Question

VGG16 网络

上图中，每一列对应一种结构配置。 重点看 D 类配置，包含：

• 13 个卷积层（Convolutional Layer）
• 3 个全连接层（Fully connected Layer）
• 5 个池化层（Pool layer）,分别用 maxpool 表示

其中，卷积层和全连接层具有权重系数，因此也被称为权重层，总数目为 13+3=16，这即是

VGG16 中 16 的来源。(池化层不涉及权重，因此不属于权重层，不被计数)。

这里我们将全连接层简化，全局池化（global pooling）后直接 softmax 输出 12 个类的概率

• “global pooling”就是 pooling 的 滑窗 size 和整张 feature map 的 size 一样大。这样，每个 W×H×C 的 feature map 输入就会被转化为 1×1×C 输出
• 全局池化取代全连接层可以大大减少我们需要训练的参数

class VGG16net(fluid.dygraph.Layer):
    def __init__(self):
        super(VGG16net,self).__init__()
        #这里可以使用循环，我没有用，是为了调整网络的时候方便，网络更直观，方便大家理解
        self.block1_conv1_3_64=fluid.dygraph.Conv2D(num_channels=3,num_filters=64,filter_size=(3,3),stride=1,padding=1,act='relu')
        self.block1_conv2_3_64=fluid.dygraph.Conv2D(num_channels=64,num_filters=64,filter_size=(3,3),stride=1,padding=1,act='relu')
        self.block1_maxpool1=fluid.dygraph.Pool2D(pool_size=(2,2),pool_stride=2,pool_padding=0,pool_type='max')
        self.block2_conv1_3_128=fluid.dygraph.Conv2D(num_channels=64,num_filters=128,filter_size=(3,3),stride=1,padding=1,act='relu')
        self.block2_conv2_3_128=fluid.dygraph.Conv2D(num_channels=128,num_filters=128,filter_size=(3,3),stride=1,padding=1,act='relu')
        self.block2_maxpool1=fluid.dygraph.Pool2D(pool_size=(2,2),pool_stride=2,pool_padding=0,pool_type='max')
        self.block3_conv1_3_256=fluid.dygraph.Conv2D(num_channels=128,num_filters=256,filter_size=(3,3),stride=1,padding=1,act='relu')
        self.block3_conv2_3_256=fluid.dygraph.Conv2D(num_channels=256,num_filters=256,filter_size=(3,3),stride=1,padding=1,act='relu')
        self.block3_conv3_3_256=fluid.dygraph.Conv2D(num_channels=256,num_filters=256,filter_size=(3,3),stride=1,padding=1,act='relu')
        self.block3_maxpool1=fluid.dygraph.Pool2D(pool_size=(2,2),pool_stride=2,pool_padding=0,pool_type='max')
        self.block4_conv1_3_512=fluid.dygraph.Conv2D(num_channels=256,num_filters=512,filter_size=(3,3),stride=1,padding=1,act='relu')
        self.block4_conv2_3_512=fluid.dygraph.Conv2D(num_channels=512,num_filters=512,filter_size=(3,3),stride=1,padding=1,act='relu')
        self.block4_conv3_3_512=fluid.dygraph.Conv2D(num_channels=512,num_filters=512,filter_size=(3,3),stride=1,padding=1,act='relu')
        self.block4_maxpool1=fluid.dygraph.Pool2D(pool_size=(2,2),pool_stride=2,pool_padding=0,pool_type='max')
        self.block5_conv1_3_512=fluid.dygraph.Conv2D(num_channels=512,num_filters=512,filter_size=(3,3),stride=1,padding=1,act='relu')
        self.block5_conv2_3_512=fluid.dygraph.Conv2D(num_channels=512,num_filters=512,filter_size=(3,3),stride=1,padding=1,act='relu')
        self.block5_conv3_3_512=fluid.dygraph.Conv2D(num_channels=512,num_filters=512,filter_size=(3,3),stride=1,padding=1,act='relu')
        self.block5_maxpool1=fluid.dygraph.Pool2D(global_pooling=True,pool_type='max')  #全局池化层
        self.fc1=fluid.dygraph.Linear(input_dim=512,output_dim=12,act='softmax')

    def normal_fward(self,x):
        '''
        用于不需要 show featuremap 的训练过程
        '''
        x=self.block1_conv1_3_64(x)
        x=self.block1_conv2_3_64(x)
        x=self.block1_maxpool1(x)
        x=self.block2_conv1_3_128(x)
        x=self.block2_conv2_3_128(x)
        x=self.block2_maxpool1(x)
        x=self.block3_conv1_3_256(x)
        x=self.block3_conv2_3_256(x)
        x=self.block3_conv3_3_256(x)
        x=self.block3_maxpool1(x)
        x=self.block4_conv1_3_512(x)
        x=self.block4_conv2_3_512(x)
        x=self.block4_conv3_3_512(x)
        x=self.block4_maxpool1(x)
        x=self.block5_conv1_3_512(x)
        x=self.block5_conv2_3_512(x)
        x=self.block5_conv3_3_512(x)
        x=self.block5_maxpool1(x)
        x=fluid.layers.squeeze(x,axes=[])    #多余的维度去除(32,1,1,512)-->(32,512)
        x=self.fc1(x)
        return x

