Why are you normalizing to 1.414 in unet.py?

[metamath1]

Why are you normalizing to 1.414 in unet.py?

class Conv3(nn.Module): ...

def forward(self, x: torch.Tensor) -> torch.Tensor:
    x = self.main(x)
    if self.is_res:
        x = x + self.conv(x)
        return x / 1.414 # <= here
    else:
        return self.conv(x)

[99991]

If you repeatedly add arrays to arrays, their magnitude increases exponentially and will eventually overflow. Therefore, it is a good idea to normalize values. A common normalization strategy is to scale arrays such that the standard deviation is 1. If you add two uncorrelated Gaussian random variables with standard deviation of 1, the standard deviation of their sum will be $\sqrt{2}$ (Proof), so you have to divide by $\sqrt{2}$ to make their standard deviation 1 again.

You can easily try this yourself with the following Python code:

import torch

# add two uncorrelated arrays
x = torch.randn(1000000) + torch.randn(1000000)

print(x.std()) # standard deviation increased from 1 to approximately 1.41
x /= 2**0.5 # divide by sqrt(2)
print(x.std()) # standard deviation is 1 again