pyNeVer 0.0.1a1

Package for the training, pruning and verification of neural networks.

pyNeVer

Neural networks Verifier (NeVer) is a tool for the training, pruning and verification of neural networks. At present it supports sequential fully connected neural networks with ReLU and Sigmoid activation functions. pyNeVer is the corresponding python package providing all the main capabilities of the NeVer tool and can be easily installed using pip. The PyPI project page can be found at https://pypi.org/project/pyNeVer/ whereas the github repository can be found at https://github.com/darioguidotti/pyNeVer.

REQUIREMENTS AND INSTALLATION

Given the characteristcs of PyTorch and ONNX we were not able to setup an auto-installation for these packages. Therefore the user is required to install the torch, torchvision and onnx packages indipendently. Guides on how to install such packages can be found at:

• https://pytorch.org/get-started/locally/
• https://github.com/onnx/onnx

After the installation of the required packages pyNeVer can be installed using the command:

• pip install pynever

DOCUMENTATION

The documentation related to the pyNeVer package can be found in the directory docs/pynever/ as html files.

SUPPORTED INPUTS

At present the pyNeVer package supports only the abstraction and verification of fully connected neural networks with ReLU and Sigmoid activation functions. The training, pruning and conversion supports also batch normalization layers. A network with batchnorm layers following fully connected layers can be converted to a "pure" fully connected neural networks using the capabilities provided in the utilities.py module.
The conversion.py provides the capabilities for the conversion of PyTorch and ONNX networks: therefore these kind of networks can be loaded using the respective frameworks and then converted to the internal representation used by pyNeVer. Examples of the loading of ONNX networks can be found in the scripts mnist_experiment.py and james_experiments.py.
At present the properties for the verification and abstraction of the networks must be defined in python code following the specification which can be found in the documentation. Examples of the specification of the properties can be found in all the scripts in the directory examples/submissions/ATVA2021/.

EXAMPLES

NB: All the scripts should be executed INSIDE the related directory!

• The directory examples/ contains some examples of application of the pyNeVer package. In particular the jupyter notebook shows a graphical example of the application of the abstraction module for the reachability of a small network with bi-dimensional input and outputs.

• The pruning_example.py script show how to train and prune some small fully connected neural networks with relu activation function. It also show how it is possible to combine batch norm layer and fully connected layers to make the networks compliant with the requirements of the verification and abstraction modules.

• The directory examples/submissions/ATVA2021 contains the experimental setup used for the experimental evaluation in our ATVA2021 paper plus a novel experiment. The experiments can be easily replicated by executing the python scripts acas_experiment.py and mnist_experiment.py from within the ATVA2021/ directory. The log files will be generated and will be saved in the logs/ directory.
  It is also possible to test the package on the script james_experiments.py which consider small fully connected neural networks with Sigmoid activation functions.