TorchVision

TorchVision is a package within the PyTorch project that provides utilities for computer vision. It offers standardized datasets, pre-trained model architectures, and image-transformation tools that help researchers and developers rapidly prototype and evaluate CV models. TorchVision is maintained by the PyTorch team with contributions from the broader community and is released under an open-source license.

Datasets: The torchvision.datasets module provides data loaders for a variety of vision datasets, including ImageFolder (a directory-based dataset), CIFAR-10 and CIFAR-100, MNIST, and COCO formats. These loaders integrate with PyTorch's DataLoader, support common train/validation splits, and facilitate downloading and preparation of data.

Models: The torchvision.models namespace offers pretrained and standard architectures for image classification, object detection, and segmentation. Examples include ResNet, VGG, DenseNet, and MobileNet for classification; Faster R-CNN, RetinaNet, and Mask R-CNN for detection; and DeepLab variants for segmentation. Weights can be loaded from a built-in repository with a consistent API.

Transforms and I/O: The torchvision.transforms module provides image preprocessing and augmentation tools such as Resize, RandomCrop, Normalize, and ToTensor, usable in a composed transform pipeline. A functional API is available for lower-level operations. The torchvision.io module handles image and video I/O, complementing other components.

Usage and ecosystem: TorchVision is designed to work seamlessly with PyTorch and complements components like torch.utils.data.DataLoader and CUDA. It is commonly used in research and industry to quickly assemble end-to-end CV pipelines. Installation is via pip or conda, with version alignment to the installed PyTorch version.