Confusion about the usage original separable conv and inverted bottleneck

edvardHua / PoseEstimationForMobile

:dancer: Real-time single person pose estimation for Android and iOS.

Apache License 2.0

1k stars 268 forks source link

Confusion about the usage original separable conv and inverted bottleneck #115

Open DanielMao2015 opened 4 years ago

DanielMao2015 commented 4 years ago

Hi, after reading the source code of cpm, I found that you use inverted bottleneck(which is mobilenet v2) and separable conv(which corresponding the original mobilenet v1) in the part of "Convolutional_Pose_Machine" scope. Why didn't you use only inverted bottleneck structure instead of original separable conv?