Reproduce “RepNeXt: A Fast Multi-Scale CNN using Structural Reparameterization” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Mingshu Zhao, Yi Luo, Yong Ouyang

RepNeXt introduces a multi-scale CNN backbone that utilizes structural reparameterization to deliver high accuracy and low latency for mobile vision tasks. The model demonstrates competitive or superior performance on ImageNet classification (e.g., RepNeXt-M2 achieving 80.1% Top-1 accuracy at 1.00ms latency on iPhone 12), object detection, and semantic segmentation benchmarks.

Key insights

  • Multi-scale depthwise convolutions, specifically the "Chunk Convolution" design, effectively combine local details and global context within an efficient CNN framework.
  • Structural reparameterization significantly boosts model representational capacity during training without increasing inference latency on target hardware.
  • A meticulously designed, simple multi-scale CNN backbone can achieve competitive or superior accuracy-latency trade-offs for mobile vision tasks, often outperforming more intricate hybrid or NAS-optimized models.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.