Dec 2018 to Mar 2019
I tried rewriting the alpha_zero_general project as zero-play, but I
couldn’t get the TensorFlow models to work.
6 May 2019
Minigo sounds promising. Maybe it has all the features I need for the Shibumi games project, and I don’t need to build my own. It was based on MuGo, so that might also be useful.
The shap project looks like it might be useful for explaining the models, either
for zero-play or for Minigo.
28 May 2019
I’ve been trying to get Minigo to run for a while. I had to install CUDA 10.0
using the instructions for multiple CUDA libraries, and then I had to
upgrade the Nvidia driver to 410. It was a little weird, because they changed
the apt package name from nvidia-390 to nvidia-driver-410. Now I get this error,
and I’m done for the day.
Loaded runtime CuDNN library: 7.3.1 but source was compiled with: 7.4.2.
29 May 2019
I upgraded CuDNN, and now I can make MiniGo play against itself!
export MODEL_NAME=000000-bootstrap
python3 bootstrap.py \
--work_dir=estimator_working_dir \
--export_path=outputs/models/$MODEL_NAME
python3 selfplay.py \
--load_file=outputs/models/$MODEL_NAME \
--num_readouts 10 \
--verbose 3 \
--selfplay_dir=outputs/data/selfplay \
--holdout_dir=outputs/data/holdout \
--sgf_dir=outputs/sgf
1 Jun 2019
It looks like Minigo is actually two projects, a Python version and a C++ version. I thought there were just some C++ components in a Python proect. The Python version also looks like it’s just a proof of concept or documentation for the C++ version. I can’t see how to make the Python version run a complete training pipeline.
I also don’t see any obvious extension points for making Minigo play other board games, so it might just be useful as a more readable template than alpha-zero-general to base my zero-play project on.
Next, I’ll try to get MuGo running.