vgg-192-define-model.py

# define the model

# init a new model with 192x192 CNN layers
# padding='same' will downsample to 96x96
# # which is the expected input size for pretrain
model = Sequential()
model.add(Conv2D(64, kernel_size=(3, 3), input_shape=(192, 192, 3), activation='relu', padding='same'))
model.add(Conv2D(64, kernel_size=(3, 3), activation='relu', padding='same'))
model.add(MaxPooling2D(pool_size=(2, 2)))

# load the pretrained model
prior = load_model('resnet96_128/model-96.h5')

# add all but the first two layers of VGG16 to the new model
# strip the input layer out, this is now 96x96
# also strip out the first convolutional layer, this took the 96x96 input and convolved it but
# this is now the job of the three new layers.
for layer in prior.layers[1:]:
    layer.name += '_prior'  # set layer names to avoid name collisions
    model.add(layer)

# the pretrained CNN layers are already marked non-trainable
# mark off the top layers as well
for layer in model.layers[-4:]:
    layer.trainable = False
    
# set layer names (otherwise names may collide)
for i, layer in enumerate(model.layers):
    layer.name = 'layer_' + str(i)

# compile the model
model.compile(
    optimizer=RMSprop(),
    loss='categorical_crossentropy',
    metrics=['accuracy']
)