TorchMorph: CUDA-accelerated Morphological Transforms
TorchMorph ports SciPy-style morphology into CUDA kernels that can stay inside PyTorch training loops.
The extension exposes 22 operators for binary and greyscale morphology, distance transforms, and entropy-regularised optimal transport on CUDA tensors with up to eight spatial dimensions. Its API mirrors SciPy’s morphology arguments so existing code can be moved over largely by changing the import. The paper reports up to 1.1e3 times SciPy throughput for greyscale morphology, up to 350x for exact Euclidean distance transforms, and up to 42x over POT for Sinkhorn. HF Daily Papers' note
The extension exposes 22 operators for binary and greyscale morphology, distance transforms, and entropy-regularised optimal transport on CUDA tensors with up to eight spatial dimensions. Its API mirrors SciPy’s morphology arguments so existing code can be moved over largely by changing the import. The paper reports up to 1.1e3 times SciPy throughput for greyscale morphology, up to 350x for exact Euclidean distance transforms, and up to 42x over POT for Sinkhorn. HF Daily Papers' note
score 5