Leela Zero

User avatar
Dan.A
Posts: 518
Joined: Mon Jan 19, 2015 12:58 pm
Location: Bucuresti
Rang: 3Kyu

Re: Leela - software bazat pe retele neurale

Postby Dan.A » Wed Dec 27, 2017 1:25 pm

Mic Update:

- Sa ajuns la 458 elo pe scara AlphaGo cu 27 elo peste Gnugo 3.8 (din scara AlphaGo) deci in sfarsit exista o "ancora".

- Banuiesc, ca Gnugo este o ancora stabila pentru ca joaca la acelas nivel indiferent de configuratia calculatorului.

- Deci o estimare cat decat medie este ca a ajuns cam la 10k-9k. Am zis medie din cauza ca , inca face gafe majore in special in jocul cu oamenii.
- Oricum depinde foarte mult de puterea placii grafice , de aceea sunt rezultate diferite.

- Inca exista suspiciunea ca se transmit jocuri proaste din cand in cand dar se merge din ce ince mai mult pe varianta ca ar fi din cauza unor drivere corupte de OpenCL si nu de atacuri intentionate. Acest lucru este rezolvat in versiunea "next" (care va avea si o versiune windows curand) care va inchide clientii cu drivere corupte de OpenCL

"There are times when the only choices you have left are the bad ones."

User avatar
Admin
Site Admin
Posts: 406
Joined: Sat Jan 17, 2015 6:30 pm
Rang: 1d

Re: Leela - software bazat pe retele neurale

Postby Admin » Wed Dec 27, 2017 11:09 pm

Puterea de joc a LeelaZeroT pe kgs pare sa creasca. Vad ca a inceput sa castige partide si la roboti de 1dan. Am vazut si eu destule partide in care pare sa aiba lacune mari insa nu am reusit sa analizez partidele pe indelete. Poate face si LeelaZ ca si AlphaGo mutari ciudate cand intra in panica.

