Handling batch normalization layers during fine-tuning (trainable vs training)

Question

I have created a CNN in Keras as shown below. I\'m using the CNN for the classification of three classes. The inputs are heat map images of several users showing the browsin