This is the code corresponding to the article "Learning Timescales in MTRNNs" by Fabien C. Y. Benureau, and Jun Tani, submitted for review. If the paper is published, this code will be released publicly.
Tested under Python 3.6.5.
First install PyTorch 0.4 (0.3 is not supported) by following the instructions on http://pytorch.org/, in accordance to your configuration.
To install other dependencies:
pip install -r requirements.txt
Then run:
python run_training.py
Computation takes roughly one hour at ~500GFLOPS.
Finally, to produce the figures, launch jupyter:
jupyter notebook figures.ipynb
The figures probably won't be the same as the article. We produced those figures
using multithreaded, non-deterministic code. The current code should be
deterministic, and use only one thread per computation (the run_training.py
code launches seven processes in parallel).
Some generic documentation, generated from the code comments, is available. But
the best is probably to go into the code. To compile the documentation,
sphinx and a few dependencies need to be installed:
pip install sphinx sphinx-autobuild sphinx-rtd-theme sphinx_autodoc_typehints
cd docs/
make html
Then open the ./build/html/index.html file.