[+] DCNN1dan03
(;CA[utf-8]FF[4]ST[2]RU[Chinese]OT[5x30 byo-yomi]C[DCNN1d03 [1d\]\: GTP Engine for DCNN1d03 (white)\: Hiratuka version 10_38A-No32-2-DCNN-VECTOR_CPU
LeelaZeroT [-\]\: GTP Engine for LeelaZeroT (black)\: Leela Zero version 0.9
]
AP[CGoban:3]SZ[19]GM[1]DT[2017-12-27]PC[The KGS Go Server at http\://www.gokgs.com/]
PB[LeelaZeroT]PW[DCNN1d03]WR[1d]KM[7.5]TM[300]RE[B+3.50]MULTIGOGM[1]
;B[dd]BL[266.962];W[pp]WL[299.77];B[pd]BL[233.782];W[dq]WL[299.245];B[co]BL[199.59]
;W[dm]WL[298.981];B[eo]BL[167.18];W[fp]WL[298.724]C[Glorrandir [1d\]\: the next target is Hirabot 3d ?
]
;B[fo]BL[135.492];W[gp]WL[298.506];B[go]BL[104.398];W[hp]WL[298.096];B[nq]BL[73.452]
;W[in]WL[297.864];B[ho]BL[43.201];W[io]WL[297.65];B[ip]BL[13.416];W[jp]WL[297.428]
;OB[5]B[iq]BL[30];W[jq]WL[297.172];OB[5]B[ir]BL[30];W[jr]WL[296.945];OB[5]B[cq]
BL[30];W[dp]WL[296.713];OB[5]B[do]BL[30];W[cp]WL[296.491];OB[5]B[bp]BL[30];W[bq]
WL[296.272];OB[5]B[ep]BL[30];W[eq]WL[296.032];OB[5]B[fq]BL[30];W[fr]WL[295.807]
;OB[5]B[gq]BL[30];W[gr]WL[295.561]C[Marcin [-\]\: OK, let it be Hirabot 3d
];OB[5]
B[hq]BL[30];W[oq]WL[295.317];OB[5]B[cr]BL[30];W[bo]WL[295.095]C[Marcin [-\]\: E2
]
;OB[5]B[er]BL[30];W[ap]WL[294.843];OB[5]B[dr]BL[30];W[bp]WL[294.58];OB[5]B[bn]BL[30]
;W[br]WL[294.361];OB[5]B[np]BL[30];W[pn]WL[294.063];OB[5]B[nn]BL[30];W[ln]WL[293.824]
;OB[5]B[pm]BL[30];W[qm]WL[293.61];OB[5]B[pl]BL[30];W[qn]WL[293.383];OB[5]B[ml]BL[30]
;W[qk]WL[293.109];OB[5]B[pk]BL[30];W[qj]WL[292.883];OB[5]B[pj]BL[30];W[pi]WL[292.623]
;OB[5]B[qi]BL[30];W[qh]WL[292.292];OB[5]B[ri]BL[30];W[ql]WL[292.075];OB[5]B[ph]
BL[30];W[oi]WL[291.604];OB[5]B[qg]BL[30];W[nj]WL[291.336];OB[5]B[nk]BL[30];W[oh]
WL[291.096];OB[5]B[rh]BL[30];W[nm]WL[290.851];OB[5]B[mm]BL[30];W[on]WL[290.614]
;OB[5]B[mn]BL[30];W[lk]WL[290.39];OB[5]B[kl]BL[30];W[il]WL[290.16];OB[5]B[kk]BL[30]
;W[kj]WL[289.932];OB[5]B[jj]BL[30];W[ki]WL[289.697];OB[5]B[jk]BL[30];W[gm]WL[289.432]
;OB[5]B[ji]BL[30];W[kh]WL[289.199];OB[5]B[jh]BL[30];W[kg]WL[288.932];OB[5]B[jg]
BL[30];W[kf]WL[288.71];OB[5]B[jf]BL[30];W[ke]WL[287.728];OB[5]B[je]BL[30];W[kd]
WL[287.494];OB[5]B[cl]BL[30];W[dl]WL[287.263];OB[5]B[ck]BL[30];W[dk]WL[287.026]
C[Glorrandir [1d\]\: now H2
];OB[5]B[dj]BL[30];W[ej]WL[286.445];OB[5]B[ei]BL[30]
;W[fj]WL[286.21];OB[5]B[di]BL[30];W[fi]WL[285.966];OB[5]B[fh]BL[30];W[gh]WL[285.732]
