Abstract Neural Networks

This repository contains code related to the SAS 2020 paper Abstract Neural Networks by Matthew Sotoudeh and Aditya V. Thakur.

Using

Dependencies

This code assumes reasonably up-to-date versions of python3, numpy, and pytest in order to run correctly. If you already have Python installed, you can install the other two dependencies like so:

python3 -m pip install -r requirements.txt

Usage

The file abstract.py exposes a method abstract_layer_wise which corresponds to Algorithm 3 in our paper. An example of its uses is provided in the file test_abstract_intervals.py, corresponding to Example 6 (Section 5.1) in our paper. To run it, and any other test cases, you can use:

python3 -m pytest *.py

in this directory.

Bazel Usage

Optionally, we support Bazel for reproducible runs and testing. After setting up bazel_python, you can run

bazel test //...

to run all test cases, then

bazel run coverage_report

to produce an HTML coverage report in a new htmlcov directory.

Citing

@inproceedings{anns:sas20,
    author = {Sotoudeh, Matthew and Thakur, Aditya V.},
    title = {Abstract Neural Networks},
    booktitle = {27th Static Analysis Symposium (SAS)},
    year = {2020},
    note = {To appear}
}