
- Monthly visits
- 7,789
- 2026-09
- Monthly growth +7.5%
Captum - Interpret PyTorch models across vision, text, and more
Captum is an open-source library for interpreting PyTorch models. It provides algorithms that attribute model outputs to inputs across modalities including vision and text. Developers can install it with conda or pip, apply methods such as Integrated Gradients, and extend the library to implement and benchmark interpretability algorithms.
- LLM Developer Tools
- LLM Evaluation
- PyTorch model interpretability
- neural network attribution
- Integrated Gradients

Capabilities
Model input attribution
Attribute a model's output to its inputs and return attribution values. Supply a PyTorch model, input tensors, and, when applicable, baseline tensors and a target; select an algorithm such as Integrated Gradients.
Multi-modal interpretation
Apply interpretability methods to models across supported modalities, including vision and text. Use Captum with supported PyTorch models; the library is designed to work with minimal modification to the original neural network.
Extensible algorithm library
Implement and benchmark new interpretability algorithms using an open-source, generic library. Researchers and developers can extend Captum rather than being limited to only the algorithms already included.
PyTorch model support
Use interpretability methods with most types of PyTorch models. Provide a prepared PyTorch neural network and selected method; the library can be used with minimal changes to the original model.
Pricing
Captum is an open-source library, and the supplied official pages list no paid plans or billing terms. There are no listed plan prices, quotas, or plan-based features to compare.
Use cases
Attribute a model prediction to its inputs
When examining a PyTorch model's output, calculate which input values contribute to a selected result. Prepare the model and input tensors, optionally specify a baseline and target, then apply an attribution method such as Integrated Gradients to obtain attribution values.
Implement and benchmark an interpretability algorithm
When an existing method does not cover an interpretability experiment, extend the generic library with a new algorithm. Captum is open source and designed to let users implement and benchmark interpretability algorithms for PyTorch models.
Who it is for
Developers working with PyTorch neural network models
Researchers implementing or benchmarking model interpretability algorithms
FAQ
Which machine learning framework does Captum support?
Captum supports PyTorch models. The homepage says it supports most types of PyTorch models and can be used with minimal modification to the original neural network.
Which model modalities can Captum interpret?
Captum supports interpretability across modalities including vision and text, as well as more. The supplied pages do not enumerate every supported modality.
How can I install Captum?
Install Captum with conda or pip. The homepage recommends conda and provides the commands conda install captum -c pytorch and pip install captum.
Can I add interpretability algorithms to Captum?
Yes. Captum is described as an extensible, open-source generic library for interpretability research, and users can implement and benchmark new algorithms.
Comments