基于知识蒸馏的两阶段去雨去雪去雾模型学习记录（一）

彭祥.2023-09-25 12:03

前面完成了基于知识蒸馏的去雨去雪去雾模型大的部署与训练，下面则进行代码的学习。

使用debug的方式进行代码的学习。

首先是网络结构展示：轻易不要打开，这个模型太复杂了。说到底倒不是多复杂，就是层数太多了

python 复制代码 
Net(
  (conv_input): ConvLayer(
    (reflection_pad): ReflectionPad2d((5, 5, 5, 5))
    (conv2d): Conv2d(3, 16, kernel_size=(11, 11), stride=(1, 1))
  )
  (dense0): Sequential(
    (0): ResidualBlock(
      (conv1): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(16, 16, kernel_size=(3, 3), stride=(1, 1))
      )
      (conv2): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(16, 16, kernel_size=(3, 3), stride=(1, 1))
      )
      (relu): PReLU(num_parameters=1)
    )
    (1): ResidualBlock(
      (conv1): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(16, 16, kernel_size=(3, 3), stride=(1, 1))
      )
      (conv2): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(16, 16, kernel_size=(3, 3), stride=(1, 1))
      )
      (relu): PReLU(num_parameters=1)
    )
    (2): ResidualBlock(
      (conv1): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(16, 16, kernel_size=(3, 3), stride=(1, 1))
      )
      (conv2): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(16, 16, kernel_size=(3, 3), stride=(1, 1))
      )
      (relu): PReLU(num_parameters=1)
    )
  )
  (conv2x): ConvLayer(
    (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
    (conv2d): Conv2d(16, 32, kernel_size=(3, 3), stride=(2, 2))
  )
  (conv1): RDB(
    (dense_layers): Sequential(
      (0): make_dense(
        (conv): Conv2d(16, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      )
      (1): make_dense(
        (conv): Conv2d(32, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      )
      (2): make_dense(
        (conv): Conv2d(48, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      )
      (3): make_dense(
        (conv): Conv2d(64, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      )
    )
    (conv_1x1): Conv2d(80, 16, kernel_size=(1, 1), stride=(1, 1), bias=False)
  )
  (fusion1): Encoder_MDCBlock1(
    (up_convs): ModuleList(
      (0): DeconvBlock(
        (deconv): ConvTranspose2d(16, 8, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1))
        (act): PReLU(num_parameters=1)
      )
    )
    (down_convs): ModuleList(
      (0): ConvBlock(
        (conv): Conv2d(8, 16, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1))
        (act): PReLU(num_parameters=1)
      )
    )
  )
  (dense1): Sequential(
    (0): ResidualBlock(
      (conv1): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(32, 32, kernel_size=(3, 3), stride=(1, 1))
      )
      (conv2): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(32, 32, kernel_size=(3, 3), stride=(1, 1))
      )
      (relu): PReLU(num_parameters=1)
    )
    (1): ResidualBlock(
      (conv1): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(32, 32, kernel_size=(3, 3), stride=(1, 1))
      )
      (conv2): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(32, 32, kernel_size=(3, 3), stride=(1, 1))
      )
      (relu): PReLU(num_parameters=1)
    )
    (2): ResidualBlock(
      (conv1): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(32, 32, kernel_size=(3, 3), stride=(1, 1))
      )
      (conv2): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(32, 32, kernel_size=(3, 3), stride=(1, 1))
      )
      (relu): PReLU(num_parameters=1)
    )
  )
  (conv4x): ConvLayer(
    (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
    (conv2d): Conv2d(32, 64, kernel_size=(3, 3), stride=(2, 2))
  )
  (conv2): RDB(
    (dense_layers): Sequential(
      (0): make_dense(
        (conv): Conv2d(32, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      )
      (1): make_dense(
        (conv): Conv2d(64, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      )
      (2): make_dense(
        (conv): Conv2d(96, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      )
      (3): make_dense(
        (conv): Conv2d(128, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      )
    )
    (conv_1x1): Conv2d(160, 32, kernel_size=(1, 1), stride=(1, 1), bias=False)
  )
  (fusion2): Encoder_MDCBlock1(
    (up_convs): ModuleList(
      (0): DeconvBlock(
        (deconv): ConvTranspose2d(32, 16, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1))
        (act): PReLU(num_parameters=1)
      )
      (1): DeconvBlock(
        (deconv): ConvTranspose2d(16, 8, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1))
        (act): PReLU(num_parameters=1)
      )
    )
    (down_convs): ModuleList(
      (0): ConvBlock(
        (conv): Conv2d(16, 32, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1))
        (act): PReLU(num_parameters=1)
      )
      (1): ConvBlock(
        (conv): Conv2d(8, 16, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1))
        (act): PReLU(num_parameters=1)
      )
    )
  )
