This repository is an official pytorch implementation of the paper - Systematic Generalization for Prediction Control in MultiVariate Time Series (arxiv).
- python 3
- pytorch >= 1.1
- numpy
- matplotlib
You can use anaconda (python library manager) or pip. We used anaconda, and downloaded all the dependencies in an anaconda environment.
We trained our models using High Performance Computing (HPC) facility provided by IIT Delhi.
Source code to run our experiments are present in src.
master_script.sh in both narma and pmsm folders would indicate actual runs of the experiments for 10 random seeds through run.py. This script differs slightly for these datasets as certain experiments (such as intervention testing (noise sensitivity here)) are exclusive to those datasets.
Below we present the raw time series files along with their data generating sources. The files are used to create temporal segments that are created on-the-fly while running the code.
In the paper, we describe four NARMA scenarios.
You can find the raw data to train and test (systematic generalization testing) the models in data/narma/6.
narma_2d_i_n0.csv contains the data for ith Scenario.
TimeSynthCombGen-spurious_corr.zip in data/narma/6 contains the data generating scripts, which have been taken from TimeSynth repository.
Raw data corresponding to PMSM described in the paper is present in data/pmsm/.
GEM_PMSM.ipynb is the data-generating notebook which is taken from GEM repository.
As this paper is a result of a lot of experiments and architectural tweaks, you might observe that the codes contain a lot more material than presented in the paper. We shall either remove those pieces with time, or add relevant explanations for such pieces in the README file.