;OB[5]B[fg]BL[30];W[cc]WL[285.499];OB[5]B[cd]BL[30];W[dc]WL[285.258];OB[5]B[ec]
BL[30];W[eb]WL[285.026];OB[5]B[fc]BL[30];W[fb]WL[284.797];OB[5]B[gc]BL[30];W[bd]
WL[284.564];OB[5]B[be]BL[30];W[bc]WL[283.895];OB[5]B[og]BL[30];W[ng]WL[283.661]
;OB[5]B[hk]BL[30];W[gk]WL[283.425];OB[4]B[hl]BL[30];W[hm]WL[283.201];OB[4]B[jm]
BL[30];W[jn]WL[282.972];OB[4]B[im]BL[30];W[hn]WL[282.748];OB[4]B[qq]BL[30];W[or]
WL[282.513];OB[4]B[nr]BL[30];W[qp]WL[282.288];OB[4]B[ko]BL[30];W[kn]WL[282.065]
;OB[4]B[lo]BL[30];W[jo]WL[281.833];OB[4]B[oc]BL[30];W[hj]WL[281.602];OB[4]B[ik]
BL[30];W[hi]WL[281.357];OB[4]B[nf]BL[30];W[of]WL[281.14];OB[4]B[pg]BL[30];W[mg]
WL[280.912];OB[4]B[ne]BL[30];W[bf]WL[280.675];OB[4]B[ce]BL[30];W[ae]WL[280.442]
;OB[4]B[cf]BL[30];W[bg]WL[280.215];OB[4]B[cg]BL[30];W[bh]WL[279.993];OB[4]B[mj]
BL[30];W[mi]WL[279.767];OB[4]B[gl]BL[30];W[fl]WL[279.552];OB[4]B[fm]BL[30];W[fn]
WL[279.315];OB[4]B[rq]BL[30];W[rp]WL[279.092];OB[4]B[os]BL[30];W[ps]WL[278.864]
;OB[4]B[ns]BL[30];W[pr]WL[278.606];OB[4]B[op]BL[30];W[pq]WL[278.372];OB[4]B[ad]
BL[30];W[ac]WL[278.15];OB[4]B[oj]BL[30];W[ni]WL[277.918];OB[4]B[gs]BL[30];W[cs]
WL[277.664];OB[4]B[fs]BL[30];W[cn]WL[277.447];OB[4]B[bm]BL[30];W[bj]WL[277.238]
;OB[4]B[cj]BL[30];W[bk]WL[277.024];OB[4]B[bi]BL[30];W[ai]WL[276.775];OB[4]B[ci]
BL[30];W[aj]WL[276.555];OB[4]B[lj]BL[30];W[li]WL[276.303];OB[4]B[mk]BL[30];W[om]
WL[276.076];OB[4]B[lc]BL[30];W[kc]WL[275.847];OB[4]B[kb]BL[30];W[jb]WL[275.613]
;OB[4]B[lb]BL[30];W[ic]WL[275.396];OB[4]B[gb]BL[30];W[hd]WL[275.129];OB[4]B[db]
BL[30];W[cb]WL[274.854];OB[4]B[jd]BL[30];W[jc]WL[274.632];OB[4]B[da]BL[30];W[ca]
WL[274.387];OB[4]B[af]BL[30];W[ag]WL[274.156];OB[4]B[bl]BL[30];W[fe]WL[273.914]
;OB[4]B[gg]BL[30];W[hg]WL[273.692];OB[4]B[hf]BL[30];W[ge]WL[273.469];OB[4]B[ed]
BL[30];W[ef]WL[273.23];OB[4]B[he]BL[30];W[gd]WL[272.973];OB[4]B[gf]BL[30];W[eg]
WL[272.738];OB[4]B[eh]BL[30];W[ch]WL[272.508];OB[4]B[dg]BL[30];W[dh]WL[272.294]
;OB[4]B[hr]BL[30];W[md]WL[272.049];OB[4]B[ld]BL[30];W[me]WL[271.803];OB[4]B[nd]
BL[30];W[mc]WL[271.576];OB[4]B[mb]BL[30];W[nc]WL[271.35];OB[4]B[nb]BL[30];W[mf]
WL[271.121];OB[4]B[lq]BL[30];W[lr]WL[270.87];OB[4]B[mr]BL[30];W[lp]WL[270.642];OB[4]
B[mp]BL[30];W[kp]WL[270.42];OB[4]B[mo]BL[30];W[rj]WL[270.14];OB[4]B[em]BL[30];W[en]
WL[269.88];OB[4]B[dn]BL[30];W[el]WL[269.65];OB[4]B[ar]BL[30];W[cm]WL[269.388];OB[4]
