saxena-mayur / weight-uncertainty-in-neural-networks Goto Github PK

Hi
Thank you for sharing this implementation on weight uncertainty in neural networks. While I was reading your code, I have some questions regarding the equation for rho.grad. In your code, rho.grad = eps*(p.grad - beta/sigma)/(1+torch.exp(-rho)). However, based on my understanding on that paper, I get rho.grad = （eps*p.grad - beta/sigma）/（1+torch.exp(-rho)）. For the second term in the bracket, I don't think we need to multiple it by eps. I appreciate it if you have a check on whether I am correct or wrong here. Thank you very much.