I have the following Siamese Network (Xception-Based) that is composed as follows:
from tensorflow.keras.models import Model, model_from_json, Sequential
from tensorflow.keras.preprocessing.image import ImageDataGenerator, array_to_img, img_to_array, load_img
from tensorflow.keras.callbacks import TensorBoard
from tensorflow.keras.layers import Dense, Dropout, Activation, Flatten, Conv2D, MaxPooling2D, SeparableConv2D, UpSampling2D, BatchNormalization, Input, GlobalAveragePooling2D
from tensorflow.keras.regularizers import l2
from tensorflow.keras.optimizers import SGD, RMSprop
from tensorflow.keras.utils import to_categorical
from tensorflow.keras.utils import plot_model
def entry_flow(inputs) :
x = Conv2D(32, 3, strides = 2, padding='same')(inputs)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = Conv2D(64,3,padding='same')(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
previous_block_activation = x
for size in [128, 256, 728] :
x = Activation('relu')(x)
x = SeparableConv2D(size, 3, padding='same')(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = SeparableConv2D(size, 3, padding='same')(x)
x = BatchNormalization()(x)
x = MaxPooling2D(3, strides=2, padding='same')(x)
residual = Conv2D(size, 1, strides=2, padding='same')(previous_block_activation)
x = tensorflow.keras.layers.Add()([x, residual])
previous_block_activation = x
return x
def middle_flow(x, num_blocks=1) :
previous_block_activation = x
for _ in range(num_blocks) :
x = Activation('relu')(x)
x = SeparableConv2D(728, 3, padding='same')(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = SeparableConv2D(728, 3, padding='same')(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = SeparableConv2D(728, 3, padding='same')(x)
x = BatchNormalization()(x)
x = tensorflow.keras.layers.Add()([x, previous_block_activation])
previous_block_activation = x
return x
def exit_flow(x) :
previous_block_activation = x
x = Activation('relu')(x)
x = SeparableConv2D(728, 3, padding='same')(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = SeparableConv2D(1024, 3, padding='same')(x)
x = BatchNormalization()(x)
x = MaxPooling2D(3, strides=2, padding='same')(x)
residual = Conv2D(1024, 1, strides=2, padding='same')(previous_block_activation)
x = tensorflow.keras.layers.Add()([x, residual])
x = Activation('relu')(x)
x = SeparableConv2D(728, 3, padding='same')(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = SeparableConv2D(1024, 3, padding='same')(x)
x = BatchNormalization()(x)
x = GlobalAveragePooling2D()(x)
x = Dense(48, activation='linear')(x)
return x
To build that Siamese Network, I just need to call the following code:
def build_siamese_model(inputShape):
# specify the inputs for the feature extractor network
inputs = Input(shape=(inputShape))
outputs = exit_flow(middle_flow(entry_flow(inputs)))
xception = Model(inputs, outputs)
# return the model to the calling function
return exception
I have already trained this model in a given dataset, and all I want to do now is to build the model again, load the previous weights, freeze the layers (except the Dense layer with 48 neurons) and train the Dense (fully-connected) layer with a new dataset. How can I do that with Siamese Networks? I could only do that with Imagenet CNNs, but I don't know how to do the same with custom Siamese Networks.
