No " Applying two sequential convolutional operations"

[lenz3000]

Hi, I was just looking at your paper and code. And I don't see any sign of applying the convolutional operation two times, only once here: https://github.com/kumar-shridhar/PyTorch-BayesianCNN/blob/d73da93fde64fa93e933f88c90cce54a4d618589/Image%20Recognition/utils/BBBlayers.py#L48

It seems like you did that in earlier versions of your code as in this commented out bit in BBBLayers_.py :

        # conv_qw_mean = F.conv2d(input=input, weight=weight, stride=self.stride, padding=self.padding,
        #                              dilation=self.dilation, groups=self.groups)
        # conv_qw_std = torch.sqrt(1e-8 + F.conv2d(input=input.pow(2), weight=torch.exp(self.log_alpha)*weight.pow(2),
        #                                          stride=self.stride, padding=self.padding, dilation=self.dilation, groups=self.groups)) 

Why is that?

[kumar-shridhar]

Yes, it is applied in the previous code (still there inside utils). However, it takes more time to converge. Hence, another version that is faster and more efficient has been provided that uses just one convolution.

[ludwigwinkler]

See ... that's what I told you :D

[TMullerSG]

is it true you just revert back to the which is used in

[kumar-shridhar]

I would say it is the LRT trick used in the code that works well.