B[an]BL[30];W[al]WL[269.168];OB[4]B[ib]BL[30];W[ia]WL[268.948];OB[4]B[hb]BL[30]
;W[ee]WL[267.695];OB[4]B[fk]BL[30];W[ek]WL[267.462];OB[4]B[ig]BL[30];W[hh]WL[267.239]
;OB[4]B[gj]BL[30];W[gi]WL[267.017];OB[4]B[kr]BL[30];W[js]WL[266.78];OB[4]B[ls]BL[30]
;W[lm]WL[266.556];OB[4]B[am]BL[30];W[ll]WL[266.331];OB[4]B[ak]BL[30];W[ok]WL[266.09]
;OB[4]B[ah]BL[30];W[ai]WL[265.846];OB[4]B[ol]BL[30];W[nl]WL[265.583];OB[4]B[ok]
BL[30];W[oo]WL[265.353];OB[4]B[aj]BL[30];W[ad]WL[265.109];OB[4]B[ah]BL[30];W[pe]
WL[264.884];OB[4]B[af]BL[30];W[od]WL[264.519];OB[4]B[qd]BL[30];W[qe]WL[264.273]
;OB[4]B[re]BL[30];W[rf]WL[263.968];OB[4]B[rd]BL[30];W[qf]WL[263.658];OB[4]B[rg]
BL[30];W[sf]WL[263.42];OB[4]B[sg]BL[30];W[si]WL[263.196];OB[4]B[id]BL[30];W[hc]
WL[262.966];OB[4]B[oe]BL[30];W[od]WL[262.709];OB[4]B[oe]BL[30];W[ne]WL[262.477]
;OB[4]B[pf]BL[30];W[le]WL[262.258];OB[4]B[gn]BL[30];W[fm]WL[262.028];OB[4]B[se]
BL[30];W[oe]WL[261.802];OB[4]B[ka]BL[30];W[ja]WL[261.522];OB[4]B[pb]BL[30];W[is]
WL[261.272];OB[4]B[sj]BL[30];W[sk]WL[261.047];OB[4]B[ha]BL[30];W[fa]WL[260.739]
;OB[4]B[ga]BL[30];W[ea]WL[260.493];OB[4]B[km]BL[30];W[kq]WL[260.277];OB[4]B[ks]
BL[30];W[no]WL[260.061];OB[4]B[ff]BL[30];W[df]WL[259.838];OB[4]B[bs]BL[30];W[as]
WL[259.606];OB[4]B[hs]BL[30];W[ds]WL[259.382];OB[4]B[rb]BL[30];W[bs]WL[259.149]
;OB[4]B[ao]BL[30];W[fr]WL[258.89];OB[4]B[es]BL[30];W[ij]WL[258.538];OB[4]B[aq]BL[30]
;W[ii]WL[258.295];OB[4]B[rr]BL[30];W[rs]WL[258.054];OB[4]B[sc]BL[30];W[sj]WL[257.789]
;OB[4]B[de]BL[30];W[sh]WL[257.552];OB[4]B[fd]BL[30];W[qh]WL[257.331];OB[4]B[ih]
BL[30];W[qs]WL[257.087];OB[4]B[pc]BL[30];W[bf]WL[256.854];OB[4]B[bg]BL[30];W[db]
WL[256.629];OB[4]B[jl]BL[30];W[cp]WL[256.392];OB[4]B[bp]BL[30];W[dq]WL[256.167]
;OB[4]B[hp]BL[30];W[eq]WL[255.952];OB[4]B[dp]BL[30];W[nh]WL[255.73]C[Marcin [-\]\: Hirabot 3d will be next
]
;OB[4]B[ie]BL[30];W[eq]WL[255.515]C[Marcin [-\]\: let me pass
];OB[4]B[ob]BL[30]
;W[ss]WL[254.754];OB[4]B[]BL[30];W[]WL[254.753]TB[eq][sb][la][ma][na][oa][pa][qa]
[ra][sa][qb][qc][rc][sd][bf][if][ag][bh][ch][dh][ai][bj][bk][al][il][bo][ap][cp]
[fp][gp][bq][dq][mq][br][fr][gr][lr][as][bs][cs][ds][ms]TW[aa][ba][da][ab][bb][nd]
[lf][nf][lg][sr][og][pg][qg][rg][sg][lh][mh][ph][rh][qi][ri][gj][fk][rk][rl][sl]
[em][rm][sm][rn][sn][po][qo][ro][so][sp][qq][rq][sq][qr][rr][pf])
Iulian Lungu

