Back to Gap Map Infrastructure / Classic ML & computer vision

torchvision

open source / Overall score: 4.2(strong)

PyTorch's computer-vision library, with popular datasets, image and video transforms, vision operators, and model architectures with pretrained weights for classification, semantic and instance segmentation, object detection, keypoint detection, video classification and optical flow. Reference training scripts reproduce each model. It is part of the PyTorch project under the PyTorch Foundation.

Openness

5 high confidence

5.0

license: BSD-3-Clause(OSI
source: public(github.com/pytorch/vision)
core features withheld: no — PyTorch Foundation project

The library is BSD-3-Clause and builds from the public repository. Its pretrained weights can carry their own terms from the data they were trained on, and the SWAG weights are non-commercial; they are optional downloads, not the library.

https://docs.pytorch.org/docs/2.14/community/governance.html recorded 2026-09-27
PyTorch governance: the project "clearly separates technical governance from business governance"; "Technical governance is strictly separated from business governance."
https://raw.githubusercontent.com/pytorch/vision/main/LICENSE recorded 2026-09-27
LICENSE is the "BSD 3-Clause License", "Copyright (c) Soumith Chintala 2016".
https://raw.githubusercontent.com/pytorch/vision/main/README.md recorded 2026-09-27
README: "The torchvision package consists of popular datasets, model architectures, and common image transformations for computer vision."; "The pre-trained models provided in this library may have their own licenses or terms and conditions derived from the dataset used for training"; "SWAG models are released under the CC-BY-NC 4.0 license."
https://ungh.cc/repos/pytorch/vision/files/main recorded 2026-09-27
Repository tree of 707 paths with no ee, enterprise, commercial or pro directory; training scripts under references/.

Adoption

5 medium confidence

5.0

Measured on monthly PyPI downloads of torchvision. Many of them come from installing it alongside PyTorch rather than from choosing it, so the figure overstates deliberate use.

https://pypistats.org/api/packages/torchvision/recent recorded 2026-09-27
last_month 15,814,091 downloads of torchvision

Capability

3 high confidence

3.0

torchvision is a model zoo with the datasets, transforms and training recipes to fine-tune it, like timm, whose catalog of image backbones is larger. Its pretrained models predict the classes of the benchmarks they were trained on, so a new task still means a training run.

https://docs.pytorch.org/vision/stable/models.html recorded 2026-09-27
Models page: "TorchVision offers pre-trained weights for every provided architecture"; sections for classification, quantized models, semantic segmentation, object detection, instance segmentation, keypoint detection, video classification and optical flow.

Verified 2026-09-27