{"full_name":"meta-pytorch/captum","owner":"meta-pytorch","name":"captum","shard":"1","url":"https://github.com/meta-pytorch/captum","description":"Model interpretability and understanding for PyTorch","homepage":"https://captum.ai","topics":["interpretability","interpretable-ai","interpretable-ml","feature-importance","feature-attribution"],"tag_pages":[],"language":"Python","license":"BSD-3-Clause","avatar":"https://avatars.githubusercontent.com/u/107212512?v=4","created_at":"2019-08-27T15:34:41Z","pushed_at":"2026-11-13T10:13:13Z","metrics":{"stars":5721,"forks":576,"watchers":274,"issues":588,"prs":1371,"releases":11,"issues_open":39,"issues_closed":549,"prs_open":58,"prs_closed":1155,"prs_merged":158,"discussions":0},"growth":{"day":{"stars":-1,"forks":0,"base_day":"2026-11-15"}},"history":{"dates":["2026-11-10","2026-11-11","2026-11-12","2026-11-13","2026-11-14","2026-11-15","2026-11-16"],"stars":[5721,5721,5722,5722,5722,5722,5721],"forks":[574,574,574,574,575,576,576],"watchers":[null,274,274,274,274,274,274],"issues_open":[null,36,36,36,38,39,39],"prs_open":[null,56,56,56,57,58,58]},"boards":{},"same_owner":[{"full_name":"meta-pytorch/torchtune","name":"torchtune","shard":"1","boards":{}}],"issues":[{"number":572,"title":"CCA/CKA/OPD support","comments":15,"created_at":"2020-12-24T20:35:04Z"},{"number":657,"title":"DeepLift RevealCancel Rule Support","comments":5,"created_at":"2021-04-27T13:12:25Z"},{"number":350,"title":"Support for Score-CAM","comments":5,"created_at":"2020-04-14T03:36:33Z"},{"number":1284,"title":"Adding IDGI to Captum","comments":4,"created_at":"2024-05-24T09:33:26Z"},{"number":933,"title":"Supporting example-specific baselines","comments":4,"created_at":"2022-05-04T23:07:10Z"},{"number":359,"title":"evaluation metric","comments":4,"created_at":"2020-04-24T15:24:44Z"},{"number":1101,"title":"need to implement the CKA function for model interpretability and understanding","comments":3,"created_at":"2023-01-12T09:51:49Z"},{"number":1103,"title":"Allow for unnormalized attribution heatmap","comments":2,"created_at":"2023-01-19T15:03:58Z"}],"discussions":[],"latest_release":{"name":"Captum v0.9.0 Release","tag":"v0.9.0","published_at":"2026-04-17T06:36:54Z","body":"The v0.9.0 release of Captum **adds NumPy 2.x support**, retires the long-deprecated Captum Insights, introduces new multimodal image-mask attribution primitives, adds remote LLM attribution via vLLM, and promotes cross-tensor attribution to a default-on feature for perturbation-based methods. This release also raises minimum supported Python and PyTorch versions to `3.10` and `1.13`, respectively.\r\n\r\nUpgrading from v0.8 should be drop-in for most users. The exceptions to watch for: any code that imports from `captum.insights` or installs the `[insights]` extra (removed — see below), environments on Python < 3.10 or PyTorch < 1.13 (no longer supported), and any code relying on the previous default behavior of `FeatureAblation` / `FeaturePermutation` treating input tensors independently (cross-tensor attribution is now default-on).\r\n\r\n## LLM & Multimodal Attribution:\r\n\r\n### Multimodal Image-Segment Attribution\r\n\r\n<img width=\"1800\" height=\"743\" alt=\"Screenshot 2026-04-16 at 11 35 25 PM\" src=\"https://github.com/user-attachments/assets/85aee6a8-34b4-445c-b0dd-27d5caeca6fe\" />\r\n\r\nThis release adds a new [interpretable input](https://captum.ai/api/utilities.html#interpretable-input) type for image-segment attribution, [ImageMaskInput](https://github.com/meta-pytorch/captum/blob/master/captum/attr/_utils/interpretable_input.py#L509), useful for vision and vision-language models where features correspond to user-defined image regions is now exposed from `captum.attr` and is compatible with perturbation-based attribution methods and with `LLMAttribution`.\r\n\r\nKey functionality:\r\n- Support for `mask_list` to describe multiple segment sets with overlapped pixels (PR [#1749](https://github.com/meta-pytorch/captum/pull/1749))\r\n- Mask-segmentation visualization utilities, with legends on the overlay `plot_mask_overlay` (PRs [#1752](https://github.com/meta-pytorch/captum/pull/1752), [#1753](https://github.com/meta-pytorch/captum/pull/1753), [#1758](https://github.com/meta-pytorch/captum/pull/1758), [#1756](https://github.com/meta-pytorch/captum/pull/1756))\r\n- New plot helper `plot_image_heatmap` for rendering pixel-level attributions on an `ImageMaskInput` (PR [#1739](https://github.com/meta-pytorch/captum/pull/1739), [#1684](https://github.com/meta-pytorch/captum/pull/1684))\r\n\r\nA new tutorial, [`Multimodal_Image_Segment_Attribution.ipynb`](https://captum.ai/tutorials/Multimodal_Image_Segment_Attribution), demonstrates end-to-end usage leveraging [SAM](https://ai.meta.com/research/sam2/) (Meta Segment Anything Model) to interpret a multimodal LLM (Gemma-4) (PR [#1811](https://github.com/meta-pytorch/captum/pull/1811), [#1812](https://github.com/meta-pytorch/captum/pull/1812)).\r\n\r\n### LLM Attribution Improvements\r\n\r\nBuilding on the LLM attribution work from v0.7 and v0.8, this release substantially expands what `LLMAttribution` can attribute over and how attributions are consumed.\r\n\r\n- New `RemoteLLMAttribution` wrapper and `VLLMProvider` allow perturbation-based attribution to be computed against a remotely hosted LLM serving endpoint, removing the requirement that the model fit in the client process (PR [#1544](https://github.com/meta-pytorch/captum/pull/1544)). This enables attribution over much larger models than was previously feasible.\r\n- New boolean `forward_in_tokens` argument to `LLMAttribution.attribute`, to choose between replicating token-by-token output decoding or directly forwarding the output sequence in one-shot (PRs [#1740](https://github.com/meta-pytorch/captum/pull/1740), [#1741](https://github.com/meta-pytorch/captum/pull/1741), [#1742](https://github.com/meta-pytorch/captum/pull/1742), [#1744](https://github.com/meta-pytorch/captum/pull/1744))\r\n- Dict-like `model_input` support (PR [#1698](https://github.com/meta-pytorch/captum/pull/1698))\r\n- `skip_tokens` is no longer accepted as a target argument, and target encoding no longer adds special tokens (PRs [#1685](https://github.com/meta-pytorch/captum/pull/1685), [#1686](ht","is_prerelease":false},"readme":true,"readme_zh":false,"news":[],"stale":false,"snapshot_at":"2026-11-16T09:23:14Z","history_start":"2026-09-06"}