zjfheart / robust-resnet Goto Github PK

Hi,
thanks for your work. I will to try on custom datasets. I have a question regarding the choice of h value.

Do you think could be useful to have different value of h, due to proposition 1, at different residual layer? My concern is that in the last residual block the upper bound create a very strong constraint on the variations of the weights. While in the first residual block maybe could be beneficial to add a smaller value of h to reduce variations of weights. I would try to set h with different values. Have you already tried this option? If yes did you get better results?