alex67a wrote: ↑Thu Sep 24, 2020 8:51 pm
Minic don't work on Win 7 and Arena gui...
Maybe you can try an older release to double check the error is not on your side.
Please be sure to use the proper executable to fit your cpu to avoid illegal instruction issue.
(I checked in my Win10 VM in cutechess, no issue, but not the same as your Win7 Arena of course)
In a (too) small home test, it looks like MinicNNUE 4 threads + SV nn-03744f8d56d8 can nearly equalize versus SF10 single threaded on my hardware. So much impact of evaluation ! (this 4threads versus 1 thread implies same nps for both engines).
xr_a_y wrote: ↑Fri Sep 25, 2020 8:55 pm
In a (too) small home test, it looks like MinicNNUE 4 threads + SV nn-03744f8d56d8 can nearly equalize versus SF10 single threaded on my hardware. So much impact of evaluation !
Not a single word about "Napping Nexus" ? A beautiful premiere. Congratulations !
xr_a_y wrote: ↑Fri Sep 25, 2020 8:55 pm
In a (too) small home test, it looks like MinicNNUE 4 threads + SV nn-03744f8d56d8 can nearly equalize versus SF10 single threaded on my hardware. So much impact of evaluation !
Not a single word about "Napping Nexus" ? A beautiful premiere. Congratulations !
Well "Napping Nexus" is an experimental first not-too-bad network I train from SF data, using Nodchip SF leaner. It knows how to play, kills Minic standard eval easily but it still very very far from best SV nets ! I'm glad it works, this is a first step in building nets by myself but I expect to release a pure Minic net soon, based on Minic data and using the learner merged inside Minic. Fingers crossed.
xr_a_y wrote: ↑Fri Sep 25, 2020 9:16 pm
(...)
Well "Napping Nexus" is an experimental first not-too-bad network I train from SF data, using Nodchip SF leaner. It knows how to play, kills Minic standard eval easily but it still very very far from best SV nets ! I'm glad it works, this is a first step in building nets by myself but I expect to release a pure Minic net soon, based on Minic data and using the learner merged inside Minic. Fingers crossed.
Bravo Vivien for your spectacular progress with Minic, it's really impressive ! I hope you will succeed with your own training attempts, I find this terribly difficult... But I keep hope !
xr_a_y wrote: ↑Fri Sep 25, 2020 9:16 pm
(...)
Well "Napping Nexus" is an experimental first not-too-bad network I train from SF data, using Nodchip SF leaner. It knows how to play, kills Minic standard eval easily but it still very very far from best SV nets ! I'm glad it works, this is a first step in building nets by myself but I expect to release a pure Minic net soon, based on Minic data and using the learner merged inside Minic. Fingers crossed.
Bravo Vivien for your spectacular progress with Minic, it's really impressive ! I hope you will succeed with your own training attempts, I find this terribly difficult... But I keep hope !
Regards from... France
Merci David !
Building nets still look more like a cooking recipe than science to me for now but I'll keep trying to better understand this.
xr_a_y wrote: ↑Fri Sep 25, 2020 8:55 pm
In a (too) small home test, it looks like MinicNNUE 4 threads + SV nn-03744f8d56d8 can nearly equalize versus SF10 single threaded on my hardware. So much impact of evaluation ! (this 4threads versus 1 thread implies same nps for both engines).
Here are some results MinicNNUE + nn-97f742aaefcd.nnue using various number of threads versus SF various versions (single threaded).