This library contains a set of modules that can be used to analyse recurrent neural networks. In particular, it contains functionality for:
- Extracting activations from different types of (language) models
- Running diagnostic classifiers on extracted activations
- Analysing word embeddings
- Performing contextual decomposition (Murdoch et al., 2018) on a model
- Running a broad linguistic downstream task suite
Our library is officially registered with pip and can be installed by running pip install diagnnose.
We will shortly update this README with explanations for the different scripts provided in the library.
This library runs with Pytorch >=1.1. The preferred version of Python is >=3.7.
- Jumelet, Zuidema & Hupkes (2019): Analysing Neural Language Models: Contextual Decomposition Reveals Default Reasoning in Number and Gender Assignment
If you intend on using diagnnose for your research, please cite us as (and feel free to reach out, we'd love to help!):
@article{diagnnose,
title={diagnnose: A Neural Net Analysis Library},
DOI={10.5281/zenodo.3445477},
publisher={Zenodo},
author={Jaap Jumelet and Dieuwke Hupkes},
year={2019},
}