search

xiaolai-sqlai / mobilenetv3

mobilenetv3 with pytorch，provide pre-train model
MIT License
1.6k stars 340 forks source link

About SE Reduction ratio? #33

Open H-YunHui opened 4 years ago

H-YunHui commented 4 years ago

In the original Squeeze-and-Excitation Networks, the author found through experiments that when reduction = 16, precision and complexity are well balanced, and in your module, reduction is set to the current number of bneck output channels , what is the reason for this setting? Did it come from the experiment?

wangh09 commented 4 years ago

In chapter5.3, the original mobilenetv3 paper, it says that the reduction ratio is set to 1/4 (namely 4), as well as in the current implementation