  (dense2): Sequential(
    (0): ResidualBlock(
      (conv1): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1))
      )
      (conv2): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1))
      )
      (relu): PReLU(num_parameters=1)
    )
    (1): ResidualBlock(
      (conv1): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1))
      )
      (conv2): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1))
      )
      (relu): PReLU(num_parameters=1)
    )
    (2): ResidualBlock(
      (conv1): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1))
      )
      (conv2): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1))
      )
      (relu): PReLU(num_parameters=1)
    )
  )
  (conv8x): ConvLayer(
    (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
    (conv2d): Conv2d(64, 128, kernel_size=(3, 3), stride=(2, 2))
  )
  (conv3): RDB(
    (dense_layers): Sequential(
      (0): make_dense(
        (conv): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      )
      (1): make_dense(
        (conv): Conv2d(128, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      )
      (2): make_dense(
        (conv): Conv2d(192, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      )
      (3): make_dense(
        (conv): Conv2d(256, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      )
    )
    (conv_1x1): Conv2d(320, 64, kernel_size=(1, 1), stride=(1, 1), bias=False)
  )
  (fusion3): Encoder_MDCBlock1(
    (up_convs): ModuleList(
      (0): DeconvBlock(
        (deconv): ConvTranspose2d(64, 32, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1))
        (act): PReLU(num_parameters=1)
      )
      (1): DeconvBlock(
        (deconv): ConvTranspose2d(32, 16, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1))
        (act): PReLU(num_parameters=1)
      )
      (2): DeconvBlock(
        (deconv): ConvTranspose2d(16, 8, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1))
        (act): PReLU(num_parameters=1)
      )
    )
    (down_convs): ModuleList(
      (0): ConvBlock(
        (conv): Conv2d(32, 64, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1))
        (act): PReLU(num_parameters=1)
      )
      (1): ConvBlock(
        (conv): Conv2d(16, 32, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1))
        (act): PReLU(num_parameters=1)
      )
      (2): ConvBlock(
        (conv): Conv2d(8, 16, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1))
        (act): PReLU(num_parameters=1)
      )
    )
  )
  (dense3): Sequential(
    (0): ResidualBlock(
      (conv1): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1))
      )
      (conv2): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1))
      )
      (relu): PReLU(num_parameters=1)
    )
    (1): ResidualBlock(
      (conv1): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1))
      )
      (conv2): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1))
      )
      (relu): PReLU(num_parameters=1)
    )
    (2): ResidualBlock(
      (conv1): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1))
      )
      (conv2): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1))
      )
      (relu): PReLU(num_parameters=1)
    )
  )
  (conv16x): ConvLayer(
    (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
    (conv2d): Conv2d(128, 256, kernel_size=(3, 3), stride=(2, 2))
  )
  (conv4): RDB(
    (dense_layers): Sequential(
      (0): make_dense(
        (conv): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      )
      (1): make_dense(
        (conv): Conv2d(256, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      )
      (2): make_dense(
        (conv): Conv2d(384, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      )
      (3): make_dense(
        (conv): Conv2d(512, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      )
    )
    (conv_1x1): Conv2d(640, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)
  )
  (fusion4): Encoder_MDCBlock1(
    (up_convs): ModuleList(
      (0): DeconvBlock(
        (deconv): ConvTranspose2d(128, 64, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1))
        (act): PReLU(num_parameters=1)
      )
      (1): DeconvBlock(
        (deconv): ConvTranspose2d(64, 32, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1))
        (act): PReLU(num_parameters=1)
      )
      (2): DeconvBlock(
        (deconv): ConvTranspose2d(32, 16, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1))
        (act): PReLU(num_parameters=1)
      )
      (3): DeconvBlock(
        (deconv): ConvTranspose2d(16, 8, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1))
        (act): PReLU(num_parameters=1)
      )
    )
    (down_convs): ModuleList(
      (0): ConvBlock(
        (conv): Conv2d(64, 128, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1))
        (act): PReLU(num_parameters=1)
      )
