MobileNet

Depthwise Separable Convolution

Suppose we have a input image of size $[n,n,n_c]$ on which we apply $n_c$ filters of size $[f,f]$ to get a output image of size $[n-f+1,n-f+1,n_c]$. This output image then put through point-wise convolution, that is we use $n_c^{\prime }$ filters of size $[1,1,n_c]$ to get a output image of size $[n-f+1,n-f+1,n_c^{\prime }]$.

Now this could have been done directly also by straight away using $n_c^{\prime }$ filters of size $[f,f]$. But this would mean we would have to perform more computations. In general using a Depthwise Separable Convolution reduces the computation by a factor of

$$\frac{1}{n_c^{\prime }}+\frac{1}{f^2}$$

MobileNet

In the original MobileNet paper the Depthwise Separable layer is used 13 times. In the v2 of this paper a residual connection was used.

In the MobileNet v2 the following architecture is used :

• We apply a expansion layer in which we expand the input image using multiple 1 x 1 filters.
• Then a Depthwise layer is applied and then a Pointwise one.

This is used 17 times in the original paper.

Now let us look at Transfer Learning

For the code implementation of this network go to - Code implementation of MobileNet with Transfer Learning.