New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Leelaz of 9 board topped cgos with a score of 3235. #1291

Open
sbf2000 opened this Issue Apr 30, 2018 · 41 comments

Comments

Projects
None yet
@sbf2000

sbf2000 commented Apr 30, 2018

http://www.yss-aya.com/cgos/9x9/cross/CLeela9Test12.html.
One 1080ti GPU was used to generate more than 3.2 million selfplay in more than 3 months.

@l1t1

This comment has been minimized.

l1t1 commented Apr 30, 2018

where download

@gcp

This comment has been minimized.

Owner

gcp commented Apr 30, 2018

where download

This is the important question :-)

Is it compatible with normal Leela Zero if one just changes #define BOARD_SIZE 19?

@Alderi-Tokori

This comment has been minimized.

Contributor

Alderi-Tokori commented Apr 30, 2018

Nice :) What architecture (number of blocks and number of filters) did you use?

@sbf2000

This comment has been minimized.

sbf2000 commented Apr 30, 2018

Is it compatible with normal Leela Zero if one just changes #define BOARD_SIZE 19?

Yes,totally compatible.

@sbf2000

This comment has been minimized.

sbf2000 commented Apr 30, 2018

What architecture (number of blocks and number of filters) did you use?

The weight of the tested is 20block,128filters;The selfplay weight is now 6block,128filter.

@Alderi-Tokori

This comment has been minimized.

Contributor

Alderi-Tokori commented Apr 30, 2018

I see, thank you!

@Hersmunch

This comment has been minimized.

Contributor

Hersmunch commented Apr 30, 2018

So no change to training to support draws with a Komi of 7?

@Marcin1960

This comment has been minimized.

Marcin1960 commented Apr 30, 2018

where is the download?

@sbf2000

This comment has been minimized.

sbf2000 commented Apr 30, 2018

So no change to training to support draws with a Komi of 7?

Two weights were used,play white according to weight of training with Komi 6.5,and paly black according to weight of training with Komi 7.5.

@hydrogenpi

This comment has been minimized.

hydrogenpi commented Apr 30, 2018

no bragging rights without published weights

@sbf2000

This comment has been minimized.

sbf2000 commented Apr 30, 2018

The weight of training with Komi 6.5 is announced in QQ group.

@hydrogenpi

This comment has been minimized.

hydrogenpi commented Apr 30, 2018

Announced or published? in any case if published, a link would be best, not everyone has a qq account

@hydrogenpi

This comment has been minimized.

hydrogenpi commented Apr 30, 2018

and 13 by 13?

@Marcin1960

This comment has been minimized.

Marcin1960 commented Apr 30, 2018

@sbf2000 "The weight of training with Komi 6.5 is announced in QQ group."

Link, please?

@gcp

This comment has been minimized.

Owner

gcp commented Apr 30, 2018

Two weights were used,play white according to weight of training with Komi 6.5,and paly black according to weight of training with Komi 7.5.

Thanks. Any particular reason for doing it this way instead of using komi = 7.0?

@sbf2000

This comment has been minimized.

sbf2000 commented Apr 30, 2018

Download Link:
https://pan.baidu.com/s/1rwndU38j3CZfxj4obGnQoQ
password:d97g

I haven't started training 13x13 yet.

@sbf2000

This comment has been minimized.

sbf2000 commented Apr 30, 2018

Any particular reason for doing it this way instead of using komi = 7.0?

1.komi = 7.0 is rarely used in reality, and komi = 7.5 or komi = 6.5 weights are more useful.

2.komi = 7.5 or komi = 6.5 does not require modification of the source code and is easy to implement.

3. By training the two weights of komi = 7.5 and komi = 6.5, it can be basically proved that the most balanced komi of 9x9 board is between 6.5 and 7.5.

4. Too many draw games are not conducive to achieving high scores on CGOS

@gcp

This comment has been minimized.

Owner

gcp commented Apr 30, 2018

Download Link:
https://pan.baidu.com/s/1rwndU38j3CZfxj4obGnQoQ

Thank you, I managed to download it.

Question: is this the weight file for white (6.5) komi, or for both players?

Second question: would you be OK if I mirror this file on my server, and add instructions for 9x9 to the README?

@sbf2000

This comment has been minimized.

sbf2000 commented Apr 30, 2018

First:The weight file is for both players.
Second:It's OK.
@gcp

@gcp

This comment has been minimized.

Owner

gcp commented Apr 30, 2018

@billyswong

This comment has been minimized.

billyswong commented Apr 30, 2018

Two weights were used,play white according to weight of training with Komi 6.5,and paly black according to weight of training with Komi 7.5.

The weight file is for both players.

So I guess the self-play is set such that while they used the same weight file, white plays as if the game is in komi 6.5, while black plays as if the game is in komi 7.5? Else you saying the weight file work for both sides doesn't make sense to me.

@Hersmunch

This comment has been minimized.

Contributor

Hersmunch commented Apr 30, 2018

@gcp How much effort would it be to add 9x9 releases now that there's some weights available?

@Ishinoshita

This comment has been minimized.

Ishinoshita commented Apr 30, 2018

@sbf2000 To train two value heads with different komi, I assume you either had each selfplay game to be played till end and so to disable resign feature, or you have modify the resign code to take into account both heads ?

