Convolution Kernel and input tensor is int(though it is float64 type ), why the convolution output is a float number?

vision
1 
class ConvNet(torch.nn.Module):
    def __init__(self):
        super(ConvNet, self).__init__()
        self.cv1 = torch.nn.Conv2d(1, 2, (3, 3), padding=(0, 0))
        self.init_edge_detect_weight()
    
    def forward(self, x):
        return self.cv1(x)
    
    def init_edge_detect_weight(self):
        sob_kernel1 = (torch.tensor([[-1, 0, 1],
                                    [-2, 0, 2],
                                    [-1, 0, 1]], dtype=torch.float64)).unsqueeze(0).unsqueeze(0)
        
        sob_kernel2 = torch.tensor([[1, 2, 1],
                                    [0, 0, 0],
                                    [-1, -2, 1]], dtype=torch.float64).unsqueeze(0).unsqueeze(0)
        
        self.cv1.weight.data = torch.cat([sob_kernel1, sob_kernel2], dim=0).requires_grad_(False)

net = ConvNet()

raster_ts = torch.ones(1, 1, 128, 128, dtype=torch.float64)

print("unique value weight: ", torch.unique(net.cv1.weight.data))
print("unique value input: ", torch.unique(raster_ts))

out_ts = net(raster_ts)

print(out_ts.shape)
print("unique value output: ", torch.unique(out_ts))
>>>
>>>
>>>

unique value weight:  tensor([-2., -1.,  0.,  1.,  2.], dtype=torch.float64)
unique value input:  tensor([1.], dtype=torch.float64)
torch.Size([1, 2, 126, 126])
unique value output:  tensor([-0.0541,  2.2535], dtype=torch.float64, grad_fn=<Unique2Backward>)

Is this caused by the precision of the float number?

Because I used 2 sobel kernel to convolve a image with all values equals to 1.0, there should not be any edge in the image, so I expected the ouput feature map should with all values equals to 0

2

Your conv layer has also a bias, which you would see in the output.
Use self.cv1 = torch.nn.Conv2d(1, 2, (3, 3), padding=(0, 0), bias=False) and you would get closer to the expected results.
Once this is solved, your output would still not contain all zeros, as you have a typo here:

        sob_kernel2 = torch.tensor([[1, 2, 1],
                                    [0, 0, 0],
                                    [-1, -2, 1]], dtype=torch.float64).unsqueeze(0).unsqueeze(0)

and would need to fix the sign of the last element.