pangafu / leelamasterweight Goto Github PK

hi,

just wanted to send message saying thank you for these leelaMaster weights.
They are great to play against and learn from.

Do you plan on releasing more when leelaZero gets stronger with more training games.?

What human games do you use, pro, amateur?

Thank you.

ELF 20B weights won both 30B networks on 30 seconds (2x1080 ti) Should I retest on 60 seconds, or these networks are just weaker (not trained)?
Thanks

What is the difference between LeelaZero_Z01 and LeelaZero_Z02 ?

网盘中，有 GX24说明.txt 和 G24.zip 。
这个后者是不是就是 GX24 权重？

似乎没有 GX42 ？

Like Fox Go Server (a.k.a Yehu Weiqi in China or HanQ Baduk in Korea), there are also online Go playing platforms in Korea. Hangame Baduk is one of them.

In the Korean Go playing platform Hangame Baduk, they run HanDol 24 hours a day there.

Many Korean Go players play with HanDol at there. Normal computer Go programs like MiniGo v17 or Leela Zero are difficult to win HanDol consistently. It seems HanDol on there is weaker than its competition version (maybe due to its slower running machine), but it is strong enough to win against MiniGo or Leela Zero.

I tested OX24 on it with 3 Titan V GPUs. OX24 tends to lose when it plays as black, but OX24 consistently beats HanDol when it plays as white.

For your reference, I attach 4 games here. Check it out.

Game 1 SGF

(;AP[PureBoard:1.0]CA[utf-8]FF[4]GM[1]ST[2]GN[]PC[]RE[W+R]PW[OX24]PB[HanDol]SZ[19]HA[0]KM[7.5];B[qd];W[dd];B[pp];W[dp];B[nd];W[qq];B[qp];W[pq];B[op];W[nr];B[pj];W[mq];B[fc];W[hc];B[jc];W[fd];B[ed];W[ee];B[ec];W[gd];B[dc];W[cd];B[cc];W[bc];B[bb];W[bd];B[ef];W[fe];B[cf];W[cg];B[gc];W[hd];B[bf];W[ab];B[ba];W[de];B[df];W[gb];B[fb];W[ib];B[ga];W[hb];B[db];W[ci];B[ei];W[dk];B[gg];W[kd];B[mc];W[kc];B[jf];W[fj];B[lf];W[lb];B[cq];W[cp];B[dq];W[ep];B[eq];W[fq];B[fr];W[mb];B[nb];W[oc];B[nc];W[qh];B[rj];W[pe];B[qe];W[oi];B[pf];W[oj];B[pk];W[ok];B[ol];W[nl];B[nm];W[ml];B[gi];W[of];B[pg];W[ob];B[oa];W[gj];B[hi];W[hj];B[ji];W[bq];B[br];W[er];B[bp];W[gr];B[aq];W[om];B[pl];W[nn];B[ij];W[ik];B[jk];W[jl];B[mp];W[lp];B[nq];W[mo];B[or];W[oq];B[np];W[mr];B[rq];W[pr];B[fp];W[gq];B[il];W[hk];B[cm];W[pm];B[qm];W[qn];B[rm];W[kk];B[em];W[fo];B[jj];W[dn];B[dm];W[fi];B[fh];W[gm];B[gl];W[fl];B[kl];W[jm];B[km];W[jn];B[lk];W[mm];B[jd];W[je];B[ie];W[ke];B[id];W[jb];B[if];W[gf];B[ff];W[hf];B[he];W[kf];B[lg];W[kg];B[kh];W[ka];B[na];W[me];B[le];W[ld];B[mh];W[oe];B[pc];W[pb];B[qb];W[od];B[pa];W[md])

