def _check_vggish_ckpt_exists():
"""check VGGish checkpoint exists or not."""
util.maybe_create_directory(FLAGS.vggish_ckpt_dir)
if not util.is_exists(vggish_ckpt_path):
url = 'https://storage.googleapis.com/audioset/vggish_model.ckpt'
util.maybe_download(url, params.VGGISH_CHECKPOINT_DIR)
if not util.is_exists(vggish_pca_path):
url = 'https://storage.googleapis.com/audioset/vggish_pca_params.npz'
util.maybe_download(url, params.VGGISH_CHECKPOINT_DIR)
def arange_urban_sound_file_by_class():
"""Arange urban sound file by it's class."""
src_paths = '/data1/data/UrbanSound8K-16bit/audio'
dst_dir = '/data1/data/UrbanSound8K-16bit/audio-classfied'
CLASSES = [
'air conditioner', 'car horn', 'children playing', 'dog bark',
'drilling', 'engine idling', 'gun shot', 'jackhammer', 'siren',
'street music'
]
CLASSES_STRIPED = [c.replace(' ', '_') for c in CLASSES]
for src in src_paths:
lbl = urban_labels([src])[0]
dst = '{dir}/{label}'.format(dir=dst_dir, label=CLASSES_STRIPED[lbl])
maybe_create_directory(dst)
maybe_copy_file(
src, '{dst}/{name}'.format(dst=dst, name=os.path.split(src)[-1]))
def convert_urban_pcm24_to_pcm16():
"""Convert urban sound codec from PCM_24 to PCM_16."""
src_dir = [
'/data1/data/UrbanSound8K/audio/fold{:d}'.format(i + 1)
for i in range(10)
]
dst_dir = [
'/data1/data/UrbanSound8K-16bit/audio/fold{:d}'.format(i + 1)
for i in range(10)
]
converted_wav_paths = []
for dsrc, ddst in zip(src_dir, dst_dir):
maybe_create_directory(ddst)
wav_files = filter(lambda FP: FP if FP.endswith('.wav') else None,
[FP for FP in os.listdir(dsrc)])
for wav_file in wav_files:
src_wav, dst_wav = os.path.join(dsrc, wav_file), os.path.join(
ddst, wav_file)
convert_wav(src_wav, dst_wav, subtype='PCM_16')
converted_wav_paths.append(dst_wav)
print('converted count:', len(converted_wav_paths))
print(converted_wav_paths, len(converted_wav_paths))
def _create_records():
"""Create audio `train`, `test` and `val` records file."""
tf.logging.info("Create records..")
util.maybe_create_directory(FLAGS.records_dir)
_check_vggish_ckpt_exists()
wav_files, wav_labels = _wav_files_and_labels()
tf.logging.info('Possible labels: {}'.format(set(wav_labels)))
train, test, val = util.train_test_val_split(wav_files, wav_labels)
with VGGishExtractor(vggish_ckpt_path, vggish_pca_path,
params.VGGISH_INPUT_TENSOR_NAME,
params.VGGISH_OUTPUT_TENSOR_NAME) as ve:
train_x, train_y = train
ve.create_records(train_records_path, train_x, train_y)
test_x, test_y = test
ve.create_records(test_records_path, test_x, test_y)
val_x, val_y = val
ve.create_records(val_records_path, val_x, val_y)
tf.logging.info('Dataset size: Train-{} Test-{} Val-{}'.format(
len(train_y), len(test_y), len(val_y)))
def create_records(self, record_path, wav_files, wav_labels):
"""Create TF Records from wav files and corresponding labels."""
record_dir = os.path.dirname(record_path)
maybe_create_directory(record_dir)
writer = tf.python_io.TFRecordWriter(record_path)
N = len(wav_labels)
n = 1
for (wav_file, wav_label) in zip(wav_files, wav_labels):
tf.logging.info('[{}/{}] Extracting VGGish feature:'
' label: {} - {}'.format(n, N, wav_label, wav_file))
n += 1
features = self.wavfile_to_features(wav_file)
num_features = features.shape[0] # one feature for one second
if num_features == 0:
tf.logging.warning('No vggish features:'
' label: {} - {}'.format(wav_label, wav_file))
continue
cur_wav_labels = [wav_label] * num_features
for (f, l) in zip(features, cur_wav_labels):
example = encodes_example(np.float64(f), np.int64(l))
writer.write(example.SerializeToString())
writer.close()
test_records_path = os.path.join(FLAGS.records_dir,
params.TF_RECORDS_TEST_NAME)
val_records_path = os.path.join(FLAGS.records_dir, params.TF_RECORDS_VAL_NAME)
vggish_ckpt_path = os.path.join(FLAGS.vggish_ckpt_dir,
params.VGGISH_CHECKPOINT_NAME)
vggish_pca_path = os.path.join(FLAGS.vggish_ckpt_dir,
params.VGGISH_PCA_PARAMS_NAME)
tensorboard_dir = os.path.join(params.TENSORBOARD_DIR, FLAGS.train_name)
audio_ckpt_dir = os.path.join(FLAGS.audio_ckpt_dir, FLAGS.train_name)
util.maybe_create_directory(tensorboard_dir)
util.maybe_create_directory(audio_ckpt_dir)
def _add_triaining_graph():
with tf.Graph().as_default() as graph:
logits = define_audio_slim(training=True)
tf.summary.histogram('logits', logits)
# define training subgraph
with tf.variable_scope('train'):
labels = tf.placeholder(tf.float32,
shape=[None, params.NUM_CLASSES],
name='labels')
cross_entropy = tf.nn.softmax_cross_entropy_with_logits_v2(
logits=logits, labels=labels, name='cross_entropy')
loss = tf.reduce_mean(cross_entropy, name='loss_op')