/COMET

Counterexample-Guided Learning of Monotonic Networks

Primary LanguagePythonMIT LicenseMIT

COMET

Counterexample-guided techniques to provably enforce and train a neural network with monotonicity constraints.


Installation     |     Onboarding a Neural Network     |     Monotonicity Verification  ·  Monotonic Envelope Predictions  ·  Training with Monotonicity Constraints     |     Paper  ·  License (MIT)

Installation

  1. Install python >= 3.8.3, Install pip >= 20.1.1

Using Python Virtual Environment

  1. Setup a Virtual Environment pip3 install virtualenv && python3 -m venv comet && source comet/bin/activate

  2. Install Packages pip install matplotlib pandas pillow tensorflow scikit-learn sexpdata tensorflow

  3. Install Solvers pip install z3-solver Download optimathsat (http://optimathsat.disi.unitn.it/pages/download-js.html) and add it to your environment path. Check the installation by running optimathsat -version

Onboarding a Neural Network

To run on example networks skip ahead. To onboard your neural network

  1. You need a trained baseline ReLU neural network model (model.h5), weights, and bias as csv files.

  2. cp templates/template.txt configurations/<dataset_name>.txt and configure COMET for your dataset.

  3. cp templates/DeepModel_Template.py src/Models/DeepModel_<dataset_name>.py and fill in the code blocks (#TODO: CODE BLOCK) in make_data function for COMET to interact with your model.

Monotonicity Verification

Once you have set up your neural network, you can verify if it is monotonic using:

python src/COMET.py configurations/<config_file>.txt --mode verifier.

This could take a while depending on the size of the network.

Monotonic Envelope Predictions

If you are interested in using the monotonic envelope for predictions, set up the configuration file, fill in the code blocks in DeepModel_<dataset_name>.py and run:

python src/COMET.py configurations/<config_file>.txt --mode envelope --test_file <path to test data csv>

Training with Monotonicity Constraints

To carry out counterexample-guided training, set up the configuration file, fill in the code blocks in DeepModel_<dataset_name>.py and run:

python src/COMET.py configurations/<config_file>.txt --mode learner

CLICK for an example

Training Auto-MPG dataset Monotonicity Constraints

  1. You can find pre-trained model, test/train.csv and weight/bias files in examples/Auto-MPG/

  2. You can find the configured file in configurations/auto-mpg.txt

  3. You can find the DeepModel_AutoMPG.py in src/Models/

  4. The output after running python src/COMET.py configurations/auto-mpg.txt --mode verifier is:

CLICK for output
  1. The output after running python src/COMET.py configurations/auto-mpg.txt --mode envelope --test_file ./examples/Auto-MPG/test.csv is:
CLICK for output
  1. The output after running python src/COMET.py configurations/auto-mpg.txt --mode learner is:
CLICK for output