Captum
captum.ai
- Free plan
- Windows
- macOS
- Linux
- Self-hosted
7.2 Club score Visit captum.ai
Plans
Every plan and priceCaptum
Free
Free plan
What the maker publishes
- Algorithms and metrics
- Beyond attribution methods, Captum includes infidelity and sensitivity metrics to estimate explanation trustworthiness. source
- Attribution algorithms
- Its algorithms help identify features contributing to a model’s output, including Integrated Gradients. source
- DeepLIFT limitation
- The documented DeepLIFT implementation supports the Rescale rule; RevealCancel is not supported in the described implementation. source
- Distribution
- The site documents installation using either conda or pip. source
- Maintainer
- The site footer identifies Meta Platforms, Inc. in its copyright notice. source
- Model compatibility caveat
- Some hook-based methods do not work with functional nonlinearities; JIT models do not support hook-dependent methods. source
- Parallel models
- The FAQ states that DataParallel and DistributedDataParallel models are supported with all model types. source
- PyTorch models
- Captum supports most types of PyTorch models and is intended to require minimal changes to the original neural network. source
- Research use
- Researchers can implement interpretability algorithms and benchmark them against methods in the library. source
- Supported modalities
- The product site says Captum supports interpretability across vision, text, and other modalities. source
- What it does
- Captum is an open-source library for interpreting machine-learning models built on PyTorch. source
- Who it is for
- The stated primary audiences are model developers and interpretability researchers; application engineers are also identified as users. source
Checked against the maker's pages on 14 Nov 2026.