Most of this code uses
haibrid_chess_utils which can be installed with
setup.py. Theres a few dependancies that should also be installed, and Pytorch and Tensorflow are needed too.
Also, you may also want
lczero_tools which does low level Leela stuff. The linked repo has been forked from Leela and updated for the new version of
I've only tested this code with Python3.6+ on Unix machines.
Our main results are in
results which has a series of Jupyter notebooks showing how the final plots were created, along with some plots. These are usually cleaned up versions of other notebooks.
Most of the raw data is stored on Ada the lab's server and it is converted with the scripts in
data_generators. Then some is transferred to the Azure VM,
haibridgpu2.eastus2.cloudapp.azure.com which I can add RSA keys to if you want.
The maias are trained in
training and use lots for code from lczero-training. The blunder prediction stuff mean while is in
blunder_prediction and is mostly home grown, thus it's in Pytorch instead of Tensorflow.
Most of the plots were generated by Jupyter notebooks found in
notebooks, although there's a bunch more I haven't added to the repo so there's probably a plot or two with no source there.
There's a few other directories I don't think there's anything useful in any of them though.