      (1): ConvBlock(
        (conv): Conv2d(32, 64, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1))
        (act): PReLU(num_parameters=1)
      )
      (2): ConvBlock(
        (conv): Conv2d(16, 32, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1))
        (act): PReLU(num_parameters=1)
      )
      (3): ConvBlock(
        (conv): Conv2d(8, 16, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1))
        (act): PReLU(num_parameters=1)
      )
    )
  )
  (dehaze): Sequential(
    (res0): ResidualBlock(
      (conv1): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1))
      )
      (conv2): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1))
      )
      (relu): PReLU(num_parameters=1)
    )
    (res1): ResidualBlock(
      (conv1): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1))
      )
      (conv2): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1))
      )
      (relu): PReLU(num_parameters=1)
    )
    (res2): ResidualBlock(
      (conv1): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1))
      )
      (conv2): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1))
      )
      (relu): PReLU(num_parameters=1)
    )
    (res3): ResidualBlock(
      (conv1): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1))
      )
      (conv2): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1))
      )
      (relu): PReLU(num_parameters=1)
    )
    (res4): ResidualBlock(
      (conv1): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1))
      )
      (conv2): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1))
      )
      (relu): PReLU(num_parameters=1)
    )
    (res5): ResidualBlock(
      (conv1): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1))
      )
      (conv2): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1))
      )
      (relu): PReLU(num_parameters=1)
    )
    (res6): ResidualBlock(
      (conv1): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1))
      )
      (conv2): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1))
      )
      (relu): PReLU(num_parameters=1)
    )
    (res7): ResidualBlock(
      (conv1): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1))
      )
      (conv2): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1))
      )
      (relu): PReLU(num_parameters=1)
    )
    (res8): ResidualBlock(
      (conv1): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1))
      )
      (conv2): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1))
      )
      (relu): PReLU(num_parameters=1)
    )
    (res9): ResidualBlock(
      (conv1): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1))
      )
      (conv2): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1))
      )
      (relu): PReLU(num_parameters=1)
    )
    (res10): ResidualBlock(
      (conv1): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1))
      )
      (conv2): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1))
      )
      (relu): PReLU(num_parameters=1)
    )
    (res11): ResidualBlock(
      (conv1): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1))
      )
      (conv2): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1))
      )
      (relu): PReLU(num_parameters=1)
    )
    (res12): ResidualBlock(
      (conv1): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1))
      )
      (conv2): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1))
      )
      (relu): PReLU(num_parameters=1)
    )
    (res13): ResidualBlock(
      (conv1): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1))
      )
      (conv2): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1))
      )
      (relu): PReLU(num_parameters=1)
    )
    (res14): ResidualBlock(
      (conv1): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1))
      )
      (conv2): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1))
      )
      (relu): PReLU(num_parameters=1)
    )
    (res15): ResidualBlock(
      (conv1): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1))
      )
      (conv2): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1))
      )
      (relu): PReLU(num_parameters=1)
    )
    (res16): ResidualBlock(
      (conv1): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1))
      )
      (conv2): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1))
      )
      (relu): PReLU(num_parameters=1)
    )
    (res17): ResidualBlock(
      (conv1): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1))
      )
      (conv2): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1))
      )
      (relu): PReLU(num_parameters=1)
    )
  )
  (convd16x): UpsampleConvLayer(
    (conv2d): ConvTranspose2d(256, 128, kernel_size=(3, 3), stride=(2, 2))
  )
  (dense_4): Sequential(
    (0): ResidualBlock(
      (conv1): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1))
