brianprichardson

Something isn't right. Suggest going back to an earlier commit. At least this far: https://github.com/Zeta36/chess-alpha-zero/commit/ef474860214b7f65ede341e5d827774b97e9dc88 Please let me know how it goes. Caution: Self play takes a very long time.

I haven't run this is many months. I should be able to look at it tomorrow when I'm not working. In the mean time, see if you can just run...

OK, something is hinky. I did a fresh clone and it did not work for me. So, next I renamed the data directory and ran with "self". After the data...

Good idea. I'm still a github noob, but I'll try to figure it out. BTW I haven't looked at this repo in many months. Most have moved on to https://github.com/LeelaChessZero/

There already is a forked repo that works for me. See: https://github.com/andynik/deep-move PS Be careful with the Keras and tensorflow-gpu versions. I have to use either 1.4 or 1.5 on...

For a python Leela option see: https://github.com/dkappe/leela_lite

Just tried it this morning and it worked fine. Did uci, self, and opt commands and all worked for me on Win7 with Keras 2.1.4 and TF gpu 1.4

More information and context is needed. What run.py command, some screen output, name of the dll.

32 policy and value head filters looks fine. LR finder and 1cycle also looking good. Now trying 2 epochs per batch. Takes about a week to tune and test for...

Larger 20x256 nets take considerably longer, so still crunching on training. Also, the cutechess run to measure any improvement takes much longer to play the games.