PyTorch-Encoding

Why we included this project

For people building or reproducing semantic segmentation models, this toolkit packages the encoding modules from several well-known papers, including Context Encoding, Deep TEN, and the ResNeSt split-attention backbone. That means you can pull those building blocks into your own experiments without reimplementing them. It also includes synchronized batch normalization and pretrained weights for ImageNet classification and ADE20K/Pascal Context segmentation, so you don't have to train from scratch or wire up checkpoints yourself. The documentation covers training scripts and recipes, which is handy when you want to extend an existing architecture rather than learn a new framework.