Sorry, for the moment, I can't make it work...
| Life In 19x19 http://www.lifein19x19.com/ |
|
| AQ v4.0.0 http://www.lifein19x19.com/viewtopic.php?f=18&t=17479 |
Page 1 of 2 |
| Author: | And [ Mon May 11, 2020 9:57 am ] |
| Post subject: | AQ v4.0.0 |
https://github.com/ymgaq/AQ/releases Both the Japanese and Chinese rules are now supported. Lizzie display is now supported. Network files learned in the GLOBIS-AQZ project. Network optimization is now possible with TensorRT. |
|
| Author: | And [ Mon May 11, 2020 11:00 am ] |
| Post subject: | Re: AQ v4.0.0 |
only CUDA version? |
|
| Author: | And [ Mon May 11, 2020 11:12 am ] |
| Post subject: | Re: AQ v4.0.0 |
Vargo will you make a test match? |
|
| Author: | Vargo [ Tue May 12, 2020 7:23 am ] |
| Post subject: | Re: AQ v4.0.0 |
Sorry, for the moment, I can't make it work...
|
|
| Author: | And [ Tue May 12, 2020 7:48 am ] |
| Post subject: | Re: AQ v4.0.0 |
thank! maybe a simpler version will be released 
|
|
| Author: | Vargo [ Wed May 13, 2020 10:25 pm ] |
| Post subject: | Re: AQ v4.0.0 |
Some very quick AQ v4.0.0 tests : gogui-twogtp 1.5.1 1xGTX 1080 1s/move for AQ, no pondering chinese rule, komi 7.5 AQ v. GnuGo AQ wins 10-0 (6 times by resignation, 4 times by 30-40 points, average length 269 moves, average time AQ=143s/game, Gnu=50s) AQ v. KG 1.4.2 g170e-b20c256x2-s4384473088-d968438914 (limited to 1po) AQ wins 10-0 (10 times by resignation, 1 duplicate game, average length 89 moves, average time AQ=48s/game, KG=0.3s) AQ v. KG 1.4.2 g170e-b20c256x2-s4384473088-d968438914 (1s/move) KG wins 20-0 (20 times by resignation, 0 duplicate game, 0 error, average length 174 moves, average time AQ=94s/game, KG=84s) But there's something strange: In AQ v. Gnu, the time looks ok : AQ has 100% gpu time and 0% cpu time, GnuGo has 100% cpu time KG has sometimes 0% cpu time, and sometimes not ??? So... Idon't know  
|
|
| Author: | Vargo [ Thu May 14, 2020 3:24 am ] | ||||
| Post subject: | Re: AQ v4.0.0 | ||||
Three AQv4.0.0 games (gpu: 1x1080, Sabaki 0.51.1 from the AI-baduk megapack) AQv4.0.0 : no pondering, 5s/move, chinese rule, komi 7.5, resign threshold 0.02 (and 0.00 for the handicap game) AQ wins by resignation against GnuGo at H5 (Maybe AQ isn't designed to play handicap go ?) AQ wins several games (here as B) by resignation against Zen7 9d (zen 9d parameters from the baduk megapack, about time parity) AQ loses several games (here as W) against LZ best network (#271 I guess, LZ parameters from the baduk megapack, about time parity)
|
|||||
| Author: | And [ Thu May 14, 2020 5:22 am ] |
| Post subject: | Re: AQ v4.0.0 |
Vargo wrote: KG has sometimes 0% cpu time, and sometimes not ??? I also have. task manager shows CPU utilization when KataGo thinks. somewhere error. thanks for the tests! |
|
| Author: | lightvector [ Thu May 14, 2020 5:49 am ] |
| Post subject: | Re: AQ v4.0.0 |
I think I know the issue. Please ignore the cputime value for KataGo right now and just pay attention only to the normal total time - I'll fix the issue next release (which might be a little while, unless you think this error is important). |
|
| Author: | Vargo [ Thu May 14, 2020 8:46 pm ] |
| Post subject: | Re: AQ v4.0.0 |
it's not important at all, thanks. |
|
| Author: | Dragon [ Fri May 15, 2020 4:05 am ] |
| Post subject: | Re: AQ v4.0.0 |
Is the project version of the program for android closed? |
|
| Author: | And [ Sat May 16, 2020 12:48 pm ] |
| Post subject: | Re: AQ v4.0.0 |
looks like this |
|
| Author: | And [ Sun May 17, 2020 6:31 am ] |
| Post subject: | Re: AQ v4.0.0 |
Vargo with what parameters do you run AQ? |
|
| Author: | Vargo [ Sun May 17, 2020 9:04 am ] |
| Post subject: | Re: AQ v4.0.0 |
config: --num_gpus=1 --num_threads=4 --rule=0 --repetition_rule=0 --komi=7.5 --main_time=0.0 --byoyomi=5.0 --byoyomi_margin=0.0 --batch_size=2 --search_limit=-1 --node_size=32768 --use_ponder=off --resign_value=0.02 --save_log=off (it's mainly the default values) Against Zen, I guess AQZ should have used --rule=1 --komi=6.5, because Zen is a japanese program, but the results would probably have been the same. |
