An amazing net for SF-NNUE.

Discussion of anything and everything relating to chess playing software and machines.

Moderators: Harvey Williamson, Dann Corbit, hgm

Forum rules
This textbox is used to restore diagrams posted with the [d] tag before the upgrade.
User avatar
Sylwy
Posts: 3481
Joined: Fri Apr 21, 2006 2:19 pm
Location: IASI (Romania) - the historical capital of MOLDOVA

Re: An amazing net for SF-NNUE.

Post by Sylwy » Wed Aug 05, 2020 7:51 am

Guenther wrote:
Tue Aug 04, 2020 3:15 pm
Good judgement...

As expected the cloning on NNUE nets has begun.
People can now test clones of engines with clones of nets w/o knowing they play A+A vs. A+A.

This so called qh-fine-0803-5 net is 99.99999% identical to Sergio Vieris net 20200729-0109.
Only at the very end of millions of hex entries 'someone' changed a few hex values ;-)
From Discord channel:

Image

Guenther
Posts: 3555
Joined: Wed Oct 01, 2008 4:33 am
Location: Regensburg, Germany
Full name: Guenther Simon
Contact:

Re: An amazing net for SF-NNUE.

Post by Guenther » Wed Aug 05, 2020 8:51 am

Sylwy wrote:
Wed Aug 05, 2020 7:51 am
Guenther wrote:
Tue Aug 04, 2020 3:15 pm
Good judgement...

As expected the cloning on NNUE nets has begun.
People can now test clones of engines with clones of nets w/o knowing they play A+A vs. A+A.

This so called qh-fine-0803-5 net is 99.99999% identical to Sergio Vieris net 20200729-0109.
Only at the very end of millions of hex entries 'someone' changed a few hex values ;-)
From Discord channel:

Image
You should have mentioned that he simply 'tried' to fine-tune an exisisting SV net (which failed miserably),
before jumping here and just post a link to something, praised as something new by you, w/o further informations.

After checking this discord channel, I see that he openly stated he used already a SV net for his previous attempt,
but not expliccitely for this new one, though.
Obviously his fine-tuning attempts failed either due to bugs or errors in his procedure.
The result remains the same. At least it seems not intentionally.
qhapaq_rope_6
03.08.2020
@linox qh 0726 is fine-tuned function from 20200721-1907.bin. It uses approximatedly 20 million positions.
https://discord.com/channels/4359437104 ... 6523195462

There is no such message about qh-fine-0803-5.

User avatar
Ovyron
Posts: 4405
Joined: Tue Jul 03, 2007 2:30 am

Re: An amazing net for SF-NNUE.

Post by Ovyron » Wed Aug 05, 2020 11:00 am

Guenther wrote:
Tue Aug 04, 2020 3:15 pm
As expected the cloning on NNUE nets has begun.
People can now test clones of engines with clones of nets w/o knowing they play A+A vs. A+A.
Welcome back Guenther, hope you have gotten a thicker anti-troll skin this time around :D

maxdeg
Posts: 222
Joined: Thu Mar 10, 2016 5:17 pm
Location: Canada

Re: An amazing net for SF-NNUE.

Post by maxdeg » Wed Aug 05, 2020 12:44 pm

Ovyron wrote:
Wed Aug 05, 2020 11:00 am
Guenther wrote:
Tue Aug 04, 2020 3:15 pm
As expected the cloning on NNUE nets has begun.
People can now test clones of engines with clones of nets w/o knowing they play A+A vs. A+A.
Welcome back Guenther, hope you have gotten a thicker anti-troll skin this time around :D
:D :D .. if I recall you're still on his list .. #100 .. :D :roll:

User avatar
Sylwy
Posts: 3481
Joined: Fri Apr 21, 2006 2:19 pm
Location: IASI (Romania) - the historical capital of MOLDOVA

Re: An amazing net for SF-NNUE.

Post by Sylwy » Wed Aug 05, 2020 4:27 pm

Guenther wrote:
Wed Aug 05, 2020 8:51 am
You should have mentioned that he simply 'tried' to fine-tune an exisisting SV net (which failed miserably),
before jumping here and just post a link to something, praised as something new by you, w/o further informations.
I took the link and the short presentation from the discord channel (the place where SF-NNUE was originally developed). I just tested the net (which is very strong). For the rest ....... I am not a programmer to know - a priori - the technical details.

Image

Zenmastur
Posts: 914
Joined: Sat May 31, 2014 6:28 am

Re: An amazing net for SF-NNUE.

Post by Zenmastur » Thu Aug 06, 2020 7:32 am

pohl4711 wrote:
Tue Aug 04, 2020 10:55 am
Zenmastur wrote:
Mon Aug 03, 2020 7:33 pm
Sylwy wrote:
Mon Aug 03, 2020 6:49 pm
Trained on Reinforcement Learn and fine tuned using CCRL games:

https://www.dropbox.com/s/71qy3j8gtgch2 ... 5.zip?dl=0
How does this net fair against some of Sergio's more recent nets?
I tested it vs. SF 200717, like all S.Vieri nets. Look here:
https://www.sp-cc.de/files/ergebnisse_nnue_schnell.txt

The net is very strong, but not the best in that tests. The best is still sv200730_1934
That's a nice list you have there...

While not enough games to be definitive it certainly gives one an idea of which nets to try and which to stay away from. I liked it enough to save a copy in my NNUE nets directory for future reference.

Regards,

Zenmastur
Only 2 defining forces have ever offered to die for you.....Jesus Christ and the American Soldier. One died for your soul, the other for your freedom.

pohl4711
Posts: 1414
Joined: Sat Sep 03, 2011 5:25 am
Location: Berlin, Germany
Contact:

Re: An amazing net for SF-NNUE.

Post by pohl4711 » Thu Aug 06, 2020 10:11 am

Zenmastur wrote:
Thu Aug 06, 2020 7:32 am

That's a nice list you have there...

While not enough games to be definitive it certainly gives one an idea of which nets to try and which to stay away from.
That was the idea. Because S.Vieri builds so much new nets, it is not possible to test them all with some thousand games using a "quite normal" thinking-time (even though I use bullet-speed, 60''+0.5''). So, 500 games have to be enough...
But this list is only a collection of pre-tests, so, I can find out, which nets I should test in a regular 7000 game testrun (with 3'+1'') for my ratinglist... Not more, not less.

Post Reply