Question

我为keras Sequential API编写了一个自定义层，但是当我用Mnist测试它时，它不能一概而论。

class ResidualLayer(Layer):

def __init__(self, conv_layer1, conv_layer2, **kwargs):
    self.__conv_layer1 = conv_layer1
    self.__conv_layer2 = conv_layer2
    super(ResidualLayer, self).__init__(**kwargs)

def build(self, input_shape):
    self.__in_shape = input_shape
    self.bn1 = BatchNormalization()
    self.bn2 = BatchNormalization()
    self.rel1 = Activation(activation='relu')
    self.rel2 = Activation(activation='relu')
    self.add = Add()
    self.scale_conv = Conv2D(filters=self.__conv_layer1.filters,
                             strides=self.__conv_layer1.strides,
                             kernel_size=(1, 1),
                             padding=self.__conv_layer1.padding)
    super(ResidualLayer, self).build(input_shape)

def call(self, x):
    input_layer = x
    tmp_layer = self.__conv_layer1(input_layer)
    tmp_layer = self.bn1(tmp_layer)
    tmp_layer = self.rel1(tmp_layer)
    tmp_layer = self.__conv_layer2(tmp_layer)
    tmp_layer = self.bn2(tmp_layer)
    if self.__conv_layer1.output_shape == self.__in_shape:
        tmp_layer = self.add([tmp_layer, input_layer])
    else:
        scaled_layer = self.scale_conv(input_layer)
        tmp_layer = self.add([tmp_layer, scaled_layer])
    tmp_layer = self.rel2(tmp_layer)
    return tmp_layer

def compute_output_shape(self, input_shape):
    return self.__conv_layer2.output_shape

模型架构不是很复杂，这引起了我的怀疑。
我使用Functional API和一个简单函数重新实现了该模型：

def reslayer(x, conv_layer1, conv_layer2):
    input_layer = x
    tmp_layer = conv_layer1(input_layer)
    tmp_layer = BatchNormalization()(tmp_layer)
    tmp_layer = Activation(activation='relu')(tmp_layer)
    tmp_layer = conv_layer2(tmp_layer)
    tmp_layer = BatchNormalization()(tmp_layer)
    if conv_layer1.output_shape == keras.backend.shape(x):
        tmp_layer = Add()([tmp_layer, input_layer])
    else:
        scale_conv = Conv2D(filters=conv_layer1.filters, strides=conv_layer1.strides, kernel_size=(1, 1), padding=conv_layer1.padding)
        scaled_layer = scale_conv(input_layer)
        tmp_layer = Add()([tmp_layer, scaled_layer])
        tmp_layer = Activation(activation='relu')(tmp_layer)
    return tmp_layer

这可以按预期工作并具有良好的泛化效果（即〜95％，而〜20％）！
但是我无法理解这两种解决方案之间的区别，并希望它们能得出相同的结果。
我想我对定制层有误解，有人可以解释一下错误以及如何正确实现定制层吗？

ResidualNetwork的keras自定义层比功能性API差很多

0 个答案: