通过keras中的特定层停止梯度反向传播

Question

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)

如何使用 K.stop_gradient() 单独通过回归层 (x_reg) 停止梯度反向传播？

Answer 1

您需要 Lambda 图层才能使用自定义函数。

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

通过keras中的特定层停止梯度反向传播

Stopping Gradient back prop through a particular layer in keras

gradient

object-detection

keras

tensorflow