scoutsaachi
commented
2 years ago Hi!
The full line here is
ones_mask = torch.where(ones_mask.view(-1) > 0)[0] + 1
which does first change the range to [1,196] and then prepends the 0. The code thus already performs exactly as you described.
ljb121002
Hi, very solid work. But I think there is a typo in the class "MuskProcessor" of /src/utils/custom_models/preprocess.py. "ones_mask = torch.where(ones_mask.view(-1) > 0)[0]" this line choose those patches which intersect with ablation columns or blocks, and they belong to [0, 195]. To keep the class token always, the code just adds a 0 at the beginning of ones_mask in the next line. But I think we should first change the range from [0, 195] to [1, 196], then add the 0 :)