result reproduction

[chmxu]

Hi, I am trying to reproduce the results on VOC using your code with default configuration. I met several problems during training:

1. Is the accuracy reported in Tab.1 in the paper the validation acc in the training log? It seems the data only has training and validation split without a testing set.
2. Train acc and val acc keeps decreasing during training with CSG path. The best validation acc is achieved in the 8th epoch without any CSG training.
3. When I try CSG on other datasets, the train acc in CSG epoches converges quickly to 100, while the ce loss in standard epoches converges much slower than training without CSG. I wonder if the model can only learn the mapping from specific filter index to the class labels, without learning any actual visual knowledge, i.e., only part of filters can activate on images of each class.
   I train the model using torch1.9.0, single A6000 GPU. Hope you can help me with these problems. Thank you!