def predict(args):
assets = AssetManager(args.base_dir)
storage = PredictionStorage(
assets.create_prediction_storage(args.model, args.data_name))
# Change path of args.model manually
network = SpeechEnhancementNetwork.load(
assets.get_model_cache_path(args.model))
with open(assets.get_normalization_cache_path(args.model),
'rb') as normalization_fd:
video_normalizer = pickle.load(normalization_fd)
try:
video_normalizer.normalize(extract_frames("data"))
predicted_speech_spectrograms = network.predict(
extract_audio()[1], extract_frames("data"))
predicted_speech_signal = data_processor.reconstruct_speech_signal(
extract_audio()[0], predicted_speech_spectrograms, 30)
predicted_speech_signal.save_to_wav_file("enhanced.wav")
except Exception:
logging.exception("failed to predict %s. skipping" % "test")
def predict(args):
assets = AssetManager(args.base_dir)
storage = PredictionStorage(
assets.create_prediction_storage(args.model, args.data_name)
)
network = SpeechEnhancementNetwork.load(assets.get_model_cache_path(args.model))
with open(
assets.get_normalization_cache_path(args.model), "rb"
) as normalization_fd:
video_normalizer = pickle.load(normalization_fd)
samples = load_preprocessed_blob(assets.get_preprocessed_blob_path(args.data_name))
for sample in samples:
try:
print(
"predicting (%s, %s)..."
% (sample.video_file_path, sample.noise_file_path)
)
video_normalizer.normalize(sample.video_samples)
loss = network.evaluate(
sample.mixed_spectrograms,
sample.video_samples,
sample.speech_spectrograms,
)
print("loss: %f" % loss)
predicted_speech_spectrograms = network.predict(
sample.mixed_spectrograms, sample.video_samples
)
predicted_speech_signal = data_processor.reconstruct_speech_signal(
sample.mixed_signal,
predicted_speech_spectrograms,
sample.video_frame_rate,
)
storage.save_prediction(sample, predicted_speech_signal)
except Exception:
logging.exception("failed to predict %s. skipping" % sample.video_file_path)
def predict(self, predict_queue, play_queue, lock):
with lock:
print("*Start Predict*\n")
counter = 0
while True:
video_data, mixed_signal, mixed_spectrograms, video_slice_number, audio_slice_number = predict_queue.get(
)
# Spectrogram Test
# self.collect_frames_for_saving(mixed_signal, self.save_original, object_flag=False)
# after_spectrograms = reconstruct_speech_signal(mixed_signal, mixed_spectrograms, self.fps)
# self.collect_frames_for_saving(after_spectrograms, self.save_signal, object_flag=True)
counter += 1
try:
predicted_speech_spectrograms = self.network.predict(
mixed_spectrograms, video_data)
predicted_speech_signal = reconstruct_speech_signal(
mixed_signal, predicted_speech_spectrograms, self.fps)
self.collect_frames_for_saving(predicted_speech_signal,
self.save_predicted,
object_flag=True)
play_queue.put((predicted_speech_signal, video_slice_number,
audio_slice_number))
if (video_slice_number == 0 and predict_queue.empty()) or\
(audio_slice_number == 0 and predict_queue.empty()):
with lock:
print("*Predict queue is empty*\n")
print("Predict: " + str(counter))
predict_queue.close()
break
except Exception:
logging.exception("Failed to predict")