search

lightvector / KataGo

GTP engine and self-play learning in Go
https://katagotraining.org/
Other
3.61k stars 569 forks source link

About Self-play Training #265

Open Dreamkeeper66666 opened 4 years ago

Dreamkeeper66666 commented 4 years ago

Hi, thanks for your amazing work!!! It helps me a lot. Recently, I have been trying to rerun the model of KataGo(19x19 only) on my own mechine.However, I don't really have a powerful GPU (GTX 1070). It took me about 5 hours to finish running a synchronous loop. So I am just wondering how long it will take to get some obvious improvements. It would be better if there are some suggestions on self-play training. Thanks a lot!!!

Friday9i commented 4 years ago

With a single average GPU, it'll be quite long to get a decent network on 19x19 (several weeks / months probably). Unless you are a high Dan player, it'll beat you in days, but pro level is significantly harder. What's your main aim? Maybe you should try smaller boards (eg up to 9x9 or 13x13) to see what it gives. Don't forget to set up the maxlen parameter (it's way faster on small boards)

Dreamkeeper66666 commented 4 years ago

Thanks for your information!!! I just want to observe the evolution of KataGo. My final aim is just reaching the level of high dan amateur. So do you have some suggestions on the set-up of the parameters(like the number of self-play games in a cycle or the number of visits in a game, etc. ) The model I am using is b10c128. @Friday9i

Friday9i commented 4 years ago

Try relatively low visits, such as 50 cheap visits and 300 normal visits, it's probably enough. Default parameters for number of games are good enough to start.

Dreamkeeper66666 commented 4 years ago

Try relatively low visits, such as 50 cheap visits and 300 normal visits, it's probably enough. Default parameters for number of games are good enough to start.

Thank you! That is exactly what I am doing right now. By the way, I happened to see your previous comments about self-play training. It seems that you used 5 cheap visits and 30 normal visits. Did that work well?

Friday9i commented 4 years ago

For small boards (up to 9x9) it was ok up to a quite good level, but then it stagnated and it was not easy to improve thereafter. Hence I would say it's probably too low, especially for 19x19. But you may test, who knows! Note: it's somewhat incompatible with PDA, as training with 8x less visits is problematic of course (5 divided by 8 is below 1 visit...). With previous KataGo version, it was crashing the process, as visits need to be strictly above 1: I didn't test 2 visits but I know 3 cheap visits works. So it means at least 24 cheap visits by default (which gives 24/8=3 visits in PDA=3 games), and for exemple 120 normal visits. So you could test for exemple 24 cheap & 120 normal visits (and possibly 16 and 80?).