CCRL Benchmark Net

[dkappe]

I can never lay my hands on a CCRL data trained net when I want one. I've made a 128x10 classical value head and convolutional policy head net available. https://github.com/dkappe/leela-chess-w ... i/CCRL-Net

It looks to be about 200 elo weaker than Bad Gyal 7, which is surprising to me.

[Max]

Hi Dietrich,
any explanation why your small girls/gyals are so much faster than the distilled networks at the same size?

On CPU: ./lc0 benchmark --minibatch-size=8 --max-prefetch=0

Code: Select all

 nps	size     net
--------------------------
7850  	32x4 meangirl-7	
4700  	48x5 bad-gyal-1
      	48x5 bad-gyal-2
      	48x5 evilgyal-6
4000	16x2 11258-16x2-se
3500	24x3 11258-24x3-se
3020	32x4 11258-32x4-se
2800	64x6 bad-gyal-3
2475	48x5 11258-48x5-se
1780	64x6 11258-64x6-se

[dkappe]

The architecture is different, but other than that, no idea.

[dkappe]

It looks to be about 200 elo weaker than Bad Gyal 7, which is surprising to me.

Make that 250 elo. I think that this is a good piece of evidence that the SL tools surrounding lc0 yield suboptimal results.