TorchVision is the official computer vision library of the PyTorch ecosystem, providing the datasets, model architectures, pre-trained weights and image transformations that vision projects are built from. If you train an image classifier, detector or video model in PyTorch, chances are TorchVision supplies the backbone and the data pipeline. Maintained since 2016, it has earned about 18,000 stars on GitHub and is released in lockstep with PyTorch itself.
The package covers the essentials of a vision workflow: popular datasets with automatic download and preparation, reference implementations of architectures from ResNet, Faster R-CNN and RetinaNet to Vision Transformers and modern ConvNeXt-style models, along with pre-trained weights for transfer learning. Its transforms API handles everything from classic augmentation to newer v2 transforms that work consistently across images, bounding boxes, masks and video. TorchVision also includes utilities for detection, segmentation, optical flow, video classification and keypoint detection, plus C++ and TorchScript support for deployment, making it the default foundation under countless research papers and production vision systems.
TorchVision is completely free and open source under the BSD-3-Clause license. Some individual pre-trained models and datasets carry their own licenses, which the documentation flags where relevant.
TorchVision is for PyTorch developers and researchers building computer vision systems, from quick transfer-learning prototypes to production detection and segmentation pipelines.