torchvision Documentation

What torchvision is

torchvision is the computer vision library that sits alongside PyTorch. In its own words it provides popular datasets, model architectures and common image transformations for computer vision. These docs are its official reference, and they are where you check which pretrained weights exist, what preprocessing they expect and how each transform behaves.

What the docs cover

• Transforms: transforming images, videos, bounding boxes and masks, with the newer v2 API that applies the same random crop or flip to an image and its labels together.
• TVTensors: tensor types such as Image, Video, BoundingBoxes, KeyPoints and Mask, which let transforms know what they are working on.
• Models and pretrained weights: classification, semantic segmentation, object detection, instance segmentation, video classification and optical flow.
• Datasets: built-in datasets and base classes for your own.
• Utils: drawing boxes, masks and keypoints for checking results.
• Operators: detection and segmentation operations such as non-maximum suppression, box conversions and IoU.
• Feature extraction: tools to pull intermediate features out of a model.

Is it worth reading?

Yes, whenever you fine-tune a vision model in PyTorch. The most common silent bug in transfer learning is preprocessing that does not match the pretrained weights. torchvision attaches the right transforms to each weights object, and the docs show how to use them. The transforms v2 pages save a lot of pain when you augment detection or segmentation data, because boxes and masks move with the image.

How to use it

Keep it open next to your code rather than reading it end to end. Start with the transforms v2 getting started example, then the models page for the architecture you want, and the operators page when you need NMS or box maths.

Prerequisites

PyTorch basics: tensors, datasets and data loaders, and a training loop you understand.

Free and paid

Free and open source under the BSD 3-Clause licence. Individual pretrained weights come from datasets with their own terms, which the model pages note.

Common questions

v1 or v2 transforms? New code should use v2, which handles boxes, masks and videos as well as images.

Best for: anyone training or fine-tuning image models in PyTorch.