经典卷积网络之DenseNet的论文解读及代码实现

文章背景

最近的工作显示如果在靠近输入输出的层数之间加入短连接（short connections）能大幅度的提升训练的深度、准确度与效率。基于这种思想，这篇论文提出了一种全新的网络结构Dense Convolutional Network(DenseNet)：将每一层与出层之外的所有层连接起来。也就是说对于有L层的传统卷积网络而言，有L个连接，但是对于DenseNet而言有L(L+1)/2个连接。

思想

全文有两个公式第一个公式是关于ResNet的，其中$\ {l}$ l表示层数，$\ {x_l}$ xl​表示第$\ {l}$ l的输出， $\ {H}$ H表示非线性变换。
$\ {x_l= H_l(x_{l-1}) + x_{l-1}}$ xl​=Hl​(xl−1​)+xl−1​、
第二个公式是关于DenseNet的
$\ x_l =H_l([x_0,x_1,x_2,...,x_{l-1}])$ xl​=Hl​([x0​,x1​,x2​,...,xl−1​])
从这两个公式可以很清楚的看出两种网络思路上的区别，resnet是通过将这层的输出与前一层的输出相加；densenet是将前面所有的输出全部concatenat。

网络结构

这是densenet中的一个dense block

这是densenet网络的结构由多个denseblock组成

代码实现

class bottleneck_layer(layers.layer):
 def __init__(self, x, filters, filters= “自定义”, stride=1):
  super(bottleneck_layer, self).__init__()
  
  #有denseblock的结构图可以看出bn+relu+conv为一个单元
  self.bn = layers.BatchNormalization()
  self.relu = layers.Activation('relu')
  self.conv1 = layers.Con2D(x, use_bias=False, filters = 4*filters, kernel_size=[1,1], strides=stride, padding='same')
  self.dp = layers.Dropout(0.2)
  
  self.conv2 = layers.Con2D(x, use_bias=False, filters = filters, kernel_size=[3,3], strides=stride, padding='same')
 
 def call(self, input, training = None):
  out = self.bn(input)
  out = self.relu(out)
  out = self.conv1(out)
  out = self.dp(out)
  
  out = self.bn(input)
  out = self.relu(out)
  out = self.conv2(out)
  out = self.dp(out)
  
  return out
class transition_layer(layers.layer):
 def __init__(self, x, filters, stride=1):
  super(transition_layer, self).__init__()
  
  self.bn = layers.BatchNormalization()
  self.relu = layers.Activation('relu')
  in_channel = x.shape[-1]
  self.conv1 = layers.Con2D(x, use_bias=False, filters = 0.5*in_channel, kernel_size=[1,1], strides=stride, padding='same')
  self.dp = layers.Dropout(0.2)
  self.avgpool=layers.GlobalAveragePooling2D()
 
 def call(self, input, training = None):
  out = self.bn(input)
  out = self.relu(out)
  out = self.conv(out)
  out = self.dp(out)
  out = self.avgpool(out)
  return out

def dense_block(input, num_blocks):
	layers_concat = list()
	layers_concat.append(input)
	x = bottleneck_layer(input)
	layers_concat.append(x)
	for i in range(nb_layers - 1):
        	x = Concatenation(layers_concat)
        	x = self.bottleneck_layer(x)
        	layers_concat.append(x)
        x = Concatenation(layers_concat)
        return x
class Dense_net(keras.Model):
	  def __init__(self, input, filters="自定义", stride=1):
	  super(Dense_net, self).__init__()
	  
	  self.conv1 = layers.Con2D(x, use_bias=False, filters = 2*filters, kernel_size=[7,7], strides=2, padding='same')
	  self.block1 = dense_block(input, nb_layers=6)
	  self.trans1 = transition_layer(input)
	  
	  self.block2 = dense_block(input, nb_layers=12)
	  self.trans2 = transition_layer(input)
	  
	  self.block3 = dense_block(input, nb_layers=48)
	  self.trans3 = transition_layer(input)
	  
	  self.block4 = dense_block(input, nb_layers=32)
	  
	  self.bn = layers.BatchNormalization()
	  self.relu = layers.Activation('relu')
	  self.avgpool=layers.GlobalAveragePooling2D()
	  
	  self.dense = layers.Dense(input, units = "自定义", name = linear)
	def call(self, input, training = None):
	  out = self.conv1(input)
	  out = self.block1(out)
	  out = self.trans1(out)
	  out = self.block2(out)
	  out = self.trans2(out)
	  out = self.block3(out)
	  out = self.trans3(out)
	  out = self.block4(out)
	  out = self.bn(out)
	  out = self.relu(out)
	  out = self.avgpool(out)
	  out = flatten(out)
	  out = self.dense(out)
	  return out
		```