zen7 30 second : leela-zero 30 second, Zen7 won consistently

[kyotosato]

game setting
zen7: uses gtp4zen-0.3.2 for sabaki
arguments: -z 7 -t 4 -T 30 -s 200000
Leeala-zero 13:
arguments: -g -noponder -t 4 -w leelaz-model-swa-4-32000.txt -p 500000
initial commands: time_settings 0 31 1;
(note: the weight text file is dated 4/9/2018, and the 31 is for constant 30 seconds thinking time)
computer: Acer One 10 ; cpu: Intel® Atom™ x5-Z8350 (32 bit, NOT 64 bit)
1.4 GHz; Quad-core; OS: windows 10

[kyotosato]

Note1: gtp4zen-0.3.2 is a program designed for Zen7 or Zen6 to be used as a bot in sabaki. I got it by searching the web thoroghly. It is a user-developed program (by someone), which or whom has nothing to do with or no relation to original Zenith author or co.
Note2: weight file matters greatly. I used the file downloaded from http://zero.sjeng.org/ which updated daily.
Note3: The result may be inconclusive due to the weight file updated daily and CPU and GPU grade , the two factors

[pnprog]

This is interesting, do you have any idea how much play-outs are performed on that hardware during those 30s thinking time?

[Vargo]

Very interesting ! I have Zen 6, so, I would be happy to test it against AQ or LZ, but where can I find this gtp4zen ? (I don't speak japanese or chinese)

[kyotosato]

my note3 said it is inconclusive. at least not in statistical sense of "consistancy," since i posted the game setting, I would like people to be able continuously testing leela-zero 13. I will do the same because, i can set the game over night, and let it run by itself. when time goes by i will be able to say something about these two bots more definitively. right now just a feeling by looking game's winrate hisory of the only two games i experienced. the leelazero once dropped down steadily, then could never recover later on.
by the way, about the arguments of leelazero I might need some folks who have experienced of LZ to give me some advice so i can make adjustment.

[kyotosato]

as for gtp4zen-0.3.2.rar:
since it is a shared and password protected zipped file in the baidu cloud, which is a cloud company of china.
Without violating copyright law of any country, i only could give you a website address,
https://tieba.baidu.com/p/5447958378?pi ... 5553152055
in it, third post, dated 2017-11-26 15:35, has a pair of things:
1bo----z,---4
the first is the baidu shared file ID, second is the password when asked.
that is all I could post, other than the above info you might have to find someone knows chinese and how to operate and access baidu cloud disk from US or other area.
I read chinese and japanese by using Google translate to get approximate idea. you may try that way first.
The author is a chinese and he/she keeps on updating the program as far as i can tell.

[Vargo]

Thanks, I'll look into it.
There was THAT,
which works, but has issues with dual GPUs, it tends to crash the system.

[kyotosato]

it is not the same i believe, because the author in my post is the original author, and the file has been updated to a newer version (it does have 2 prior version 0.2.3 and 0.3.1, the one i used is 0.3.2. i tell by the date, 6/2017 and 11/2017. on the other hand, may be you are right. I cannot talk about crash 'cause I do not have dual GPUs.

[kyotosato]

out of curiosity, i myself did tests to see how strong Zen7 commercial; i registered in wild fox site, and temgen site. Zen7 reaches 9dan in temgen and 8dan in fox with auto/random matching. Both tests took me quite a time, from 3dan default to 8dan and 9dan. if i have time i will try to play with ZenDeep bot in Temgen (only play with 9dans). For wild fox, i have to go further to reach 9dan (which is easy for Zen7) to play with Jue-Yi(name? the china's strongest AI go bot.) and the computer i used is a pentium 4, with Windows XP.
BTW, to use Zen7 (or Zen6) other than dan setting or time setting through out a game, third way is to employ Zen7's "Analysis" feature along with the other two ways to play with an opponent, manually.
The analysis feature is not only good for after-game analysis, but also good for playing a game in realtime. I will explain how to do it later in case you do not yet know how.

[kyotosato]

To play with an opponent realtime by Zen7 "analysis" feature:
1. setup by choosing view->pane->analysis (1st pic)
2. also choose view->hotspots (1st pic)
3. when on-line play begins (you click for both players, yourself and your opponent), click "analysis" or edit->analysis (2nd pic)
4. play by clicking the suggested the spot (blinking one among hotspots), and afterward it will show your opponent's hotspots (3rd pic)
note1: sometimes you have to wait longer before you choose the suggested spot, if there are a few competitive spots judging from winrates. wait longer the suggested spot may change.
note2: by this way, zen7 is about 8-9dan in wildfox and temgen, and you could save play time for some trivial moves.

screenshot_zen701.JPG (121 KiB) Viewed 21540 times

screenshot_zen702.JPG (125.57 KiB) Viewed 21540 times

screenshot_zen703.JPG (124.97 KiB) Viewed 21540 times

[Bas]

arguments: -g -noponder -t 4 -w leelaz-model-swa-4-32000.txt -p 500000
initial commands: time_settings 0 31 1;
(note: the weight text file is dated 4/9/2018, and the 31 is for constant 30 seconds thinking time)
computer: Acer One 10 ; cpu: Intel® Atom™ x5-Z8350 (32 bit, NOT 64 bit)
1.4 GHz; Quad-core; OS: windows 10

Is this an unofficial network? How many blocks and filters does it have, how strong is it, where is it from?

But the result is not unexpected. Zenith Go 7 (which uses Zen 15, not 7) is very strong even it uses the CPU only, whereas LZ needs a GPU to play strong. With a 15b network I have 850n/s on my GTX 1080, but 75n/s when using the CPU with the 64bit version. The 32bit version is even worse, only 25n/s.

[kyotosato]

benchmark evaluation is not my concern, as a go player, and an AI practitioner. i concern about the availability, accessibility, and controlled envi testing of various computer go software. before discussing CPU prone, and/or GPU prone, we do need to know the GPU factor itself weighted how heavy to the baduk/go/weiqi game as a whole. Some "guesses" 1 or 2 stones differences, i cannot judge it without research on it, or statistic data, or benchmark evaluation. it can only be sure that GPU matters. matter how much?
as to temgen deepzen bots, i wonder if it is offered by zenith company. if it is , then that is another story. those bots maybe become test-run for next version Zen8 (or whatever name) commercial.
BTW, any optimization method has overshooting problem. from my own observation, Zen7 may be less overshooting prone than leela-zero, how to tell? from the trend of winrates history of a game. just my own guess.

[Vargo]

I've finally managed to download gtp4zen, but it doesn't work with zen6...
The initialization phase seems ok, at the end it says zen.dll(zen6) initialize success, and then ... nothing.
Anyway, thanks all the same.

I think there's something missing in gtp4zen's commands : you can't disable pondering. It doesn't seem fair to the other program, as a lot of the CPU could be used by zen's pondering.

[kyotosato]

I use sabaki; Zen6 uses argument -z 6 differs from zen7's -z 7. I have zen6, and i will try it out, see if it works for zen6. as for -noponder, i will try it out. as for the game shown above, it didn't ponderring by my checking CPU meter from time to time.
might i ask what is your gtp4zen version?

[Vargo]

I've tried with the 3 available versions (0.2.3, 0.3.1, 0.3.2)
Sabaki 0.33.4 with parameters : -z 6 -t 4 -T 10 -s 10000