patrik-ha / explainable-minichess Goto Github PK

The part of this project that deals with the training pipeline was originally tailored for a mid-range personal workstation. (Meaning a small amount of relatively fast CPU threads, with a low-range GPU) This affected some of the decisions when it came to how the interfacing with the neural nets during training should be done, i.e. giving each thread access to a copy of the neural net on the CPU, essentially only using the GPU for training between epochs. (This is also less of a big deal when the models are small)

However, for larger models, or larger variants of chess, it would be very useful to better utilise the potential of GPUs, when available. This would mean implementing prediction-batching through virtual loss. This would greatly reduce training time, making it possible to experiment more with different architectures for different goals.

I am currently working on this, as it would be very useful to be able to test such architectures more rapidly.

When I run "python -m minichess.rl.fast_mcts configs/base.json" (without the quotes), I get a FileNotFoundError.

FileNotFoundError: [Errno 2] No such file or directory: '/explainable-minichess/minichess/chess/magics/5x4/straights.npz'.

I am trying to reproduce your training setup for 5x4, though I'd eventually like to get 8x8 running as well.

It is somewhat cluttersome to have these in the main repo. Could be moved to another repo, and referenced as a submodule?