Game 2 SGF

(;AP[PureBoard:1.0]CA[utf-8]FF[4]GM[1]ST[2]GN[]PC[]RE[W+R]PW[OX24]PB[HanDol]SZ[19]HA[0]KM[7.5];B[pc];W[dd];B[pp];W[dp];B[pf];W[qq];B[pq];W[qp];B[po];W[rn];B[cq];W[cp];B[dq];W[eq];B[er];W[fq];B[fr];W[gq];B[fc];W[cf];B[qm];W[ic];B[kc];W[ie];B[ke];W[ec];B[fd];W[ff];B[gf];W[fg];B[hd];W[jc];B[kd];W[je];B[gg];W[jg];B[ig];W[if];B[fe];W[ih];B[fh];W[dg];B[kf];W[hg];B[gh];W[eb];B[lh];W[qd];B[re];W[pd];B[oc];W[od];B[mc];W[ng];B[og];W[qn];B[pn];W[oh];B[nf];W[ph];B[rg];W[mf];B[nh];W[mg];B[mh];W[qf];B[qe];W[qg];B[pe];W[ne];B[rh];W[rf];B[sf];W[pj];B[rj];W[qc];B[of];W[rd];B[pb];W[qb];B[se];W[sd];B[rm];W[md];B[lb];W[qa];B[mk];W[rk];B[qk];W[qj];B[ri];W[ql];B[pk];W[dj];B[gk];W[ek];B[ik];W[kh];B[lg];W[kj];B[fk];W[hi];B[fj];W[hl];B[hk];W[pr];B[or];W[qr];B[kk];W[gb];B[sn];W[ro];B[fm];W[mj];B[nj];W[iq];B[bp];W[bo];B[bq];W[cn];B[kp];W[nq];B[oq];W[nr];B[ns];W[lr];B[lq];W[kr];B[np];W[mp];B[qo];W[rp];B[mo];W[mr];B[mq];W[lk];B[ll];W[lj];B[ml];W[jk];B[kl];W[el];B[hp];W[ip];B[gr];W[hq];B[hn];W[jj];B[di];W[ci];B[jl];W[in];B[ej];W[dk];B[im];W[ho];B[jn];W[gn];B[io];W[go];B[gm];W[gc];B[gd];W[jo];B[hr];W[in];B[fb];W[fa];B[io];W[ps];B[os];W[in];B[hc];W[hb];B[io];W[pl];B[ok];W[in];B[jb];W[ib];B[io];W[pm];B[rl];W[in];B[id];W[jd];B[io];W[jp];B[ko];W[en];B[de];W[ce];B[dh];W[ch];B[ir];W[kq];B[lp];W[pa];B[nb];W[ja];B[ka];W[kb];B[jr];W[jq];B[jb];W[fl];B[gl];W[kb];B[em];W[dm])

Game 3 SGF

(;AP[PureBoard:1.0]CA[utf-8]FF[4]GM[1]ST[2]GN[]PC[]RE[W+R]PW[OX24]PB[HanDol]SZ[19]HA[0]KM[7.5];B[cd];W[dp];B[pd];W[pp];B[fd];W[qc];B[qd];W[pc];B[od];W[nb];B[qq];W[pq];B[qp];W[qn];B[qo];W[po];B[rn];W[qm];B[rm];W[ql];B[cn];W[fq];B[cp];W[cq];B[bq];W[dr];B[bo];W[do];B[di];W[ph];B[em];W[mc];B[hp];W[kp];B[gq];W[br];B[jo];W[ko];B[jm];W[jp];B[io];W[lm];B[fr];W[fp];B[er];W[aq];B[bp];W[eq];B[fn];W[en];B[dn];W[eo];B[jk];W[ck];B[dl];W[ch];B[ci];W[bi];B[bj];W[bh];B[cj];W[ce];B[dd];W[be];B[de];W[df];B[dg];W[dh];B[ef];W[eh];B[cf];W[bf];B[gh];W[ei];B[dk];W[gi];B[hi];W[cg];B[kc];W[gj];B[md];W[lc];B[kd];W[ld];B[le];W[ke];B[lf];W[je];B[kb];W[nd];B[ne];W[me];B[mf];W[ic];B[ib];W[hc];B[hb];W[gc];B[jd];W[id];B[ie];W[kg];B[if];W[kf];B[md];W[lb];B[jb];W[me];B[lh];W[nf];B[md];W[lg];B[mg];W[me];B[ng];W[oe];B[og];W[kh];B[ki];W[ji];B[kj];W[mh];B[li];W[pg];B[of];W[pf];B[ne];W[hh];B[pe];W[hg];B[gb];W[gf];B[he];W[pr];B[rl];W[qk];B[ml];W[mm];B[bd];W[rr];B[rk];W[rd];B[re];W[rc];B[oj];W[qf];B[nm];W[nn];B[qe];W[pj];B[ok];W[nl];B[jr];W[hr];B[iq];W[kr];B[om];W[rj];B[ri];W[rf];B[qi];W[qj];B[qr];W[qs];B[rq];W[ol];B[pl];W[nk];B[pi];W[nj];B[oi];W[se];B[oc];W[ob];B[ii];W[ih];B[jj];W[jh];B[hk];W[ff];B[ni];W[mj];B[pk];W[pm];B[sj];W[on];B[ks];W[ls];B[js];W[lr];B[gr];W[ej];B[aj];W[ad];B[ac];W[ae];B[bc];W[jq];B[ir];W[dj];B[bk];W[go];B[sr];W[ho];B[ip];W[rs];B[eg];W[fg];B[ar];W[as];B[ap];W[ar];B[kn];W[ln])

