Discit (lat.: it learns) is a set of tools that were developed for various learning tasks, from supervised to multi-agent reinforcement learning research.
It focuses on traceable tensor operations so that they can be significantly accelerated with CUDA graphs.
Start by downloading or cloning this repository.
Setup will check for the following packages:
numpy
: vectorised processing,torch >= 2.0
: main processing and AI integration with CUDA graphs,tensorboard
: tracking training progress.
It is strongly advised that PyTorch is installed separately beforehand. See the PyTorch instructions and ensure that the installation targets your CUDA device.
Finally, install Discit in editable/development mode from the main directory with:
pip install -e .
Discit is being developed and used as part of the following works: