Shift Learning implementation

[ZhiyuanDang]

Hi, thanks for your works. However, in your paper, the implementation of shift learning has not been described detail.

I guess that the BN parameters are re-trained in Stage-II, since the different means and variances. Is that true?

[zs-zhong]

Hello! The BN running parameters (running means and variances) are updated in Stage-II. We fix the BN affine parameters (alpha and beta) in Stage-II.

[ZhiyuanDang]

Thanks for your reply.

'update' means reset the running means and variances in Stage-I or save them? And, which BN running parameters you adopted at the test phase?

[zs-zhong]

Hi, Zhiyuan, thanks for your interest! Concretely, in Stage-1, BN is trained on instance-balanced data (normal). The statistic running means and variances are , and the learnable affine parameters are . In Stage-2, we fix the learnable affine parameters and momentum update on classes-balanced data (re-sample). Finally, we get new statistic running means and variances . During the test phase, we adopt and for BN.

[ZhiyuanDang]

Thanks a lot for your detailed reply. I think it addresses my confusion.