Life In 19x19
http://www.lifein19x19.com/

HayLee vs Leela Zero game, review from LZ point of view
http://www.lifein19x19.com/viewtopic.php?f=4&t=15520		 Page 1 of 1
Author:  pnprog [ Mon Mar 19, 2018 1:23 am ]
Post subject:  HayLee vs Leela Zero game, review from LZ point of view
Hi all!

Following the HayLee vs Leela Zero game on OGS on 2018-03-15, I reconstructed the RSGF file using the log created by Leela Zero, as black during the match, and completing with an analysis for white move with same network.

So the RSGF file includes all variations plus data (win rate, policy network, play-outs).

Here is the win rate delta graph for HayLee:
Attachment:
White Value Network win rate delta graph.png
White Value Network win rate delta graph.png [ 10.72 KiB | Viewed 7061 times ]

As you can see, no clear mistake from HayLee (which is in fact very scary). As white, Leela Zero would have resign at move 92.

Here is the RSGF file: http://ns1297.imingo.net/Haylee-LeelaZero.rsgf
You will need that development version of GoReviewPartner to display it properly: http://ns1297.imingo.net/goreviewpartner-dev20170319.zip

If you prefer to read it from a standard SGF editor (good luck) here is a conversion from RSGF to SGF: http://ns1297.imingo.net/Haylee-LeelaZero.rsgf.sgf

Screen-shot for move 100:
Attachment:
move100.png
move100.png [ 57.95 KiB | Viewed 7047 times ]


(not so) Fun fact: On my very modest computer (integrated GPU from 2014) it took me ~42 hours to perform the analysis of HayLee moves using the same network, and same number of visits as during the match :mrgreen:
Author:  Vargo [ Mon Mar 19, 2018 3:58 am ]
Post subject:  Re: HayLee vs Leela Zero game, review from LZ point of view
Quote:
(not so) Fun fact: On my very modest computer (integrated GPU from 2014) it took me ~42 hours to perform the analysis of HayLee moves using the same network, and same number of visits as during the match :mrgreen:
Thank you for this program.
I found it not so easy to install, but once it's up and running, it's a strong pro teacher at home, showing you exactly what you did wrong... Beautiful !

I tried to run the same analysis you did, checking
"all the 287 moves", and "white moves only".
Command line :
--gtp --noponder -w ed26f.txt --gpu 0 --gpu 1 --visits=150'000

It took 43 minutes, with an additional 5-10 minutes at the end of the analysis, when GRP was "closing"

So, with good GPU(s) it must take between 30 minutes and 2 hours to go through a whole game. It seems reasonable for such an in-depth analysis.
Author:  johnsmith [ Mon Mar 19, 2018 4:24 am ]
Post subject:  Re: HayLee vs Leela Zero game, review from LZ point of view
Thank you very much for this! I like this form of reviewing very much and I feel very comfortable being able to easily scroll through variations.

P.S. Is there a website or even a thread where one can share their pro game analysis (with GRP and especially LZ)? I would be very interested into sharing mine and looking at other people's rsgf files. I think it would also be a good PR step :)
Author:  pnprog [ Mon Mar 19, 2018 9:20 am ]
Post subject:  Re: HayLee vs Leela Zero game, review from LZ point of view
Vargo wrote:
I tried to run the same analysis you did, checking "all the 287 moves", and "white moves only".
Command line :
--gtp --noponder -w ed26f.txt --gpu 0 --gpu 1 --visits=150'000


In your case, around 10 seconds to reach 150000 visits. In my case, it takes around 1000 seconds :D
I guess your hardware is on par with what was used during the match with HayLee. May I as ask what GPU is that?

In case some other users with a slow computer like mine want to try, there is another concern: if not told otherwise, Leela Zero thinks for a maximum of 50 seconds. So in top of the command line parameter --visits=150000 I also had to give a maximum time per move of 10 000 seconds (I added plenty of margin just in case). The time per move parameter (in GRP settings) uses the GTP time_settings command telling Leela to use 0 second main time, plus 1 period of 10 000 seconds bioyomi. One should check that the sum of the number of play-outs for all variations of the first move adds up to around 150 000 to be sure.

johnsmith wrote:
P.S. Is there a website or even a thread where one can share their pro game analysis (with GRP and especially LZ)? I would be very interested into sharing mine and looking at other people's rsgf files. I think it would also be a good PR step :)
Haha, not at the moment. I think a website like eidogo for RSGF would be nice, it would allow to review games on a smartphone.

In fact, I started to "standardize" the way information is stored into RSGF file (so I am working to define a set of SGF properties used by GRP) to make it possible to implement RSGF viewer (without the analysis part) for that raison.
Author:  Vargo [ Mon Mar 19, 2018 10:05 pm ]
Post subject:  Re: HayLee vs Leela Zero game, review from LZ point of view
My computer is : 2x1080Ti, i9 7920X.
I tried the same analysis with the dev. version of GRP, it's a bit quicker, with better graphs.
Although I checked "all the 283 moves", the finished analysis only has 281 moves...
https://drive.google.com/open?id=10-KFxL8p-DLJuSiE-gavizW4TtJN9WQi
Author:  pnprog [ Wed Mar 28, 2018 8:10 pm ]
Post subject:  Re: HayLee vs Leela Zero game, review from LZ point of view
Vargo wrote:
Although I checked "all the 283 moves", the finished analysis only has 281 moves...
This is certainly because the two last moves are "pass" and GRP does not analyse them. In the older versions, I had trouble determining the color to play in those cases, so I would skip them.

Now that I found a way around, I should add them again, to see if there are some missed end game tesuji or whatever.
Page 1 of 1 All times are UTC - 8 hours [ DST ]
Powered by phpBB © 2000, 2002, 2005, 2007 phpBB Group
http://www.phpbb.com/