Game 4 SGF

(;AP[PureBoard:1.0]CA[utf-8]FF[4]GM[1]ST[2]GN[]PC[]RE[W+R]PW[OX24]PB[HanDol]SZ[19]HA[0]KM[7.5];B[dq];W[pp];B[dd];W[pd];B[dn];W[cc];B[dc];W[cd];B[de];W[bf];B[qc];W[qd];B[pc];W[nc];B[oc];W[od];B[nb];W[mc];B[mb];W[lc];B[nq];W[qn];B[pq];W[qq];B[qr];W[oq];B[pr];W[np];B[or];W[op];B[mq];W[pj];B[mp];W[cg];B[hd];W[fq];B[hp];W[cq];B[cr];W[dp];B[cp];W[bq];B[ep];W[dr];B[do];W[eq];B[pl];W[dp];B[qm];W[pn];B[nl];W[mo];B[ph];W[nj];B[nh];W[lo];B[ne];W[rc];B[rb];W[qb];B[pb];W[ra];B[qa];W[sb];B[sa];W[og];B[oh];W[ra];B[ce];W[be];B[sa];W[hc];B[sc];W[id];B[ie];W[gd];B[he];W[fc];B[jd];W[ic];B[kc];W[je];B[ke];W[jf];B[kf];W[jg];B[kg];W[jh];B[gf];W[ff];B[fg];W[ef];B[kh];W[ji];B[hh];W[lj];B[eg];W[ge];B[hg];W[dg];B[ll];W[hi];B[gi];W[hj];B[gj];W[gk];B[ej];W[eh];B[fh];W[fk];B[dh];W[im];B[oj];W[ok];B[ol];W[mf];B[qg];W[oi];B[jl];W[il];B[kn];W[rq];B[jp];W[mm];B[ml];W[ki];B[mi];W[mj];B[kb];W[dl];B[cj];W[ko];B[rn];W[kq];B[ro];W[qp];B[lr];W[jo];B[qj];W[qi];B[ri];W[pi];B[rh];W[kr];B[lq];W[kp];B[cl];W[rr];B[ns];W[fp];B[dq];W[qk];B[rj];W[dp];B[cb];W[bb];B[dq];W[qs];B[ps];W[dp];B[ha];W[fb];B[dq];W[pk];B[rl];W[dp];B[fe];W[df];B[dq];W[rk];B[sk];W[dp];B[bc];W[bd];B[dq];W[sj];B[si];W[dp];B[ba];W[ab];B[dq];W[dk];B[ck];W[dp];B[eo];W[bp];B[dq];W[fo];B[er];W[dp];B[nn];W[nm];B[dq];W[ds];B[om];W[lm];B[cs];W[dp];B[on];W[no];B[dq];W[ld];B[kd];W[dp];B[ac];W[eb];B[dq];W[me];B[es];W[nd];B[of];W[pf];B[ng];W[nf];B[pg];W[oe];B[fn];W[gn];B[hr];W[dr];B[ds];W[fr];B[fs];W[co];B[dp];W[cn];B[cm];W[br];B[bs];W[dm])

Did they prove bad in some way, are they deliberately not made public yet or is it just mistake? Thanks

Ran this for 2 millions playouts per move, both sides the GXAA

http://eidogo.com/#3GwR01KlJ
https://gokibitz.com/kifu/r1V-rty_H

The link seems to point at a deleted file
https://drive.google.com/drive/folders/1bB8ee1wFuRWL9nPhsl4_BPUhcWSBuxO0?usp=sharing

maybe @wonderingabout still has them

http://foxwq.com/qipu/search.html?title=%E7%AB%A0%E9%B1%BC%E5%9B%B4%E6%A3%8B
did you use same hardware and config on above matchs ?

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.

Hi, I enjoy the strength of GX78, GX 79 and GX 88, GX 89. I would like to make 80% human games and 20% computer games, but with the new updated Leelazero training games How do I do this?

Otherwise, I would love if you were able to make an updated GX8* series on a 20b net.

I like the weights you have trained very much.
Recently no new weights have been uploaded, have you stopped training the weights?
If you are still training for new weights, I wonder if it is possible for you to release them.
I always thank you for your hard work.

Hi,

I noticed that G13 and E8 miss a simple tesuji move.

I tried standard leela zero weights 15x192 and it finds the move T7 at move 268 for white.
G13 and E8 dont find the move to win the game.
I use only 1000 playouts.

Mybe over trained and lose a lot of reading/tesuji moves?

