chandar-lab / iirc Goto Github PK

Thank you for your great work!!!
It is very interesting to build the association between superclasses and subclasses. I think some label propagation methods can improve our model's performance.
However, when I run your code, I meet some problems:

• Why is the experiment result different every time run your code?
• Why icarl_cnn and icarl_norm can't reach the accuracy in the paper? the accuracy in your paper is about 0.15 on cifar100，but i get 0.08.