def validate_VGG16(cls):
# Create a VGG16 Model without top layers
model_base = VGG16(include_top=False, input_shape=CNN.input_shape())
# Create top layers from own specifications
model_top = CNN.get_top_layers(input_shape=model_base.output_shape[1:])
model_top.load_weights(config.URL_TOP_MODEL)
# Combine the 2 models into a full model
model_full = Sequential()
for layer in model_base.layers:
model_full.add(layer)
model_full.add(model_top)
# Lock the pre-trained VGG16 layers
for layer in model_full.layers[:25]:
layer.trainable = False
# Compile the full model
model_full.compile(
loss='categorical_crossentropy',
optimizer=Adam(),
metrics=['accuracy']
)
# Run the validation
cls.run_validation(model_full)
def train_top_layers(cls):
top_layers = CNN.get_top_layers(input_shape=CNN.input_shape())
top_layers_compiled = CNN.compile_model(top_layers)
cls.fit_model(top_layers_compiled, save_full=False)