It is currently Sun Feb 28, 2021 5:28 pm

All times are UTC - 8 hours [ DST ]




Post new topic Reply to topic  [ 97 posts ]  Go to page Previous  1, 2, 3, 4, 5  Next
Author Message
Offline
 Post subject: Re: KataGo Distributed Training and new networks
Post #41 Posted: Fri Jan 29, 2021 10:44 am 
Lives in gote
User avatar

Posts: 670
Liked others: 139
Was liked: 124
I thought that it was always, but after looking at the last log, I found that there were 2 playouts 2 times per game. thanks for the info!
2 times per game is not much, but now, of course, I'll change numSearchThreads


This post by And was liked by: go4thewin
Top
 Profile  
 
Offline
 Post subject: Re: KataGo Distributed Training and new networks
Post #42 Posted: Fri Jan 29, 2021 11:21 am 
Lives with ko

Posts: 138
Liked others: 180
Was liked: 28
Rank: 25 kyu
I wonder if using this getconfig.cpp when compiling would fix the issue
https://github.com/portkata/KataGo/tree ... pp/program
probably not, but it uses the line
Code:
if(maxPlayouts = 1)                  replace("$$NUM_SEARCH_THREADS", "numSearchthreads = 1")
at line 217, that's the only difference from the master branch getconfig.cpp . It feels like maybe that command is not in the right syntax or in the right place?

You're right, not much difference between 1 & 2 playouts
s580 1 po won 5/6 against gtp4zen7.
Attachment:
Screenshot_20210130-205017.png
Screenshot_20210130-205017.png [ 366.63 KiB | Viewed 1085 times ]


Edit: Quick testing (6 games) with new optimized net, it seems to be atleast as strong as the last 40b net of the non-distributed run. optimized s580 policy (t1 p1 ncachesize = 0) against 20b net 5 playouts (t1 p5) was 4 wins - 2 losses, non optimized s503 40b was 2 wins - 4 losses . The optimized net is almost 1/4 the size and 4 times faster. Really amazing stuff.

Top
 Profile  
 
Offline
 Post subject: Re: KataGo Distributed Training and new networks
Post #43 Posted: Mon Feb 01, 2021 9:42 am 
Lives in gote
User avatar

Posts: 670
Liked others: 139
Was liked: 124
go4thewin wrote:
You're right, not much difference between 1 & 2 playouts

I did not state this

Top
 Profile  
 
Offline
 Post subject: Re: KataGo Distributed Training and new networks
Post #44 Posted: Mon Feb 01, 2021 10:08 am 
Lives with ko

Posts: 138
Liked others: 180
Was liked: 28
Rank: 25 kyu
My mistake, I misread your post. In the nets I have tested though, usually over sample sizes of 50 or a 100 games, at the larger net sizes, there is a very minimal difference between 1 and 2 playouts. Sometimes, because of noise the 2 playouts may seem weaker.

But what I'm really curious about is whether the optimized net is stronger than the s503 net. I haven't run enough games yet, but I'm guessing it is? That would really be incredible given it's size and speed. It is almost the same size as Leela Zero 157! As well as the 15b katago nets

Top
 Profile  
 
Offline
 Post subject: Re: KataGo Distributed Training and new networks
Post #45 Posted: Mon Feb 01, 2021 10:18 am 
Lives in gote
User avatar

Posts: 670
Liked others: 139
Was liked: 124
look at here https://lifein19x19.com/viewtopic.php?p=263170#p263170
I don't know what the accuracy of the stored weights is, but it looks like 99 or even 100%!

