This was my final project for CS6114, Software-Defined Networking at Cornell with Nate Foster. I implemented the linear perceptron described in the paper "Binarized Neural Networks" (Hubara et al., 2016) in the networking devices programming language P4.
The main files are:
perceptron.py: creates the initial packet and passes it to the switch over and over until thepkt.finishedbit is set. Then prints out the answer from the switch, and whether or not it matches the answer given by numpy.perceptron.p4: implements the perceptron in P4. Most of the hard work is done in the ingress control.
To run this, you need to have Petr4 and mininet installed. Run make controller in one window and make run in another. In the second window, type h1 python perceptron.py to generate random matrices and do the perceptron computation on them. It will print out to let you know if the switch got the same answer as calculated using numpy/linear algebra (it should be the same!). You can run the progam multiple times to see that it works with different randomly-generated values.
For more information, see report.pdf.
