Neural Network Transpiler

Convert a model from tflite to a C++ source code using Android Neural Network API.

Prerequisites

• Boost library
  • sudo apt-get install libboost-all-dev #ubuntu
  • sudo pacman -S boost #arch-linux
  • CMake ver. 3.8 or higher
  • compiler with C++17 support

Build

In the project, directory create a build directory

$ mkdir build
$ cd build

In the build directory, call cmake to generate the makefile, and call the make program

$ cmake .. -DCMAKE_VERBOSE_MAKEFILE=true
$ make

Test the compiled program

$ ./nnt -h

How to use

Assume that I have a mobilenet_quant_v1_224.tflite flatbuffer model file in build directory, the same directory from where I am executing the nnt executaeble.

Model info

./nnt -m mobilenet_quant_v1_224.tflite -i

This outputs:

::Inputs::
 Placeholder<UINT8> [1, 224, 224, 3] (quantized)
   └─ Quant: {min:[0], max:[1], scale: [0.00392157], zero_point:[0]}

::Outputs::
 MobilenetV1/Predictions/Softmax<UINT8> [1, 1001] (quantized)

Generating dot file

./nnt -m mobilenet_quant_v1_224.tflite -d mobnet.dot

The file mobnet.dot os generated in the same directory

Generating NNAPI files to use on Android

./nnt -m mobilenet_quant_v1_224.tflite -j com.nnt.nnexample -p mobnet_path

This creates a directory named mobnet_path with files: [jni.cc, nn.h, nn.cc, weights_biases.bin] where the Java package is com.nnt.nnexample