Thank you for sharing your code!
Can you resolve my confusion?
When I observed network architectures in models.py in this github, it didn't match the description in ICML paper that you wrote.
LeNet5 that was used here is not like normal LeNet5 I know,
CNN used for svhn in models.py is only 5-layer when it is said to use 6-layer in paper.
CNN used for cifar10 in models.py is only 8-layer when it is said to use 12-layer in paper.
Thank you!