Top
 Profile  
 
Offline
 Post subject: Re: KataGo Distributed Training and new networks
Post #46 Posted: Mon Feb 01, 2021 10:23 am 
Lives in gote
User avatar

Posts: 670
Liked others: 139
Was liked: 124
@akigo
could you evaluate the accuracy of the optimized version of KG 40b? maybe about?
EDIT I mean, how does the level of strength of this network correspond in relation to the original


Last edited by And on Tue Feb 02, 2021 8:18 am, edited 1 time in total.
Top
 Profile  
 
Offline
 Post subject: Re: KataGo Distributed Training and new networks
Post #47 Posted: Mon Feb 01, 2021 10:53 am 
Lives with ko

Posts: 138
Liked others: 180
Was liked: 28
Rank: 25 kyu
Who cares about accuracy? The most important thing are the tests that you have posted. The nets keep getting stronger. He published the nets, so someone can figure out how to do the same thing. A few months from now, if someone optimizes the strongest confidently rated net at that time, it will be as strong as the nets being released now. That's a really great thing. Before he published the nets no one knew how to do this. There is only a 3-4 month lag in strength for an enormous increase in speed. He has actually made the net stronger at time parity than the unoptimized net s580, but katago is strong enough that the strength differences are irrelevant as long as the nets keep getting stronger. Even if the distributed run stalls, the 40x384 net can take over and then be used to train the 40x256 net. Then the smaller net can be optimized in the end. You will end up with an optimized net stronger than any currently available. Probably with a policy network that will beat the vast majority of amateurs. In the end we have a 48 mb net that is as strong as 175 mb nets and as fast as other 48 mb nets.


This post by go4thewin was liked by: Maharani
Top
 Profile  
 
Offline
 Post subject: Re: KataGo Distributed Training and new networks
Post #48 Posted: Mon Feb 01, 2021 12:34 pm 
Lives in gote
User avatar

Posts: 670
Liked others: 139
Was liked: 124
if the optimized network matches the strength of, for example, 99%, then the games results are suitable for it too

Top
 Profile  
 
Offline
 Post subject: Re: KataGo Distributed Training and new networks
Post #49 Posted: Mon Feb 01, 2021 3:04 pm 
Lives with ko

Posts: 138
Liked others: 180
Was liked: 28
Rank: 25 kyu
I see what you're saying, after playing through some games with the optimized net (2 wins 3 losses against s503, white won all the games) I think maybe the best way to get a net that strong would be if a katago run trained a 50x320 net, and then that net was optimized by either akigo or someone on github. But hopefully I'm wrong. maybe better for the end user to just use the optimized net until they can beat it at kyu rank -8, then switch to s580 non optimized. At the lower kyu ranks, the optimized net is really good because it doesn't like to tenuki or play unconnected to either it's own or opponent's stones. So in the kyu rank bots, which tend to tenuki too much with the normal nets, with the optimized net it is lessoned and plays very sensibly all the way down to kyu_rank 10. I think because of the extremely small size of the final net, it would be hard or impossible to make it much more accurate than it already is. I think that's how we get the amazing speed.

Top
 Profile  
 
Offline
 Post subject: Re: KataGo Distributed Training and new networks
Post #50 Posted: Tue Feb 02, 2021 4:46 am 
Lives in gote
User avatar

Posts: 670
Liked others: 139
Was liked: 124
is there a position where the moves of the original s580 and the optimized one differ?


This post by And was liked by: go4thewin
Top
 Profile  
 
Offline
 Post subject: Re: KataGo Distributed Training and new networks
Post #51 Posted: Tue Feb 02, 2021 1:37 pm 
Dies with sente

Posts: 92
Liked others: 12
Was liked: 113
And wrote:
@akigo
could you evaluate the accuracy of the optimized version of KG 40b? maybe about?
EDIT I mean, how does the level of strength of this network correspond in relation to the original


Whenever I release an optimized network (either within the app or outside), I run a series of test matches with 300 playouts between the original net and the optimized one. All these series ended roughly even (+/- 1 game).

When I read your last posts, I thought that perhaps it's different with 1 playout. So I ran another series between original s580 and optimized s580 with 1 playout. After 60 games the score was 31:29 in favour of the original s580.

All these results strongly suggest that the strength difference between original and optimized s580 is pretty small (though you would need many more matches to quantify it).

But what I learnt from this last match is that statistical variance is much bigger with 1 playout: The original s580 started with an 8:2 run in the first ten matches ...

The kyu rank bot further increases the variance, since even at high levels it still considers only a random subset of moves, so there is another element of chance here. So keep in mind that you would need a really huge number of matches to confidently judge its strength.


