Depthwise Separable Convolution: MobileNet's Efficiency Trick

Every convolutional neural network for computer vision spends most of its compute budget sliding filters across feature maps. This simulator renders that sliding-window mechanic in 3D and puts a real efficiency trick next to it: MobileNet's depthwise-separable convolution, which factors a standard K×K×C_in→C_out filter into a cheap per-channel depthwise pass and a 1×1 pointwise pass that mixes channels. Change the kernel size and the input/output channel counts and watch the two cost towers — standard convolution versus depthwise-plus-pointwise — update from the exact multiply-accumulate and parameter-count formulas used in the MobileNet paper, with a live speedup readout showing exactly how much cheaper the factored version is.