xiaohua-chen / risda Goto Github PK

I really appreciate your work and am grateful for your open-sourcing of the code.
I was able to successfully reproduce the experimental results on CIFAR100-LT using your code, but I was wondering if you could update the code for the larger datasets, ImageNet-LT and iNaturalist.

I download code in your repository and read it carefully and thoroughly.

I notice that the code keeps the manual random seed the same and sets cudnn.deterministic True. All of this makes sure the result is asthe same as you report in the paper. Specifically, when Imbalance Factor is 20, the result is 41.33.

However, according to the code, 41.33 is calculated as accuracy rate, while in your paper, 41.33 is reported as error rate. The 2 metrics are opposite aspect of an algo, which makes me puzzled.

Could you give an explanation?