This project is intended to be a framework for testing different architectures and features for how good different Tensorflow graphs are at predicting the next few moves and the who will win the game.
The core parts of the framework is written in Cython, and the machine learning parts are written in Tensorflow so performance should be good.
This project is intended to be a sandbox, and is not intended for public use beyond proof of concept.
The framework will load all files matching the glob data/*.sgf and train a
neural network to predict the next few moves, as well as the winner of the game.
The script will run for 819,200 steps, exactly how long this will take depends
on the hardware of the computer but it should take about 8 hours on an NVIDIA
GTX 1080 Ti:
make runThe framework will periodically write the trained models and logs to models/.
You can monitor these logs using tensorboard:
tensorboard --logdir models/See the wiki for test results as it is cumbersome to have to do a commit for every update to the article.