User avatar
Dan.A
Posts: 518
Joined: Mon Jan 19, 2015 12:58 pm
Location: Bucuresti
Rang: 3Kyu

Re: Leela - software bazat pe retele neurale

Postby Dan.A » Tue Jan 02, 2018 6:12 pm

Mic Update:

Ultimul Network la data postarii este http://zero.sjeng.org/networks/58da6176 ... 06bdef3.gz

- Incepe sa joace multe scari (sper sa si invete ceva)

- Are codat sa nu abandoneze in mai putin de 93 de mutari (setare care cica e folosita pentru a nu da abandon prematur la prtidele cu handicap)

- Inca am senzatia ca joaca mai bine cu Negru.

- Sabaki a facut un update care se pare ca a rezolvat mai multe buguri legate de robotii care folosesc istoricul ultimelor mutari jucate. Se pare ca at cand afisa mutarile le lua in alta ordine rezultand printre altele si mutari ilegale. Boti afectati : LeelaZero si AQ.

- La peste 20k Playouts se pare ca joaca in jur de 1k. Desi are un rang mare pe KGS multe partide sunt jucate la misto (punctul meu de vedere) de catre jucatorii umani.

- Cineva pe serverul de computere CGOS a reusit sa castige cu ultimul network la Zen 11. Se pare ca foloseste 1M (milion) playouts cu o placa Titan V (estimata la 3000$) Detalii despre placa puteti gasi aici : https://www.nvidia.com/en-us/titan/titan-v/

- Partida cu zen puteti sa o vedeti aici : http://www.yss-aya.com/cgos/viewer.cgi? ... 344177.sgf Sunt mici suspiciuni ca nu ar fi Leela sau ar fi o leela modificata dar pentru asta cineva tre sa aibe 1M playouts sa testeze. :)

- Inca abandoneza partide castigate. Mi sa intamplat testand un meci cu ea vs Gnugo. Dupa ce gnugo a dat pas evaluarea tablei a scazut de la 100% la 0% si a dat abandon desi era in fata pe tabla. Nu stiu daca e un bug de Sabaki (se pare ca are buguri cand bagi roboti noi) sau un bug de Leelaz.

- In curand va fi lansata versiunea Next ceea ce va imbunatatii multe din problemele curente.

- Ca deobicei la fiecare stagnare mai mare de 8 ore unii contributori intra in panica si sugereaza tot felul de modificari. :) Acest network rezista schimbarii de vreo 2 zile. :) La fel si GCP sugestiilor de deviere de la publicatie. De ce sai faci , multi specialisti pareri impartite. Unii vor sa testeze lucruri (profitand de contributori) care nu ajuta neaparat la acest proiect dar aj in domeniul lor de activitate.
"There are times when the only choices you have left are the bad ones."

User avatar
Admin
Site Admin
Posts: 406
Joined: Sat Jan 17, 2015 6:30 pm
Rang: 1d

