def test_no_effect_on_layers(self):
utils.setup_trainable_layers(self.model)
# extract layers trainable boooleans in to separate list
affected_layers_bool = []
for layer in self.model.layers:
affected_layers_bool.append(layer.trainable)
self.assertListEqual(self.layers_bool, affected_layers_bool)
def test_no_layers_are_freezing_when_number_of_specified_layers_is_0(self):
nb_layers = 0
utils.setup_trainable_layers(self.model, nb_layers)
# extract layers trainable boooleans in to separate list
affected_layers_bool = []
for layer in self.model.layers:
affected_layers_bool.append(layer.trainable)
# check layer trainable booleans are all set to True
self.assertTrue(all(affected_layers_bool))
def test_layers_after_specified_number_of_layers_are_not_freezed(self):
nb_layers = len(self.model.layers) // 2
utils.setup_trainable_layers(self.model, nb_layers)
# extract layers trainable boooleans in to separate list
affected_layers_bool = []
for layer in self.model.layers:
affected_layers_bool.append(layer.trainable)
# check layer trainable booleans prior to specified index are changed to False
self.assertTrue(all(affected_layers_bool[nb_layers:]))
def test_first_specified_number_of_layers_are_freezed(self):
nb_layers = len(self.model.layers) // 2
utils.setup_trainable_layers(self.model, nb_layers)
# extract layers trainable boooleans in to separate list
affected_layers_bool = []
for layer in self.model.layers:
affected_layers_bool.append(layer.trainable)
# check layer trainable booleans prior to specified index are changed to False
self.assertListEqual(affected_layers_bool[:nb_layers],
[False for i in range(nb_layers)])
def test_all_layers_are_freezing_when_number_of_layers_is_same_as_model_length(
self):
nb_layers = len(self.model.layers)
utils.setup_trainable_layers(self.model, nb_layers)
# extract layers trainable boooleans in to separate list
affected_layers_bool = []
for layer in self.model.layers:
affected_layers_bool.append(layer.trainable)
# check layer trainable booleans prior to specified index are changed to False
self.assertListEqual(affected_layers_bool,
[False for i in range(nb_layers)])
def build_finetuned_model(args, input_shape, fc_size):
"""
Builds a finetuned VGG model from VGGFace implementation
with no weights loaded and setting up new fresh prediction layers at last
Args:
args: necessary args needed for training like train_data_dir, batch_size etc...
input_shape: shape of input tensor
fc_size: number of nodes to be used in last layers will be based on this value i.e its multiples may be used
Returns:
finetuned vgg model
"""
# setup model
vgg, base_vgg = VGGWithCustomLayers(args.nb_classes, input_shape, fc_size)
# setup layers to be trained or not
setup_trainable_layers(vgg, args.layers_to_freeze)
# compiling the model
vgg.compile(optimizer=SGD(lr=0.0001, momentum=0.9),
loss='categorical_crossentropy',
metrics=['accuracy'])
return vg
def build_finetuned_model(args, input_shape, fc_size):
"""
Builds a finetuned InceptionV3 model from tensorflow implementation
with imagenet weights loaded and setting up new fresh prediction layers at last
Args:
args: necessary args needed for training like train_data_dir, batch_size etc...
input_shape: shape of input tensor
fc_size: number of nodes to be used in last layers will be based on this value i.e its multiples may be used
Returns:
finetuned inceptionV3 model
"""
# setup model
iv3, base_iv3 = InceptionV3WithCustomLayers(args.nb_classes, input_shape,
fc_size)
setup_trainable_layers(iv3, args.layers_to_freeze)
# compiling the model
iv3.compile(optimizer='RMSprop',
loss='categorical_crossentropy',
metrics=['accuracy'])
return iv3
def test_raises_valueError_when_number_of_layers_specified_is_negative(
self):
nb_layers = -1
with self.assertRaises(ValueError) as ve:
utils.setup_trainable_layers(self.model, nb_layers)
def test_raises_valueError_when_number_of_layers_specified_is_greater_than_available_layers(
self):
nb_layers = len(self.model.layers) + 1
with self.assertRaises(ValueError) as ve:
utils.setup_trainable_layers(self.model, nb_layers)