/implicit-flow-networks

Official code for the paper "Physics-Informed Implicit Representations of Equilibrium Network Flows"

Primary LanguagePythonApache License 2.0Apache-2.0

Implicit Flow Networks

This repository is the official implementation of Implicit Flow Networks, from the paper "Physics-Informed Implicit Representations of Equilibrium Network Flows", to appear in NeurIPS 2022.

Command Line Usage

Use the following command to train and test IFN:

python transductive_flow_estimation.py <dataset_name> <network_name> <model_name>

The script performs the following steps:

  1. randomly splits the edges of the specified network into train and test edges,
  2. trains a new IFN model on the train edges,
  3. computes the RMSE loss from the test edges, and
  4. saves all of the relevant data and results to a pickle file in /results/TransductiveFlowEstimation/.

Arguments

dataset_name

Name of the dataset from which to generate experiments. Currently implemented datasets are power (AC power flows) and water (municipal water network flows).

network_name

Name of a particular network within the dataset. Options for power are case57, case118, case145, case300, case_ACTIVSg200, and case_ACTIVSg500. Options for water are bellingham, fairfield, Net3, northpenn, and oberlin.

model_name

IFN model to train and test. Options are ifn (IFN using edge weights directly from the dataset) and ifn+loglinweight (IFN with a log-linear model to learn edge weights from edge attributes).

Additional arguments

Hyperparameter settings and other options can be provided via optional flags:

Flag Description Default
--frac-labeled Fraction of edges to include in training set 0.1
--hidden-size Hidden layer size for inverse flow function model 128
--dmin Min slope of the flow function (NOT its inverse) 0.5
--dmax Max slope of the flow function (NOT its inverse) 2.0
--mitr Max number of iterations for forward / backward pass 100
--eps Convergence tolerance for forward / backward pass 1e-2
--lr Learn rate 1e-2
--epochs Max number of training epochs 300
--early-stop Number of epochs for early stopping 10
--dev-name PyTorch device "cuda"
--seed Random seed 1234
--output-fname Name of results file None

Examples

python transductive_flow_estimation.py water fairfield ifn --dmin 0.2 --dmax 20.0 --eps 1e-4 --epochs 2000

python transductive_flow_estimation.py power case300 ifn+loglinweight --dmin 0.5 --dmax 2.0

Reference

@inproceedings{smith2022ifn,
    author={Smith, Kevin D. and Seccamonte, Francesco and Swami, Ananthram and Bullo, Francesco},
    booktitle={Advances in Neural Information Processing Systems},
    title={Physics-Informed Implicit Representations of Equilibrium Network Flows},
    year={2022}
}