EfficientNet is a procedural method to designing 👁️ Convolutional Neural Network architectures; specifically, while landmark models like VGG or ResNeXt scaled depth and width, EfficientNet proposes to scale depth, width, and input resolution with a constant ratio.
The key observation is that we can scale depth , width , and resolution proportional to their effect on the network’s FLOPS—, , and respectively. If we let control our computational budget (with FLOPS proportional to ), the proposed compound scaling method sets:
are constants that can be found by grid search on the base network for some temporarily fixed . Once they’re set, we can easily scale up the hyperparameters by increasing to meet our budget.