Sample always = True in forward method of the BBBConv2D class?

[rahulrathnakumar]

I am trying to replicate the results of the paper, and also extend it for segmentation.
Why is "sample" always true for both training and validation passes in the forward method of the network modules? Shouldn't validation just use the acquired weights from the training phase? Also, I am unable to understand why the validation run uses "net.train()" instead of "net.eval()".
Any clarifications would be super helpful. Thanks!