Life In 19x19

KataGo v1.12.0 is released!

(+ some bugfix releases, bringing us to v1.12.3)
https://github.com/lightvector/KataGo/r ... ag/v1.12.3

If you've been following this other thread viewtopic.php?f=18&t=17944 you might have seen some posts about a new neural net "b18". This version officially adds support for a new neural net architecture, of which that net is the first notable new net trained. The new architecture replaces the old "resnet" architecture with a "nested bottleneck resnet", as well as incorporating some major policy training improvements suggested by a contributor "lionfenfen". (thanks to them and all others who have contributed to development or to the public run!).

Once enough active contributors to https://katagotraining.org/ update, I hope we can also use the new architecture for the public run, hopefully it can lead to overall improvements there as well. And of course, if you have a decent machine and are interested in helping KataGo train, https://katagotraining.org/ would always be happy to have new people donating their spare GPU cycles.

Edit: A bug was found and fixed in the TensorRT version. The fixed release v1.12.1 is here.
Edit: And one more quick round of fixes, including a major fix for OpenCL users. New release v1.12.2.
Edit: And after another week or so, some more bugfixes bringing us to v1.12.3:
https://github.com/lightvector/KataGo/r ... ag/v1.12.3

Thanks to "hyln9" for finding and quickly fixing a bug in the TensorRT version of KataGo that would cause it to compute incorrect values, which I had accidentally made when trying to update that backend for the new architecture. A fix has been released (top post edited accordingly).