diagnnose ·

This library contains a set of modules that can be used to analyse recurrent neural networks. In particular, it contains functionality to:

• Extracting activations from different types of (language) models
• Running diagnostic classifiers [1] on extracted activations
• Doing intervention studies [2] with language models
• Analysing word embeddings
• Doing dimensionality reduction and plotting state space trajectories of trained models

Quickstart

Our library is not (yet) officially registered with pip. You can use the library by cloning it and do an editable install with pip:

git clone git@github.com:i-machine-think/diagnnose.git $custom_path
pip3 install -e $custom_path

We will shortly update this README with explanations for the different scripts provided in the library.

Requirements

This library runs with Pytorch 1.0. We refer to the PyTorch website to install the right version for your environment. The preferred version of python is >=3.7.

To install additional requirements, run:

pip install -r requirements.txt