Captum

What is Captum?

Captum is a PyTorch library for model interpretability. It provides tools and techniques to understand and attribute the predictions of PyTorch models across various modalities, including vision and text. Captum supports most PyTorch models and allows for easy implementation and benchmarking of new interpretability algorithms.

How to use Captum?

Install Captum via conda or pip. Import necessary libraries like numpy, torch, and IntegratedGradients from captum.attr. Define and prepare your PyTorch model. Instantiate an interpretability algorithm (e.g., IntegratedGradients). Apply the algorithm to your input data and baseline to obtain attributions and convergence delta.

Captum’s Core Features

Multi-modal interpretability support Built on PyTorch Extensible and open source

Captum’s Use Cases

• Understanding feature importance in image classification models
• Analyzing the influence of words in text classification models