This was match E8 vs G13

(;FF[4]CA[UTF-8]AP[GoGui:1.4.9]
KM[6.5]PB[Leela Zero E8]DT[2018-06-27]
AB[pd][dp]PL[B];B[pq];W[dd];B[qk];W[fq];B[ip];W[cq];B[dq];W[cp];B[do];W[dr]
;B[er];W[cr];B[eq];W[cn];B[jd];W[qc];B[qd];W[pc];B[nc];W[nb]
;B[mc];W[rd];B[re];W[rc];B[od];W[mb];B[ok];W[nq];B[op];W[kq]
;B[np];W[ir];B[mq];W[ko];B[fp];W[gp];B[go];W[fo];B[fn];W[ep]
;B[eo];W[fp];B[gn];W[hp];B[io];W[fr];B[fs];W[gs];B[es];W[hr]
;B[dm];W[cm];B[dl];W[rf];B[se];W[pf];B[qe];W[oh];B[mh];W[qi]
;B[mj];W[rk];B[rl];W[rj];B[ql];W[rh];B[cf];W[fc];B[ef];W[bd]
;B[gd];W[gc];B[hd];W[dj];B[cl];W[dg];B[df];W[fj];B[ei];W[ej]
;B[hj];W[fh];B[ch];W[ci];B[dh];W[fi];B[hh];W[bh];B[bg];W[bj]
;B[bl];W[eh];B[ah];W[ai];B[bi];W[hi];B[ii];W[bh];B[ed];W[ec]
;B[bi];W[hk];B[aj];W[ij];B[gi];W[gj];B[hi];W[gl];B[il];W[ik]
;B[jl];W[ji];B[jh];W[kj];B[kh];W[lk];B[lm];W[mk];B[nk];W[mm]
;B[ln];W[nl];B[ll];W[ml];B[jk];W[jj];B[fl];W[ol];B[pl];W[nj]
;B[oj];W[ni];B[mi];W[pn];B[qn];W[qo];B[po];W[rn];B[on];W[pm]
;B[qm];W[om];B[ro];W[nn];B[oo];W[mn];B[lo];W[lp];B[mp];W[di]
;B[lb];W[oc];B[ma];W[ob];B[nf];W[jb];B[kb];W[jc];B[kc];W[hc]
;B[lr];W[kr];B[an];W[bo];B[ao];W[ap];B[al];W[ar];B[be];W[cc]
;B[gg];W[id];B[ie];W[ic];B[fd];W[of];B[ng];W[oi];B[qf];W[qg]
;B[sf];W[rg];B[iq];W[hq];B[lq];W[kp];B[kn];W[eg];B[de];W[cd]
;B[gm];W[hl];B[fk];W[gk];B[hm];W[ho];B[jo];W[hn];B[in];W[fg]
;B[ff];W[gh];B[hg];W[cg];B[bh];W[ae];B[af];W[ad];B[sl];W[pj]
;B[pk];W[ce];B[bf];W[ek];B[el];W[dn];B[en];W[bn];B[am];W[ls]
;B[ms];W[ks];B[jq];W[jr];B[ds];W[cs];B[og];W[pg];B[qj];W[sk]
;B[pi];W[ph];B[sg];W[sh];B[sd];W[sc];B[oe];W[na];B[la];W[kk]
;B[kl];W[li];B[lh];W[bk];B[ak];W[co];B[ja];W[ia];B[kd];W[ka]
;B[ck];W[cj];B[ja];W[qp];B[rp];W[ka];B[bp];W[bq];B[ja];W[qq]
;B[rq];W[ka];B[gr];W[gq];B[ja];W[qr];B[rr];W[ka];B[hs];W[is]
;B[ja];W[pj];B[sn];W[ka];B[pi];W[ja];B[pj];W[mo];B[pr];W[pe]
;B[nh];W[ki];B[lj];W[no];B[jp];W[dk];B[bm];W[];B[pp];W[]
;B[qs];W[cb];B[rm];W[];B[])

if I give you access to a server remetloy can you log in on my behalf and set up training for me?

I wish to net2net Leela zero current net (#116) to a 256f 20b network (new net) and then backtrain the last 2 million games.

Can you help

Hi.

I am an author of iOS app "A Master of Go".
Is it possible to use your weights for my app?
Please tell me know conditions which you require.
(I am not quite sure what GPL means for trained weights...)

Thank you!

Results from http://zero.sjeng.org

Network hash b03= 006772ff

Start date 2018-07-04 02:22

I am just asking out of interest because OX24 seems to be just playing zero-style opening most of the time.

Do the nets still follow naming scheme with G nets where "4" in "OX24" mean 40% oh human games? If so any chance we might get higher % of human games with reinforcement learning later? Thank you

Hello, are you still training any new weights? It's been more than a month since you uploaded some leelamaster weights. I wonder if this project is still alive or not?

Readme.md shows that GX5B has been released but I cant find it from your Google Drive folder.
I guess you've forgot to upload new files.
Check it please.

Thank you always...:)

Making Leela Master the three (10-20-30%) version is not a big difference and it is not unique. Rather, I think it would be better to make two versions (15-85 or 20-80%) wider.

Is it possible to get the games of Leela Master OX series in sgf? No matter, is it selfplay from Reinforcement Learning or evaluation games for promotion the best Network.
Would be great to take a look on it and study OX's fuseki choices during its evolution.