What is this?
Torchvision is PyTorch's computer vision package: popular datasets, ready-made model architectures, and common image transformations for vision work.
Torchvision is PyTorch's computer vision package: popular datasets, ready-made model architectures, and common image transformations for vision work.Is it a standalone framework or part of PyTorch?It is a companion library, not a framework: it installs alongside PyTorch with matched versions (torchvision 0.24 pairs with torch 2.9) and builds on the torch tensor ecosystem.
It is a companion library, not a framework: it installs alongside PyTorch with matched versions (torchvision 0.24 pairs with torch 2.9) and builds on the torch tensor ecosystem.Do I need a specific Python or torch version?Yes — releases are pinned: each torchvision version maps to a torch version and a Python range, e.g. 0.24 goes with torch 2.9, and main/nightly builds require Python 3.10+.
Yes — releases are pinned: each torchvision version maps to a torch version and a Python range, e.g. 0.24 goes with torch 2.9, and main/nightly builds require Python 3.10+.Does torchvision host the datasets it gives me access to?No — the library only downloads and prepares public datasets. Checking that you are allowed to use a dataset under its license is entirely your responsibility.
No — the library only downloads and prepares public datasets. Checking that you are allowed to use a dataset under its license is entirely your responsibility.Is the license permissive, or are there catches?The code is BSD-3-Clause, but pre-trained models can carry their own terms from the data they were trained on — for example, SWAG weights are CC-BY-NC 4.0, non-commercial only.
The code is BSD-3-Clause, but pre-trained models can carry their own terms from the data they were trained on — for example, SWAG weights are CC-BY-NC 4.0, non-commercial only.What does a typical workflow with it look like?Load a public dataset, run images through the transform pipeline for preprocessing and augmentation, then train or fine-tune one of the included pre-trained model architectures.
Load a public dataset, run images through the transform pipeline for preprocessing and augmentation, then train or fine-tune one of the included pre-trained model architectures.What image formats can I feed it?It works with torch tensors and PIL images. For faster decoding you can swap Pillow for Pillow-SIMD, a drop-in replacement accelerated with SIMD.
It works with torch tensors and PIL images. For faster decoding you can swap Pillow for Pillow-SIMD, a drop-in replacement accelerated with SIMD.How do I cite it in a paper?The README includes a BibTeX entry for citing TorchVision in academic work, and the full API documentation is hosted on the PyTorch website.
The README includes a BibTeX entry for citing TorchVision in academic work, and the full API documentation is hosted on the PyTorch website.Is the project still actively developed?Yes — it tracks PyTorch releases with a published compatibility table (torchvision 0.24 for torch 2.9), ships nightly builds, and the repo has around 18k stars and is not archived.
Yes — it tracks PyTorch releases with a published compatibility table (torchvision 0.24 for torch 2.9), ships nightly builds, and the repo has around 18k stars and is not archived.What are similar projects?pointdit, albumentations, facenet, techniques, SolarWM, lama, ControlNet, yolov3
pointdit, albumentations, facenet, techniques, SolarWM, lama, ControlNet, yolov3Which topics is it in?augmentation, model-training, datasets, image-classification, neural-networks, pretrained-models, dataset, image-processing, bsd, cpp, computer-vision, pytorch
augmentation, model-training, datasets, image-classification, neural-networks, pretrained-models, dataset, image-processing, bsd, cpp, computer-vision, pytorch