Search found 285 matches

by dkappe
Thu Jan 31, 2019 8:25 am
Forum: Computer Chess Club: General Topics
Topic: 11258-104x9-se distilled network released
Replies: 4
Views: 900

11258-104x9-se distilled network released

11258-104x9-se distilled network released. See here: https://github.com/dkappe/leela-chess-w ... d-Networks

32x4-se is next. The hunt continues.
by dkappe
Tue Jan 29, 2019 4:48 pm
Forum: Computer Chess Club: General Topics
Topic: Distilled Networks for Lc0
Replies: 21
Views: 3703

Re: Distilled Networks for Lc0

The 24x3-se network has been released. See the distilled networks page. https://github.com/dkappe/leela-chess-weights/wiki/Distilled-Networks 24x3-se is ~300 elo stronger than 16x2-se, just around winter in strength. If someone has a very slow machine, like a raspberry pi, this might be the one for ...
by dkappe
Mon Jan 28, 2019 2:13 pm
Forum: Computer Chess Club: General Topics
Topic: Distilled Networks for Lc0
Replies: 21
Views: 3703

Re: Distilled Networks for Lc0

OK, but these papers do not tell us your own works and the results depend on what you actually did. So I think an interpretation of your works would be needed. Naturally LC0 has the both head. But there are NN with separated value and policy head and there are NN in which the both are in the same s...
by dkappe
Mon Jan 28, 2019 1:11 am
Forum: Computer Chess Club: General Topics
Topic: Distilled Networks for Lc0
Replies: 21
Views: 3703

Re: Distilled Networks for Lc0

OK, thanks. But about the lost of information during the process there is no any data/opinion. I think in the case of GPUs the playing power also a trade off between speed and knowledge. The recent 20x256 net size is not enough to overcharge an RTX 2080Ti. Maybe a 40x256 can do it. I think you know...
by dkappe
Sun Jan 27, 2019 5:58 pm
Forum: Computer Chess Club: General Topics
Topic: Distilled Networks for Lc0
Replies: 21
Views: 3703

Re: Distilled Networks for Lc0

When you "distill" a network to get a smaller and faster NN it may lost some information. What is your experience about it? Did you make tests between original and "distilled" NN? How faster the "distilled" NN-s are? See the distilled network page for an extensive tournament testing various sizes. ...
by dkappe
Sat Jan 26, 2019 10:26 pm
Forum: Computer Chess Club: General Topics
Topic: Distilled Networks for Lc0
Replies: 21
Views: 3703

Re: Distilled Networks for Lc0

Hi Dietrich, thanks a lot for your work. When I compare your list with our CEGT results I see a difference with Crafty. Does it run on 1CPU? 1 ethereal : 3262 35 (CEGT 3187) 2 ID11258-112x9-se : 2965 3 crafty25.2 : 2948 26 (CEGT 2790) Everything runs on 1 CPU. Although on CCRL, crafty 25.2 is at 30...
by dkappe
Sat Jan 26, 2019 10:18 pm
Forum: Computer Chess Club: General Topics
Topic: Distilled Networks for Lc0
Replies: 21
Views: 3703

Re: Distilled Networks for Lc0

Werewolf wrote:
Fri Jan 18, 2019 9:08 pm
What do you mean by “distilled”?
You use one network to train another (usually smaller) network, rather than being trained by self-play or supervised learning data.
by dkappe
Sat Jan 26, 2019 9:54 pm
Forum: Computer Chess Club: General Topics
Topic: Distilled Networks for Lc0
Replies: 21
Views: 3703

Re: Distilled Networks for Lc0

Also worth mentioning is that I am almost done with a 24x3 net. Not sure where that will come in handy, but I'll make them and let other people figure out what they're good for.
by dkappe
Sat Jan 26, 2019 9:45 pm
Forum: Computer Chess Club: General Topics
Topic: Distilled Networks for Lc0
Replies: 21
Views: 3703

Re: Distilled Networks for Lc0

I've also added a link to a pb.gz version of the 598 old main line net. They are probably available elsewhere, but hard to track down.
by dkappe
Sat Jan 26, 2019 9:31 pm
Forum: Computer Chess Club: General Topics
Topic: Distilled Networks for Lc0
Replies: 21
Views: 3703

Re: Distilled Networks for Lc0

Added another.

https://github.com/dkappe/leela-chess-w ... d-Networks

This is 128x10. I know we had another, but not based on the same net and distillation process. Testing ongoing. If it slots in below 112x9, then we'll see how 104x9, 112x10, 120x9 and so on perform.