@AlexandreBriot The easiest thing to do would be to add tensorflow/examples/android/proto/box_coder.proto to CORE_PROTO_SRCS in tensorflow/core/BUILD (you may need to copy the file), and then you should be able to rebuild TF with support for reading it. It's not there already simply because it's a very model-specific proto only currently used for the demo.

@andrewharp Hello Andrew and thank you for your precious help !
I followed your instructions and here is where I am.

• I modified CORE_PROTO_SRCS in tensorflow/core/BUILD adding "framework/box_coder.proto",

• I copied box_coder.proto in the corresponding core/framework/ directory

• I rebuild tensorflow from source but get some errors related to proto version 2 vs 3

• I modified box_coder.proto by doing these changes

  • Change syntax = "proto2"; to syntax = "proto3";
  • Changepackage org_tensorflow_demo;to package tensorflow;
  • Removing all optional and all default value
    ex : optional float mean = 1 [default = 0.0]; becomes float mean = 1;
  • Suppressing all ; after message section

  Is that correct ? Or is there some inexact/unappropriate changes ?

• Now I am able to import this freshly built tensorflow in my iPythonNotebook
  The new proto definition is supposed to decode my box-priors.pb file
  I want to do so using the lines as I can do for GraphDef but I don't know which Object I should create to replace GraphDef one and then loading data with ParseFromString method ...

with tf.gfile.FastGFile("box-priors.pb", 'rb') as f:
    priors = tf._?Object?_()
    priors.ParseFromString(f.read())

Could you give me your advice on that part ?
Thanks in advance for your help
Alex

@AlexandreBriot We'll probably update the code soon to just load from a txt file, as separately supporting a proto build in the Android example adds disproportionate overhead.

In the meantime, I've exported the entire box prior proto to text, which you should be able to easily load into your Python code:

https://gist.github.com/andrewharp/136b07af1f221e9c95b383889f9911e0

edit: leaving this issue open to track until the discussed conversion is complete

@andrewharp Thanks for your answer !

@andrewharp I have a trained spam detection model trained using keras and have got the .pb I am trying it to implement it into a mobile app. Can you please guide me to the right documentation please. It will be of great help.