tx2 1.3.0

Transformer eXplainability and eXploration

TX2

Welcome to TX2! This library is intended to aid in the explorability and explainability of transformer classification networks, or transformer language models with sequence classification heads. The basic function of this library is to take a trained transformer and test/train dataset and produce an ipywidget dashboard as seen in the screenshot below, which can be displayed in a jupyter notebook or jupyter lab.

NOTE: Currently this library's implementation is partially torch-dependent, and so will not work with tensorflow/keras models - we hope to address this limitation in the future!

Installation

You can install this package from pypi:

pip install tx2

NOTE: depending on the environment, it may be better to install some of the dependencies separately before pip installing tx2, e.g. in conda:

conda install pytorch pandas scikit-learn=1.1 numpy=1.22 -c conda-forge

Examples

Example jupyter notebooks demonstrating and testing the usage of this library can be found in the examples folder.

Note that these notebooks can take a while to run the first time, especially if a GPU is not in use.

Packages you'll need to install for the notebooks to work (in addition to the conda installs above):

pip install tqdm transformers==4.21 datasets==2.4

Running through each full notebook will produce the ipywidget dashboard near the end.

The tests in this repository do not depend on transformers, so raw library functionality can be tested by running pytest in the project root.

Documentation

The documentation can be viewed at https://ornl.github.io/tx2/.

The documentation can also be built from scratch with sphinx as needed.

Install all required dependencies:

pip install -r requirements.txt

Build documentation:

cd docs
make html

The docs/build/html folder will now contain an index.html

Two notebooks demonstrating the dashboard and how to use TX2 are included in the examples folder, highlighting the default and custom approaches as discussed in the Basic Usage page of the documentation.

Citation

To cite usage of TX2 in a publication, the DOI for this code is https://doi.org/10.21105/joss.03652

bibtex:

@article{Martindale2021,
  doi = {10.21105/joss.03652},
  url = {https://doi.org/10.21105/joss.03652},
  year = {2021},
  publisher = {The Open Journal},
  volume = {6},
  number = {68},
  pages = {3652},
  author = {Nathan Martindale and Scott L. Stewart},
  title = {TX$^2$: Transformer eXplainability and eXploration},
  journal = {Journal of Open Source Software}
}