si4attention

This package provides a statistical test for attention in Transformers for images and time series. The tequnical details are described in the paper "Statistical Test for Attention in Transformers for Images and Time Series"

Installation & Requirements

This package has the following dependencies:

• Python (version 3.10 or higher, we use 3.12.5)
  • torch (version 2.5.0 or higher, we use 2.5.0)
  • sicore (version 2.2.0 or higher, we use 2.2.0)
  • tqdm (version 4.66.5 or higher, we use 4.67.1)

To install this package, please run the following commands (dependencies will be installed automatically):

$ pip install si4attention

Usage

This package provides a main function test_attention_map that allows you to conduct a statistical test for attention in Transformers. For API details and usage examples, please refer to the docstring of the function and the example notebook in the demonstration.ipynb file.