def inference_wav(wav_file):
"""Test audio model on a wav file."""
label = util.urban_labels([wav_file])[0]
graph = tf.Graph()
with tf.Session(graph=graph, config=SESS_CONFIG) as sess:
with VGGishExtractor(VGGISH_CKPT,
VGGISH_PCA,
audio_params.VGGISH_INPUT_TENSOR_NAME,
audio_params.VGGISH_OUTPUT_TENSOR_NAME) as ve:
vggish_features = ve.wavfile_to_features(wav_file)
assert vggish_features is not None
labels = [label] * vggish_features.shape[0]
# restore graph
# _restore_from_meta_and_ckpt(sess, META, CKPT)
_restore_from_defined_and_ckpt(sess, CKPT)
# get input and output tensor
# graph = tf.get_default_graph()
inputs = graph.get_tensor_by_name(audio_params.AUDIO_INPUT_TENSOR_NAME)
outputs = graph.get_tensor_by_name(audio_params.AUDIO_OUTPUT_TENSOR_NAME)
predictions = sess.run(outputs, feed_dict={inputs: vggish_features})
idxes = np.argmax(predictions, 1)
probs = np.max(predictions, 1)
print(predictions)
print(idxes)
print(labels)
print(probs)
acc = accuracy_score(labels, idxes)
print('acc:', acc)
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 close(self):
self.sess.close()
if __name__ == '__main__':
import audio_params
import vggish_params
import timeit
from audio_util import urban_labels
wav_file = 'F:/3rd-datasets/UrbanSound8K-16bit/audio-classified/siren/90014-8-0-1.wav'
wav_dir = 'F:/3rd-datasets/UrbanSound8K-16bit/audio-classified/siren'
wav_filenames = os.listdir(wav_dir)
wav_files = [os.path.join(wav_dir, wav_filename) for wav_filename in wav_filenames]
wav_labels = urban_labels(wav_files)
# test VGGishExtractor
time_start = timeit.default_timer()
with VGGishExtractor(audio_params.VGGISH_CHECKPOINT,
audio_params.VGGISH_PCA_PARAMS,
vggish_params.INPUT_TENSOR_NAME,
vggish_params.OUTPUT_TENSOR_NAME) as ve:
vggish_features = ve.wavfile_to_features(wav_file)
print(vggish_features, vggish_features.shape)
ve.create_records('./vggish_test.records', wav_files[:10], wav_labels[:10])
time_end = timeit.default_timer()
# print('cost time: {}s, {}s/wav'.format((time_end-time_start), (time_end-time_start)/(i+1)))