改进后的代码与原文res stack有一些区别

[silentsaber]

您在复现residual stack模块时，代码如下：

def residual_stack(Xm,kennel_size,Seq,pool_size):
    #1*1 Conv Linear
    Xm = Conv2D(32, (1, 1), padding='same', name=Seq+"_conv1", kernel_initializer='glorot_normal',data_format=data_format)(Xm)
    #Residual Unit 1
    Xm_shortcut = Xm
    Xm = Conv2D(32, kennel_size, padding='same',activation="relu",name=Seq+"_conv2", kernel_initializer='glorot_normal',data_format=data_format)(Xm)
    Xm = Conv2D(32, kennel_size, padding='same', name=Seq+"_conv3", kernel_initializer='glorot_normal',data_format=data_format)(Xm)
    Xm = layers.add([Xm,Xm_shortcut])
    Xm = Activation("relu")(Xm)
    #Residual Unit 2
    Xm_shortcut = Xm
    Xm = Conv2D(32, kennel_size, padding='same',activation="relu",name=Seq+"_conv4", kernel_initializer='glorot_normal',data_format=data_format)(Xm)
    !!!X = Conv2D(32, kennel_size, padding='same', name=Seq+"_conv5", kernel_initializer='glorot_normal',data_format=data_format)(Xm)!!!
    Xm = layers.add([Xm,Xm_shortcut])
    Xm = Activation("relu")(Xm)
    #MaxPooling
    Xm = MaxPooling2D(pool_size=pool_size, strides=pool_size, padding='valid', data_format=data_format)(Xm)
    return Xm

在我加"!!!"的语句处（即建立unit2的第3行语句中），返回值为X而不是Xm。这会导致您的Res stack模块缺少一个卷积层。这一点我在您打印出的网络结构中(summary)也得到了验证。

我不清楚您这里是有意改动，以便使得性能更好；还是无心之失。当我将X更正为Xm时性能反而有退化（学习率1e-3），因此或许您的此处改动反而可能使得性能提升？