TorchVision

TorchVision is PyTorch's official library for computer vision, providing pretrained models (ResNet, Faster R-CNN, ViT, and others), common datasets, and image/video transform utilities designed to plug directly into PyTorch training loops. It is maintained by the PyTorch team as part of the core ecosystem, not a third-party wrapper.

Most projects use it for two things: loading pretrained backbones for transfer learning instead of training from scratch, and standardized data augmentation pipelines (resize, crop, normalize) that match what those pretrained models expect. It also includes reference implementations for classification, detection, and segmentation that serve as a practical starting point rather than a from-scratch build.

Being free, open source, and version-locked to PyTorch releases, TorchVision is the default choice for any team already training vision models in PyTorch; teams on TensorFlow would instead reach for TF-native equivalents like Keras Applications or TF Model Garden.