This post by akigo was liked by 3 people: And, ez4u, go4thewin
Top
 Profile  
 
Offline
 Post subject: Re: KataGo Distributed Training and new networks
Post #52 Posted: Wed Feb 03, 2021 4:26 pm 
Lives with ko

Posts: 138
Liked others: 180
Was liked: 28
Rank: 25 kyu
@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!

Top
 Profile  
 
Offline
 Post subject: Re: KataGo Distributed Training and new networks
Post #53 Posted: Thu Feb 04, 2021 3:20 am 
Lives in gote
User avatar

Posts: 670
Liked others: 139
Was liked: 124
go4thewin wrote:
@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!

many @akigo solutions are convenient and simple! there is probably a special version of BadukAI that does this


This post by And was liked by: go4thewin
Top
 Profile  
 
Offline
 Post subject: Re: KataGo Distributed Training and new networks
Post #54 Posted: Thu Feb 04, 2021 2:31 pm 
Dies with sente

Posts: 92
Liked others: 12
Was liked: 113
go4thewin wrote:
@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).


This post by akigo was liked by: go4thewin
Top
 Profile  
 
Offline
 Post subject: Re: KataGo Distributed Training and new networks
Post #55 Posted: Thu Feb 04, 2021 5:17 pm 
Lives with ko

Posts: 138
Liked others: 180
Was liked: 28
Rank: 25 kyu
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!


Attachments:
zenithgo6_7danVSoptimized_policy.sgf [1.5 KiB]
Downloaded 13 times
Top
 Profile  
 
Offline
 Post subject: Re: KataGo Distributed Training and new networks
Post #56 Posted: Fri Feb 05, 2021 4:39 am 
Lives in gote
User avatar

Posts: 670
Liked others: 139
Was liked: 124
KataGo s603 playouts 1 - Zen 9d H3 1:1


Attachments:
KataGo - Zen.sgf [1.69 KiB]
Downloaded 110 times
Top
 Profile  
 
Offline
 Post subject: Re: KataGo Distributed Training and new networks
Post #57 Posted: Fri Feb 05, 2021 5:22 am 
Lives with ko

Posts: 138
Liked others: 180
Was liked: 28
Rank: 25 kyu
Thanks And for the zen games! I think t4 makes zen weaker than 9d though, because of parallelization. Breakwa11 uses t1 https://github.com/breakwa11/GoAIRating ... /AIcmds.md
But very grateful for the games

Top
 Profile  
 
Offline
 Post subject: Re: KataGo Distributed Training and new networks
Post #58 Posted: Fri Feb 05, 2021 6:04 am 
Lives in gote
User avatar

Posts: 670
Liked others: 139
Was liked: 124
I asked a question about this almost 2 years ago
https://github.com/breakwa11/GoAIRating ... -445428245


This post by And was liked by: go4thewin
Top
 Profile  
 
Offline
 Post subject: Re: KataGo Distributed Training and new networks
Post #59 Posted: Fri Feb 05, 2021 6:48 am 
Lives in gote
User avatar

Posts: 670
Liked others: 139
Was liked: 124
Zenith 7 cpu usage


Attachments:
Zenith 7 9d.jpg
Zenith 7 9d.jpg [ 148.16 KiB | Viewed 567 times ]
Top
 Profile  
 
Offline
 Post subject: Re: KataGo Distributed Training and new networks
Post #60 Posted: Fri Feb 05, 2021 7:16 am 
Lives in gote
User avatar

Posts: 670
Liked others: 139
Was liked: 124
it is interesting that on a Ryzen 5, which has 12 threads, Zenith 7 uses only 4

Top
 Profile  
 
Display posts from previous:  Sort by  
Post new topic Reply to topic  [ 97 posts ]  Go to page Previous  1, 2, 3, 4, 5  Next

All times are UTC - 8 hours [ DST ]


Who is online

Users browsing this forum: lovego and 1 guest


You cannot post new topics in this forum
You cannot reply to topics in this forum
You cannot edit your posts in this forum
You cannot delete your posts in this forum
You cannot post attachments in this forum

Search for:
Jump to:  
Powered by phpBB © 2000, 2002, 2005, 2007 phpBB Group