meta-pytorch/captum is a library dealing with the interpretability and understanding of PyTorch models. It sits on the side of explaining why a model produced a given prediction. The latest release recorded in the repository is v0.9.0. However, the provided record contains no changelog for that version, so what was added or fixed cannot be confirmed.
The README states that Captum helps users grasp how features, training examples, and concepts contribute to model predictions through Integrated Gradients, TCAV, TracIn influence functions, and more. It covers three axes: features of the input, training examples, and concepts. Contribution is the notion of how much each element was involved in the prediction. Adversarial attacks and minimal input perturbation are supported as well.
The README also says it helps with benchmarking against existing algorithms and that it integrates easily with models built with domain-specific libraries such as torchvision or torchtext. Domain-specific libraries refer to handling models tailored to a particular field. It is written in Python and distributed under the BSD-3-Clause license. The repository was created on August 27, 2019, and has 5,712 stars.
However, the published record contains no performance figures or benchmark results. The momentum score is recorded only as 0.0, while the freshness, technology, and quality scores are blank. The kinds of interpretability algorithms and the license terms can be confirmed, but whether useful interpretations come out of a particular model is hard to judge from this record.