Graph Neural Solver applied to a Linear System

DISCLAIMER : I am currently modifying and refactoring this repo, so there are many things that won't work perfectly.

This repo provides some code for:

• Building a dataset of random linear systems
• Building and training a Graph Neural Solver that solves these linear systems
• Reloading a previous model, and perform inference
• A visualization notebook

The Graph Neural Solver algorithm has been introduced in Graph Neural Solver for Power Systems and Neural Networks for Power Flow : Graph Neural Solver. It relies on Graph Neural Networks. More info about this work can be found here.

Installation

Firstly, I recommend that you create a virtual environment. (You need to have virtualenv installed for this step, see https://virtualenv.pypa.io/en/latest/)

cd ./GraphNeuralSolver_LinearSystem/
virtualenv -p python3 ENV

Then you can activate it

source ENV/bin/activate

All the instructions below assume that the virtual environment is activated.

Now install the requirements. If you have a GPU (which you have setup to work smoothly with Tensorflow 1.14), then use requirements-gpu.txt instead of requirements.txt

pip install -r requirements.txt

Building the dataset

Instead of providing you with a dataset, you can find in data/ the piece of code I used to generate a set of random linear systems. To generate a dataset:

cd ./data
python build_dataset.py

If you wish to modify the generated data, you can modify the parameters N_NODES, N_EDGES, etc.

Now go back to the previous directory

cd ..

Training a model

Here is a typical command to start a training

python main.py --max_iter=1000 --minibatch_size=100

If you wish to have more informations about some args, try:

python main.py --help

Monitoring a training process

First of all, there is a logger that keeps track of the training and validation losses. Moreover you can use tensorboard to visualize the loss of the different layers of the Graph Neural Solver.

tensorboard --logdir=./

Hyperparameter search

To perform a grid search, with for instance a learning rate in [1e-3, 1e-2], and a latent dimension in [5, 10]:

python grid_search.py --learning_rate 1e-3 1e-2 --latent_dimension 5 10

Once again, more info about all args are available using:

python grid_search.py --help

Reloading a trained model

There is a trained model provided with the code : you just need to unzip the file that keeps track of the training loss:

unzip results/1575991681/train.zip

Visualization notebook

Take a look at the visualization.ipynb notebook.

ENV/bin/jupyter notebook

Infos and Contact

Check out my blog bdonon.github.io for more info. If you have any questions or observations, contact me at balthazar.donon@rte-france.com