      )
      (conv2): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1))
      )
      (relu): PReLU(num_parameters=1)
    )
    (1): ResidualBlock(
      (conv1): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1))
      )
      (conv2): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1))
      )
      (relu): PReLU(num_parameters=1)
    )
    (2): ResidualBlock(
      (conv1): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1))
      )
      (conv2): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1))
      )
      (relu): PReLU(num_parameters=1)
    )
  )
  (conv_4): RDB(
    (dense_layers): Sequential(
      (0): make_dense(
        (conv): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      )
      (1): make_dense(
        (conv): Conv2d(128, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      )
      (2): make_dense(
        (conv): Conv2d(192, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      )
      (3): make_dense(
        (conv): Conv2d(256, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      )
    )
    (conv_1x1): Conv2d(320, 64, kernel_size=(1, 1), stride=(1, 1), bias=False)
  )
  (fusion_4): Decoder_MDCBlock1(
    (down_convs): ModuleList(
      (0): ConvBlock(
        (conv): Conv2d(64, 128, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1))
        (act): PReLU(num_parameters=1)
      )
    )
    (up_convs): ModuleList(
      (0): DeconvBlock(
        (deconv): ConvTranspose2d(128, 64, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1))
        (act): PReLU(num_parameters=1)
      )
    )
  )
  (convd8x): UpsampleConvLayer(
    (conv2d): ConvTranspose2d(128, 64, kernel_size=(3, 3), stride=(2, 2))
  )
  (dense_3): Sequential(
    (0): ResidualBlock(
      (conv1): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1))
      )
      (conv2): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1))
      )
      (relu): PReLU(num_parameters=1)
    )
    (1): ResidualBlock(
      (conv1): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1))
      )
      (conv2): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1))
      )
      (relu): PReLU(num_parameters=1)
    )
    (2): ResidualBlock(
      (conv1): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1))
      )
      (conv2): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1))
      )
      (relu): PReLU(num_parameters=1)
    )
  )
  (conv_3): RDB(
    (dense_layers): Sequential(
      (0): make_dense(
        (conv): Conv2d(32, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      )
      (1): make_dense(
        (conv): Conv2d(64, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      )
      (2): make_dense(
        (conv): Conv2d(96, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      )
      (3): make_dense(
        (conv): Conv2d(128, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      )
    )
    (conv_1x1): Conv2d(160, 32, kernel_size=(1, 1), stride=(1, 1), bias=False)
  )
  (fusion_3): Decoder_MDCBlock1(
    (down_convs): ModuleList(
      (0): ConvBlock(
        (conv): Conv2d(32, 64, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1))
        (act): PReLU(num_parameters=1)
      )
      (1): ConvBlock(
        (conv): Conv2d(64, 128, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1))
        (act): PReLU(num_parameters=1)
      )
    )
    (up_convs): ModuleList(
      (0): DeconvBlock(
        (deconv): ConvTranspose2d(64, 32, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1))
        (act): PReLU(num_parameters=1)
      )
      (1): DeconvBlock(
        (deconv): ConvTranspose2d(128, 64, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1))
        (act): PReLU(num_parameters=1)
      )
    )
  )
  (convd4x): UpsampleConvLayer(
    (conv2d): ConvTranspose2d(64, 32, kernel_size=(3, 3), stride=(2, 2))
  )
  (dense_2): Sequential(
    (0): ResidualBlock(
      (conv1): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(32, 32, kernel_size=(3, 3), stride=(1, 1))
      )
      (conv2): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(32, 32, kernel_size=(3, 3), stride=(1, 1))
      )
      (relu): PReLU(num_parameters=1)
    )
    (1): ResidualBlock(
      (conv1): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(32, 32, kernel_size=(3, 3), stride=(1, 1))
      )
      (conv2): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(32, 32, kernel_size=(3, 3), stride=(1, 1))
      )
      (relu): PReLU(num_parameters=1)
    )
    (2): ResidualBlock(
      (conv1): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(32, 32, kernel_size=(3, 3), stride=(1, 1))
      )
      (conv2): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(32, 32, kernel_size=(3, 3), stride=(1, 1))
      )
      (relu): PReLU(num_parameters=1)
    )
  )
  (conv_2): RDB(
    (dense_layers): Sequential(
