Available to download here.
Release candidate 1 was mostly bug free, but there were still things to tweak:
- The Rule50 encoding bug was fixed.
- Default batch size for openCl changed to 16.
Up to 5x speedup, promised in RC1, should be visible with default settings. - Time management constants were tweaked a bit.
Feel free to use this version for training, but it's not necessary. It is expected that the "rule50" fix will not have either positive or negative effect on networks in the test10 run.
(Reason: all weights related to that plane are equal to 0 since long ago due to regularization, and it's not really possible to recover from that state.)
We hope that no further changes will be needed and this release candidate will become the v0.17.0.
We've sent this version to CCCC organizers, and it's quite possible that they will have another test of Lc0 playing before the main event, so follow their news if you are interested! Links to CCCC:
is possible to put short engine name? in fritz gui i have now The Lc0 chess engine v0.17.0-rc1cuda which is really long, fritz gui cant work with so long name and if i make same engine for self play its not possible to have different name. I reccomend to make shorter name like Lc0 v0.17
ReplyDeleteWill do.
Delete"all weights related to that plane are equal to 0 since long ago due to regularization, and it's not really possible to recover from that state"
ReplyDeleteIs it bad that we can not recover from that state?
It's bad in a sense that the network cannot start using that information, while this information could be useful in some cases.
DeleteWe'll need a reset for it. There are other reasons for reset too, so it will happen some time in a future.
Will this regularization significantly reduce Leela's strength?
DeleteI don't think they can estimate how much stronger the fix will make Leela on future training runs.
DeleteIs there somewhere that explains why the plane cannot just randomly fill up with non-zero values during training? Then start shifting towards useful values? Is this capability not possible in training?
DeleteCurrent learning rate is 0.0002, it is too little to be able to drag random values to something useful, it will be stuck in some very local maximum. If we still were on the initial learning rate, it would be worth trying.
DeleteI see. Thank you for the informative answer. 😊 - atanh
DeleteAs of right now the CCCC UI has only 23 participants excluding Leela. Are they going to include Leela anytime soon?
ReplyDeleteThey had finished up testing with the NN server and I think they are renting it from Amazon so it's pricey. Thus they don't run it 24/7 during test play. I don't know if more testing is planned before the event starts.
DeleteThey are renting it from Amazon? How do you know this?
DeleteAnd isn't this event continuous, meaning it will run a long time anyway and they would have to pay a lot of money anyway?
There is a big thread on talkchess.com about this. http://talkchess.com/forum3/viewtopic.php?f=2&t=68253&start=50#top
Delete