About Batch Normalization Statistics When using Domain Adaptation Setting

google-research / adamatch

Apache License 2.0

66 stars 5 forks source link

About Batch Normalization Statistics When using Domain Adaptation Setting #1

Open jiangzhengkai opened 3 years ago

jiangzhengkai commented 3 years ago

Hi, thanks for your wonderful work. From the paper, the batch normalization statistics updates for combined source and target images, and keep fixed when computing source images for the second time. However, from https://github.com/google-research/adamatch/blob/main/domain_adaptation/adamatch.py#L63 and L67, the model keeps training states. I am not familiar with Jax. Would you mind giving some explanations?

shreejalt commented 2 years ago

Good point. @jiangzhengkai I am also confused with the same. Did you figured it out?