|
|
Post by matejst on Apr 8, 2022 18:39:46 GMT
Is there a NNUE based only on Leela data? I searched among the SF nets, but I did not find one. I probably missed it.
|
|
|
|
Post by Ozymandias on Apr 14, 2022 9:40:01 GMT
matejst Fat Fritz 2, as long as you accept that Fat Fritz was lc0  sorry for the delay, I didn't see this post
|
|
|
|
Post by matejst on Apr 14, 2022 16:40:06 GMT
It's OK, Ozy, I am at work myself, with an intermittent internet connection.
I am not sure that FF2 was based on FF1 -- which was Lc0. I understood that the net was developed on human games, but I am surely wrong. There was a GPLed version of FF2 and I tried it, but, like I repeated several times, I somehow don't like SF, and I just tried it for a day, than deleted it. I have somewhere a backup.)
An interesting experiment was Texel with Giraffe's net, and I still use it from time to time. Not too strong, the technology was at his beginning, but it was a net built from scratch and the style is original.
|
|
|
|
Post by Ozymandias on Apr 14, 2022 17:20:51 GMT
Here's the summary, don't be fooled by the parody, it's quite exact:
Looks like both versions were trained on GM data.
|
|
|
|
Post by matejst on Apr 14, 2022 17:49:07 GMT
Looks like both versions were trained on GM data. 1. To tell you the truth, I tend to think that it was trained, like you said, on Leela data. Imho, AS probably took GM games, analyzed them with Lc0, than learned from the evaluation and outcome. But it still is not enough data: we see that you need 1b-3b positions for a good 30 MB net, and I doubt that he found enough played by GMs. It was pure Lc0 all the time.
2. Where did they found this guy? |
|
|
|
Post by Ozymandias on Apr 14, 2022 18:00:42 GMT
|
|
