Have you tried using the quantized MobileNet from Caffe 2? Does it work?

[yxchng]

https://github.com/caffe2/models/tree/master/mobilenet_v2_quantized#model

[cedrickchee]

I have tried running the quantized MobileNet v2 from your link. According to the docs there, I have ensure my Caffe2 has QNNPACK based Int8 operators before running inference on it.

But, the Android app crashed with the following error:

A/native: [F int8_fc_op.h:43] Check failed: K == W.t.size(1) 
    terminating.
A/libc: Fatal signal 6 (SIGABRT), code -6 in tid 7829 (Camera Backgrou), pid 7790 (facebook.f8demo)

I fixed that by normalizing the input images in the same way the pre-trained model expect by modifying native-lib.cpp. This time, the app didn't crash but stuck at the "LOADING" view and Logcat is telling me:

2019-05-11 14:07:09.378 10561-10594/facebook.f8demo I/Adreno: Invalid colorspace 1

I guess I have to modify native-lib.cpp for the image preprocessing to work with the different mean and std.