Number of filters in comparison to official Keras implementation

Bisonai / mobilenetv3-tensorflow

Unofficial implementation of MobileNetV3 architecture described in paper Searching for MobileNetV3.

Apache License 2.0

225 stars 81 forks source link

Number of filters in comparison to official Keras implementation #18

Closed Eichhof closed 3 years ago

Eichhof commented 3 years ago

Hello

As far as I can see the large model conatins filters of size 16, 24, 24, 40, 40, 40, 80, 80, 80, 80, 112, 112, 160, 160, 160. On the other hand, the official Keras MobileNetV2 implementation contains filters of size 16, 24, 24, 32, 32, 32, 64, 64, 64, 64, 96, 96, 96, 160, 160, 160, 320.

Why does this implementation here differ in terms of the number of filters?

martinkersner commented 3 years ago

Hi @Eichhof,

Our implementation is based on MobileNetV3 paper. MobileNetV3 architecture is different from MobileNetV2.

Cheers, Martin