RubiChess NNUE player implemented

[Rebel]

Just please, do not use "Minic" as name when a net is given, please display MinicNNUE.

I will add MinicNNUE to the OKE test, curious if it will give the early opening a boast as well.

3th place right after SF12 and Lc0 at the 2-move set.

http://rebel13.nl/html/2-moves.html

[xr_a_y]

With such good results and such a huge similarity, I definitly won't use SF-based-net and NNUE as "official" Minic.

But I'm working on a home made net currently, based only on Minic eval ... and it seems to be already +80elo versus standard Minic.

By the way, you should probably give credits to the net being use in the name of NNUE engines. So here probably "Minic 2.48 NNUE + SV XXYY"

[dkappe]

Makes me wonder what a (depth=1) similarity test will reveal.

A perfect acid test for your similarity test.

Yes, I can understand the frustration of Andrew and Andreas.

You have a number of my Toga and ICE based nets available to you. You can run them on a number of different NNUE supporting engines and compare them against the originals. I’d love to see the results of this test.

Help me out with some links, I haven't followed everything.

Also, I tried Igel, but there is no popcount version on github.

Minic is a different story, I can run it with or without the NNUE file, the result remains the same.

Code: Select all

Positions 10000 Minic Minic Minic Minic sf-nn 
Minic 2.48      ----- ----- ----- ----- 39.01 
Minic 2.48x     ----- ----- ----- ----- 39.01 
Minic 2.48y     ----- ----- ----- ----- 39.01 
Minic 2.48z     ----- ----- ----- ----- 39.01 
sf-nnue         39.01 39.01 39.01 39.01 ----- 

Toga https://www.patreon.com/posts/toga-iii-0-4-41320965
Frosty (ice) https://www.patreon.com/posts/frosty-0-2-40668615

Unless you use something like Cfish or NetRunner, you’re going to get the hybrid eval, where stockfish eval is used past some threshold.

[Albert Silver]

Minic is a different story, I can run it with or without the NNUE file, the result remains the same.

Probably the NNUEFile is not read, because of bad path.
You can check that easily looking at speed. NNUE version shall be slower.
If you're on Windows I guess there is some question about how to write the path (with / or \).
Also horrible things can probably happen with spaces in path.

How does the default path+name of the NNUE file look in your engine?

I assumed : eval\nn-97f742aaefcd.nnue

1) Place the NN file in the root directory of the executable and change that EvalFile reference to its basic name, the extension is unimportant. So if Minic.bin was the filename of its NN, that is what you would write.

2) Remember that if you are using any SF build it will almost certainly have hybrid activated. This means that after a certain material imbalance is reached it starts using plain SF as the engine for those nodes, and of course this will not only have a big impact on similarity but the overall evaluation. While it has tested better for SF with SV, it is not better for all other nets.

3) There is an unmentioned aspect to keep in mind too. There was a fix to the training code which augments data during the training. The original code was incorrect and was fixed in a patch, but this patch has not been merged into SF12. The reason is that a new NN would need to be generated as the binary would be unable to read it. Since SF had no replacement they left the old support in place for the SF12 release, but it means that any net made with the fixed code will either not run at all in SF12 or be broken with very bad results. The reverse is true if you use a fixed binary trying to use SV.

[Rebel]

With such good results and such a huge similarity, I definitly won't use SF-based-net and NNUE as "official" Minic.

Well, if I am not mistaken the SF net is a Sergio one, thus public domain.

But I'm working on a home made net currently, based only on Minic eval ... and it seems to be already +80elo versus standard Minic.

You are a gentleman.

[Rebel]

Makes me wonder what a (depth=1) similarity test will reveal.

A perfect acid test for your similarity test.

Yes, I can understand the frustration of Andrew and Andreas.

You have a number of my Toga and ICE based nets available to you. You can run them on a number of different NNUE supporting engines and compare them against the originals. I’d love to see the results of this test.

Help me out with some links, I haven't followed everything.

Also, I tried Igel, but there is no popcount version on github.

Minic is a different story, I can run it with or without the NNUE file, the result remains the same.

Code: Select all

Positions 10000 Minic Minic Minic Minic sf-nn 
Minic 2.48      ----- ----- ----- ----- 39.01 
Minic 2.48x     ----- ----- ----- ----- 39.01 
Minic 2.48y     ----- ----- ----- ----- 39.01 
Minic 2.48z     ----- ----- ----- ----- 39.01 
sf-nnue         39.01 39.01 39.01 39.01 ----- 

Toga https://www.patreon.com/posts/toga-iii-0-4-41320965
Frosty (ice) https://www.patreon.com/posts/frosty-0-2-40668615

Thanks.

But unless I need new glasses I am missing the executables of Toga and Frosty.

Unless you use something like Cfish or NetRunner, you’re going to get the hybrid eval, where stockfish eval is used past some threshold.

[voffka]

Tried, did not work, no avx support on my PC.

Ahh, it is a matter of avx2 and not popcount. Feel free to contact me in PM with exact version of your CPU, I will provide you a build best suited for it so that you can try it.

[dkappe]

Makes me wonder what a (depth=1) similarity test will reveal.

A perfect acid test for your similarity test.

Yes, I can understand the frustration of Andrew and Andreas.

You have a number of my Toga and ICE based nets available to you. You can run them on a number of different NNUE supporting engines and compare them against the originals. I’d love to see the results of this test.

Help me out with some links, I haven't followed everything.

Also, I tried Igel, but there is no popcount version on github.

Minic is a different story, I can run it with or without the NNUE file, the result remains the same.

Code: Select all

Positions 10000 Minic Minic Minic Minic sf-nn 
Minic 2.48      ----- ----- ----- ----- 39.01 
Minic 2.48x     ----- ----- ----- ----- 39.01 
Minic 2.48y     ----- ----- ----- ----- 39.01 
Minic 2.48z     ----- ----- ----- ----- 39.01 
sf-nnue         39.01 39.01 39.01 39.01 ----- 

Toga https://www.patreon.com/posts/toga-iii-0-4-41320965
Frosty (ice) https://www.patreon.com/posts/frosty-0-2-40668615

Thanks.

But unless I need new glasses I am missing the executables of Toga and Frosty.

Unless you use something like Cfish or NetRunner, you’re going to get the hybrid eval, where stockfish eval is used past some threshold.

Grab Cfish and use the uci options to load the nets. Also “Use NNUE” takes a “pure” value to turn off hybrid.

[Laskos]

I won't upload binaries so you have to compile yourself.

Any compile available yet?

I have put one here: https://drive.google.com/drive/folders/ ... IreoNOB2xX

Requires BMI2 CPU.

Not compressed as that causes a warning that I will be denied Google services if I continue to spread malware.

For the same reason, I will delete it within a couple of hours or at most one day.

Can someone re-upload it? Thanks.

[Gabor Szots]

Can someone re-upload it? Thanks.

Done.