MobileNet V3

MobileNet-v3 is the result of an explicit neural architecture search layered on top of the inverted-residual block of v2. A platform-aware NAS (MnasNet-style) selects the per-stage expansion ratio, kernel size, and channel count to minimise a latency-and-accuracy joint objective on a real mobile CPU, after which a NetAdapt pass fine-tunes the channel counts of each layer individually under a hard latency budget. The output is two hand-tuned variants — Large and Small — targeting different parts of the latency/accuracy frontier.

Two architectural refinements give v3 its accuracy gain over v2. First, a lightweight squeeze-and-excitation module is inserted inside selected bottleneck blocks; it pools the spatial dimensions and re-weights the channel responses, providing a cheap channel-attention mechanism. Second, a new activation function called hard-swish,

replaces swish/SiLU in the deeper half of the network. Hard-swish matches swish closely in shape but uses only piecewise-linear primitives, making it dramatically cheaper to evaluate on mobile-friendly fixed-point hardware.

A further trick is rewriting the expensive final layers of v2. The $1 \times 1$ expansion before the global average pool is moved after the pool, where it operates on a single spatial location, saving roughly $10\%$ of total latency. Combined, these changes give MobileNet-v3-Large $3.2\%$ higher ImageNet top-1 accuracy than v2 at $20\%$ lower latency on a Pixel-1 CPU.