The essence of dynamic convolution

The parameter quantity of dynamic convolution will increase exponentially , The amount of calculation will increase less

The amount of calculation is mainly in $Wx+b$, For convolution W yes $3\times 3$ Convolution kernel of through step stride adopt $H,W$ When the characteristic diagram of , A lot of computation is generated .

For dynamic convolution $K$ A core , The main reason is that the parameter quantity is increased , But the amount of calculation increases very little , because K With different parameters $W$ The amount of computation of weighted aggregation is related to the size of the nucleus , However, usually the size of convolution kernel is mainly $3\times 3$ So the computational cost is very small .