Classic segmentation architectures, frozen in PyTorch 0.2 amber

What it does Implements a laundry list of classic semantic segmentation architectures—FCN variants built on VGG, ResNet, and DenseNet, plus U-Net, SegNet, PSPNet, GCN, and DUC/HDC—and wraps them in shared training and evaluation scaffolding. It is essentially a survey paper that actually runs.

The interesting bit The repo’s value is partly archaeological. Because it targets PyTorch 0.2.0, it preserves the exact idioms and tensor gymnastics used to express these networks in one of PyTorch’s earliest releases.

Key highlights

• Six model families, including FCN32/16/8 with three encoder backbones.
• U-Net, SegNet, PSPNet, GCN, and DUC/HDC all in one place.
• Unified training and testing pipeline around the implementations.
• Roadmap openly admits DeepLab v3, RefineNet, and ADE dataset support are still missing.

Caveats

• Locked to PyTorch 0.2.0, so modern environments will likely complain.
• README documentation is minimal; the author notes you should “find what you miss when running the code.”

Verdict Useful if you are studying how classic segmentation architectures were expressed in early PyTorch, but anyone building production pipelines today will want something more current.