LoginSign Up
← Back to all sparks
T

torchvision

ANALYTICS
Velocity0.0

R interface to torchvision. Contribute to mlverse/torchvision development by creating an account on GitHub.

github.com/mlverse/torchvision ↗

R's torchvision is porting PyTorch's vision stack one task at a time — instance segmentation just landed.

computer-visionr-languageinstance-segmentationpytorch-paritydataset-loaders
◆Current state
torchvision for R has moved past being a thin tensor-transform helper into a task-complete vision library. The last three releases added dataset loaders by the dozen, then face detection and recognition, and now Mask R-CNN for instance segmentation. The 0.9.0 release also splits the COCO detection loader from a new segmentation loader, cutting memory use roughly in half for detection-only work.
◆Where it's heading
The pattern is a deliberate walk through PyTorch's torchvision feature matrix: datasets first, then model architectures, then the visualization and transform utilities that make each task usable end to end. Each release breaks a little API to align R naming with upstream PyTorch conventions — `$categories` became `$classes`, `coco_classes()` now matches the 90-class sparse PyTorch layout. Community contributors are doing most of the volume, with maintainers arbitrating the API shape.
◆Prediction
Expect the next release to fill in the remaining segmentation and detection model families and continue aligning class and label handling with upstream PyTorch, given that every release so far has paired new models with a matching dataset loader.

◆Recent moves

  1. 3mo ago

    Mask R-CNN brings instance segmentation to R torchvision

    ⚡ SPARK

    The release that adds a whole task rather than more coverage of an existing one: Mask R-CNN models, dedicated segmentation datasets, and mask transform functions arrive together. It also pays down API debt, splitting the COCO loader and renaming attributes to match upstream PyTorch.

    View source ↗
  2. 9mo ago

    Face detection models and 35 RoboFlow datasets land

    A wide coverage release on the same porting arc: MTCNN and FaceNet for face detection and recognition, MaxViT and MobileNet V3 architectures, plus six thematic RoboFlow 100 dataset collections totalling 35 datasets. Broad rather than directional — it deepens the catalog the 0.9.0 task work then builds on.

    View source ↗
  3. 1y ago

    Fashion-MNIST, COCO, and a dozen more dataset loaders

    The dataset-first phase of the port: EuroSAT, QMNIST, EMNIST, FGVC-Aircraft, COCO detection and captions, Caltech 101/256, Flowers102, Flickr8k/30k and Oxford-IIIT Pet all gain loaders. Necessary groundwork — models are only useful once the standard benchmark data is one function call away.

    View source ↗