Sacha Chernyavskiy

@fuyuan-li, I erroneously had an odd relu in the network. Sorry! Now, both impls should converge.

@lanctot both versions now converge for kuhn, look at the https://github.com/google-deepmind/open_spiel/pull/1406 However, without additional code improvements (e.g. jitting the buffer which causes constant recompilation due to changing size thereof), the...

@fuyuan-li , thank you for your testing. Good that the results are reproducible. The newest commit should lower the comp. time like twice (on my mac, at least). The further...

@fuyuan-li check it out now. `jax` version should be only 4-6 times slower than pytorch It's partly because of https://github.com/jax-ml/jax/issues/16587 (at `append_to_reservior` function) and because I allocate the whole buffer...

Pytorch implementation hasn't really changed in terms of readability. In `jax` implementation, I replaced the buffer with a set of functions and made a jittable training loop. Should still stay...

@fuyuan-li thank you for your updates. Yes, I will update it locally. When everything is confirmed, we can merge. Just for the sake of interest, can you compare GPU performances?...

@fuyuan-li, thank you, will back tomorrow with the fixed results. Let's say that 0.66 is what we're looking for. Thank you for the testing, couldn't have done it without you

My final results with the latest commit for `leduc_poker` and default hyperparameters: 1) w/ `pytorch` ``` I1221 22:48:52.097683 8456724608 deep_cfr_pytorch.py:76] Final policy loss: 4.595390796661377 I1221 22:48:52.189984 8456724608 deep_cfr_pytorch.py:82] Deep CFR...

Yes, I just trimmed a couple of lines, it doesnt affect the results

@lanctot I am sorry I checked the tests they didn't fail but I forgot(?) to rerun before the commit. Also, I am not sure that they will pass on python3.10...