def valid(self):
chopper = Chopper()
params = eval(chopper.params)
params["upper"] = False
chopper.name = "sliding_full"
params["step"] = 32
params["scale"] = 128
chopper.params = str(params)
chop = chopper.get()
x_valid, y_valid = self.prepare_data(chop, self.validation_tracks)
x_valid = remove_track_boundaries(x_valid)
y_valid = remove_track_boundaries(y_valid)
return x_valid, y_valid
def random(self):
chopper = Chopper()
# only one slices at a time is needed
if chopper.params:
chopparams = eval(chopper.params)
else:
chopparams = {}
chopparams["slices"] = 1
chopper.params = str(chopparams)
# get full or partial depending on chopper
if "full" in chopper.name:
chopper.name = "random_full"
else:
chopper.name = "random"
chop = chopper.get()
def generator(features, labels, batch_size):
shape = self._calculate_shape(features[0].shape)
# Create empty arrays to contain batch of features and labels
batch_features = np.zeros((batch_size, *shape))
batch_labels = np.zeros((batch_size, *shape))
n_tracks = len(features)
while True:
for i in range(batch_size):
# get random track
t = random.randrange(n_tracks)
x_track = features[t]
y_track = labels[t]
feature, label = chop(x_track, y_track)
batch_features[i] = feature[0]
batch_labels[i] = label[0]
yield batch_features, batch_labels
return generator
def process_spectrogram(self, spectrogram, channels=1):
chopper = Chopper()
chopper.name = "infer"
chopper.params = "{'scale': %d}" % self.config.inference_slice
chop = chopper.get(both=False)
slices = chop(spectrogram)
normalizer = Normalizer()
normalize = normalizer.get(both=False)
denormalize = normalizer.get_reverse()
new_spectrogram = np.zeros((spectrogram.shape[0], 0, channels))
for slice in slices:
# normalize
slice, norm = normalize(slice)
epanded_spectrogram = conversion.expand_to_grid(
slice, self.peakDownscaleFactor, channels)
epanded_spectrogram_with_batch_and_channels = \
epanded_spectrogram[np.newaxis, :, :]
predicted_spectrogram_with_batch_and_channels = self.model.predict(
epanded_spectrogram_with_batch_and_channels)
# o /// o
predicted_spectrogram = \
predicted_spectrogram_with_batch_and_channels[0, :, :, :]
local_spectrogram = predicted_spectrogram[:slice.shape[0], :slice.
shape[1], :]
# de-normalize
local_spectrogram = denormalize(local_spectrogram, norm)
new_spectrogram = np.concatenate(
(new_spectrogram, local_spectrogram), axis=1)
console.log("Processed spectrogram")
return spectrogram, new_spectrogram