pytorch NFSP doesn't match tensorflow NFSP

[VitamintK]

I ran examples/leduc_nfsp.py with the given tensorflow NFSP, and I also ran it with the pytorch NFSP. For both runs, I used the same command-line arguments (I used hyperparameters from Walton & Lisy, 2021. The Tensorflow run seems to work well, matching the results in the aforementioned paper. The pytorch run does drastically worse.

Filing this issue here for someone to look into this. I may or may not look into it myself soon.

[lanctot]

Thanks, that does look quite different.

@Asugawara I thought we checked this when you first submitted the pytorch NFSP. Do you remember?

I know this goes way back.. possibly a bug was introduced when we updated the files while we upgraded it to support the newer version of pytorch.

[lanctot]

Hi @Asugawara, just checking in to see if you have any idea?

[Root970103]

I ran examples/leduc_nfsp.py with the given tensorflow NFSP, and I also ran it with the pytorch NFSP. For both runs, I used the same command-line arguments (I used hyperparameters from Walton & Lisy, 2021. The Tensorflow run seems to work well, matching the results in the aforementioned paper. The pytorch run does drastically worse.

Filing this issue here for someone to look into this. I may or may not look into it myself soon.

Hi, have you found out the reason? I met the same problem.

[lanctot]

No, this would require looking through the implementation line by line and we just don't have the time.

I tried contacting the original contributor and did not hear back from them. We will remove it shortly.

[Root970103]

No, this would require looking through the implementation line by line and we just don't have the time.

I tried contacting the original contributor and did not hear back from them. We will remove it shortly.

ok, thank you for your reply. I will try to find the difference.

[lanctot]

I appreciate that! It would be great to fix & keep it than have to delete it.

[VitamintK]

Hey all. Sorry, I never looked into it more. I did play around more with the tensorflow-based NFSP though, and discovered that in some settings (not Leduc), using Adam would vastly outperform SGD. So maybe the same would help for pytorch?

[lanctot]

ok, thank you for your reply. I will try to find the difference.

Any luck? If not, I think we should remove this soon. Then we could list fixing it in the call for contributions.