One: [guide] How to create a tflite file easily from Python TF program

[hyunsik-yoon]

for https://github.com/Samsung/ONE/issues/2275#issuecomment-645103277

This issue is a guide to share how to create a tflite file easily from Python TF program for whom may want to know. :-)

Assume that we want to test cast op with int64 so let's call it cast_int64.

• create a program with python using import tensorflow as tf ....
  
  
  
  • please write like res/TensorFlowPythonExamples/examples/cast_01/__init__.py
  
  
• create pbtxt file
  
  
  
  • run the following (in my case). (You may want to use 1.13 for official test
    
    $ export MODEL='cast_01' $ docker run -v /home/eric/:/home/eric -w /home/eric/work/1/ONE/res/TensorFlowPythonExamples \ tensorflow/tensorflow:1.15.0-py3 python3 tfpem.py ${MODEL}
  
  
  • Then you will get cast_01.pbtxt file under res/TensorFlowPythonExamples
  
  
• create `tflite file
  
  
  
  • run the following (in my case). I used nightly but it would be better to use official version.
    
    docker run -it -w "$(pwd)" -v /home/eric:/home/eric tensorflow/tensorflow:1-15-0-py \ python3 ../ONE/compiler/tf2tfliteV2/tf2tfliteV2.py --v1 ... <params>
  
  
  • please refer to ONE/compiler/tf2tfliteV2/readme.md for params
  
  
  • let's assume that cast_int64.tflite is created.
  
  
• Then open the created cast_int64.tflite with, e.g., netron to see if conversion is done as you wish. tflite converter does some optimization so tflite file sometime contains what you don't wish to be.
• ask @wateret and @hseok-oh to upload the tflite file to test file server
• create cast_int64 dir under tests/scripts/framework/tests and create config.sh.

on odroid,
run

$ LD_LIBRARY_PATH=./Product/out/lib ./tests/scripts/test-driver.sh --artifactpath=. --frameworktest

[hyunsik-yoon]

Does anyone know how to test locally without the following?

ask @wateret and @hseok-oh to upload the tflite file to test file server

Does anyone know how to run only cast_int64?

[hyunsik-yoon]

/cc @intom @ragmani for https://github.com/Samsung/ONE/issues/2275#issuecomment-645103277

[ragmani]

Does anyone know how to test locally without the following?
Does anyone know how to run only cast_int64?

$ BACKENDS="cpu" ./Product/armv7l-linux.debug/out/bin/nnpackage_run --nnpackage ${cast_int64_dir} -l ${cast_int64_dir}/metadata/tc/input.h5 -d output.h5

$ h5diff -v ${cast_int64_dir}/metadata/tc/expected.h5 output.h5

But int64 type has a issue

[intom]

Does anyone know how to test locally without the following?
Does anyone know how to run only cast_int64?

$ BACKENDS="cpu" ./Product/armv7l-linux.debug/out/bin/nnpackage_run --nnpackage ${cast_int64_dir} -l ${cast_int64_dir}/metadata/tc/input.h5 -d output.h5

$ h5diff -v ${cast_int64_dir}/metadata/tc/expected.h5 output.h5

But int64 type has a issue

I've tested with cast model that you converted.
Is it right?
If the result is right, I will create the draft PR for int64 test

intom@intom-PMBSB09A-Samsung-DeskTop:~/working_directory/2020/one/myPublicGit/ONE/Product/x86_64-linux.debug/out/bin$ ./nnpackage_run ~/working_directory/2020/one/models/cast_01 -l ~/working_directory/2020/one/models/cast_01/metadata/tc/input.h5 -d output.h5
... warmup 1 takes 0.652 ms
... run 1 takes 0.096 ms
===================================
MODEL_LOAD   takes 34.344 ms
PREPARE      takes 73.387 ms
EXECUTE      takes 0.096 ms
- MEAN     :  0.096 ms
- MAX      :  0.096 ms
- MIN      :  0.096 ms
- GEOMEAN  :  0.096 ms
===================================
intom@intom-PMBSB09A-Samsung-DeskTop:~/working_directory/2020/one/myPublicGit/ONE/Product/x86_64-linux.debug/out/bin$ h5diff -v ~/working_directory/2020/one/models/cast_01/metadata/tc/expected.h5 output.h5
file1     file2
---------------------------------------
    x      x    /              
    x           /name          
    x      x    /value         
    x      x    /value/0       

group  : </> and </>
0 differences found
group  : </value> and </value>
0 differences found
dataset: </value/0> and </value/0>
0 differences found
intom@intom-PMBSB09A-Samsung-DeskTop:~/working_directory/2020/one/myPublicGit/ONE/Product/x86_64-linux.debug/out/bin$ 

[ragmani]

@intom It was right.

[hyunsik-yoon]

let me close this since it is a guide and i guess it's been shared to people.