< Network > Understanding Shufflenet_v2

The relationship with Densenet

Since I use shufflent_V2 on the task of Face Antispoofing for quit a time, I found that shufflenet_V2　is efficient yet also could provide high accuracy.

I found that for each block of shufflent_V2, the block itself could be treated as special case of densenet.

Different from densenet, each block of shufflent_V2 gets more information from the recent conv feature and get less information from the far-away conv feature.

As we know, densely connected structure could be a tool to reuse feature from different layer, thus could be treated as a multi scale feature pyramid for deep learning.

The shufflenet_V2 structure put more emphasis on the feature near-by and put less emphasis on the feature far away. It is probably because the feature from the layer near-by is more correlated.