    def Reverse_forward(self,x):
        '''
        正向传播之后，反向卷积并且显示特征图
        '''
        x=self.block1_conv1_3_64(x)
        x=self.block1_conv2_3_64(x)
        block1_maxpool1_temp=x
        x=self.block1_maxpool1(x)
        x=self.block2_conv1_3_128(x)
        x=self.block2_conv2_3_128(x)
        block2_maxpool1_temp=x
        x=self.block2_maxpool1(x)
        x=self.block3_conv1_3_256(x)
        x=self.block3_conv2_3_256(x)
        x=self.block3_conv3_3_256(x)
        block3_maxpool1_temp=x
        x=self.block3_maxpool1(x)
        x=self.block4_conv1_3_512(x)
        x=self.block4_conv2_3_512(x)
        x=self.block4_conv3_3_512(x)
        block4_maxpool1_temp=x
        x=self.block4_maxpool1(x)
        x=self.block5_conv1_3_512(x)
        x=self.block5_conv2_3_512(x)
        x=self.block5_conv3_3_512(x)

        #反向卷积，池化
        rblock5_conv3 = Reverse_Conv('block5_conv3',512,self.block5_conv3_3_512,512,512,3)
        rx = rblock5_conv3.reverse_conv(x)
        rblock5_conv2 = Reverse_Conv('block5_conv2',512,self.block5_conv2_3_512,512,512,3)
        rx = rblock5_conv2.reverse_conv(rx)
        rblock5_conv1 = Reverse_Conv('block5_conv1',512,self.block5_conv1_3_512,512,512,3)
        rx = rblock5_conv1.reverse_conv(rx)
        rblock4_maxpool = Reverse_Pooling(block4_maxpool1_temp,2,2,2,0)
        rx = rblock4_maxpool.backward(rx)
        rblock4_conv3 = Reverse_Conv('block4_conv3',512,self.block4_conv3_3_512,512,512,3)
        rx = rblock4_conv3.reverse_conv(rx)
        rblock4_conv2 = Reverse_Conv('block4_conv2',512,self.block4_conv2_3_512,512,512,3)
        rx = rblock4_conv2.reverse_conv(rx)
        rblock4_conv1 = Reverse_Conv('block4_conv1',256,self.block4_conv1_3_512,512,256,3)
        rx = rblock4_conv1.reverse_conv(rx)
        rblock3_maxpool = Reverse_Pooling(block3_maxpool1_temp,2,2,2,0)
        rx = rblock3_maxpool.backward(rx)
        rblock3_conv3 = Reverse_Conv('block3_conv3',256,self.block3_conv3_3_256,256,256,3)
        rx = rblock3_conv3.reverse_conv(rx)
        rblock3_conv2 = Reverse_Conv('block3_conv2',256,self.block3_conv2_3_256,256,256,3)
        rx = rblock3_conv2.reverse_conv(rx)
        rblock3_conv1 = Reverse_Conv('block3_conv1',256,self.block3_conv1_3_256,256,128,3)
        rx = rblock3_conv1.reverse_conv(rx)
        rblock2_maxpool = Reverse_Pooling(block2_maxpool1_temp,2,2,2,0)
        rx = rblock2_maxpool.backward(rx)
        rblock2_conv2 = Reverse_Conv('block2_conv2',128,self.block2_conv2_3_128,128,128,3)
        rx = rblock2_conv2.reverse_conv(rx)
        rblock2_conv1 = Reverse_Conv('block2_conv1',128,self.block2_conv1_3_128,128,64,3)
        rx = rblock2_conv1.reverse_conv(rx)
        rblock1_maxpool = Reverse_Pooling(block1_maxpool1_temp,2,2,2,0)
        rx = rblock1_maxpool.backward(rx)
        rblock1_conv2 = Reverse_Conv('block1_conv2',64,self.block1_conv2_3_64,64,64,3)
        rx = rblock1_conv2.reverse_conv(rx)
        rblock1_conv1 = Reverse_Conv('block1_conv1',64,self.block1_conv1_3_64,64,3,3)
        rx = rblock1_conv1.reverse_conv(rx)

        x=self.block5_maxpool1(x)
        x=fluid.layers.squeeze(x,axes=[])
        x=self.fc1(x)
        return x

    def forward(self,x,is_display_feature=False):
        if is_display_feature:
            x = self.Reverse_forward(x)
        else:
            x = self.normal_fward(x)
        return x