RegNetY

RegNetY is a family of convolutional neural networks developed by Facebook AI Research (FAIR) as part of the RegNet design space. It adheres to a simple, regular design philosophy intended to make high-performance networks easier to design, tune, and scale across different compute budgets. The Y variant is one of the RegNet configurations that emphasizes regularity and efficiency in its block design and network progression.

Architecturally, RegNetY networks are built from repeating bottleneck blocks arranged in a multi-stage pyramid that downscales spatial resolution while increasing channel capacity. Each block uses a bottleneck structure with a sequence of convolutions and employs grouped convolutions to improve parameter efficiency. Width, depth, and the group width follow a regular progression across stages, enabling predictable scaling of accuracy with compute. The architecture favors a uniform, modular block type across stages, which supports stable training and hardware efficiency.

RegNetY configurations span a range of compute budgets, from compact models suitable for real-time or mobile applications to larger models aimed at higher accuracy on large-scale datasets. The family is designed so that accuracy scales predictably as model size grows, making it straightforward to select a model that fits a given resource constraint. RegNetY has been used as a backbone for image classification and is also adopted in larger computer vision pipelines, including object detection and segmentation systems, where modularity and efficiency are valued.

In practice, RegNetY models are released with pretrained weights in common deep learning frameworks and have been cited as reliable backbones in both research and industry settings. Related variants in the RegNet family, such as RegNetX, illustrate the broader design philosophy of regular, scalable network architectures.