This repository provides code to run experiments from,
Learning Procedural Abstractions and Evaluating Discrete Latent Temporal Structure
Karan Goel and Emma Brunskill
ICLR 2019
Specifically, this repository reproduces experiments with the new evaluation criteria proposed in the paper. The results are reproduced from logged runs of the methods being compared, since running them from scratch can be very slow.
(1) Install the requirements into your virtual environment using pip install -r requirements.txt.
(2) Unzip logs.zip and datasets.zip which should yield logs/ and datasets/.
(3) Then, reproduce results on the bees_0 dataset by running
> python evaluate.py --dataset bees_0
which will reproduce all the plots and visualizations from the first sequence of the Bees dataset in plots/.
Run python evaluate.py to see a list of the available datasets.
Please contact Karan Goel kgoel <at> cs <dot> stanford <dot> edu for questions!