cmvn['stddev_labels'] + \
cmvn['mean_labels']
save_result = np.vstack(sequence)
writer.write_next_utt(write_ark_path, utt_id[0],
save_result)
print("[{}/{}] Write inferred {} to {}".format(
batch + 1, num_batch, utt_id[0], write_ark_path))
sys.stdout.flush()
except tf.errors.OutOfRangeError:
tf.logging.error(
"Out of range error occured when decoding.")
break
sess.close()
end = datetime.datetime.now()
duration = (end - start).total_seconds()
print("Decoding time is {}s".format(duration))
sys.stdout.flush()
tf.logging.info("Decoding Done.")