13x13 and 9x9 nets

[ihavnoid]

9x9 - https://drive.google.com/open?id=19QnzpENWRWk-c97HyV4NC2P-VEYE9OGd 13x13 - https://drive.google.com/open?id=1knpQd7wSC8Slqnz_Jro9VnhqQMT5sovE

These are the nets I created using three GTX 1080s - 9x9 took roughly 3 weeks, 13x13 took roughly 8 weeks and still running.

[alreadydone]

Your nets are 128x10 and 192x15 respectively. These are trained with 7.5 komi, right? @hzyhhzy tested (with --noponder -p50 -r10 -t1 -m10) the 9x9 net with his series of nets and found that it's roughly on par with (slightly stronger than) his 20th net (64x5, trained with -v 50 on 1x1080Ti for 70 hours). Here is the 48th net (96x5, 2 weeks) for comparison. https://drive.google.com/file/d/1s9WLzt2lfuWbJfS_0TAyj08A6wX7CMpU/view?usp=sharing He's now at 149th net (224x16) with -p 50. https://drive.google.com/file/d/1ViWarBvxhzMdvyC8IACQ-jahL03LBD7m/view?usp=sharing He advises using a smaller window at the beginning to speed things up and increasing noise significantly to allow opening moves to change.

[ihavnoid]

Yeah 7.5 komi. I haven't paid a lot of attention on the parameters so it is possible that it ended up being stuck somewhere suboptimal.

To me it's a bit surprising that 50 playoffs are enough. That definitely will speed up things by a lot. I used 400.

[fantianwen]

Could you give a proper playouts number when training 13x13. I am now using 1000 for self-play, but it was so slow, I am not sure which number for playout vell be better.

[alreadydone]

I don't personally know but at least two persons in this QQ group have trained 13x13 nets.

[fantianwen]

@alreadydone which group, and what is the QQ group number? I want to join in.

HAha, actually, I started from last week.

[alreadydone]

Leelaz7路&9路&13路 225292625 A link was posted above actually.

[fantianwen]

HAha, thank you very much!

[ihavnoid]

FYI, the 13x13 was trained using 800 playouts. I didn't pay a lot of attention to the parameters, so again, what I did may not be anything close to optimal. :) It at least beats me 10-0 very easily.

[roy7]

Is there a Leela Zero for Windows binary someplace that can be used for 13x13?

[ihavnoid]

@roy7 I don't even have a Windows development machine, but my last experience is that even the leela-zero master branch works perfectly after rebuilding with the board size (in config.h) is changed. Won't be too hard to build one.

[trunterzx]

I don't even have a Windows development machine, but my last experience is that even the leela-zero master branch works perfectly after rebuilding with the board size (in config.h) is changed. Won't be too hard to build one.

@ihavnoid @roy7 I managed to compile one using VS2017, but I'm not sure how to incorporate that into Lizzie...

[roy7]

Thanks, someone on Discord provided one to the person on Discord who was asking (I posted here to ask for him). :)

[trunterzx]

@roy7

Is it possible to share the link? Mine doesn't seem to be working very well due to some reasons...

Thanks, someone on Discord provided one to the person on Discord who was asking (I posted here to ask for him). :)

[alreadydone]

lz13.zip

[roy7]

@trunterzx The one shared on Discord:

I compiled a lz exe for 13×13, but it might be a little bit outdated. You need to download the leelaz 0.15 and replace the leelaz.exe with this one. https://share.weiyun.com/516RHoY And here is a weight for 13×13. I have tested it with a pro player and it won 2-0. https://share.weiyun.com/5f8mdIA

[hzyhhzy]

My latest 9x9 weight(komi 7.5,30b320f). Black's winrate is about 25% in self-play games now.Maybe white will never lose someday. https://drive.google.com/open?id=1I7LtMcuwyLUZDNjV5VL9eUwTDdyA2nIT

[trunterzx]

@alreadydone @roy7

Thanks for sharing!

Is this compatible with Lizzie0.6? I tried replacing the leelaz.exe in Lizzie0.6 with this one, but it didn't seem to work... Not sure whether it's just my computer...

[roy7]

I assume you'd need to recompile Lizzie for alternate board sizes. You may need to use Sabaki for a GUI.

