Question: Fig 3 accuracy differences?

mr-eggplant / EATA

Code for ICML 2022 paper — Efficient Test-Time Model Adaptation without Forgetting

Other

106 stars 9 forks source link

Question: Fig 3 accuracy differences? #8

Closed justinkay closed 7 months ago

justinkay commented 7 months ago

Hi there, thanks for the interesting work.

I have a question about Figure 3 of the paper: For the Gaussian noise setting, shouldn't the numbers be the same between the left and right plots (since it is the first episode of the lifelong adaptation scenario)? Why do you instead obtain ~68% corruption accuracy for gaussian noise in the left plot but ~35% in the right plot?

mr-eggplant commented 7 months ago

Hi, thanks for your interest! Figure 3 has two y-axes and the "corruption accuracy" corresponds to the right y-axis (in blue color). So, they are the same ~35%

justinkay commented 7 months ago

Ah, I see! Thanks for the clarification!