Fig. 3

Overview of macro-architectural and micro-architectural design decisions leading to a simple, efficient, and easily customizable neural network architecture. (A) Schematic of the macro-architectural design process illustrating the implementations of hierarchical and isotropic neural networks. Red indicates input images, yellow indicates intro blocks, green indicates downsampling blocks, blue indicates upsampling blocks, purple indicates outro blocks, and gray indicates residual blocks. H: input height; W: input width; I: input channels; M: middle channels; O: output channels; N: residual blocks. (B) Schematic of the micro-architectural design process illustrating the implementations of the MobileNetV3 inverted bottleneck residual block and subsequent modifications. Small arrows indicate the flow of information through the macro- and micro-architectures, while large arrows indicate the flow of the design process