Re: Leela - software bazat pe retele neurale

Postby Admin » Wed Jan 03, 2018 9:11 pm

Eu incep sa cred ca joaca la nivel de 1 dan amator in partida de mai jos, cu negrul (am observat cateva greseli care nu au ce cauta la nivel 1 dan, dar per total as face o medie de 1 dan). Si network-ul 83015869 pare sa fi invatat ce e scara:
[+] Partida
(;GM[1]FF[4]RU[Chinese]DT[2018-01-03]SZ[19]KM[7.5]PB[Leela Zero 0.9 83015869]PW[Leela Zero 0.9 58da6176]RE[B+Resign] ;B[pd];W[dp];B[dd];W[pp];B[pf];W[pc];B[qc];W[oc];B[qb];W[od];B[pe];W[df];B[fd];W[lc];B[jc];W[ph];B[ne];W[le];B[md];W[ld];B[mc];W[mb];B[nb];W[lb];B[nc];W[hc];B[hd];W[id];B[ic];W[jd];B[hb];W[gc];B[gb];W[gd];B[ge];W[fc];B[ec];W[he];B[fb];W[hd];B[fe];W[dh];B[cj];W[gf];B[ff];W[fg];B[gg];W[hf];B[eg];W[fh];B[eh];W[ei];B[fi];W[gh];B[di];W[ej];B[dg];W[ch];B[cg];W[bh];B[bg];W[dj];B[ci];W[ck];B[bi];W[fq];B[dk];W[dl];B[ek];W[fj];B[gi];W[hh];B[el];W[pn];B[cl];W[pk];B[kq];W[mq];B[hq];W[qg];B[gj];W[fk];B[fl];W[re];B[rd];W[eb];B[db];W[hp];B[ip];W[io];B[gp];W[ho];B[gq];W[gk];B[jo];W[jn];B[ko];W[kn];B[hm];W[jp];B[iq];W[lo];B[kp];W[hk];B[il];W[gm];B[gl];W[hl];B[im];W[gn];B[ik];W[hj];B[in];W[ln];B[go];W[cq];B[lk];W[ij];B[kl];W[ml];B[mk];W[nk];B[nj];W[ok];B[nh];W[qr];B[co];W[do];B[cn];W[dm];B[bk];W[fr];B[bq];W[bp];B[cp];W[br];B[dq];W[cr];B[dn];W[ep];B[bo];W[aq];B[mr];W[nr];B[lr];W[na];B[ob];W[oi];B[nq];W[np];B[oq];W[or];B[mp];W[pq];B[lq];W[ni];B[rf];W[qf];B[qe];W[rg];B[se];W[op];B[en];W[mh];B[jj];W[ji];B[ki];W[jh];B[kj];W[gr];B[hr];W[mo];B[kh];W[kg];B[lg];W[mg];B[jg];W[fo];B[kf];W[ef];B[mf];W[fn];B[fp];W[eo];B[eq];W[dr];B[hn];W[er];B[ig];W[hg];B[ih];W[ea];B[da];W[cd];B[fa];W[ce];B[cc];W[bc];B[bb];W[ed];B[ee];W[de];B[ed];W[eb];B[ea];W[em];B[cm];W[fm];B[ck];W[hi];B[ii];W[mj];B[ng];W[mi];B[lf];W[mq];B[nq];W[ll];B[km];W[lm];B[jm];W[ab];B[ba];W[ja];B[jb];W[ka];B[ia];W[ke];B[kb];W[la];B[ah];W[lj];B[kk];W[kc];B[ib];W[pg];B[og];W[oh];B[ao];W[ap];B[sg])
Iulian Lungu

User avatar
Dan.A
Posts: 518
Joined: Mon Jan 19, 2015 12:58 pm
Location: Bucuresti
Rang: 3Kyu

Leela Zero news

Postby Dan.A » Thu Jan 04, 2018 8:10 pm

Mic Update:

- Se pare ca problema cu partide proaste s-a intensificat. Cine e interesat de teorii si solutii poate sa urmareasca la acest link: https://github.com/gcp/leela-zero/issues/458

- O sa fac un mic rezumat :

- Good people: 271, bad people: 35 Good games: 5078, bad games: 3316 13% of clients, 35% of games are anomalous

- Caz ca nu se rezolva pot aparea intarzieri deoarece " The top 2 IP's are highly significantly different. They're also "somehow" generating 2-4x as much games as the next best client. The next clients look perfectly average."

Not sure how to quantity it, but it makes perfect sense to me that the learning is very robust to unsophisticated bad actors (and I said so a while ago).
If your junk is random, it will just get averaged out. If it is structured, the network will learn to recognize the bad data and won't spend much weights on it. Structuring the data so that it is guaranteed to "break" the learning may be unobvious and computationally intensive enough to be unattractive.

Asta da o nota de optimism.
The bad games seem to be mostly coming off of random universities in the USA, which makes it more likely that those people are experimenting and trying to be clever. Or they're running on some university clusters with OpenCL drivers that haven't been updated in years, etc.

Iar asta ca nu ar fi persoane cu intentie ceea ce e un lucru bun.

- Multe din aceste probleme neintentionate se vor rezolva odata cu aparitia noului client.

