keras教程【3】使用ResNet

由于ResNet模型结构 比之前几篇介绍基础模型 复杂，这里使用Functional模型(即函数式模型)来替代 Sequential模型，因为Functional模型使用起来更加灵活，可以处理分支网络等情况。

1.构建ResNet Block.

ResNet Block包含了1个卷积层，1个BN层（对输出参数做归一化，使网络更易训练，一般地，加了BN层的网络，可以不必再用Droput层），1个激活层。

def resnet_block(inputs, num_filters = 16, 
                 kernel_size = 3,
                 strides = 1,
                 activation = 'relu'):
    x = Conv2D(num_filters, 
               kernel_size = kernel_size,
               strides = strides,
               padding = 'same',
               kernel_initializer = 'he_normal',
               kernel_regularizer = l2(1e-4))(inputs)
    x = BatchNormalization()(x)
    if (activation):
        x = Activation('relu')(x)
    return x

其中，在卷积层中加入L2正则化，可以提高模型泛化能力。

2.使用ResNet Block搭建ResNet网络

注意，在Functional模型中，必须初始化模型，才真正完成模型的搭建。

def resnet_simple(input_shape):
    inputs = Input(shape = input_shape)
    
    # 第1层
    x = resnet_block(inputs)
    
    # 第2-7层
    for i in range(6):
        a = resnet_block(inputs = x)
        b = resnet_block(inputs = a, activation=None)
        x = keras.layers.add([x, b])
        x = Activation('relu')(x)
     
     
     # 第8-13层
     for i in range(6):
         if i == 0:
             a = resnet_block(inputs = x, 
                 strides = 2,
                 num_filters = 32)
         else:
             a = resnet_block(inputs = x,
                 num_filters = 32)
         b = resnet_block(inputs = a, 
             activation = None,
             num_filters = 32)
         if i == 0:
             x = Conv2D(32, kernel_size=3,
                 strides = 2,
                 padding = 'same',
                 kernel_initializer = 'he_normal',
                 kernel_regularizer = l2(1e-4))(x)
         x = keras.layers.add([x, b])
         x = Activation('relu')(x)
         
     # 第14-19层  
     for i in range(6):
         if i == 0:
             a = resnet_block(inputs = x,
                 strides = 2,
                 num_filters = 64)
          else:
             a = resnet_block(inputs = x,
                 num_filters = 64)
          b = resnet_block(inputs = a,
              activation = None,
              num_filters = 64)
          if i == 0:
              x = Conv2D(64, kernel_size=3,
                  strides = 2,
                  padding = 'same',
                  kernel_initializer = 'he_normal',
                  kernel_regularizer = l2(1e-4))(x) 
          x = keras.layers.add([x, b])
          x = Activation('relu')(x)
       # 第20层
       x = AveragePooling2D(pool_size=2)(x)
       
       y = Flatten()(x)
       outputs = Dense(10, activation = 'softmax',
                 kernel_initializer = 'he_normal')(y)
       # 初始化模型
       model = Model(inputs = inputs,
               outputs = outputs)
       return model