Difference between DSBN and AdaBN about dsbn HOT 1 CLOSED

To add more to the differences you mentioned, the main difference is that AdaBN doesn't have separate affine parameters for each domain. DSBN has separate affine parameters for each domain.
Also, AdaBN propose a online algorithm to re-estimate the mean and variance of BN layer for target domain, which is different from the common moving update scheme. The algorithm simply re-estimates batch mean and variance on the target domain, while DSBN jointly updates batch statistics(mean and variance) using moving average scheme of BN paper during training.
When we wrote the paper, the source code of AdaBN was not available, so we couldn't verify the source code and reproduce it. Also, they didn't use ResNet and didn't provide experimental results on VisDA dataset.

from dsbn.