- Momentan singura metota de a vedea daca clientul genereaza partide proaste este sa se urmareasca partida generata dupa primele 30 de mutari (care sunt random din cod). Caz ca partida genereaza mutari random (fara sens) pe intreaga durata e posibil sa fie o problema o problema. Daca dupa 30 de mutari partida decurge normal at e ok. Pentru a pastra pe calculator partidele generate se poate folosi leelaz -k kifu (kifu fiind directorul unde salveaza sgful inainte sa-l trimita pe server)

- Randomizarea la inceput este doar pentru partidele de antrenament. 1.Pentru a nu uita ce a invatat in trecut (cica uita surprinzator) 2. Pentru a invata sa se descurce si din alte pozitii decat cele invatate pe care are reflexul sa le tot joace.

Eu incep sa cred ca joaca la nivel de 1 dan amator

Si eu mai ales daca foloseste mai mult de 10k playouts. Ce ma distreaza cel mai mult este ca de la network la network se joaca mult cu scara incercand sa gaseasca solutii. :) . Alt lucru observat in unele jocuri este ca invtand treptat formele moarte a invatat cateva pe care si Leela 11 le mai greseste.

Asta e avantajul invatatului de la zero spre deosebire de invatatul cu partide de om (unde de Ex rar vezi pe cineva sa duca scara pana la capat sau sa joace dintr-un grup mort)

P.S. Se pare ca unul a fost gasit. Cica ar fi generat 10537 games in the past 24 hours (eu abia generez 2-3/h). A postat de cand a inceput sa foloseasca supercalculatorul si cu ce setari si se incearca o solutie. E bine ca e bine intentionat. Inca 2-3 ca el cu setari corecte si cine stie cat poate sa salte proiectul.:)
"There are times when the only choices you have left are the bad ones."

User avatar
Dan.A
Posts: 518
Joined: Mon Jan 19, 2015 12:58 pm
Location: Bucuresti
Rang: 3Kyu

Re: Leela - software bazat pe retele neurale

Postby Dan.A » Sun Jan 07, 2018 11:39 pm

Mic Update:

A fost lansata versiunea 10 care poate fi gasita la acest link: https://github.com/gcp/leela-zero/releases

- Sunt 2 versiuni. Una normala si una pentru persoane care vor sa foloseasca Cpu sau nu au o placa grafica destul de buna.

- Daca nu apar buguri majore in urmatoarele zile aceasta versiune va deveni minim necesara.

Planuri de viitor in general vor fi postate la acest link: https://github.com/gcp/leela-zero/issues/591

Cea mai buna parte din discurs si planuri :

My wife would also like to see me again!

De asa se intampla cand esti "om de stiinta".:)

P.S. Se pare ca aceasta versiune este folosita si pentru a testa networkuri , nu numai de generat partide. Din acest motiv numarul networkurilor din director ar putea sa creasca mult mai repede. Un network are cam 8MB.
"There are times when the only choices you have left are the bad ones."

User avatar
Admin
Site Admin
Posts: 406
Joined: Sat Jan 17, 2015 6:30 pm
Rang: 1d

Re: Leela - software bazat pe retele neurale

Postby Admin » Fri Jan 12, 2018 2:05 pm

LeelaZero cu network-ul 5x64 se pare ca a ajuns in ultima faza a evolutiei. Nu a mai produs de aproape 4 zile un network mai bun si in momentul asta se incearca ultima evaluare (sunt 146k partide self-play jucate de ultimul network promovat - 46ff4b94). Daca nu reuseste sa promoveze o noua retea, GCP va reduce learning rate intr-o ultima incercare de a continua antrenarea modelului actual.
Din ce inteleg, cel mai probabil daca nici asa nu va reusi sa promoveze un nou network, se va trece la network de 6x128. Cum zicea si Dan, asta va incetini de cateva ori generarea de partide la clientii care contribuie cu hardware.
Iulian Lungu

