Problems when reproducing pensieve with PyTorch

[PkuDavidGuan]

Dear Hongzi,

I am sorry to bother you, but I went through a few problems when reproducing Pensieve with PyTorch. I found the critic loss is very high (about 30), but the total QoE seems normal(about 40.) I wonder the range of the critic loss after Pensieve's convergence.

[hongzimao]

Very nice that you reproducing it in pytorch! I don't remember the range of the loss numbers on top of my head. Also, how do you compute the loss, did you sum it over a batch or you average it? I would recommend you run our code and compare the scale of those numbers.

[qsibmini-khu]

Dear Hongzi,

I am sorry to bother you, but I went through a few problems when reproducing Pensieve with PyTorch. I found the critic loss is very high (about 30), but the total QoE seems normal(about 40.) I wonder the range of the critic loss after Pensieve's convergence.

Hello, did you finish making the PyTorch version by chance?
It would be a great help.

Thank you.