StockFiNN Release and Scaling

[dkappe]

Just for fun I decided to train a net with Komodo 14 evals at depth 8. Only 4 million positions. I was expecting something pretty weak, but it’s not half bad. Here with 30 threads vs sf10 (also with 30 threads). So far +3=9-0.

[pgn]
[Event "lizard1-sf10-11-254, Blitz 1.0min+1.0se"]
[Site "Rio de Janeiro, Brazil"]
[Date "2020.07.15"]
[Round "5.1"]
[White "SF NNUE LizardFish1"]
[Black "Stockfish 10 64 POPCNT"]
[Result "1-0"]
[ECO "A38"]
[Annotator "0.10;-0.08"]
[PlyCount "115"]
[EventDate "2020.07.15"]
[EventType "tourn"]
[SourceTitle "Fritz Engine Tournament"]
[Source "Silver"]

{AMD Ryzen Threadripper 1950X 16-Core Processor 3394 MHz W=23.9 plies; 15,
037kN/s; 358,973 TBAs B=23.3 plies; 32,783kN/s; 1,070,588 TBAs} 1. Nf3 Nf6 2.
g3 g6 3. Bg2 Bg7 4. O-O O-O 5. d3 d6 6. c4 c5 {-0.08/22 2} 7. Nc3 {0.10/23 4}
Nc6 {-0.09/21 1} 8. a3 {0.12/25 4 (Rb1)} Qd7 {-0.08/21 2 (a6)} 9. Rb1 {0.22/20
2} b6 {-0.21/22 1} 10. b4 {0.20/23 3} Bb7 {-0.05/21 2} 11. Qa4 {0.19/24 2 (e3)}
e6 {0.02/20 4} 12. e3 {0.14/23 2 (Rd1)} Ng4 {0.02/20 4 (h6)} 13. Bb2 {0.23/21 2
} Qe7 {0.00/22 1} 14. Ne1 {0.13/23 3 (Nd2)} Rfc8 {-0.18/21 2 (Nce5)} 15. Qb3 {
0.22/22 3} Rab8 {-0.31/21 2 (f5)} 16. b5 {0.16/21 3} Na5 {-0.21/22 1} 17. Qc2 {
0.16/24 4} f5 {-0.02/22 4} 18. Rd1 {0.05/20 2 (Bxb7)} Rf8 {-0.29/20 3 (Re8)}
19. Bxb7 {0.00/25 3 (h3)} Qxb7 {-0.21/24 1} 20. h3 {0.09/23 1} Nf6 {-0.39/23 1}
21. Nb1 {0.00/23 1} Nh5 {-0.44/22 1 (Qd7)} 22. Bxg7 {0.00/27 2} Qxg7 {-0.26/23
3} 23. Nf3 {0.00/27 7 (Nc3)} g5 {-0.17/24 10} 24. Kg2 {0.00/25 4 (Rde1)} h6 {
-0.16/22 4 (Rbe8)} 25. Nc3 {0.00/24 3 (Rde1)} Rbd8 {-0.22/24 2 (Rf7)} 26. a4 {
0.00/22 1 (Rde1)} Nf6 {-0.17/25 4} 27. Rde1 {0.00/28 1} Qf7 {-0.20/23 1} 28.
Qe2 {0.00/25 2} Nd7 {-0.21/24 1} 29. Qc2 {0.00/27 1} Kh7 {-0.25/24 2 (Nf6)} 30.
Nd2 {0.10/22 3 (Ne2)} Rde8 {0.00/24 8 (Ne5)} 31. f4 {0.55/20 1} Qg6 {0.33/25 9}
32. e4 {0.74/23 2} gxf4 {0.51/26 6 (fxe4)} 33. Rxf4 {0.69/28 3} Rg8 {0.51/26 2
(fxe4)} 34. Re3 {0.63/24 4 (Ne2)} Ne5 {0.33/20 1} 35. Ne2 {0.89/23 1} Qh5 {
0.44/22 2} 36. exf5 {0.78/24 4} exf5 {0.31/23 1} 37. Rf2 {0.76/23 1} Qg6 {
0.62/26 6 (Qg5)} 38. Nf4 {1.34/25 4 (Nf1)} Qg5 {0.73/20 1} 39. Kf1 {1.43/23 1}
Nb7 {0.67/22 1} 40. Qc3 {1.28/28 8 (Nd5)} Rgf8 {0.79/24 4} 41. Nd5 {1.23/25 2}
Qh5 {0.88/23 1} 42. Kg2 {1.20/25 1} Qg6 {0.86/24 1} 43. Nf4 {1.33/23 1 (Nb3)}
Qf6 {1.03/25 1 (Qg7)} 44. Nf3 {1.27/26 4} Nd8 {0.97/25 0} 45. Qe1 {1.29/23 0}
Qg7 {1.14/23 1} 46. Kh2 {1.47/23 2} Ne6 {1.08/25 1 (Nxf3+)} 47. Nxe5 {1.53/21 0
} Nxf4 {1.03/25 0} 48. gxf4 {1.63/19 0} Rg8 {1.03/27 1 (dxe5)} 49. Rfe2 {
1.64/21 1} dxe5 {0.63/24 1} 50. Qf2 {1.29/23 1} e4 {0.67/26 1} 51. dxe4 {
1.86/23 1} Qf7 {0.61/25 0 (Rxe4)} 52. exf5 {1.78/23 1} Rxe3 {1.07/26 2} 53.
Rxe3 {2.61/24 1 (Qxe3)} Qxc4 {0.26/20 1 (Qxf5)} 54. f6 {3.63/22 1 (a5)} Qxa4 {
3.76/24 3 (Rf8)} 55. Re7+ {10.20/27 1} Kh8 {12.84/26 2} 56. f7 {11.83/25 1} Rf8
{16.69/24 1} 57. Qb2+ {12.30/27 1} Kh7 {21.79/24 1 (Qd4)} 58. Qb1+ {12.60/24 1}
1-0

[/pgn]

[matejst]

Dietrich,

could you make the NN.bin available? Thanks in advance!

[dkappe]

Dietrich,

could you make the NN.bin available? Thanks in advance!

https://www.dropbox.com/sh/glywvt4qf4x ... JB_ma?dl=0

Let me know how it goes.

Code: Select all

Score of LizardFish1 vs sfdev150720: 5 - 8 - 30  [0.465] 43

1+1, 6 man, 1 thread.

[matejst]

Thanks again, Dietrich. I will test it against Shashchess and Komodo 11 this week-end and post the results here or on Discord.

[Ovyron]

Makes it hard to convince people when in fact the entire evaluation function has no shared essence at all.

Correct me if I'm wrong, but isn't it doing the normal Stockfish evaluation and THEN modifying it with the contents of nn.bin?

EDIT - "This evaluation is created from scratch" (Opening post)

Oh, so that's the difference between Stockfish NNUE and StockFiNN...