def create_tfrecords(params, data_dir, debug=False):
"""Create a dataset of TFRecords for VAD.
Args:
params (dict): dataset parameters
data_dir (str): path to data directory
debug (bool, optional): debug with a small amount of data. Defaults to False.
"""
tf.logging.set_verbosity(tf.logging.INFO)
np.random.seed(0)
output_path = os.path.join(data_dir, "tfrecords/")
if not tf.gfile.IsDirectory(output_path):
tf.gfile.MakeDirs(output_path)
input_size = params["input_size"]
data_split = params["data_split"]
num_shards = params["num_shards"]
data_type = params["data_type"]
# Data & label directories
label_dir = os.path.join(data_dir, "labels/")
data_dir = os.path.join(data_dir, "test-clean/")
# Split data on files
train, val, test = split_data(label_dir, data_split, random_seed=0)
tot_files = len(train) + len(val) + len(test)
logger.info(f"Total files: {tot_files}")
logger.info(f"Train/val/test split: {len(train)}/{len(val)}/{len(test)}")
train_it = slice_iter(data_dir, label_dir, train, input_size)
val_it = slice_iter(data_dir, label_dir, val, input_size)
test_it = slice_iter(data_dir, label_dir, test, input_size)
# Write data to tfrecords format
nmax = 100 if debug else -1
if "train" in data_type:
logger.info("Writing train tfrecords ...")
train_path = os.path.join(output_path, "train/")
if not tf.gfile.IsDirectory(train_path):
tf.gfile.MakeDirs(train_path)
write_tfrecords(train_path, train_it, num_shards, nmax=nmax)
if "val" in data_type:
logger.info("Writing val tfrecords ...")
val_path = os.path.join(output_path, "val/")
if not tf.gfile.IsDirectory(val_path):
tf.gfile.MakeDirs(val_path)
write_tfrecords(val_path, val_it, num_shards, nmax=nmax)
if "test" in data_type:
logger.info("Writing test tfrecords ...")
test_path = os.path.join(output_path, "test/")
if not tf.gfile.IsDirectory(test_path):
tf.gfile.MakeDirs(test_path)
write_tfrecords(test_path, test_it, num_shards, nmax=nmax)
def create_tfrecords(
data_dir,
seq_len=1024,
split="0.7/0.15",
num_shards=256,
debug=False,
data_type="trainval",
):
np.random.seed(0)
output_path = os.path.join(data_dir, "tfrecords/")
if not tf.gfile.IsDirectory(output_path):
tf.gfile.MakeDirs(output_path)
# Data & label directories
label_dir = os.path.join(data_dir, "labels/")
data_dir = os.path.join(data_dir, "test-clean/")
# Split data on files
train, val, test = split_data(label_dir, split, random_seed=0)
tot_files = len(train) + len(val) + len(test)
logger.info(f"Total files: {tot_files}")
logger.info(f"Train/val/test split: {len(train)}/{len(val)}/{len(test)}")
train_it = slice_iter(data_dir, label_dir, train, seq_len)
val_it = slice_iter(data_dir, label_dir, val, seq_len)
test_it = slice_iter(data_dir, label_dir, test, seq_len)
# Write data to tfrecords format
nmax = 100 if debug else -1
if "train" in data_type:
logger.info("Writing train tfrecords ...")
train_path = os.path.join(output_path, "train/")
if not tf.gfile.IsDirectory(train_path):
tf.gfile.MakeDirs(train_path)
write_tfrecords(train_path, train_it, num_shards, nmax=nmax)
if "val" in data_type:
logger.info("Writing val tfrecords ...")
val_path = os.path.join(output_path, "val/")
if not tf.gfile.IsDirectory(val_path):
tf.gfile.MakeDirs(val_path)
write_tfrecords(val_path, val_it, num_shards, nmax=nmax)
if "test" in data_type:
logger.info("Writing test tfrecords ...")
test_path = os.path.join(output_path, "test/")
if not tf.gfile.IsDirectory(test_path):
tf.gfile.MakeDirs(test_path)
write_tfrecords(test_path, test_it, num_shards, nmax=nmax)
def create_tfrecords(data_dir,
seq_len=1024,
split='0.7/0.15',
num_shards=256,
debug=False,
data_type='trainval'):
np.random.seed(0)
output_path = os.path.join(data_dir, 'tfrecords/')
if not tf.gfile.IsDirectory(output_path):
tf.gfile.MakeDirs(output_path)
# Data & label directories
label_dir = os.path.join(data_dir, 'labels/')
data_dir = os.path.join(data_dir, 'test-clean/')
# Split data on files
train, val, test = split_data(label_dir, split, random_seed=0)
print('\nTotal files: {}'.format(len(train) + len(val) + len(test)))
print('Train/val/test split: {}/{}/{}'.format(len(train), len(val),
len(test)))
train_it = slice_iter(data_dir, label_dir, train, seq_len)
val_it = slice_iter(data_dir, label_dir, val, seq_len)
test_it = slice_iter(data_dir, label_dir, test, seq_len)
# Write data to tfrecords format
nmax = 100 if debug else -1
if 'train' in data_type:
print('\nWriting train tfrecords ...')
train_path = os.path.join(output_path, 'train/')
if not tf.gfile.IsDirectory(train_path):
tf.gfile.MakeDirs(train_path)
write_tfrecords(train_path, train_it, num_shards, nmax=nmax)
if 'val' in data_type:
print('\nWriting val tfrecords ...')
