OpenSBT MNIST - Fashion-MNIST

This repository integrates the MNIST and Fashion-MNIST case study into OpenSBT and is based on the implementation in DeepJanus.

Preliminaries

Follow the installation instruction of OpenSBT in the main repository. Install MNIST related dependencies by following the instructions here.

Problem Definition

The mutation of digits is performed by selection of control points/vertices of the corresponding svg model of the digit and performing a displacment in x and y direction of that point. The index of the point and extent represent the search parameters of the search based testing problem. Right now, two configs are supported

• 3D Problem. (diplacement_1, displacement_2, index_vertex). Displaces vertex with given index and the following vertex with provided displacement.
• 6D Problem. (diplacement_1, displacement_2, displacement_3, diplacement_4, index_vertex_1, index_vertex_2). Displaces vertex_1 and vertex_2 with provided displacement.

There are multiple fitness functions implemented integrating static image properties (num turns, saturation, etc.).

Run example for MNIST

Example OpenSBT experiments can be found in default_experiments.py. To run an experiment via console use syntax from OpenSBT:

python run.py -e 1000

To see an example experiment definition without flag based execution check out run_mnist.py.

Run example for Fashion-MNIST

Example OpenSBT experiments can be found in default_experiments_fmnist.py. To run an experiment via console use syntax from OpenSBT:

python run.py -e 201

To see an example experiment definition without flag based execution check out run_fmnist.py.

Authors

Lev Sorokin ([email protected])