def do_inference(num_tests, concurrency=1):
channel = implementations.insecure_channel(host, int(port))
stub = prediction_service_pb2.beta_create_PredictionService_stub(channel)
coord = _Coordinator(num_tests, concurrency)
for _ in range(num_tests):
# dummy audio
duration, sr, n_fft, win_length, hop_length, n_mels, max_db, min_db = 4, 16000, 512, 512, 128, 80, 35, -55
filename = librosa.util.example_audio_file()
wav = read_wav(filename, sr=sr, duration=duration)
mel = wav2melspec_db(wav, sr, n_fft, win_length, hop_length, n_mels)
mel = normalize_db(mel, max_db=max_db, min_db=min_db)
mel = mel.astype(np.float32)
mel = np.expand_dims(mel, axis=0) # single batch
n_timesteps = sr / hop_length * duration + 1
# build request
request = predict_pb2.PredictRequest()
request.model_spec.name = 'voice_vector'
request.model_spec.signature_name = 'predict'
request.inputs['x'].CopyFrom(
tf.contrib.util.make_tensor_proto(mel,
shape=[1, n_timesteps, n_mels]))
coord.throttle()
# send asynchronous response (recommended. use this.)
result_future = stub.Predict.future(request, 10.0) # timeout
result_future.add_done_callback(_create_rpc_callback(coord))
# send synchronous response (NOT recommended)
# result = stub.Predict(request, 5.0)
coord.wait_all_done()
def wav2melspec_db(wav,
sr,
n_fft,
win_length,
hop_length,
n_mels,
normalize=False,
max_db=None,
min_db=None,
time_first=True,
**kwargs):
# Mel-spectrogram
mel_spec = wav2melspec(wav,
sr,
n_fft,
win_length,
hop_length,
n_mels,
time_first=False,
**kwargs)
# Decibel
mel_db = librosa.amplitude_to_db(mel_spec)
# Normalization
mel_db = normalize_db(mel_db, max_db, min_db) if normalize else mel_db
# Time-axis first
if time_first:
mel_db = mel_db.T # (t, n_mels)
return mel_db
def get_data(self):
while True:
speaker_id = random.choice(list(self.speaker_dict.keys()))
wav = self._load_random_wav(speaker_id)
mel_spec = wav2melspec_db(wav, hp.signal.sr, hp.signal.n_fft,
hp.signal.win_length,
hp.signal.hop_length, hp.signal.n_mels)
mel_spec = normalize_db(mel_spec,
max_db=hp.signal.max_db,
min_db=hp.signal.min_db)
yield wav, mel_spec, speaker_id