How to install Stockfish NNUE!

margaretmcpherson
drmrboss wrote:

Go and check engine rating list or go check Stockfish Official Github where SF nnue is about to merge in SF.

 

 

 

 

 

 

If you dont know these website,  or if you dont know how to test engines, google for 1 week and come back.

 

Thanks! These are fast moving developments, please be nice people ! happy.png This is all very interesting .  Are there NNUE nets trained on deep LC0 evaluation of positions?  This would then allow LC0 evaluations in an AB engine.

PerpetuallyPinned
infinitless wrote:
drmrboss wrote:
infinitless wrote:

How about performance. LC0 is pretty useless for the majority of CPU only users. I am assuming that SFNNUE uses Neural Nets so to run it only on a CPU is equally useless. The debate again comes down to how much you want to invest in a decent GPU (or GPUs) to get the optimum/maximum performance from the NN Engine and the corresponding net, whether SFNNUE or LC0.

The developers really need to concentrate on how to run NNs on CPU only machines. There is no point if the average user has to buy GPUs with tensor cores to run Chess Engines for daily use.

Dude, you need to know what is NNUE first  before talking about nonsense.

 

https://www.chess.com/forum/view/general/stockfishnnue-probably-the-future-of-stockfish

 

Many themes of puzzles are used to test various engines. Here is a collection of 213 such puzzles where engines mostly fail as the themes are really complex. I tested SF, SFNNUE and LC0 (all on CPU) and only LC0 managed to solve about 50% of these. SF and SFNNUE failed miserably. It would be great if someone can make an objective comparison and post results. @drmrboss you can keep talking rude nonsense sitting on your couch. The rest of us will get on with adding some value in the World.

#1 to #60 : https://lichess.org/study/y6sM5eAf
#61 to #120 : https://lichess.org/study/Vfe59GOH
#121 to #180 : https://lichess.org/study/0FQa6riN
#181 to #213 : https://lichess.org/study/s2mTErEZ

Can you proved the results for puzzle #1?

I found a discrepancy in the first puzzle's solution.

Kinda hard to pass a test when the incorrect answer is the right one.

drmrboss
infinitless wrote:

@drmrboss I am on Discord. Github is for foetuses like you.

How old are you?

If you are a teenager, stop posting non sense here and stop wasting my time.

This post is educational for people. 

Dont troll in this post!

Learn what is Github and come back!

 

PerpetuallyPinned

How do the engines/NNs fair against this puzzle?

Daniel2340
drmrboss wrote:

This guy, the authour of  Rebel engine, organized built in SFnnue for download in his website.

 

http://rebel13.nl/download/stockfish-nnue.html

 

Interesting... If you come across a compile for Mac, please share.

drmrboss

Stockfish NNUE become official stockfish build in abrok.

 

https://abrok.eu/stockfish-nnue/

 

 

The most interesting thing is how Neural Networks activation can be visualized.

https://hxim.github.io/Stockfish-Evaluation-Guide/?p=nnue

 

drmrboss

SFnnue passed initial test with +87 elo. ( biggest gain in Stockfish history). Probably +50 elo in regression test( not tested yet).

infinitless
PerpetuallyPinned wrote:

How do the engines/NNs fair against this puzzle?

 LC0 says Rf4

PerpetuallyPinned
infinitless wrote:
PerpetuallyPinned wrote:

How do the engines/NNs fair against this puzzle?

 LC0 says Rf4

Any context to go with it?

Nodes/depth/time or hardware setup?

infinitless

500 NPS max - just an old laptop with Intel i5 4 cores. 8 GB System RAM. No GPU. (Well... Intel Onboard graphics)

PerpetuallyPinned

Did Lc0 find it immediately?

infinitless

Yeah after about 4/5 seconds. The laptop never goes over 500 NPS as compared to if I run SF on it. SF starts with like 2500 KNPS!

PerpetuallyPinned

SF 11 on just a cheap Android...

Doesn't find it as the best move until depth 18 (still has Black with the advantage). I did a little testing a few days ago and needed to analyze number of variations equal to number of possible moves. Guessing this is prunning effect due to sacrificing material.

I'll share the evaluation in a few.

infinitless

I have a screencast video of how quickly LC0 finds it but the comments dialog box won't lemme upload it. http://gofile.me/26RAj/TL4c7CmW6 Enjoy!

PerpetuallyPinned

Lc0 still has as -0.xx forgot exact number...

Would be interesting to see when it finds the draw.

infinitless

Thing is neural nets are good at plans with full pgns but not so good with positions. It finds the best line as soon as you move Rf4