CNN学习记录： VGG16, freeze some layers, turn conv_layer5= trainable,
Overfitting problem, use regularizer rather than early stopping. For regularizer, Kernel regularizer means give penalty to W(weight) in WX+b; bias regularizer means b