Hi!
Been looking at deeppavlov as a candidate to test FGSM attack on. However, FGSM requires to differentiate the loss - Xadv=X+ϵsign(∇XJ(X,Ytrue) - and the loss is the cross-entropy between the output of the nn.module and targest.
I couldn’t find a way of accessing NeMo’s module directly nor a method activating any backward()
Any insight would be helpful