SRU_for_GCI

This repository contains the python code for our ICLR 2020 paper: Economy Statistical Recurrent Units For Inferring Nonlinear Granger Causality.

Dependencies

• Python 3.7
• CUDA 10.1
• PyTorch 1.0.1
• NumPy, SciPy

Example usage:

Model: SRU Dataset: Lorenz-96, T=250, F=10, dataset id = 1

• python main.py --dataset lorenz --dsid 1 --model sru --n 10 --T 250 --F 10 --nepochs 2000 --mu1 0.021544 --mu2 0.031623 --mu3 0 --lr 0.001 --joblog crossval/logs/test.npz

Model: eSRU_2LF (Economy SRU with two-layer MLP as feedback's second stage) Dataset: VAR, T=250, F=10, dataset id = 1

• python main.py --dataset var --dsid 1 --model eSRU_2LF --n 10 --T 500 --F 30 --nepochs 2000 --mu1 0.021544 --mu2 0.031623 --mu3 0.464159 --lr 0.001 --joblog crossval/logs/test.npz

Model: eSRU_2LF (Economy SRU with two-layer MLP as feedback's second stage) Dataset: NetSim (BOLD signals), T=200, dataset id = 1

• python main.py --dataset netsim --dsid 1 --model eSRU_2LF --n 15 --T 200 --F 0 --nepochs 2000 --mu1 0.021544 --mu2 0.031623 --mu3 0.464159 --lr 0.001 --joblog crossval/logs/test.npz

Model: eSRU_1LF (Economy SRU with single layer MLP as feedback's second stage) Dataset: Dream-3 (Yeast1), T=966

• python main.py --dataset gene --dsid 1 --model eSRU_1LF --n 100 --T 966 --F 0 --nepochs 2000 --mu1 0.021544 --mu2 0.031623 --mu3 0.464159 --lr 0.001 --joblog crossval/logs/test.npz

Input arguments

Argument	Type	Description
dataset	string	lorenz/var/netsim/gene
dsid	int	dataset id (Range: 1-5)
model	string	sru (Standard SRU) eSRU_1LF (Economy SRU with feedback's second stage implemented as single layer MLP) eSRU_2LF (Economy SRU with feedback's second stage implemented as two layer MLP) <\p>
n	int	No. of timeseries/Nodes in the Granger causal graph (n = 10 for lorenz/var, n = 15 for netsim, n = 100 for gene)
T	int	Length of input timeseries (T = 250/500/1000 for lorenz, T =500/1000 for var, T = 200 for netsim, T = 966 for gene)
F	int	External forcing in lorenz model/ percentage sparsity in var model (F = 10/40 for lorenz, F = 30 for var, F = 0 for gene/netsim)
nepochs	int	No. of training epochs
mu1	float	Bias for ridge regularization of all unregularized weights in the model
mu2	float	Bias for block sparse regularization of input layer weights
mu3	float	Bias for group sparse regularization of output feature layer weights in Economy SRU
lr	float	Learning rate for BPTT step in the proximal gradient descent algorithm

Output

• The adjacency matrix of the estimated Granger causal graph is printed as the output once the run is completed.
• A summary of model configuration and final results is stored in 'test.npz' located at: SRU_for_GCI_standalone\crossval\logs

Acknowledgements

• The Dream-3 datasets for the gene interaction network inference experiments are taken from "http://dreamchallenges.org/project-list/dream3-2008/".
• The BOLD-FMRI (NetSim) timeseries data used for the brain connectome inference experiments are taken from "https://www.fmrib.ox.ac.uk/datasets/netsim/index.html"