Leela 0.11.0 Beta version slower than 0.10.0 ?

[Pio2001]

One of our club members is 7d on Tygem, but 1 kyu FFG (EGF).

[Uberdude]

One of our club members is 7d on Tygem, but 1 kyu FFG (EGF).

France is notorious for its under-ranked players (particularly young improvers), but also FFG rank != EGF rank, they have their own parallel internal rating system as well as submitting to EGD. Is this player 1 kyu in the EGF system, and has he gone about 50/50 in a tournament, such as the EGC, not just full of under-ranked French? I could believe a solid EGF 2d good at blitz is Tygem 7d, but not a 1k.

[pookpooi]

Anyways, LEELA(B) on Tygem has 50:50 score against pros. That surely tells something about its strength. And it only plays against pros.

Very impressive. I also notice that Leela has 50% winrate against Fox 9d. Can we unanimously declare that Leela is the strongest freeware go AI? (and strongest go AI available to public right now, just before Zenith Go 7 release in upcoming months)

[johnsmith]

Anyways, LEELA(B) on Tygem has 50:50 score against pros. That surely tells something about its strength. And it only plays against pros.

Very impressive. I also notice that Leela has 50% winrate against Fox 9d. Can we unanimously declare that Leela is the strongest freeware go AI? (and strongest go AI available to public right now, just before Zenith Go 7 release in upcoming months)

What is Leela's name on Fox? I'd like to check out those games.
P.S. FineArt also started playing and is stronger than ever - only 1 loss and 35 wins. It has kinda new style, it prefers high short shimari it seems.

[pookpooi]

Leela测试

And yes, FineArt, DeepZenGo, CGI strength are another league, they're all 10d fox (with Ke Jie join as the only 10d human).
FineArt is able to maintain more than 60% winrate even giving 2-4 stones handicap to 9d. And to think that AlphaGo is likely to be 3 stones stronger than them is insane.

[Bas]

If you like numbers, here are the benchmark results of an Intel Core i7-7700K@4.2GHz and a Nvidia GeForce GTX 1080.

Leela 0.10.0:

Code: Select all

  200000 games in  7.06 seconds -> 28328 g/s (3541 g/s per thread)
  Avg Len: 451.75 Score: -1.973360

  CPU:
   2000 predictions in 21.45 seconds -> 93 p/s
  10000 evaluations in 22.74 seconds -> 439 p/s

  GPU:
   2000 predictions in  1.69 seconds -> 1183 p/s
  10000 evaluations in  3.92 seconds -> 2551 p/s

Leela 0.11.0b11:

Code: Select all

  200000 games in  8.39 seconds -> 23837 g/s (2979 g/s per thread)
  Avg Len: 418.56 Score: -2.432680

  CPU:
   2000 predictions in 15.97 seconds -> 125 p/s
  10000 evaluations in 33.16 seconds -> 301 p/s

  GPU:
   2000 predictions in  3.28 seconds -> 609 p/s
  10000 evaluations in  4.13 seconds -> 2421 p/s

And I have some games of Leela 0.11.0b11 against CSDL with and without GPU, Leela 0.11.0b11 with GPU against itself without GPU and Leela 0.11.b12 against AQ 2.0.3.1. Leela’s opening book was used. Time limit was 30 minutes in all cases (CSDL had 30 minutes + 5s byoyomi, because setting absolute times is not possible, but it had never used the additional time).

I like Leela’s aggressive and exact style; the games against CSDL are interesting. By the way: Leela with GPU won all games except the last one against AQ. Leela was leading very clearly, made a questionable move which forced it to ignore the next Ko thread, but even than it would have won but instead surprisingly gave up.

Leela vs Crazy Stone

Leela CPU vs Leela GPU

Leela vs AQ

[johnsmith]

Which OS did you use for netbench and benchmark?

[Bas]

Windows 10