第四周：卷积神经网络 part3

代码练习

HybridSN 高光谱分类网络

class HybridSN(nn.Module):
  def __init__(self):
    super(HybridSN,self).__init__()
    # conv1：（1, 30, 25, 25）， 8个 7x3x3 的卷积核 ==>（8, 24, 23, 23）
    self.conv3d1 = nn.Conv3d(1,8,kernel_size=(7,3,3),stride=1,padding=0)
    self.bn1 = nn.BatchNorm3d(8)
    #conv2：（8, 24, 23, 23）， 16个 5x3x3 的卷积核 ==>（16, 20, 21, 21）
    self.conv3d2 = nn.Conv3d(8,16,kernel_size=(5,3,3),stride=1,padding=0)
    self.bn2 = nn.BatchNorm3d(16)
    #conv3：（16, 20, 21, 21），32个 3x3x3 的卷积核 ==>（32, 18, 19, 19）
    self.conv3d3 = nn.Conv3d(16,32,kernel_size=(3,3,3),stride=1,padding=0)
    self.bn3 = nn.BatchNorm3d(32)

    #二维卷积：（576, 19, 19） 64个 3x3 的卷积核，得到 （64, 17, 17）
    self.conv2d4 = nn.Conv2d(576,64,kernel_size=(3,3),stride=1,padding=0)
    self.bn4 = nn.BatchNorm2d(64)

    #flatten
    self.fc1 = nn.Linear(18496,256)
    #256，128节点的全连接层，都使用比例为0.4的 Dropout
    self.fc2 = nn.Linear(256,128)
    #最后输出为 16 个节点，是最终的分类类别数
    self.fc3 = nn.Linear(128,16)
    self.dropout = nn.Dropout(0.4)


  def forward(self,x):
    #前向计算
    out = F.relu(self.bn1(self.conv3d1(x)))
    out = F.relu(self.bn2(self.conv3d2(out)))
    out = F.relu(self.bn3(self.conv3d3(out)))
    out = F.relu(self.bn4(self.conv2d4(out.reshape(out.shape[0],-1,19,19))))
    out = out.reshape(out.shape[0],-1)
    out = F.relu(self.dropout(self.fc1(out)))
    out = F.relu(self.dropout(self.fc2(out)))
    out = self.fc3(out)
    return out

参考：
参考作业
Dive-into-DL-PyTorch
Pytorch 从0开始学
网络中BN层的作用

SENet 实现

在HybridSN中加入SENet模块

#这代码我还得再看看TAT
class_num = 16

class SEBlock(nn.Module):
  def __init__(self,in_channels,r=16):
    super(SEBlock,self).__init__()
    #平均池化
    #Squeeze操作，将一个channel上整个空间特征编码为一个全局特征
    self.globalAvgPool = nn.AdaptiveAvgPool2d((1,1))
    #Excitation操作
    self.fc1 = nn.Linear(in_channels,round(in_channels/r))
    self.fc2 = nn.Linear(round(in_channels/r),in_channels)

  def forward(self,x):
    out = self.globalAvgPool(x)
    out = out.view(out.shape[0],-1)
    out = F.relu(self.fc1(out))
    out = F.sigmoid(self.fc2(out))
    out = out.view(x.shape[0],x.shape[1],1,1)
    out = x * out
    return out

class HybridSN(nn.Module):
  def __init__(self):
    super(HybridSN,self).__init__()
    self.conv3d1 = nn.Conv3d(1,8,kernel_size=(7,3,3),stride=1,padding=0)
    self.bn1 = nn.BatchNorm3d(8)
    self.conv3d2 = nn.Conv3d(8,16,kernel_size=(5,3,3),stride=1,padding=0)
    self.bn2 = nn.BatchNorm3d(16)
    self.conv3d3 = nn.Conv3d(16,32,kernel_size=(3,3,3),stride=1,padding=0)
    self.bn3 = nn.BatchNorm3d(32)
    self.conv2d4 = nn.Conv2d(576,64,kernel_size=(3,3),stride=1,padding=0)
    self.SElayer = SEBlock(64,16)
    self.bn4 = nn.BatchNorm2d(64)
    self.fc1 = nn.Linear(18496,256)
    self.fc2 = nn.Linear(256,128)
    self.fc3 = nn.Linear(128,16)
    self.dropout = nn.Dropout(0.4)

  def forward(self,x):
    out = F.relu(self.bn1(self.conv3d1(x)))
    out = F.relu(self.bn2(self.conv3d2(out)))
    out = F.relu(self.bn3(self.conv3d3(out)))
    out = F.relu(self.bn4(self.conv2d4(out.reshape(out.shape[0],-1,19,19))))
    out = self.SElayer(out)
    out = out.reshape(out.shape[0],-1)
    out = F.relu(self.dropout(self.fc1(out)))
    out = F.relu(self.dropout(self.fc2(out)))
    out = self.fc3(out)
    return out

参考：
参考作业
SENet实现
pytorch实现SeNet

视频学习

语义分割中的自注意力机制和低秩重重建

参考：语义分割中的Attention和低秩重建

图像语义分割前沿进展