Paper Reading: MobileNets: Effective Convolutional Neural Networks for Mobile Vision Applications

MobileNet v1
MobileNets: Effective Convolutional Neural Networks for Mobile Vision Applications

Section 1 & 2

To build small and efficient neural network

• compressing pretrained networks
• train small networks directily

MobileNet is a class of network that works on restricted resources like latency and size.

Section 3

Depthwise Separable Convolution

Let’s assume that we have a input data $F$ of shape $D_F\times D_F\times M$, a expected output data $G$ of shape $D_G\times D_G\times N$, and a convolution kernel $K$ of shape $D_K\times D_K\times M$.
For a standard convolution, to achieve the result, we have $N$ kernels names $K_1,K_2,...,K_N$, and the computation cost is:
$$\boxed{D_K\times D_K\times M}\times \boxed{D_G\times D_G}\times N$$Since $D_G\le D_F$, the maximum computation cost is:
$$\boxed{D_K\times D_K\times M}\times \boxed{D_F\times D_F}\times N$$

name	shape
input $F$	$D_F\times D_F\times M$
output $G$	$D_G\times D_G\times N$
kernel $K_1,...,K_N$	$\boxed{D_K\times D_K\times M}\times N$
maximum computation cost	$\boxed{D_K\times D_K\times M}\times \boxed{D_F\times D_F}\times N$

For depthwise separable convolution, the kernel is separated into 2 parts: $M$ depthwise convolution kernel of shape $D_K\times D_K\times 1$ and $N$ pointwise convolution kernels of shape $1\times 1\times M$. Thus, the computation cost is:
$$\boxed{D_K\times D_K\times 1}\times \boxed{D_G\times D_G}\times M+\boxed{1\times 1\times M}\times \boxed{D_G\times D_G}\times N$$Since $D_G\le D_F$, the maxinum computation is:
$$\boxed{D_K\times D_K\times 1}\times \boxed{D_F\times D_F}\times M+\boxed{1\times 1\times M}\times \boxed{D_F\times D_F}\times N$$
which is:
$$D_K\times D_K\times D_F\times D_F\times M+D_F\times D_F\times N\times M$$

name	shape
input $F$	$D_F\times D_F\times M$
output $G$	$D_G\times D_G\times N$
kernel $K$	$D_K\times D_K\times M$
pointwise convolution	$\boxed{1\times 1\times M}\times N$
maximum computation cost — depthwise part	$\boxed{D_K\times D_K\times 1}\times \boxed{D_F\times D_F}\times M$
maximum computation cost — pointwise part	$\boxed{1\times 1\times M}\times \boxed{D_F\times D_F}\times N$

Network Structure

Shrinking hyper-parameters

1. Width multiplier $\alpha$
   This hyper-parameter changes the input channel number to $M^{\prime }=\alpha M$ and output channel number $N^{\prime }=\alpha N$. Thus, the maximum computation cost becomes:
   $$D_K\times D_K\times D_F\times D_F\times \alpha M+D_F\times D_F\times \alpha N\times \alpha M$$which is:
   $$\alpha \times D_K\times D_K\times D_F\times D_F\times M+{\alpha }^2\times D_F\times D_F\times N\times M$$
2. Resolution multiplier $\rho$
   This hyper-parameter changes the input resolution to $\rho D_F\times \rho D_F$and output resolution to $\rho D_G\times \rho D_G$, and maximum compution cost becomes:
   $$\rho D_K\times \rho D_K\times \rho D_F\times \rho D_F\times M+\rho D_F\times \rho D_F\times N\times M$$which is:
   $${\rho }^4\times D_K\times D_K\times D_F\times D_F\times M+{\rho }^2\times D_F\times D_F\times N\times M$$
3. When combine $\alpha$ and $\rho$ together, weget the final maximum conputation cost as follows:
   $$\rho D_K\times \rho D_K\times \rho D_F\times \rho D_F\times \alpha M+\rho D_F\times \rho D_F\times \alpha N\times \alpha M$$where $0<\alpha \le 1,0<\rho \le 1$.

Section 4 & 5

From the shown results in Table 4, the MobileNet do reduce a great number of parameters with a little accuracy sacrificed. And the hyper-parameters $\alpha$ and $\rho$ performs well when larger than 0.5.
According the paper, MobileNet also perform well in other tasks.