User avatar
Admin
Site Admin
Posts: 406
Joined: Sat Jan 17, 2015 6:30 pm
Rang: 1d

Re: Leela - software bazat pe retele neurale

Postby Admin » Wed Jan 17, 2018 10:55 am

Observ ca Gian Carlo testeaza mai multe variante in perioada asta si apar primele vesti bune. Reteaua de 6 blocuri si cu 128 de filtre, antrenata de la zero (deci fara integrarea retelei 5x64) dar cu partidele de la 850k incoace jucate de reteaua 5x64, a inceput sa o depaseasca pe cea de 5x64. Din ce inteleg se urmareste evolutia retelei de 6x128 o perioada si daca evolutia e sustinuta si arata bine, se va trece la retea de 10x128. Intrucat rezultatele de pana acum sunt rezonabile, cel mai probabil nu se va incerca un bootstrap in sensul ca noua retea (fie ca e 6x128, fie ca va fi 10x128) nu va fi antrenata doar cu partide generate de ea.
Si reteaua de 5x64 continua, incet dar continua, sa isi imbunatateasca puterea de joc (asta e o veste buna pentru cei care vor incerca sa isi faca un robot pentru mobile).
Iulian Lungu

User avatar
Dan.A
Posts: 518
Joined: Mon Jan 19, 2015 12:58 pm
Location: Bucuresti
Rang: 3Kyu

Re: Leela - software bazat pe retele neurale

Postby Dan.A » Sat Jan 20, 2018 9:04 pm

Mic Update:

A aparut versiunea 11 cu autogtp 13 care poate fi gasita la acest link: https://github.com/gcp/leela-zero/releases

- De notat ca la prima rulare va lua ceva timp pentru a face niste teste

In sfarsit s-a trecut de la arhitectura de 5x64 la 6x128

- Se pare ca transferul a reusit intr-un fel desi am inteles ca nu e chiar asa de usor de trecut de la o arhitectura mai mica la una mai mare.

- Un efect secundar din transfer este ca are un "Policy Net" mai bun dar un "Value Net" mai slab decat ultima versiune 5x64.

The problem 7fde81e8 has right now is that it remembers the outcome of all the training games in the current window, as it was trained on them, it as a lot of capacity the previous net didn't have, and there's very few new games.


Deci si pentru ei este teren nou. :)
I assume this will fix itself when the new data starts weighting more in the training window.


- Fisierul de acum in colo va avea 23M in loc de 8M si e posibil sa mearga mai incet.

Este in test si versiunea de 10x128.
"There are times when the only choices you have left are the bad ones."

User avatar
Dan.A
Posts: 518
Joined: Mon Jan 19, 2015 12:58 pm
Location: Bucuresti
Rang: 3Kyu

Re: Leela - software bazat pe retele neurale

Postby Dan.A » Fri Jan 26, 2018 10:37 am

Mic Update:

Current plan:

Continue 6x128 until it stalls for ~200k ish games.
Drop the learning rate and see what happens.
If this has also stalled, try the AZ always-promote technique.
Use net2net to go to 10x128.
If AZ worked, use it, else maybe try it now (?)


Imediat ce GCP a postat planul B au aparut imbunatatiri in antrenament. :)

- Se pare ca acest experiment a dat roade : https://github.com/gcp/leela-zero/pull/747

- Desi numarul de contributori se mentine constant , a scazut numarul de partide/h trimise. Se pare ca trecerea la 6x128 a scazut un pic rata de generat partide. Ma intreb cat o sa scada cand se trece la 10x128 sau chiar mai mult.

- Ca tarie inca nu stiu cum se descurca, dar cu placi grafice gen Titan V functioneaza ca 6D pe FoxWeiqi. Ultima versiune a reusit sa bata Hirabot42 care este 6D pe KGS.
"There are times when the only choices you have left are the bad ones."


Return to “GO-ul si IT-ul/AI-ul”

Who is online

Users browsing this forum: No registered users and 1 guest

cron