Why RegNet Is the Most Flexible Architecture for Computer Vision

RegNet is a scalable and highly efficient model design space for computer vision.

Traditional convolutional neural network architectures are optimized for specific goals: ResNet series for top ImageNet accuracy, MobileNets for mobile inference, and EfficientNet for visual recognition.

Radosavovic et al. (2020) set a goal to explore and design a highly flexible network architecture that can adapt to high efficiency or mobile deployment. Flexibility is achieved by controlling network width and depth through a quantized linear function with a set of parameters.

AnyNet Design Space

The authors first define a universal design space called AnyNet , which enumerates all possible models from different combinations of parameters such as width and depth. Every model in AnyNet is trained with a consistent training regime (epochs, optimizer, weight decay, learning‑rate scheduler) on the ImageNet dataset.

Analysis of which parameters yield the best performance leads to an iterative simplification of AnyNet into a compact RegNet design space. Improvements include sharing bottleneck ratio and group width, and parameterizing width and depth to increase in later stages.

RegNet Parameterization

The final RegNet design space contains only models that satisfy the quantized linear function. To instantiate a RegNet model, one must set parameters d (depth), w0 (initial width), wa (slope), wm (width multiplier), b (bottleneck ratio), and g (group width).

RegNet uses a single type of block: a standard residual bottleneck with grouped convolution. The network consists of a stem, a body made of multiple stages (each stage contains several identical‑width blocks), and a head.

Block width is computed by a quantized linear function: uj = w0 + wa·j where j starts at 0 and runs to the network depth. After introducing the additional parameter wm, an intermediate value is calculated: sj = (uj / w0)^{1/wm} The final block width is obtained by rounding sj:

width = round(sj)

Blocks with the same width are grouped into a stage, so each stage has a uniform width.

RegNet Variants

By setting different values for the parameters, the authors create RegNet variants optimized for:

Mobile deployment

High efficiency

High accuracy

Empirical Results

At equal FLOPs, RegNet variants outperform other mobile‑optimized networks. Compared with EfficientNet, RegNetX‑F8000 is about five times faster while delivering comparable accuracy. When configured for high precision, RegNet matches or exceeds the performance of ResNet and ResNe(X)t.

RegNet is not a single architecture but a design space defined by quantized linear functions, offering a highly flexible way to balance accuracy, efficiency, and latency for computer‑vision tasks.

Reference: Radosavovic, Ilija, et al. “Designing Network Design Spaces.” Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. https://openaccess.thecvf.com/content_CVPR_2020/papers/Radosavovic_Designing_Network_Design_Spaces_CVPR_2020_paper.pdf