MobileNet

MobileNet is built around the depthwise separable convolution, a factorisation that decomposes a standard convolution into two much cheaper stages: a per-channel depthwise $3 \times 3$ spatial filter followed by a $1 \times 1$ pointwise convolution that mixes channels. For an input of $M$ channels and an output of $N$ channels with kernel size $D_K$, the cost drops from $D_K \cdot D_K \cdot M \cdot N \cdot D_F \cdot D_F$ to

a reduction of roughly $1/N + 1/D_K^2$. For $D_K = 3$ and the typical $N$ of a few hundred, that is an 8–9× FLOPs saving with only a small accuracy loss on ImageNet.

Two scalar hyper-parameters expose a smooth accuracy–latency trade-off curve at inference time without retraining the whole family. The width multiplier $\alpha \in (0, 1]$ uniformly thins every layer (input and output channels become $\alpha M$ and $\alpha N$), giving four standard variants at $\alpha \in \{1.0, 0.75, 0.5, 0.25\}$. The resolution multiplier $\rho$ rescales the input spatial size — it does not change the model itself, only the feature maps flowing through it, and is therefore not enforced inside the network.

Together these two knobs let MobileNet-v1 cover a wide spectrum of mobile/embedded budgets from sub-mW edge devices up to desktop-class deployment, while keeping a single training recipe and a single MobileNetV1Config schema.