search

yxgeee / MMT

[ICLR-2020] Mutual Mean-Teaching: Pseudo Label Refinery for Unsupervised Domain Adaptation on Person Re-identification.
https://yxgeee.github.io/projects/mmt
MIT License
469 stars 73 forks source link

Stage-1 pre-train with target feature #17

Closed JosephPai closed 4 years ago

JosephPai commented 4 years ago

Hi, Thanks for releasing the code. I've a question about pre-train stage. In PreTrainer, I cannot make sense why there is a forward for target input, since I did not find the usage of target_feature in training.

Thanks.

yxgeee commented 4 years ago

Hi, thanks for the question. The operation of "forward for target input" in pre-training can be thought of as a trick to improve the pre-training performance on the target domain. The insight is to roughly adapt the BN layers to the target domain, since the running mean&var in BN can take each input into the computation even there is no backpropagation. However, I found that whatever the performance of pre-training model is, MMT can always achieve competitive performance.

JosephPai commented 4 years ago

I see. Thanks for your reply!