Regnet

RegNet refers to a family of convolutional neural networks designed for image classification. Introduced by researchers at Facebook AI Research (FAIR) in 2020, RegNet emphasizes a regular, scalable design space that aims to simplify architecture search and improve training efficiency.

The design principle centers on parameterizing a small set of architectural choices, including width (the number of channels), depth (the number of blocks), and the progression of widths across stages. Blocks are formed from regularized bottlenecks using simple, repeatable patterns and, in some variants, grouped convolutions. The resulting networks are built from a sequence of stages with consistent downsampling and channel-growth rules, which promotes stable training and efficient hardware mapping.

Two widely cited variants are RegNetX and RegNetY. RegNetX emphasizes regularized convolutional blocks with standard convolutions and groupings, while RegNetY augments the blocks with squeeze-and-excitation (SE) attention to improve accuracy on visual tasks.

In practice, RegNets offer competitive image-classification performance across a range of model sizes and compute budgets, and their regular design makes them favorable backbones for broader vision systems, including object detection and segmentation when integrated into modern frameworks. They are implemented and pretrained in several deep learning libraries, enabling researchers and practitioners to scale up or down with predictable behavior.

RegNet contributes to the discourse around model scaling by showing that a carefully regularized design space can yield strong accuracy with straightforward engineering, reducing the need for large, bespoke search processes.