ResNeSt

Why we included this project

ResNeSt reworks the familiar ResNet backbone by adding split-attention blocks, and the repo ships pretrained checkpoints in four depths, from 50 to 269 layers, that load into PyTorch or Gluon in a few lines. Teams building object detectors or segmentation pipelines get a drop-in swap: the backbone already plugs into Detectron2, MMDetection, the PyTorch Encoding toolkit, and GluonCV, so existing training code keeps working. It is a research release rather than maintained software, so plan on adapting code and rerunning the reproducibility scripts yourself. For vision work where vanilla ResNet accuracy is the bottleneck, the pretrained weights and downstream configs make it a practical upgrade.