@sbf2000

This comment has been minimized.

sbf2000 commented May 1, 2018

So I guess the self-play is set such that while they used the same weight file, white plays as if the game is in komi 6.5, while black plays as if the game is in komi 7.5? Else you saying the weight file work for both sides doesn't make sense to me.

In the course of self-play training,only replace the komi from 7.5 to 6.5,and the resulting weight is the same for both black and white,which is 6.5.
I was just testing on cgos,passing an GTP proxy,using two different weights at the same time(one komi 6.5,another komi 7.5),opening two Leelaz instances,and the proxy passing the black to an Leelaz,Pass the white to another.

@Ishinoshita
Same weight,only one value heads.

@KyleBiedermann

This comment has been minimized.

KyleBiedermann commented May 1, 2018

I noticed the lack of Zen on CGOS hopefully this didn't affect the rating accuracy too much considering the previous versions reached about 3000 range with it playing.

@sbf2000

This comment has been minimized.

sbf2000 commented May 1, 2018

@KyleBiedermann
http://www.yss-aya.com/cgos/9x9/cross/CLeelaz9Test8_20b.html
This is a test used weight of komi 6.5 alone,the result is that the white won all the time,and the black didn't lose.

@KyleBiedermann

This comment has been minimized.

KyleBiedermann commented May 1, 2018

@sbf2000 Yes I noticed I usually work some games from CGOS in my database to compare data, I am just saying that that the rating may be a little inflated due to lack of zen programs as it draws against it quite a few times.

@Hersmunch

This comment has been minimized.

Contributor

Hersmunch commented May 1, 2018

It is now top of bayes elo table too http://www.yss-aya.com/cgos/9x9/bayes.html

@hydrogenpi

This comment has been minimized.

hydrogenpi commented May 1, 2018

perhaps an officual 13 by 13 can be started by the project now that experience had been gained with 9 and 19

@PhilipFRipper

This comment has been minimized.

PhilipFRipper commented May 1, 2018

@hydrogenpi let us know how it goes

@diadorak

This comment has been minimized.

diadorak commented May 1, 2018

I am still confused. How do I use this? Can I load the network using leelaz -w 20_128-6.5.txt?

Does LZ know the board size automatically ?

@alreadydone

This comment has been minimized.

Contributor

alreadydone commented May 1, 2018

#1291 (comment)
I think you need to change the line and recompile.

@LetterRip

This comment has been minimized.

LetterRip commented May 1, 2018

Is the raw data going to be made available as well? Also do you have the older networks available or just the best network?

@sbf2000

This comment has been minimized.

sbf2000 commented May 2, 2018

The raw data is in the precess of sorting and will be released to the net disk later.Most of the old networks have been deleted.
@LetterRip

@Alderi-Tokori

This comment has been minimized.

Contributor

Alderi-Tokori commented May 3, 2018

It's beating minus 35 to 5 so it's quite a bit stronger. I've set minusGo to use these weights on OGS until sunday if anyone wants to have a go at it.

@rexl2018

This comment has been minimized.

rexl2018 commented May 8, 2018

Looking forward to your good news!

The raw data is in the precess of sorting and will be released to the net disk later.Most of the old networks have been deleted.

@y-ich

This comment has been minimized.

y-ich commented May 11, 2018

Hi.

Please let me advertise my outcome.
I implemented JavaScript version of Pyaq (https://github.com/ymgaq/Pyaq) using WebDNN (https://mil-tokyo.github.io/webdnn/) and have equipped it with the weights here.

Now you can enjoy to play Go with strong AI on your browser!

https://y-ich.github.io/A9.js/index.html
(Sorry, it is Japanese page.)

https://new3rs.github.io/AZ.js/index.html
(updated: I changed the repository of my app and now it supports 19-size board with ELF OpenGo weights as well as 9-size.)

On my old iMac(2012), the time of one evaluation is about 50ms, not so fast but it is strong enough.

I enjoyed my little project and thank authors of Pyaq(@ymgaq), WebDNN(@Kiikurage, @milhidaka), @sbf2000, and Leela Zero people so much!

@Tsun-hung

This comment has been minimized.

Tsun-hung commented May 25, 2018

LZ has been bugfixed several time since you opened this issue, do you catch up and continue to train the new networks?

@l1t1

This comment has been minimized.

l1t1 commented May 28, 2018

if the program can works as a engine for android gridmaster? #779

@KyleBiedermann

This comment has been minimized.

KyleBiedermann commented Jun 1, 2018

@y-ich cheers on the project of yours it looks good and in english now. I was curious if their was a way in the future to maybe implement custom positions (edit the board) to test various responds the program does in certain situations due to once a player knows the winning lines rarely does the program deviate from the previous game, as I have noticed. Keep up the great work!

@y-ich awesome on the new version allowing 9x9 and 19x19 play also the score percentage in positions.

@alreadydone alreadydone referenced this issue Jul 4, 2018

Open

Batching #1601

@jokkebk

This comment has been minimized.

jokkebk commented Jul 13, 2018

@y-ich great work! Games looked nice, and I lost against it, which is to be expected.

Special thanks for using jGoBoard, nice to see my library in use!

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment