Request for network.

SeniorChess · Post by **SeniorChess** » 09/02/2021, 19:18

Hello friends. Could one of you please make a nnue network based on KOMODO DRAGON engine at the best depth and performance ??
Best Regards...

Skynet · Post by **Skynet** » 10/02/2021, 4:35

SeniorChess wrote:Hello friends. Could one of you please make a nnue network based on KOMODO DRAGON engine at the best depth and performance ??
Best Regards...

Hello. Your post has been moved here in Request Section.

SeniorChess · Post by **SeniorChess** » 10/02/2021, 7:14

Ok bro

deeds · Post by **deeds** » 12/02/2021, 10:55

So you mean to take all the games of komodo dragon, extract the data in order to convert them into plain text and train a net. Maybe it's possible without rescoring...

SeniorChess · Post by **SeniorChess** » 12/02/2021, 11:27

Yes deeds , if u can please train a net of Komodo Dragon ( Games from Ccrl Blitz , Komodo Dragon 8 CPU - rating 3671 ).
I'm sure 1300 games ( this is number of games disponible of Dragon in CCRL BLITZ ) of Dragon Komodo it's enough...
I don't know to make a nnue net...So please make a 20 MB nnue net... BEST REGARDS...

SeniorChess · Post by **SeniorChess** » 12/02/2021, 11:29

Or maybe you can take the commented pgn games of Komodo Dragon 64 bit 4 CPU ( 40/15 CCRL ) to extract nnue ...

deeds · Post by **deeds** » 12/02/2021, 15:42

1st problem : these games aren't some "selfplay" because they contain the opponent's scores/depths.

2nd problem : 1000 games @ 100 positions/game minus 10% or 15% of duplicated EPDs, do we really talk about training a net with only 80k EPDs !?

3rd problem : i haven't a komodo dragon licence. Furthermore why train a net from an engine which embed even a net !?

I have even offer my own toolkit to (for example) extract data from existing commented games. Without rescoring, no need a powerful rig !

Outskirts CheSS ForuM

Outskirts CheSS ForuM