Stopping Gradient back prop through a particular layer in keras

问题

x = Conv2D(768, (3, 3), padding='same', activation='relu', kernel_initializer='normal', 
           name='rpn_conv1',trainable=trainable)(base_layers)

x_class = Conv2D(num_anchors, (1, 1), activation='sigmoid', kernel_initializer='uniform', 
                 name='rpn_out_class',trainable=trainable)(x)

    # stop gradient backflow through regression layer
x_regr = Conv2D(num_anchors * 4, (1, 1), activation='linear', kernel_initializer='zero', 
                name='rpn_out_regress',trainable=trainable)(x)

How to use K.stop_gradient() to stop gradient back-prop via the regression layer (x_reg) alone?

回答1:

You need a Lambda layer for using custom functions.

x_regr_constant = Lambda(
                          lambda x: K.stop_gradient(x), 
                          output_shape=notNecessaryWithTensorflow
                        )(x_regr)

来源：https://stackoverflow.com/questions/47546100/stopping-gradient-back-prop-through-a-particular-layer-in-keras