Stockfish NN release (NNUE)

[ChickenLogic]

It is now easily possible to use any pgn as training data! First, use this pgn-extract command: pgn-extract: pgn-extract --fencomments -Wlalg --nochecks --nomovenumbers --noresults -w500000 -N -V -o output_file_name input_file_name

and then: learn convert_bin_from_pgn-extract output_file_name output_file_name input_file_name

Use this binary (bmi2) with the latest commit from nodchip/ttak (current release doesn't have that change last time I checked): https://cdn.discordapp.com/attachments/ ... ev_blas.7z

[matejst]

Guys, I just want to thank you for this interesting project! I think it is the right way to go in engines' development.

[SheikhYerbouti]

There's a guy on Leela Discord who generated 1 million Stockfish v Leela games for NN training. So these too could be used for NNUE and maybe produce a better result than using SF games or leela games alone.

[Engin]

where is the discord link ?

[kranium]

Is that the one you want?
https://discord.com/channels/4359437104 ... 6266188890

It's for Stockfish-NNUE (not Leela)

[Engin]

yes thats it thx

[matejst]

Norm,

You uploaded popc binaries on the 26th, 27th and 29th of this month. But something is strange: the binaries from the 27th are three times faster than the binaries from the other days on my laptop.

[kranium]

Norm,

You uploaded popc binaries on the 26th, 27th and 29th of this month. But something is strange: the binaries from the 27th are three times faster than the binaries from the other days on my laptop.

yes, strange..

If you're using bench
I believe I did read somewhere (nnue discord?) from the author, that bench is not reliable/functional in this implementation.

I'll try to verify.

Thanks Boban

[kranium]

Norm,

You uploaded popc binaries on the 26th, 27th and 29th of this month. But something is strange: the binaries from the 27th are three times faster than the binaries from the other days on my laptop.

Thanks, I found the issue.
I'll upload new binaries soon.

[matejst]

Norm, I am not using bench: you know that it is too complicated for me. I was surprised when I noticed a surge in performance with the new binaries, two days ago -- on my old AMD A8, SF is achieves 550kn/s, the binaries from the 26th were ten times slower, about 53kn/s, and those of the 27th were achieving 155 kn/s. The new one are slower again. I also checked the number of positions after one mn, the depth, and it seems coherent with the nodes's count. Anyway, thank you for the popcount binaries -- it allowed me to finally try this new engine. It is similar to Winter, and recently Jonathan tried with two NN for the eval. I see a lot of intriguing, new possibilities arising.