      (0): make_dense(
        (conv): Conv2d(16, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      )
      (1): make_dense(
        (conv): Conv2d(32, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      )
      (2): make_dense(
        (conv): Conv2d(48, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      )
      (3): make_dense(
        (conv): Conv2d(64, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      )
    )
    (conv_1x1): Conv2d(80, 16, kernel_size=(1, 1), stride=(1, 1), bias=False)
  )
  (fusion_2): Decoder_MDCBlock1(
    (down_convs): ModuleList(
      (0): ConvBlock(
        (conv): Conv2d(16, 32, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1))
        (act): PReLU(num_parameters=1)
      )
      (1): ConvBlock(
        (conv): Conv2d(32, 64, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1))
        (act): PReLU(num_parameters=1)
      )
      (2): ConvBlock(
        (conv): Conv2d(64, 128, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1))
        (act): PReLU(num_parameters=1)
      )
    )
    (up_convs): ModuleList(
      (0): DeconvBlock(
        (deconv): ConvTranspose2d(32, 16, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1))
        (act): PReLU(num_parameters=1)
      )
      (1): DeconvBlock(
        (deconv): ConvTranspose2d(64, 32, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1))
        (act): PReLU(num_parameters=1)
      )
      (2): DeconvBlock(
        (deconv): ConvTranspose2d(128, 64, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1))
        (act): PReLU(num_parameters=1)
      )
    )
  )
  (convd2x): UpsampleConvLayer(
    (conv2d): ConvTranspose2d(32, 16, kernel_size=(3, 3), stride=(2, 2))
  )
  (dense_1): Sequential(
    (0): ResidualBlock(
      (conv1): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(16, 16, kernel_size=(3, 3), stride=(1, 1))
      )
      (conv2): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(16, 16, kernel_size=(3, 3), stride=(1, 1))
      )
      (relu): PReLU(num_parameters=1)
    )
    (1): ResidualBlock(
      (conv1): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(16, 16, kernel_size=(3, 3), stride=(1, 1))
      )
      (conv2): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(16, 16, kernel_size=(3, 3), stride=(1, 1))
      )
      (relu): PReLU(num_parameters=1)
    )
    (2): ResidualBlock(
      (conv1): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(16, 16, kernel_size=(3, 3), stride=(1, 1))
      )
      (conv2): ConvLayer(
        (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
        (conv2d): Conv2d(16, 16, kernel_size=(3, 3), stride=(1, 1))
      )
      (relu): PReLU(num_parameters=1)
    )
  )
  (conv_1): RDB(
    (dense_layers): Sequential(
      (0): make_dense(
        (conv): Conv2d(8, 8, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      )
      (1): make_dense(
        (conv): Conv2d(16, 8, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      )
      (2): make_dense(
        (conv): Conv2d(24, 8, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      )
      (3): make_dense(
        (conv): Conv2d(32, 8, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      )
    )
    (conv_1x1): Conv2d(40, 8, kernel_size=(1, 1), stride=(1, 1), bias=False)
  )
  (fusion_1): Decoder_MDCBlock1(
    (down_convs): ModuleList(
      (0): ConvBlock(
        (conv): Conv2d(8, 16, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1))
        (act): PReLU(num_parameters=1)
      )
      (1): ConvBlock(
        (conv): Conv2d(16, 32, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1))
        (act): PReLU(num_parameters=1)
      )
      (2): ConvBlock(
        (conv): Conv2d(32, 64, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1))
        (act): PReLU(num_parameters=1)
      )
      (3): ConvBlock(
        (conv): Conv2d(64, 128, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1))
        (act): PReLU(num_parameters=1)
      )
    )
    (up_convs): ModuleList(
      (0): DeconvBlock(
        (deconv): ConvTranspose2d(16, 8, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1))
        (act): PReLU(num_parameters=1)
      )
      (1): DeconvBlock(
        (deconv): ConvTranspose2d(32, 16, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1))
        (act): PReLU(num_parameters=1)
      )
      (2): DeconvBlock(
        (deconv): ConvTranspose2d(64, 32, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1))
        (act): PReLU(num_parameters=1)
      )
      (3): DeconvBlock(
        (deconv): ConvTranspose2d(128, 64, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1))
        (act): PReLU(num_parameters=1)
      )
    )
  )
  (conv_output): ConvLayer(
    (reflection_pad): ReflectionPad2d((1, 1, 1, 1))
    (conv2d): Conv2d(16, 3, kernel_size=(3, 3), stride=(1, 1))
  )
)

