adaptivetokensampling/ATS

Training on alternative datasets?

Opened this issue · 0 comments

kirk86 commented

  1. In the file libs/config/defaults.py there's an indication of ImageNet as training dataset and kinetics as test dataset, why is that?

  2. If we want to train on a different dataset say CIFAR100, would it suffice to just change the data options in defaults.py?

  3. Also, there seems to be some some PCA eigenvalues (I suppose from the dataset?) in the defaults.py, how are those gonna affect the model when training on a different dataset other than ImageNet?