Torch.cat() error

1

Hi I build this model for super ressolution and at the end of it i want to split bypass with x

x = self.output(torch.cat([x, x_bypass], dim=1)) This line
It throws error

RuntimeError: Given groups=1, weight of size [3, 64, 5, 5], expected input[3, 67, 512, 512] to have 64 channels, but got 67 channels instead

Any sugesstions how to solve this ?

class Super_ress_model(torch.nn.Module):
    def __init__(self, input_shape= (3, 128, 128), output_shape= (3, 384, 384)):
        super(Super_ress_model, self).__init__()

        self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
        self.upsample = torch.nn.Upsample(size=None, scale_factor=8, mode='nearest', align_corners=None)


        self.input = nn.Conv2d(3, 64, kernel_size=9, stride=1, padding=9 // 2)
        self.act1 = nn.ReLU()

        self.block0 = ResidualBlock(64,5)
        self.block1 = ResidualBlock(64,5)
        self.block2 = ResidualBlock(64,5)
        self.block3 = ResidualBlock(64,5)

        self.upsample0 = torch.nn.Upsample(size=None, scale_factor=2, mode='nearest', align_corners=None)
        self.block6 = ResidualBlock(64, 5)
        self.block7 = ResidualBlock(64, 5)

        self.upsample1 = torch.nn.Upsample(size=None, scale_factor=2, mode='nearest', align_corners=None)
        self.block10 = ResidualBlock(64, 5)
        self.block11 = ResidualBlock(64, 5)

        self.upsample2 = torch.nn.Upsample(size=None, scale_factor=2, mode='nearest', align_corners=None)
        self.block13 = ResidualBlock(64,5)
        self.block14 = ResidualBlock(64,5)


        self.conv1 = nn.Conv2d(64, 64, kernel_size=5, stride=1, padding=5 // 2)
        self.act2 = nn.ReLU()

        self.output = nn.Conv2d(64, 3, kernel_size=5, stride=1, padding=5 // 2)

        torch.nn.init.xavier_uniform_(self.input.weight)
        torch.nn.init.zeros_(self.input.bias)

        torch.nn.init.xavier_uniform_(self.conv1.weight)
        torch.nn.init.zeros_(self.conv1.bias)

        torch.nn.init.xavier_uniform_(self.output.weight)
        torch.nn.init.zeros_(self.output.bias)

    def forward(self,x):
        x_bypass = self.upsample(x)

        x = self.input(x)
        x = self.act1(x)
        x = self.block0.forward(x)
        x = self.block1.forward(x)
        x = self.block2.forward(x)
        x = self.block3.forward(x)
        x = self.upsample0(x)

        x = self.block6.forward(x)
        x = self.block7.forward(x)

        x = self.upsample1(x)

        x = self.block10.forward(x)
        x = self.block11.forward(x)

        x = self.upsample2(x)

        x = self.block13.forward(x)
        x = self.block14.forward(x)

        x = self.conv1(x)
        x = self.act2(x)

        x = self.output(torch.cat([x, x_bypass], dim=1))

        return x
2

It’s not torch.cat. One of convolutional layers gets input with 67 channels instead of 64.

Given groups=1, weight of size [3, 64, 5, 5]

It means layer you are looking for layer that has following parameters: Conv2d(64, 3, kernel_size=(5,5)) Since you are passing 67 channels, I guess you’re doing cat over 64+3 channels (dim=1) before passing to convolutional layer mentioned above.

1 Like