phantran/beamng-neuron-coverage

This command line tool can be used to collect image dataset from BeamNG gameplay and supports calculating neuron coverage of a DNN model. It can be extended to create a neuron coverage guided testing tool to generate test cases for simulation-based testing of DNN-based self-driving vehicles

BEAMNG NEURON COVERAGE

Introduction

The main purpose of this source code is to demonstrate the correlation between image transformations and neuron coverage. A model is trained on generated BeamNG dataset for the simple task of keeping vehicles on the road. A part of the dataset is then transformed (rotate, adjust brightness, etc...) and input to the trained model. Neuron coverage and output from the trained model on both original and transformed image data will be recorded and summarize to verify the correlation.

Setup

Clone config.yml.sample file and rename it to config.yml. Modify the configurations corresponding to your local setup.

Running

python run.py --task TASK_TO_PERFORM
Replace TASK_TO_PERFORM with one of the following options: collecting, data_cleaning, training, evaluating