首先进入训练模式，又称知识收集训练阶段：

python 复制代码 
def train_kc_stage(model, teacher_networks, ckt_modules, train_loader, optimizer, scheduler, epoch, criterions):
	print(Fore.CYAN + "==> Training Stage 1")
	print("==> Epoch {}/{}".format(epoch, args.max_epoch))
	print("==> Learning Rate = {:.6f}".format(optimizer.param_groups[0]['lr']))
	meters = get_meter(num_meters=5)	
	criterion_l1, criterion_scr, _ = criterions
	model.train()
	ckt_modules.train()
	for teacher_network in teacher_networks:
		teacher_network.eval()

声明所需要的损失函数，ckt_models（协作知识迁移模型）的训练模式
ckt_models 的详细结构如下：

python 复制代码 
ModuleList(
  (0): CKTModule(
    (teacher_projectors): TeacherProjectors(
      (PFPs): ModuleList(
        (0): Sequential(
          (0): Conv2d(64, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
          (1): ReLU(inplace=True)
          (2): Conv2d(32, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
        )
        (1): Sequential(
          (0): Conv2d(64, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
          (1): ReLU(inplace=True)
          (2): Conv2d(32, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
        )
        (2): Sequential(
          (0): Conv2d(64, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
          (1): ReLU(inplace=True)
          (2): Conv2d(32, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
        )
      )
      (IPFPs): ModuleList(
        (0): Sequential(
          (0): Conv2d(32, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
          (1): ReLU(inplace=True)
          (2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
        )
        (1): Sequential(
          (0): Conv2d(32, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
          (1): ReLU(inplace=True)
          (2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
        )
        (2): Sequential(
          (0): Conv2d(32, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
          (1): ReLU(inplace=True)
          (2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
        )
      )
    )
    (student_projector): StudentProjector(
      (PFP): Sequential(
        (0): Conv2d(64, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
        (1): ReLU(inplace=True)
        (2): Conv2d(32, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      )
    )
  )
  (1): CKTModule(
    (teacher_projectors): TeacherProjectors(
      (PFPs): ModuleList(
        (0): Sequential(
          (0): Conv2d(128, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
          (1): ReLU(inplace=True)
          (2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
        )
        (1): Sequential(
          (0): Conv2d(128, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
          (1): ReLU(inplace=True)
          (2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
        )
        (2): Sequential(
          (0): Conv2d(128, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
          (1): ReLU(inplace=True)
          (2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
        )
      )
      (IPFPs): ModuleList(
        (0): Sequential(
          (0): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
          (1): ReLU(inplace=True)
          (2): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
        )
        (1): Sequential(
          (0): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
          (1): ReLU(inplace=True)
          (2): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
        )
        (2): Sequential(
          (0): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
          (1): ReLU(inplace=True)
          (2): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
        )
      )
    )
    (student_projector): StudentProjector(
      (PFP): Sequential(
        (0): Conv2d(128, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
        (1): ReLU(inplace=True)
        (2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      )
    )
  )
  (2): CKTModule(
    (teacher_projectors): TeacherProjectors(
      (PFPs): ModuleList(
        (0): Sequential(
          (0): Conv2d(256, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
          (1): ReLU(inplace=True)
          (2): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
        )
        (1): Sequential(
          (0): Conv2d(256, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
          (1): ReLU(inplace=True)
          (2): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
        )
        (2): Sequential(
          (0): Conv2d(256, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
          (1): ReLU(inplace=True)
          (2): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
        )
      )
      (IPFPs): ModuleList(
        (0): Sequential(
          (0): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
          (1): ReLU(inplace=True)
          (2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
        )
        (1): Sequential(
          (0): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
          (1): ReLU(inplace=True)
          (2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
        )
        (2): Sequential(
          (0): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
          (1): ReLU(inplace=True)
          (2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
        )
      )
    )
    (student_projector): StudentProjector(
      (PFP): Sequential(
        (0): Conv2d(256, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
        (1): ReLU(inplace=True)
        (2): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      )
    )
  )
  (3): CKTModule(
    (teacher_projectors): TeacherProjectors(
      (PFPs): ModuleList(
        (0): Sequential(
          (0): Conv2d(256, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
          (1): ReLU(inplace=True)
          (2): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
        )
        (1): Sequential(
          (0): Conv2d(256, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
          (1): ReLU(inplace=True)
          (2): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
        )
        (2): Sequential(
          (0): Conv2d(256, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
          (1): ReLU(inplace=True)
          (2): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
        )
      )
      (IPFPs): ModuleList(
        (0): Sequential(
          (0): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
          (1): ReLU(inplace=True)
          (2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
        )
        (1): Sequential(
          (0): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
          (1): ReLU(inplace=True)
          (2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
        )
        (2): Sequential(
          (0): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
          (1): ReLU(inplace=True)
          (2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
        )
      )
    )
    (student_projector): StudentProjector(
      (PFP): Sequential(
        (0): Conv2d(256, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
        (1): ReLU(inplace=True)
        (2): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      )
    )
  )
)

