Selective builds to reduce tensorflow-lite binary size missing needed java package org.tensorflow.lite

[duudat]

Please make sure that this is a build/installation issue. As per our GitHub Policy, we only address code/doc bugs, performance issues, feature requests and build/installation issues on GitHub. tag:build_template

System information
Try to build a custom .aar file from https://raw.githubusercontent.com/tensorflow/tensorflow/master/tensorflow/lite/tools/build_aar_with_docker.sh

Describe the problem
The result aar.file missing all java-wrapper classes (Interpreter, etc) under package: org.tensorflow.lite. These classes are needed to perform inference from a model in Android.

Provide the exact sequence of commands / steps that you executed before running into the problem
Following the instruction building selective build for a given tflite model using docker: https://www.tensorflow.org/lite/guide/reduce_binary_size#selectively_build_tensorflow_lite_with_docker

curl -o build_aar_with_docker.sh \
  https://raw.githubusercontent.com/tensorflow/tensorflow/master/tensorflow/lite/tools/build_aar_with_docker.sh &&
chmod +x build_aar_with_docker.sh

and

sh tensorflow/lite/tools/build_aar.sh \
  --input_models=/a/b/my_model.tflite \
  --target_archs=x86,x86_64,arm64-v8a,armeabi-v7a \
  --checkpoint=master \

Any other info/ logs
Rename resulted aar files to .zip and attached:

1. aar file built from Mobilenet_1.0_224(float) model: tensorflow-lite-with-mobile-net-model.zip
2. aar file built from my custom model: tflite-with-my-custom-model.zip

Both missing java classes

Note: did try to copy needed java classes into the resulted_aar but this approach seems not working. Might-be the generated .so files missing C++ files which support org.tensorflow.lite java classes as well.

[thaink]

Hi @duudat What it means by "in order for it to run locally: org.tensorflow.lite.Interpreter, etc."?
Are you using it in an Android project?

[thaink]

@wangtz I found it has been this way since day 1 of this feature.
However, It seems to work fine with Android since then.
Do you have any insights on this?

[duudat]

@thaink Yes, you are correct, I use the generated aar file in my Android project to cutting the size of 'org.tensorflow:tensorflow-lite:x.x.x'.
Specifically, my assumption by follow the instructions: https://www.tensorflow.org/lite/guide/reduce_binary_size#selectively_build_tensorflow_lite_with_docker is that:
The resulted tensorflow-lite.aar will replace 'org.tensorflow:tensorflow-lite:x.x.x' dependency in my Android project.

In Android, to run on-device inference out of a tflite model, it requiresInterpreter, and other java classes under package: org.apache.tensorflow-lite which is missing in the new generatedtensorflow-lite.aar, so I got compile errors of cannot find symbol class Interpreter.

[duudat]

Minor thing I notices is that: the generated aar file seems to also exclude other files inside normal 'org.tensorflow:tensorflow-lite:x.x.x' aar file from jCenter: proguard.txt, headers.

At least proguard.txt is, for sure, needed in most aar files if the Android app, which uses the aar, enabled size optimization, resource shrinking.

[duudat]

@thaink you have suggestion to get around with the errors: cannot find symbol class Interpreter. without adding 'org.tensorflow:tensorflow-lite:x.x.x' dependency to my build.gradle file.

[thaink]

@duudat Your use is right. Let me check the cause of this.

[duudat]

@thaink any update on this issue?

[duudat]

I could locally include needed java classes by modyfing build_aar.sh as following.

(.sh is changed into .txt since not allowed to upload a .sh file)
modified_build_aar.txt

The aar file generated by the modified scripts didn't encounter the missing java classes problems.

[thaink]

thanks for the info @duudat. I was able to reproduce this issue.

[thaink]

The issue is fixed at master branch.