PyTorch 1.x gets its deformable convolution transplant

What it does Implements Deformable Convolutional Networks v2 as custom PyTorch 1.x operators, adapted from the official Deformable-ConvNets repository. The project compiles native extensions via a build script and provides test scripts to verify gradients on both CPU and GPU.

The interesting bit The author debugged offset gradient failures until all checks passed in double precision, then openly documented a lingering backward reentrancy issue with numerical error small enough that the README literally asks whether it is even a serious problem.

Key highlights

• Adapted from the official msracver/Deformable-ConvNets DCNv2 operators
• Master branch targets PyTorch 1.x; legacy pytorch_0.4 branch available
• Includes separate CPU and CUDA test scripts for gradient verification
• All gradient checks pass with double precision
• Build process handled by a single make.sh script

Caveats

• Backward pass is not reentrant; the README notes a RuntimeError with drift under 1e-7 for float and 1e-15 for double
• Documentation is limited to build instructions and known issues

Verdict Useful if you need DCNv2 layers in a PyTorch 1.x pipeline and can tolerate an unofficial adaptation with a non-reentrant backward pass. Skip it if you require an officially maintained implementation.