A trainer for Zahak's NNUE. The produced nets are compatible with both Zahak and Bit-Genie engine. In theory one should be able to load networks that is produced with this trainer to either engine, it might not always produce the optimal play though.
Input format is expected to look like the following:
1n1qkb1r/r2b1p2/4pn1p/p2p4/2pP1BpP/2P1P3/1PN2PP1/RN1QKB1R b KQk - 1 11;score:-31;eval:25;qs:25;outcome:0.5
r5kr/1p3ppp/1p6/1P1p4/3P2P1/qP5P/3NPP2/2RQK2R w K - 0 16;score:459;eval:449;qs:449;outcome:1.0
8/1r2p1k1/2Nnn1pp/R7/8/5NPP/5PK1/8 b - - 6 38;score:32;eval:12;qs:12;outcome:0.5
r3k1r1/1p6/2p1pb1p/2P2p2/p2P1P2/4P1P1/PP1B1K1P/1R2R3 w q - 2 27;score:188;eval:164;qs:164;outcome:1.0
6R1/pp6/1k4P1/r7/5K2/2p4P/8/8 w - - 2 54;score:-94;eval:-88;qs:-88;outcome:0.0
8/5k2/8/4p2p/4P3/2r2PP1/2p2K2/2R5 w - - 0 53;score:-123;eval:-112;qs:-112;outcome:0.5
8/3k1n2/1K5p/7R/6P1/8/8/8 b - - 3 55;score:315;eval:306;qs:306;outcome:0.5
8/6R1/8/5K2/5p2/2r2k2/8/8 w - - 4 67;score:-193;eval:-154;qs:-154;outcome:0.0
r2qkbr1/ppp1pb2/2n2n2/3pP2p/P2P2p1/2PQ2P1/1P1N1PB1/R1B1K1NR b KQq - 0 9;score:87;eval:-1;qs:-1;outcome:0.5
5rk1/R4pp1/5n1p/1r6/1bqp1P2/4PK1P/3P2P1/1N2Q1NR w - - 0 23;score:-98;eval:-145;qs:-145;outcome:0.0
Basically, <FEN>;score:<SCORE>;eval:<EVAL>;qs:<QUIESCENCE SEARCH SCORE>;outcome:[1.0|0.5|0.0]
Self-play games data can be generated by using cutechess-cli, for example:
cutechess-cli -tournament gauntlet -concurrency 15 \
-pgnout zahak_games/PGN_NAME.pgn "fi" \
-engine conf=zahak_latest tc="inf" depth=9 \
-engine conf=zahak_latest tc="inf" depth=9 \
-ratinginterval 1 \
-recover \
-event SELF_PLAY_GAMES \
-draw movenumber=40 movecount=10 score=20 \
-resign movecount=5 score=1000 \
-resultformat per-color \
-openings order=random policy=round file=SOME_BOOK_HERE format="epd" \
-each proto=uci option.Hash=32 option.Threads=1 \
-rounds 100000000
And then fengen can be used to convert
the games into FENs, please consult the fengen repository for more
information.
The trainer exposes many command line arguments to configure the net and its architecture:
$ ./zahak-trainer -help
Usage of ./zahak-trainer:
-epochs int
Number of epochs (default 100)
-from-net string
Path to a network, to be used as a starting point
-hiddens string
Number of hidden neurons, for multi-layer you can send comma separated numbers (default "256")
-input-path string
Path to input dataset (FENs), for multiple files send a comma separated set of files
-inputs int
Number of inputs (default 769)
-lr float
Learning Rate (default 0.009999999776482582)
-network-id int
A unique id for the network (default 1277010531)
-output-path string
Final NNUE path directory
-outputs int
Number of outputs (default 1)
-profile
Profile the trainer
-sigmoid-scale float
Sigmoid scale (default 0.0068359375)
- Aryan Parekh the author of Bit-Genie for helping me understand the basics of NN and NNUE, and sticking with me and patiently available for answering all my (stupid) qeustions until the very end.