TalkChess.com

Testrun finished: Lc0 0.22.0 with huge 384x30-t40-1097 S.Vieri Net.

https://www.sp-cc.de

(Perhaps you have to clear your browsercache or reload the website)

For me, it becomes more and more clear, that it is the wrong way to build nets, which are bigger size, than the net-size in the training run, which delievered the learning-data. IMHO the other direction is the right way:
Building nets, which are same or smaller size, than the net-size in the training run, which delievered the learning-data.

Means: T40 training was 256x20 size. So any net built out of T40-games should be 256x20 or smaller, not bigger.
T60 training is 320x24. So any net built out of T60-games should be 320x24 or smaller, not bigger.

Next NN-testrun: Lc0 0.22.0 with DarkQueen 2.0 Net (first testrun of a net built out of human (online) games, only)

I doubt, that human games are good enough to build a really strong net out of them. But I want to test that...

pohl4711 wrote: ↑Thu Nov 07, 2019 12:10 pm Testrun finished: Lc0 0.22.0 with huge 384x30-t40-1097 S.Vieri Net.

https://www.sp-cc.de

(Perhaps you have to clear your browsercache or reload the website)

How many nodes per second were you getting?

Albert Silver wrote: ↑Sun Nov 17, 2019 2:54 am

pohl4711 wrote: ↑Thu Nov 07, 2019 12:10 pm Testrun finished: Lc0 0.22.0 with huge 384x30-t40-1097 S.Vieri Net.

How many nodes per second were you getting?

On my mobile RTX 2060 (with reduced speed by afterburner-tool, in order to get a Leela-Ratio of 1.3, because the mobile Hexacore CPU i7-8750H is not so fast (with turbo-boost switched off)), lc0 reaches 11500 n/s in the starting position (wrote this on my website), with a 256x20 net. The 384x30 net is 2.25 times bigger, than a 256x20 net, so around 5100 n/s (chess starting position) are calculated by lc0 on my machine.