Aran Komatsuzaki

@afrozenator @lukaszkaiser Could you tell me the advantage of JAX over PyTorch? Does JAX provide any tool with which Transformer and Reformer becomes faster for the reason other than the...

@richemslie Your implementation using C++ looks very interesting. I'm curious how many self-play games per second or minute you can generate. By taking into account your small architecture, it may...

@mokemokechicken achieved 22 seconds per game with 256 filters, a single 1080ti and half as many sims/move. I assume it would take 22s per game with two 1080ti's and 800...

@mokemokechicken @Akababa and others: thanks for developing an open source AlphaZero. If I understand correctly, there's still a huge difference in per-move time between Leela Zero and ReversiAZ, as you...

@Akababa Thank you for your response. It's good to know that LZ uses 64 filters only and confirm that the setting tried in ReversiAZ is 'normal' one rather than 'mini.'...

Oh does RAZ have 7 res blocks? According to "norma.pyl," it has "res_layer_num = 10," though? If 5 vs 10 is the case, then RAZ is 2.4 times faster than...

I guess being 2.4 times faster exactly corresponds to the speed advantage with cudnn, so my question is solved, assuming that, when @mokemokechicken said it takes 100s for a single...

@mokemokechicken I asked @gooooloo a similar question in other thread, but what is the default ratio of the number of games per gradient update ratio of your algorithm? I guess...

Thanks for your answer. In the case of Go with AlphaZero, 700k minibatches (2048 positions each) and 21 million self-play games were performed. Assuming that each game ended with 150...

Cool. So, multi-processing successfully decreased the ratio and achieved 36s per game under 400 sims/move. Now, it suffices to elucidate the trade-off between training/selfplay ratio and sims/move. I'm excited for...