The OP has contacted the author and apparently it isn't possible to convert any plain text fens/epds/pgns into usable training data for SF NNUE even if they are evaluated. By "not possible" I actually mean "not implemented" as in: the author made this as a proof of concept and if we want it we need to code it on our own. I would volunteer. Only 2 problems: 1) I don't speak Japanese 2) I can't codecucumber wrote: ↑Wed Jun 03, 2020 9:36 pm I made a Dropbox for the ~100,000,000 FEN + game result + evals that I have. Here, for anyone who wants them.
https://www.dropbox.com/s/lri6ujmnrtm8o ... D.zip?dl=0
They've been shuffled already, so the FENs aren't in any particular order. Only problem is that the neural network will never see mate, since Fishtest adjudicates games. It might not see the fifty move rule either, but I'm not sure. SF's search is good enough that maybe that isn't an issue. I hope they can still be of some use.
What I will try to do is to get OpenMP to work and update the search from SF10 to SFdev.
If anyone is crazy enough to write a Fen to SF NNUE training-data conversion tool in their spare time let me know
There seems to be an option to generate the training data in plain text but it is commented out. That might be a good starting point.From this: https://github.com/nodchip/Stockfish/bl ... earner.cpp Line 2468 to 2500 is this snippet below:
Code: Select all
// if (fs.read((char*)&p, sizeof(PackedSfenValue))) {
// // plain textとして書き込む
// ofs << "sfen " << tpos.sfen_unpack(p.sfen) << std::endl;
// ofs << "move " << to_usi_string(Move(p.move)) << std::endl;
// ofs << "score " << p.score << std::endl;
// ofs << "ply " << int(p.gamePly) << std::endl;
// ofs << "result " << int(p.game_result) << std::endl;
// ofs << "e" << std::endl;