An inside-out convolution operator for vision backbones

What it does

The repo implements involution, a neural primitive that rethinks the standard convolution operator. It provides a memory-efficient CuPy-based implementation and a full model zoo—dubbed RedNet—covering image classification, object detection, instance segmentation, and semantic segmentation. Everything is packaged as drop-in modules for the OpenMMLab ecosystem.

The interesting bit

Instead of sharing kernels across spatial locations like convolution, or computing pairwise attention like transformers, involution generates location-specific kernels. The authors claim this hits a middle ground: leaner than self-attention in form, yet more efficient and effective than standard convolution. The model zoo tables suggest the trade-off can pay off, showing sizable parameter and FLOP reductions on ImageNet and mixed but often positive accuracy shifts on COCO.

Key highlights

• Ships with pre-trained RedNet backbones (26 to 152 layers) and downstream task configs for OpenMMLab toolkits.
• Memory-efficient operator implementation relies on CuPy; the authors note a custom CUDA kernel could accelerate it further.
• Model zoo reports reduced parameters and FLOPs compared to convolution baselines across ImageNet, COCO, and Cityscapes.
• Checkpoints and training logs are provided, though the authors caution that reimplementation results may differ slightly from the paper.

Caveats

• The code is tightly coupled to OpenMMLab (mmcls, mmdet, mmseg), so adopting it means buying into that toolchain.
• Training the ImageNet models required 64 GPUs; COCO used 8 and Cityscapes used 4, suggesting serious hardware expectations.
• The current CuPy implementation is not fully hardware-optimized, and the README invites community contributions for a custom CUDA kernel.

Verdict

Worth a look if you are building vision backbones and want an off-the-shelf alternative to ResNet within the OpenMMLab universe. Skip it if you are outside the PyTorch/OpenMMLab ecosystem or looking for a framework-agnostic, standalone operator.