MPL framework performance

[shu-bamma]

Hello! thank you for the amazing code and support I am trying MPL for one of my research problems but it seems the MPL loss does very little in this whole recipe. The performance seems to heavily depend on UDA (removing it i see a huge drop in the performance) can you confirm if that's the case you experienced while testing?

2. the paper claims that it seems intuitive that the teacher should also become better in this whole loop. Did the author's comment on whether it outperforms the student on these tasks or did you check (in my case it is true) this https://github.com/google-research/google-research/issues/623#issuecomment-800634170