Regnets

RegNets are a family of convolutional neural network architectures introduced by Facebook AI Research to provide a regular, scalable design space for image recognition. The central idea is to replace opaque, highly tuned networks with architectures built from simple, regular building blocks whose width and depth follow a predictable schedule. By constraining the growth of channel counts and the configuration of residual blocks, RegNets become highly scalable and easy to compare.

In practice, RegNets are organized into stages. Each stage contains a fixed number of residual blocks, and the width (the number of channels) grows according to a fixed rule across stages. The design uses a bottleneck-style block and often employs grouped convolutions to reduce computational cost. Two main families, RegNetX and RegNetY, differ in the specific block configuration but share the overarching philosophy of regular, regularized expansion of width. A small set of hyperparameters controls the model: initial width, width growth rate, depth per stage, and the bottleneck ratio or group width.

RegNets have demonstrated strong performance on ImageNet and other vision benchmarks, achieving competitive accuracy with lower architectural search costs than many neural architecture search approaches. Their regular structure also facilitates scaling to different compute budgets, from mobile-friendly models to high-capacity backbones used in detection and segmentation systems.

Overall, RegNets aim to systematize CNN design by enforcing regular growth in width and uniform block patterns, yielding performant models across a range of sizes without relying on expensive search methods.