criterions的结构，其定义的是损失函数，分别是L1损失，SCR损失以及HCR损失

python 复制代码 
ModuleList(
  (0): L1Loss()
  (1): SCRLoss(
    (vgg): Vgg19(
      (slice1): Sequential(
        (0): Conv2d(3, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
        (1): ReLU(inplace=True)
      )
      (slice2): Sequential(
        (2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
        (3): ReLU(inplace=True)
        (4): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
        (5): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
        (6): ReLU(inplace=True)
      )
      (slice3): Sequential(
        (7): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
        (8): ReLU(inplace=True)
        (9): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
        (10): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
        (11): ReLU(inplace=True)
      )
      (slice4): Sequential(
        (12): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
        (13): ReLU(inplace=True)
        (14): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
        (15): ReLU(inplace=True)
        (16): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
        (17): ReLU(inplace=True)
        (18): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
        (19): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
        (20): ReLU(inplace=True)
      )
      (slice5): Sequential(
        (21): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
        (22): ReLU(inplace=True)
        (23): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
        (24): ReLU(inplace=True)
        (25): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
        (26): ReLU(inplace=True)
        (27): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
        (28): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
        (29): ReLU(inplace=True)
      )
    )
    (l1): L1Loss()
  )
  (2): HCRLoss(
    (vgg): Vgg19(
      (slice1): Sequential(
        (0): Conv2d(3, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
        (1): ReLU(inplace=True)
      )
      (slice2): Sequential(
        (2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
        (3): ReLU(inplace=True)
        (4): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
        (5): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
        (6): ReLU(inplace=True)
      )
      (slice3): Sequential(
        (7): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
        (8): ReLU(inplace=True)
        (9): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
        (10): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
        (11): ReLU(inplace=True)
      )
      (slice4): Sequential(
        (12): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
        (13): ReLU(inplace=True)
        (14): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
        (15): ReLU(inplace=True)
        (16): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
        (17): ReLU(inplace=True)
        (18): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
        (19): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
        (20): ReLU(inplace=True)
      )
      (slice5): Sequential(
        (21): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
        (22): ReLU(inplace=True)
        (23): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
        (24): ReLU(inplace=True)
        (25): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
        (26): ReLU(inplace=True)
        (27): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
        (28): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
        (29): ReLU(inplace=True)
      )
    )
    (l1): L1Loss()
  )
)

可以看到教师网络就是将先前的Net网络复制了3份，只是加载不同权重而已。即三个model。

继续训练

python 复制代码 
start = time.time()
pBar = tqdm(train_loader, desc='Training')
for target_images, input_images in pBar:
	
	# Check whether the batch contains all types of degraded data
	if target_images is None: continue

	# move to GPU
	target_images = target_images.cuda()
	input_images = [images.cuda() for images in input_images]

	# Fix all teachers and collect reconstruction results and features from cooresponding teacher
	preds_from_teachers = []
	features_from_each_teachers = []
	with torch.no_grad():
		for i in range(len(teacher_networks)):
			preds, features = teacher_networks[i](input_images[i], return_feat=True)
			preds_from_teachers.append(preds)
			features_from_each_teachers.append(features)	
			
	preds_from_teachers = torch.cat(preds_from_teachers)
	features_from_teachers = []
	for layer in range(len(features_from_each_teachers[0])):
		features_from_teachers.append([features_from_each_teachers[i][layer] for i in range(len(teacher_networks))])

	preds_from_student, features_from_student = model(torch.cat(input_images), return_feat=True)   

	
	# Project the features to common feature space and calculate the loss
	PFE_loss, PFV_loss = 0., 0.
	for i, (s_features, t_features) in enumerate(zip(features_from_student, features_from_teachers)):
		t_proj_features, t_recons_features, s_proj_features = ckt_modules[i](t_features, s_features)
		PFE_loss += criterion_l1(s_proj_features, torch.cat(t_proj_features))
		PFV_loss += 0.05 * criterion_l1(torch.cat(t_recons_features), torch.cat(t_features))

