def main(mode, conv_until=None):
# setup stuff to build model
# This is it. use melgram, up to 6000 (SR is assumed to be 12000, see model.py),
# do decibel scaling
assert mode in ('feature', 'tagger')
if mode == 'feature':
last_layer = False
else:
last_layer = True
if conv_until is None:
conv_until = 4
K.set_image_dim_ordering('th')
assert K.image_dim_ordering() == 'th', ('image_dim_ordering should be "th". ' +
'open ~/.keras/keras.json to change it.')
args = Namespace(tf_type='melgram', # which time-frequency to use
normalize='no', decibel=True, fmin=0.0, fmax=6000, # mel-spectrogram params
n_mels=96, trainable_fb=False, trainable_kernel=False, # mel-spectrogram params
conv_until=conv_until) # how many conv layer to use? set it 4 if tagging.
# set in [0, 1, 2, 3, 4] if feature extracting.
model = my_models.build_convnet_model(args=args, last_layer=last_layer)
model.load_weights('weights_layer{}_{}.hdf5'.format(conv_until, K._backend),
by_name=True)
model.layers[1].summary()
model.summary()
# and use it!
return model
def main(mode):
# setup stuff to build model
# This is it. use melgram, up to 6000 (SR is assumed to be 12000, see model.py),
# do decibel scaling
assert mode in ('feature', 'tagger')
if mode == 'feature':
last_layer = False
else:
last_layer = True
args = Namespace(test=False,
data_percent=100,
model_name='',
tf_type='melgram',
normalize='no',
decibel=True,
fmin=0.0,
fmax=6000,
n_mels=96,
trainable_fb=False,
trainable_kernel=False)
model = models.build_convnet_model(args=args, last_layer=last_layer)
model.load_weights('weights_{}.hdf5'.format(K._backend), by_name=True)
model.layers[1].summary()
model.summary()