|
| Author: | And [ Sun May 17, 2020 10:13 am ] |
| Post subject: | Re: AQ v4.0.0 |
run in Sabaki like LZ? |
|
| Author: | Vargo [ Sun May 17, 2020 12:10 pm ] |
| Post subject: | Re: AQ v4.0.0 |
Yes, gogui-twogtp or Sabaki, everything is in the config file. ## 2. Requirements + OS : Windows 10, Linux (64-bit) + GPU : Nvidia's GPU ([Compute Capability](https://developer.nvidia.com/cuda-gpus) >3.0) + [CUDA Toolkit](https://developer.nvidia.com/cuda-toolkit) 10.0 or 10.2 + [TensorRT 7.0.0](https://docs.nvidia.com/deeplearning/sd ... index.html) Edit : and I have been asked to install visual studio, along the way. |
|
| Author: | And [ Sun May 17, 2020 12:17 pm ] |
| Post subject: | Re: AQ v4.0.0 |
thank! I assumed that it connects without problems, but now I can not check. I don’t understand why: https://lifein19x19.com/viewtopic.php?p=256618#p256618 in q5go engines are easily connected! |
|
| Author: | mb76 [ Sun May 24, 2020 3:00 am ] | ||||
| Post subject: | Re: AQ v4.0.0 | ||||
For fun I did a few match games in Sabaki to test the strength of GLOBIS-AQZ against different neural nets. Out of 14 games GLOBIS-AQZ won 8 games! My personal observation about the strength of GLOBIS-AQZ: At fixed playouts AQZ is stronger than Leela Zero (#272) but weaker than KataGo 40b (#b40-s3708042240-d967) and Leela Zero (#273). Does not fear ladders (game C1) and takes on ko's. Playing strength (Including the attached two bonus games): KataGo 40b > LeelaZero #273 > Globis-AQZ > Leelazero #272. Ps. Maybe the conclusion will be different with a few thousand test games instead of only fourteen games Attached: -Original games (.ZIP file) -Analyzed games A t/m K with KataGo 40b (.ZIP file) -Bonus games Leela Zero #273 vs Katago 40b (.ZIP file) Games: Game A1: -Result:AQZ wins -"White:SAI Zedirr-nn" -Playouts:Time -Comment:A double Ko starting from move 110. Game B1: -Result:AQZ wins -"White:SAI Zedirr-nn" -Playouts:3200 -Comment:AQZ captures a big group (@ D11). Game C1: -Result:AQZ wins -"White:SAI LZ 272" -Playouts: 3200 -Comment:LZ 272 starts a non-working ladder from move 46. Game D1: -Result:AQZ loses -"White:KataGo 40b" Playouts:3200 -Comment:Game looks beautifully complicated. Game E1: -Result:AQZ loses -"Black:KataGo 40b" -Playouts: 3200 -Comment:According to the analyzed sgf AQZ was winning till move 214 (loses due a ko at R15/16?). Game F1: -Result:AQZ wins -"Black:SAI LZ 272" -Playouts: 3200 -Comment: The placement of move 60! looks like a tesuji. Check the moves till 96. Move 46 necessary? Game G1: -Result:AQZ wins -"White:SAI 341" -Playouts: 3200 -Comment: To move 40 a nice looking joseki (bottem right). Game H1: -Result:AQZ wins -"Black:SAI 341" -Playouts: 3200 -Comment: How to make two eyes starting from move 60! Game I1: -Result:AQZ wins -"White:SAI LeelaMaster GX58" -Playouts: 3200 -Comment: Looks complicated! Game J1: -Result:AQZ loses -"Black:SAI LeelaMaster GX58" -Playouts: 3200 -Comment: Surprisingly AQZ loses (because of the Ko starting from move 94?) Game K1: -Result:AQZ wins -"Black:SAI LeelaMaster GX58" -Playouts: 6400 -Comment: Like the placement of move 99! is the bottem-left group captured? Game L1: -Result:AQZ loses -"White:SAI LZ 273" -Playouts:3200 -Comment: Lots of epic moves Game M1: -Result:AQZ loses -"Black:SAI LZ 273" -Playouts:3200 -Comment: Lots of epic moves Game N1: -Result:AQZ loses -"White:SAI LZ 273" -Playouts:6400 -Comment: Lots of epic moves
|
|||||
| Author: | Vargo [ Sun May 24, 2020 10:22 am ] |
| Post subject: | Re: AQ v4.0.0 |
mb76 wrote: At fixed playouts AQZ is stronger than Leela Zero (#272) On my PC, at 3200 po , AQZ takes 2.4 times longer than LZ(272), and 1.5 times longer than KG(40b) ...
|
|
| Author: | And [ Fri May 29, 2020 8:25 am ] |
| Post subject: | Re: AQ v4.0.0 |
mb76 launched in q5go? |
|
| Page 1 of 2 | All times are UTC - 8 hours [ DST ] |
| Powered by phpBB © 2000, 2002, 2005, 2007 phpBB Group http://www.phpbb.com/ |
|
