madrylab / data-transfer Goto Github PK

Do you decrease the number of logits after removing categories? Say, you remove 100 categories, then the logits of ImageNet are still 1000 as the original number of labels or changed to 900? 

It seems that I cannot find the removing logits in your source code, so I assume it is the former. 

If yes, why do you consider keeping logits as the origin rather than decreasing the number, being equal to the number of remaining categories?

I understand the indexes for npys in class_exclusion are the category indexes. But what about the ones in example_exclusion?

Now I index the examples based on the ImageFolder's indexing on ImageNet, for example:

train_dataset = datasets.ImageFolder(args.data_dir) 
exs_list = [x[0].split('/')[-1].split('.')[0] for x in train_dataset.imgs]

But it seems slow. Could you guide me to the indexing to-be-excluded examples in your source code? (If there is any). Or let me know what is your way to do it?