无法在Pytorch中实现模型

Question

I have been trying to implement this paper (https://doi.org/10.1109/ICASSP40776.2020.9054563）。现在，在过去的几天里，我遇到了模型中的错误。 我想模仿的模型如下：

纸模型

层。 该模型的代码如下：

class Net(nn.Module):
    def __init__(self, upscale_factor = 2):
        super(Net, self).__init__()
#         self.input_start = nn.(16,256,3)
        self.e1 = nn.Conv2d(16,16, (5,1),stride = (1,2),padding='valid')
        self.e1_dr = nn.Dropout(.5)
        
        self.e2 = nn.Conv2d(128,128, (5,1),stride = (1,2),padding='valid')
#         self.e2.weight
        self.e2_dr = nn.Dropout(.5)

        self.e3 = nn.Conv2d(256, 256,(5,1),stride = (1,2),padding='valid')
        self.e3_dr = nn.Dropout(.5)

        self.e4 = nn.Conv2d(512, 512, (5,1),stride = (1,2),padding='valid')
        
        self.e5 = nn.Conv2d(512,512, (5,1),stride = (2,2),padding='valid')
        
        self.e6 = nn.Conv2d(512,512, (3,1),stride = (2,2),padding='valid')
        
        self.e7 = nn.Conv2d(512,512, (3,1),stride = (2,2),padding='valid')
        
        self.e8 = nn.Conv2d(512,512, (3,1),stride = (2,2),padding='valid')
        
        self.d1 = nn.PixelShuffle(upscale_factor )
        self.d2 = nn.PixelShuffle(upscale_factor)
        self.d3 = nn.PixelShuffle(upscale_factor)
        self.d4 = nn.PixelShuffle(upscale_factor)
        self.d5 = nn.PixelShuffle(upscale_factor )
        self.d6 = nn.PixelShuffle(upscale_factor)
        self.d7 = nn.PixelShuffle(upscale_factor)
        self.d8 = nn.PixelShuffle(upscale_factor)

        

    def forward(self, x):
#         x = (F.leaky_relu(self.input_start(x)))
        x = (F.leaky_relu(self.e1(x)))

        x = (F.leaky_relu(self.e1(x)))
        x = (F.leaky_relu(self.e1_dr(x)))
        x = (F.leaky_relu(self.e2(x)))
        x = (F.leaky_relu(self.e2_dr(x)))
        x = (F.leaky_relu(self.e3(x)))
        x = (F.leaky_relu(self.e3_dr(x)))
        x = (F.leaky_relu(self.e4(x)))
        x = (F.leaky_relu(self.e5(x)))
        
        x = (F.leaky_relu(self.e6(x)))
        x = (F.leaky_relu(self.e7(x)))
        x = (self.e8(x))
        
        
        x = (F.leaky_relu(self.d1(x)))
        x = (F.leaky_relu(self.d2(x)))
        x = (F.leaky_relu(self.d3(x)))
        x = (F.leaky_relu(self.d4(x)))
        x = (F.leaky_relu(self.d5(x)))
        x = (F.leaky_relu(self.d6(x)))
        x = (F.leaky_relu(self.d7(x)))
        x = (F.leaky_relu(self.d8(x)))


          
        return x

每当我运行模型时，我都会收到以下错误： 给定的组= 1，大小的重量[128、128、5、1]，预期输入[2、16、248、1]具有128个通道，但有16个通道。

Answer 1

频道计数中存在不匹配。查看您的模型：第一卷积e1期望16频道和输出16，而以下卷积e2期望128！ OUT_CHANNELS e1和in_channels e2必须相等。