	T_loss = criterion_l1(preds_from_student, preds_from_teachers)
	SCR_loss = 0.1 * criterion_scr(preds_from_student, target_images, torch.cat(input_images))
	total_loss = T_loss + PFE_loss + PFV_loss + SCR_loss

	optimizer.zero_grad()
	total_loss.backward()
	optimizer.step()

进入评估模块：加载模型，验证集，最终输出psnr与ssim

python 复制代码 
if epoch % args.val_freq == 0:
			psnr, ssim = evaluate(model, val_loader, epoch)
			# Check whether the model is top-k model
			top_k_state = save_top_k(model, optimizer, scheduler, top_k_state, args.top_k, epoch, args.save_dir, psnr=psnr, ssim=ssim)

evaluate(model, val_loader, epoch) 函数详细代码：

随后进行结果输出：

python 复制代码 
pred = model(image)

即跳入Net的forward中进行特征提取

输入值：

输入x: 图像维度为640x480，此时初始维度：torch.Size([1, 3, 480, 640])

随后经过一系列的卷积降维，生成了如下特征图：这个过程就不赘述了。

输出值：

输出x与feature:最终的x的维度依旧为torch.Size([1, 3, 480, 640])

feature的维度，共有4个特征图，分别如下：

这里设置只输出x,所以pred的值即为x的值：

得到输出值后，即可进行损失的计算了：

python 复制代码 
psnr_list.append(torchPSNR(pred, target).item())
ssim_list.append(pytorch_ssim.ssim(pred, target).item())

具体实现：

python 复制代码 
@torch.no_grad()
def torchPSNR(prd_img, tar_img):
	if not isinstance(prd_img, torch.Tensor):
		prd_img = torch.from_numpy(prd_img)
		tar_img = torch.from_numpy(tar_img)

	imdff = torch.clamp(prd_img, 0, 1) - torch.clamp(tar_img, 0, 1)
	rmse = (imdff**2).mean().sqrt()
	ps = 20 * torch.log10(1/rmse)
	return ps

最终将19张图片全部评估完毕：

得到psnr_list值：

需要19张全部评估完，这里只进行了两张。

最终返回平均值：

python 复制代码 
return np.mean(psnr_list), np.mean(ssim_list)

该方法最终的值变为：

相关推荐
CV学术叫叫兽5 分钟前
一站式学习：害虫识别与分类图像分割
学习·分类·数据挖掘
我们的五年16 分钟前
【Linux课程学习】：进程程序替换，execl，execv，execlp，execvp，execve，execle，execvpe函数
linux·c++·学习
好看资源平台36 分钟前
网络爬虫——综合实战项目：多平台房源信息采集与分析系统
爬虫·python
一棵开花的树，枝芽无限靠近你40 分钟前
【PPTist】添加PPT模版
前端·学习·编辑器·html
果冻人工智能1 小时前
2025 年将颠覆商业的 8 大 AI 应用场景
人工智能·ai员工
代码不行的搬运工1 小时前
神经网络12-Time-Series Transformer (TST)模型
人工智能·神经网络·transformer
进击的六角龙1 小时前
深入浅出：使用Python调用API实现智能天气预报
开发语言·python
檀越剑指大厂1 小时前
【Python系列】浅析 Python 中的字典更新与应用场景
开发语言·python
VertexGeek1 小时前
Rust学习（八）：异常处理和宏编程：
学习·算法·rust
石小石Orz1 小时前
Three.js + AI：AI 算法生成 3D 萤火虫飞舞效果~
javascript·人工智能·算法
热门推荐
01(欧拉)openEuler系统添加网卡文件配置流程、(欧拉)openEuler系统手动配置ipv6地址流程、(欧拉)openEuler系统网络管理说明02PyTorch机器学习实现液态神经网络03【HarmonyOS】HUAWEI DevEco Studio 下载地址汇总04玄机平台应急响应—webshell查杀05Coze扣子平台完整体验和实践（附国内和国际版对比）06Ubuntu 20.04使用Livox mid 360 测试 FAST_LIO07【目标跟踪】相机运动补偿08Unity中PICO实现 隔空取物 和 接触抓取物体09RAG 实践- Ollama+RagFlow 部署本地知识库10量化方法怎么选？如何评估量化后的大模型LLM？