This repository contains the code that was used to conduct the analyses presented in the following workshop paper at NIPS 2016, in which we investigate how recursive and recurrent artificial neural networks can process hierarchical compositionality. We packaged the code, such that it is easy to experiment with it and the results are easily reproducible.
Installing the package
The processing_arithmetics package uses the following dependencies:
- numpy, scipy, sklearn
Furthermore, the package uses an extended version of Keras, that can be found here to which a few classes and metrics are added. You can install this version of keras by cloning the repository and then doing an editable install via pip:
git clone https://github.com/dieuwkehupkes/keras pip install -e /path/to/keras
This version of keras is frequently updated to match the master branch of the original keras repository and should behave identically (aside from the added functionality). If you already have your own copy of the keras repository on your system, consider adding the keras repository required for the processing_arithmetics package as a remote to your local installation.
You can install the processing_arithmetics package by cloning the repository and doing an editable install via:
git clone https://github.com/dieuwkehupkes/processing_arithmetics pip install -e /path/to/processing_arithmetics
Although the test suite of the package does not cover all functionality, it can be used to confirm if the most important methods still work.
To run the test suite, you will need to install
pip install pytest.
Then simply run:
The scripts folder contains a few scripts that can be used to train different kind of models.