KataGo Distributed Training and new networks

[And]

@akigo On your own PC, do you have a way to use the optimized net with a gtp program like gogui-twogtp? If so, would it be possible to upload that net? Even if it is slower that way, it would be so much fun to test it against other bots. Either way, Thanks for the program!

Currently, using the optimized nets is only possible in the android app. For automated matches you could try running BadukAI in an emulator on PC und use the tools that And described in his very interesting thread "GoAutomation" (they seem to use mouse events and image analysis for communication instead of gtp, so this might also work with the emulator).

BadukAI works with the Nox emulator (https://www.bignox.com) with GTool2 viewtopic.php?p=258492#p258492
for example BadukAI and Sabaki

[go4thewin]

oh wow thanks!

[akigo]

Ah I see, thanks for running so many games. Just two last questions. 1) For example, I played this game that the optimized net policy beat zen 6 7 dan as white with no komi. But since winrate is not measured when using just 1 playout on my device, I hit ponder alot to see who was winning. Does this give katago more visits, or does it still just use the policy output?
2) It looks like a 60b net is being trained. If hypothetically someone were to make an optimized version of that net using the exact same methods you used, what would the size and speed be? Is it oversimplication to say that since 60*256 is 1.5 times 40*256, the size will be about 48*1.5 mb, so it will be about the size and speed of the non optimized 20b net?
Thanks!

1) I'm sorry, I don't know, ask @lightvector about this.

2) Other parameters being equal, the network size will be roughly proportional to the number of blocks while speed will be inversely proportional. So your size estimate seems reasonable, the speed depends on the speedup that the optimization provides which depends on device type:
Best speedup I heard of was factor 6, then an optimized 60b network will still be twice as fast as the unoptimized 20b.
Worst speedup was factor 2, then an optimized 60b network will be 33% slower than the unoptimized 20b.

[go4thewin]

Thanks for all the replies everyone! Don't know how much of it is by random chance, but the s604 net is the first katago policy network of the run that went 4-0 was gtp4zen 9d for me. Earlier in the original run, it would take 12 playouts. Now the policy is enough. Probably stating the obvious, but I think Katago is probably stronger than Alphago.

zenVSs604.png (112.76 KiB) Viewed 55786 times

[And]

PytoGo viewtopic.php?p=263430#p263430
also supports the game between BadukAI (Nox emulator) and Sabaki

[go4thewin]

@And I cant get any of the emulators or pytogo to work. Would it be possible for you to do 4 games of the actual zen 7 program at 9 dan vs the optimized net policy? Thanks for the info either way!

[And]

@And I cant get any of the emulators or pytogo to work. Would it be possible for you to do 4 games of the actual zen 7 program at 9 dan vs the optimized net policy? Thanks for the info either way!

BadukAI 40b(optimized version s580) playouts 1 - Zenith 7 9d?
or maybe better with CS Zero? it seems to me that CS Zero is stronger. I play several games every day against CS Zero with a handicap of 2, but so far no network has won (1 playouts)

[go4thewin]

Really big thanks for this. maybe 2 games zen 7 vs optimized s580 with komi playouts 1 search thread 1? im just trying to see if the optimized net can atleast win 1 of the games since last year katago needed more playouts against zen. i know by the other tests it should, but would be really nice to have sgf, thanks!

[And]

in the first game, the Zenith resigned on move 179. tomorrow I will continue and upload the sgf

[go4thewin]

Wow thats really helpful to know, thanks!

[And]

BadukAI 40b(optimized version s580, maxPlayouts 1, numSearchThreads 1) - Zenith 7 9d 2:0

[And]

the same games in Zenith format, there is a win rate, they can be opened in Zenith 6

[VietGo]

I also try to run the Kata Go with strongest network s585 on my PC with low-mid level GPU like GTX 750, then settle a match with Golaxy pro 3X. I let Kata Go play with maximum 2000 visits per moves as white and it defeated Golaxy pro 3X (which spent 5 second per move with 700 visits/s). This was a surprised (for me) so far. I will try to run more game like this.

[go4thewin]

Wow this helped me out so much, thanks And. I can't believe how strong and fast the optimized net is.

[And]

BadukAI 40b(optimized version s580, maxPlayouts 1, numSearchThreads 1) - CS Zero 9d