ducminhkhoi / interpretablecnn Goto Github PK

The accuracy of interpretableCNN always approximates random guess when it is trained from scratch, i.e., 0.5% in the CUB200 dataset. The same question is that there is no progress in the training of pretrained CNN with 2 interpretable layers, almost 51-52% acc. with the given code.

Hi why loss_1 = function(x1)? x1 is the feature after mask, why not the feature before mask?

I found there are no files listed below under the dir of ImagenetPartDataset:

• train_test_split.txt
• classes.txt
• images.txt
• image_class_labels.txt
• bounding_boxes.txt
  thus, the preprocessing codes for ImagenetPartDataset can not run.

I wonder if there are any updates in this repo and look forward to your new commit. Thanks.