Why is MobileNetV2 faster than MobileNetV1? Given that MobileNetV1's block is basically a separable convolution, MobileNetV2 has this separable convolution but it also expands the channels, that is, there are more pieces of information to be processed. The difference is that the channel squeezing part is linear. This leads me to believe that removing an activation function makes the channel squeezing process so fast that the remaining process of expansion using ReLU6 + depthwise convolution performed in the inverted bottleneck and the addition of the residue is not very relevant in terms of processing. I don't know if I was clear, but to me, the convolution block of V2 is more complex than V1, so how does this reduce the parameters and make it faster?
Basically, everywhere I look people say that this architecture is better... but it seems like magic to me, because I can't understand the reason. Since you're making the process more robust, it doesn't make sense for it to be lighter and faster.