multiplication by sigma? about pytorch-radial-basis-function-layer HOT 3 CLOSED

Sigma has been replaced with log_sigma.

from pytorch-radial-basis-function-layer.

Traditionally, yes I think it would be norm divided by sigma. I chose to do norm multiplied by sigma instead to avoid the possibility of dividing by zero. It actually doesn't really matter. Just think of the self.sigma parameter as 1 / sigma if you like.

from pytorch-radial-basis-function-layer.

Ok, I understand your point.
There is another solution: adding a small eps to sigma.
Still, nothing prevents 1 / sigma to be < 0.

I implemented a "self.logvar" parameter (log variance) so that now:

sigma = exp(self.logvar) ^ 0.5
distance = ||x - c|| / sigma

Probably it's slower, but I feel safer because now sigma it's always strictly greater than zero.

from pytorch-radial-basis-function-layer.