[trunterzx]

@hzyhhzy

Thanks for sharing!

Maybe 7.5 is too much komi for black on 9x9... Perhaps 6.5 or 5.5 will be a fairer komi?

My latest 9x9 weight(komi 7.5,30b320f). Black's winrate is about 25% in self-play games now.Maybe white will never lose someday.

[hzyhhzy]

@hzyhhzy

Thanks for sharing!

Maybe 7.5 is too much komi for black on 9x9... Perhaps 6.5 or 5.5 will be a fairer komi?

My latest 9x9 weight(komi 7.5,30b320f). Black's winrate is about 25% in self-play games now.Maybe white will never lose someday.

No,the fairest komi must be an integer(most likely an odd number),and it is probably 7.0 according to research.

[kennyfs]

No, I don't think so. The komi was designed for a decimal because it won't make draw game.

[alreadydone]

Note that major computer Go competitions as well as CGOS uses 7.0 komi for 9x9 games, so it must be empirically clear that the fair komi is 7.0. As hzy keeps training the 6.5 net I bet that the white winrate will keep dropping. Of course there's still a possibility that not losing the game is easier for one color than for the other. Even if a non-losing strategy is found, it may take more space to store / more time to compute for one color than the other.

[mccxj]

@ihavnoid Where to download the latest weight?

[l1t1]

good

[trunterzx]

Thanks for sharing!

https://github.com/leela-zero/leela-zero/issues/2240#issue-412504088

Is the one shared above stronger than the one shared below?

https://github.com/ihavnoid/leelaz-ninenine/issues/6#issuecomment-457891216

How about the 13x13, is it stronger than the one below?

https://github.com/ihavnoid/leelaz-ninenine/issues/6#issuecomment-457890152

[hzyhhzy]

Thanks for sharing!

leela-zero/leela-zero#2240 (comment)

Is the one shared above stronger than the one shared below?

#6 (comment)

How about the 13x13, is it stronger than the one below?

#6 (comment)

13x13 modified from 205 is much stronger than that one.But 9x9 is much weaker.

[l1t1]

did you try to modified from elfv2? @hzyhhzy

[hzyhhzy]

Do it yourself

[iopq]

@hzyhhzy 149s is competitive with 152s at time parity

Maybe 30b is overkill for 9x9?

[alreadydone]

latest is 158s I think Is 152s 30b?

[iopq]

Where do you get 158s?

[mihhailnovik]

Tested 152s on GoQuest server against real people. Result - 150+ wins, few draws, 0 lose.

If anyone interested.. this is a bridge i had to implement to connect GTP and GoQuest https://github.com/mihhailnovik/questGo

[iopq]

@hzyhhzy any newer nets? Or should I just post your 169 net here as the final one?

[hzyhhzy]

My most recent one is 171. But I think it's meaningless to release it. Recently katago released 20b weights. I tested and found that the last one is almost as strong as my "171s"(about 200 games, 10000po, komi 7.5 , the winrate of katago is 43%), and it also supports other boardsizes and komis. So I stopped my LZ9 and started to train Katago9. I had several runs of Katago 9x9 on my computer. It's about 30x faster than my LZ9. It only took me about 3 days on 2*2060 to reach the level of the last Katago 20b on 9x9 board. I'm trying some parameters, several weeks later I may release a stronger Katago 9x9 weight.

[iopq]

I really need the Katago 9x9 weight, since I play on 7 komi server. Game reviews are not valid with the wrong komi

[leeazoo]

Can someone share the network.gz for lizzie to play 13x13, i found this topic, i try with the files you put here, but i cant unrar the rar, and the 13x13.txt.gz , gets engine loading, and lz13 in zip is just for windows, Can any one share a network trained for 13x13,

Also, can someone point a tutorial on how to train networks, pls dont be hard with me im just learning.

[iopq]

You need to compile the 13x13 Leela Zero yourself by changing the board size from 19 to 13, then it will be able to load the network file

[leeazoo]

That was the issue, thanks for the tip.

[dustinlacewell]

Does anyone have an updated link to 13x13 weights? It seems the links died.

[iopq]

@dustinlacewell use KataGo, the latest nets are stronger