Weight Perturbation for Conv layers/ Batch norm

[Gnomy17]

I had two questions

1. Could you give pointers on implementing the weight perturb feature for Conv layers? Should it just follow what's done in BoundLinear.py assuming a separate epsilon for each filter?
2. In the code it's stated that batch norm does not support weight perturb (as opposed to not implemented), is there a reason for this beyond simply not being implemented? Or does batch-norm weight perturb (on the linear coefficients) somehow not conform to the LiRPA framework? (which would be surprising since it's just a linear operator)

Thank you for your work

[Mounice]

Can you redirect me to the folder where it is stated that the batch norm does not support weight perturb? since I am trying to use either auto_LiRPA or alpha-beta-crown for verification of a CNN model with batch norm, and it is crashing with the error of CUDA out of memory, when applying an epsilon different than zero.

[Gnomy17]

It's in /auto_LiRPA/operators/normalization.py line 112