crem wrote: ↑Tue Jun 19, 2018 10:34 am
There was a bug in versions 20180609 and 20180614, which caused it to miss checkmates sometimes (especially mates in 1).
Update is highly recommended.
Thanks, I suppose it affected training too, right?
I checked the new net ID1039 with the old 20180614 engine in the same gauntlet against AB engines at short TC, it came at 47.5/200, which is not bad. It remains to see if by say ID1050 it comes to the levels 5th run of ID48 of 60.0/200 or ID57 of 64.5/200 (which was a record high for any 6x64 nets).
Yes, that also affects training and may explain why test1 had much better metrics than later runs.
We are restarting test training again within next few hours.
Not sure who is working on LC0 or not but it seems that since about a week there are serious issues as there was multiple restart and that each of them though I am close to clueless didn't seem to improve on the original work?!
While I am still very curious as to how things evolve, I am sure there is more frustration on the ones who are spending much more time on direct testing or improvements.
Anyhow, looking forward to see it going back to live with clean & straight progress.
crem wrote: ↑Tue Jun 19, 2018 2:54 pm
The latest lc0 is said not to work at all.
I'll take a look what's happening with it in ~6 hours, but meanwhile I guess it doesn't make sense to use lc0 at all.
Latest lc0 seems to be good after all. The training failed because of the training pipeline, not because of lc0 as I thought.
So, the latest lc0 is fine to use.
crem wrote: ↑Tue Jun 19, 2018 10:34 am
There was a bug in versions 20180609 and 20180614, which caused it to miss checkmates sometimes (especially mates in 1).
Update is highly recommended.
Thanks, I suppose it affected training too, right?
I checked the new net ID1039 with the old 20180614 engine in the same gauntlet against AB engines at short TC, it came at 47.5/200, which is not bad. It remains to see if by say ID1050 it comes to the levels 5th run of ID48 of 60.0/200 or ID57 of 64.5/200 (which was a record high for any 6x64 nets).
The latest net I tested, ID9044, is the strongest so far in a gauntlet against AB engines of all 6x64 nets, scoring 73.0/200. The previous best net was ID57 with 64.5/200. I think it's pretty obvious that the 6x64 net is close to saturation now, and I don't know why they don't bootstrap the data and go to larger size nets. Do they wait for something? They seem like a broken wheel with all these re-runs of the smallest net, especially when it performs so well.
mesilikas wrote: ↑Sun Jun 24, 2018 6:46 pm
Thanks Laskos for your infos about these small nets. Seems that all we have old computers must be limited to these small nets. What is your opinion?
No, even on weak hardware larger nets are better, at least to some point if only CPU is used. In the main branch, before I bought my GTX 1060 GPU, the reasonably trained 15x192 nets were still performing better on CPU than smaller nets, although NPS were much lower.