search

pangafu / LeelaMasterWeight

Leela Master weight is training from leela zero self-play sgf and human sgf file
https://drive.google.com/drive/folders/1bB8ee1wFuRWL9nPhsl4_BPUhcWSBuxO0?usp=sharing
GNU General Public License v3.0
49 stars 6 forks source link

Leela OZ with Dkomi is not working correctly #17

Open tartaric opened 5 years ago

tartaric commented 5 years ago

Hello,

I tried Leela Master GX37 with the komi+half-0829-3 of alreadydone on Sabaki and it works perfectly (though being weaker than Leela 11 GPU above 5 stones). It can beat Leela 11 or Zen 6 with 4 handicaps.

But for Leela OZ 18 (supposed stronger than Leela Master GX 37?) with the LeelaZero_DKomi+Filter (file of 947 KB) and with these settings: -m 12 -g -r 1 --km-player=1 --km-startmovenum=7 --km-filterstep=40 --km-s1 --km-s1-bias=82.5 --km-s1-step=3.75 --km-s1-maxwr=0.35 --km-s1-minwr=0.05 --km-s2 --km-s2-target=15 --km-s2-step=0.5 --km-s2-maxwr=0.80 --km-s2-minwr=0.55 --km-s3 --km-s3-step=0.5 --km-s3-maxwr=0.80 --km-s3-minwr=0.55 -w OZ.txt.gz it appeared really weak at 4 stones of handicaps, playing weird moves (as the slide on the 2nd line after approaching a Hoshi stone) and unable to defeat Leela 11 GPU with 4 stones.

So my question is what's wrong? Is --km-s1-bias not settled correctly? I did my test putting the komi at 0.5. Also, the winrate is strange, it appears being 0,01 or even 0,00% for the 2 first White moves and then rises to 35,5%. On the console I can see "STAGE" passing from 0 to 1 and being writen "KOMI STAGE 0 -> 1", what is the meaning of that?

One final question. It is written in the instructions: "km-s1-bias Adjust to winrate 15% when OZ(white) play first move". How to "adjust"? How does it work? It is already put at 82.5, but if I changed the handicaps to 3 for example, how can I know what is the number to put? And if I understood correctly, I should put km-startmovenum to 5 (cause it's handi 2-1 so 32 = 6-1 = 5?) and --km-filterstep to 30 cause it's handi*10?

Thanks for your help.

pangafu commented 5 years ago

Maybe you should set --km-startmovenum=4 for handicap 4 stone. In different GTP environment, the km-startmovenum maybe different, adjust to "KOMI STAGE 0 -> 1" in white first move.

tartaric commented 5 years ago

Thanks for your answer. The komi is now showing correct percentages from the first move, but it's still playing badly compared to GX37.

pangafu commented 5 years ago

Please see the test condition, in my test, oz handicap upper limit is very high, but need high po

tartaric commented 5 years ago

@pangafu Could you copy-paste or screen the commands you entered for the settings please? And I don't understand what means "wpo" in the test. Anyway I tried to run OZ18 with no limit of PO/time and still weak.

pangafu commented 5 years ago

I had write in OZ18.txt, I had test many times, it did work~