Changelog

Release notes are published on GitHub Releases.

Unreleased

Dependency floors rise to PyTorch 2.13, optional torchvision 0.28, and Pillow 13. PyTorch macOS wheels require macOS 14+ on Apple Silicon.

PredictionExplanation.save() requires the resized/cropped model view at input dimensions. Bundles add lossless input_image and optional string context to schema 1. Unsupported image modes fail before writing.

Added EigenCAM for class-independent principal-component maps of spatial features, including batched and multi-layer extraction without gradients. Half-precision activations are computed in float32.

CAM fusion now rejects resizing a single map; resize it explicitly with torch.nn.functional.interpolate.

v0.5.0 (2026-10-03)

Release note: v0.5.0

RefineCAM, FinerCAM, LeGrad, Vision Transformer reshape transforms, callable output targets, deletion/insertion metrics, torchcam.explain, and demo improvements. NumPy 2 is supported; PyTorch 2.4.1+, Matplotlib 3.8.4+, NumPy 1.26.4+, and Pillow 9.3+ are now required. The PyTorch and Matplotlib minimums ensure NumPy 2 interoperability, including on Windows. Published PyTorch wheels for this range require Apple Silicon on macOS.

v0.4.1 (2025-10-27)

Release note: v0.4.1

Public enable_hooks() / disable_hooks() API (replaces private _hooks_enabled).

v0.4.0 (2023-10-19)

Release note: v0.4.0

Evaluation metrics and PyTorch 2.0 support.

v0.3.2 (2022-08-02)

Release note: v0.3.2

CAM fixes and batch processing support.

v0.3.1 (2021-10-31)

Release note: v0.3.1

Improved demo and package reorganization.

v0.3.0 (2021-10-31)

Release note: v0.3.0

Layer-CAM and multi-layer CAM computation.

v0.2.0 (2021-04-10)

Release note: v0.2.0

3D input support and documentation improvements.

v0.1.2 (2020-12-27)

Release note: v0.1.2

Automatic target-layer resolution and IS-CAM.

v0.1.1 (2020-08-03)

Release note: v0.1.1

SmoothGradCAM++, Score-CAM, and SS-CAM.

v0.1.0 (2020-03-24)

Release note: v0.1.0

Initial release.