val_path = os.path.join(output_path, 'val/')
if not tf.gfile.IsDirectory(val_path):
tf.gfile.MakeDirs(val_path)
write_tfrecords(val_path, val_it, num_shards, nmax=nmax)
if 'test' in data_type:
print('\nWriting test tfrecords ...')
test_path = os.path.join(output_path, 'test/')
if not tf.gfile.IsDirectory(test_path):
tf.gfile.MakeDirs(test_path)
write_tfrecords(test_path, test_it, num_shards, nmax=nmax)
def main(_):
np.random.seed(0)
# Directories
data_dir = os.path.join(FLAGS.data_dir, 'test-clean/')
label_dir = os.path.join(FLAGS.data_dir, 'labels/')
_, _, test = split_data(label_dir, split='0.7/0.15', random_seed=0)
file_it = file_iter(data_dir, label_dir, files=test)
# TensorFlow inputs
features_input_ph = tf.placeholder(shape=FEAT_SIZE, dtype=tf.float32)
features_input_op = tf.transpose(features_input_ph, perm=[1, 0])
features_input_op = tf.expand_dims(features_input_op, axis=0)
# TensorFlow exported model
speech_predictor = tf.contrib.predictor.from_saved_model(export_dir=FLAGS.exported_model)
init = tf.initializers.global_variables()
classes = ['Noise', 'Speech']
# Iterate though test data
with tf.Session() as sess:
for signal, labels, fn in file_it:
sess.run(init)
print('\nPrediction on file {} ...'.format(fn))
signal_input = deque(signal[:FLAGS.seq_len].tolist(), maxlen=FLAGS.seq_len)
preds, pred_time = [], []
pointer = FLAGS.seq_len
while pointer < len(signal):
start = time()
# Preprocess signal & extract features
signal_to_process = np.copy(signal_input)
signal_to_process = np.float32(signal_to_process)
features = extract_features(signal_to_process, freq=16000, n_mfcc=5, size=512, step=16)
# Prediction
features_input = sess.run(features_input_op, feed_dict={features_input_ph: features})
speech_prob = speech_predictor({'features_input': features_input})['speech'][0]
speech_pred = classes[int(np.round(speech_prob))]
# Time prediction & processing
end = time()
dt = end - start
pred_time.append(dt)
print('Prediction = {} | proba = {:.2f} | time = {:.2f} s'.format(speech_pred, speech_prob[0], dt))
# For visualization
preds.append([pointer - FLAGS.seq_len, pointer, np.round(speech_prob)])
# Update signal segment
signal_input.extend(signal[pointer + FLAGS.stride:pointer + FLAGS.stride + FLAGS.seq_len])
pointer += FLAGS.seq_len + FLAGS.stride
print('Average prediction time = {:.2f} ms'.format(np.mean(pred_time) * 1e3))
# Smoothing & hangover
if FLAGS.smoothing:
preds = smooth_predictions(preds)
# Visualization
visualize_predictions(signal, fn, preds)
def run_inference(params, data_dir, exported_model):
"""Run Voice Activity Detection CNN inference over raw audio signals.
Args:
params (dict): dictionary of inference parameters
data_dir (str): path to raw dataset directory
exported_model (str): path to exported pre-trained TF model directory
"""
os.environ["TF_CPP_MIN_LOG_LEVEL"] = "2"
tf.logging.set_verbosity(tf.logging.INFO)
np.random.seed(0)
input_size = params["input_size"]
stride = params["stride"]
smoothing = params["smoothing"]
# Directories
test_data_dir = os.path.join(data_dir, "test-clean/")
label_dir = os.path.join(data_dir, "labels/")
_, _, test = split_data(label_dir, split="0.7/0.15", random_seed=0)
file_it = file_iter(test_data_dir, label_dir, files=test)
# TensorFlow inputs
features_input_ph = tf.placeholder(shape=FEAT_SIZE, dtype=tf.float32)
features_input_op = tf.transpose(features_input_ph, perm=[1, 0])
features_input_op = tf.expand_dims(features_input_op, axis=0)
# TensorFlow exported model
speech_predictor = tf.contrib.predictor.from_saved_model(export_dir=exported_model)
init = tf.initializers.global_variables()
classes = ["Noise", "Speech"]
# Iterate though test data
with tf.Session() as sess:
for signal, labels, fn in file_it:
sess.run(init)
logger.info(f"Prediction on file {fn} ...")
signal_input = deque(signal[:input_size].tolist(), maxlen=input_size)
preds, pred_time = [], []
pointer = input_size
while pointer < len(signal):
start = time.time()
# Preprocess signal & extract features
signal_to_process = np.copy(signal_input)
signal_to_process = np.float32(signal_to_process)
features = extract_features(
signal_to_process, freq=16000, n_mfcc=5, size=512, step=16
)
# Prediction
features_input = sess.run(
features_input_op, feed_dict={features_input_ph: features}
)
speech_prob = speech_predictor({"features_input": features_input})[
"speech"
][0]
speech_pred = classes[int(np.round(speech_prob))]
# Time prediction & processing
end = time.time()
dt = end - start
pred_time.append(dt)
logger.info(
f"Prediction = {speech_pred} | proba = {speech_prob[0]:.2f} | time = {dt:.2f} s"
)
# For visualization
preds.append([pointer - input_size, pointer, np.round(speech_prob)])
# Update signal segment
signal_input.extend(
signal[pointer + stride : pointer + stride + input_size]
)
pointer += input_size + stride
logger.info(f"Average prediction time = {np.mean(pred_time) * 1e3:.2f} ms")
# Smoothing & hangover
if smoothing:
preds = smooth_predictions(preds)
# Visualization
visualize_predictions(signal, fn, preds)