This repository contains the source code for a tool that finds optimal weights for a 3 node neural network architecture.

The training data is drawn from the unit-hypercube, and the labels are assigned to be binary.

We use PySMT with the z3 solver to find both a CNF formula and a satisfying assignment for the model.

The tool records the following quantities using tensorboard:

• Formula size (in bytes)

• Time (s) to extract formula and assignment

• Accuracy of assigned weights on the training data

• Comparative accuracy of gradient descent weights on the same data

• Time (s) to find a solution on the same data with gradient descent

• pytorch
• pysmt -- with z3 solver
• numpy
• tensorboard
• python3
• probably ubuntu

First clone the repository to the home directory (or anywhere you prefer)

git clone https://github.com/neale/pysatdnn . && cd pysatdnn

Its helpful to instantiate a virtual environment, to avoid cluttering the system

python 3 -m venv pysat_venv

Activate the virtual environment to start from a clean slate

source ./pysat_venv/bin/activate

Install the required packages for PySATDNN

pip3 install -r requirements.txt

If there is an error, you might need to update pip pip3 install --upgrade pip

Now that you have the required packages, we depend on the z3 solver for pysmt

pysmt-install --z3 [y]

Now you're free to check out the source code, or run the test suite with ./start_all.sh, which will loop over data with 2 to 32 dimensions, on training set sizes 100 to 100k

If you get the error NoSolverAvailable: No Solver is Available, then do the following:

Run pysmt-install --check

The output should show that z3 is installed as a solver. If not, then install as shown above.

If it is shown as installed, then the PYTHONPATH is likely pointing somewhere else.

Run pysmt-install --env to obtain a string that you can run to temporarily update the PYTHONPATH to the solver's location