def generate_batch(self, audio_paths, labels, batch_size, shuffled):
num_batches = len(audio_paths) // batch_size
while True:
batch_index_list = list(range(0, num_batches))
if shuffled:
shuffle(batch_index_list)
for batchIdx in batch_index_list:
start = batchIdx * batch_size
end = (batchIdx + 1) * batch_size
X = np.zeros(
shape=(
batch_size,
self.input_shape[0],
self.input_shape[1],
self.input_shape[2],
),
dtype=np.float32,
)
for i in range(start, end):
audio_path = audio_paths[i]
mg = compute_melgram(audio_path)
X[i - start, :, :, :] = mg
yield nd.array(X,
ctx=self.data_ctx), nd.array(labels[start:end],
ctx=self.data_ctx)
def compute_melgram(self, audio_path):
if audio_path in self.cache:
return self.cache[audio_path]
else:
mg = compute_melgram(audio_path)
# mg = (mg + 100) / 200 # scale the values
self.cache[audio_path] = mg
return mg
def encode_audio(self, audio_path):
mg = compute_melgram(audio_path)
mg = nd.array(np.expand_dims(mg, axis=0), ctx=self.model_ctx)
return self.model(mg).asnumpy()[0]