Closed jongchyisu closed 5 years ago
Upper bound is the output of the greedy search. You first run the greedy search then you use its value as the upper bound. Delta is also computed as ub+lb/2 as explained in the paper.
Thanks for your reply! Another question, in the paper you mentioned "activations of the final fc layers" are used as the feature. I just want to make sure that is this the logits (10 dimension as in SVHN) instead of the feature from the penultimate layer?
Sorry, we meant the penultimate layer. It is the layer before the softmax
So it is logits (10 dimension as in SVHN). Thanks.
No, it is the penultimate layer. Layer before the softmax. If you refer to here: http://torch.ch/blog/2015/07/30/cifar.html It is the 512 dimensional output before nn.Linear(512,10)
HI, I'm wondering what's the value you used for upper bound and delta in your experiments. I only found the upper bound \Chi=1e-4*n in the paper. Is this the UB used in the code? How about the delta? Thank you very much!