search

Haochen-Wang409 / U2PL

[CVPR'22] Semi-Supervised Semantic Segmentation Using Unreliable Pseudo-Labels
Apache License 2.0
426 stars 59 forks source link

Question about EMA #128

Closed yingao123 closed 1 year ago

yingao123 commented 1 year ago

I wonder how is the teacher model initialized before doing EMA? (i.e. directly copy the parameters from the student model, randomly initialized)

Haochen-Wang409 commented 1 year ago

In our code, the teacher is randomly initialized.

yingao123 commented 1 year ago

In our code, the teacher is randomly initialized.

Why not directly copy the parameters of the student model? Because at that moment, the student model has a good performance. It doesn't seem to make sense to use a randomly initialized network to supervise the student model?

Haochen-Wang409 commented 1 year ago

Seems reasonable, but we simply followed the common practice.

yingao123 commented 1 year ago

Thanks a lot!