# meta-pytorch/captum

Model interpretability and understanding for PyTorch

Repository: https://github.com/meta-pytorch/captum
Canonical: https://ross.abutalabs.com/products/captum
Homepage: https://captum.ai
Language: Python
License: BSD-3-Clause
License Family: permissive
Topics: interpretability, interpretable-ai, interpretable-ml, feature-importance, feature-attribution
Last push: 2026-08-22T02:50:58+00:00

## Health v2 (maintenance only)
Score: 81/100 (v2, computed 2026-09-02T17:46:02.011165+00:00)
- activity 99, release rhythm 48, longevity 100
- inputs: {"age_days": 2563, "days_push": 11, "days_rel": 138, "gap_med": 386, "n_releases_24m": 2}
- flags: none
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 5693, forks 563 (observed 2026-08-28T04:09:27.996833+00:00)

## What it is
Captum is a model interpretability and understanding library for PyTorch, providing implementations of algorithms like Integrated Gradients, saliency maps, SmoothGrad, TCAV, and TracIn. It helps researchers and developers understand which features, examples, or concepts contribute to a model's predictions, and includes adversarial attack and counterfactual explanation capabilities.

## Use cases
- explain why a PyTorch model made a prediction
- compute feature attributions for a neural network
- generate saliency maps for image classifiers
- identify which training examples influenced a model's output
- debug unexpected model outputs by finding important features
- run adversarial attacks to test model robustness
- benchmark new interpretability algorithms against existing ones

## When to choose
- you use PyTorch and need explainability for vision, text, or multimodal models
- you want state-of-the-art attribution methods like Integrated Gradients with minimal code changes
- you are an interpretability researcher needing an extensible benchmarking framework

## When to avoid
- your models are built with TensorFlow, JAX, or scikit-learn rather than PyTorch
- you need interpretability for non-neural models like gradient-boosted trees
- you only need simple model metrics rather than attribution or explanation methods

## Facets
- artifact type: library
- maturity: active
- function: machine-learning, data-science, developer-tools
- domain: machine-learning, deep-learning, artificial-intelligence, data-science
- platform: python, cross-platform
- tags: interpretability, explainable-ai, feature-attribution, integrated-gradients, saliency-maps, pytorch, model-debugging, adversarial-attacks, gpu

## Member repositories
- meta-pytorch/captum (main) score 81

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:09:27.996833+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-29T17:53:38.476730+00:00, confidence not recorded.
  - readme: https://github.com/meta-pytorch/captum (fetched 2026-08-28T04:09:27.996833+00:00, sha 695db3c03720)
  - homepage: https://captum.ai (fetched 2026-08-29T08:49:17.755871+00:00, sha 21ec3b5b35b2)
  - site_page: https://captum.ai/docs/introduction (fetched 2026-08-29T08:49:17.758546+00:00, sha e9eb13f3c1fd)
  - site_page: https://captum.ai/docs/introduction.html (fetched 2026-08-29T08:49:17.760313+00:00, sha e9eb13f3c1fd)
  - site_page: https://captum.ai/docs/getting_started (fetched 2026-08-29T08:49:17.762135+00:00, sha 81162cca8e2e)
- Data as of 2026-08-30T08:39:29.467469+00:00.