iwatake2222 / play_with_tflite

Sample projects for TensorFlow Lite in C++ with delegates such as GPU, EdgeTPU, XNNPACK, NNAPI

Play with tflite

• Sample projects to use TensorFlow Lite in C++ for multi-platform
• Typical project structure is like the following diagram
  • 

Target

• Platform
  • Linux (x64)
  • Linux (armv7)
  • Linux (aarch64)
  • Android (aarch64)
  • Windows (x64). Visual Studio 2019
• Delegate
  • Edge TPU
  • XNNPACK
  • GPU
  • NNAPI(CPU, GPU, DSP)

Usage

./main [input]

 - input = blank
    - use the default image file set in source code (main.cpp)
    - e.g. ./main
 - input = *.mp4, *.avi, *.webm
    - use video file
    - e.g. ./main test.mp4
 - input = *.jpg, *.png, *.bmp
    - use image file
    - e.g. ./main test.jpg
 - input = number (e.g. 0, 1, 2, ...)
    - use camera
    - e.g. ./main 0

How to build a project

0. Requirements

• OpenCV 4.x

1. Download

• Download source code and pre-built libraries

  git clone https://github.com/iwatake2222/play_with_tflite.git
  cd play_with_tflite
  git submodule update --init
  sh InferenceHelper/third_party/download_prebuilt_libraries.sh

• Download models

  sh ./download_resource.sh

2-a. Build in Linux

cd pj_tflite_cls_mobilenet_v2   # for example
mkdir -p build && cd build
cmake ..
make
./main

2-b. Build in Windows (Visual Studio)

• Configure and Generate a new project using cmake-gui for Visual Studio 2019 64-bit
  • Where is the source code : path-to-play_with_tflite/pj_tflite_cls_mobilenet_v2 (for example)
  • Where to build the binaries : path-to-build (any)
• Open main.sln
• Set main project as a startup project, then build and run!

2-c. Build in Android Studio

• Please refer to
  • https://github.com/iwatake2222/InferenceHelper_Sample#2-d-build-in-android-studio
• Copy resource directory to /storage/emulated/0/Android/data/com.iwatake.viewandroidtflite/files/Documents/resource
  • the directory will be created after running the app (so the first run should fail because model files cannot be read)
• Modify ViewAndroid\app\src\main\cpp\CMakeLists.txt to select a image processor you want to use
  • set(ImageProcessor_DIR "${CMAKE_CURRENT_LIST_DIR}/../../../../../pj_tflite_cls_mobilenet_v2/image_processor")
  • replace pj_tflite_cls_mobilenet_v2 to another
• By default, InferenceHelper::TENSORFLOW_LITE_DELEGATE_XNNPACK is used. You can modify ViewAndroid\app\src\main\cpp\CMakeLists.txt to select which delegate to use. It's better to use InferenceHelper::TENSORFLOW_LITE_GPU to get high performance.
  • You also need to select framework when calling InferenceHelper::create .

Note

Options (Delegate)

# Edge TPU
cmake .. -DINFERENCE_HELPER_ENABLE_TFLITE_DELEGATE_EDGETPU=on  -DINFERENCE_HELPER_ENABLE_TFLITE_DELEGATE_GPU=off -DINFERENCE_HELPER_ENABLE_TFLITE_DELEGATE_XNNPACK=off
cp libedgetpu.so.1.0 libedgetpu.so.1
#export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:`pwd`
sudo LD_LIBRARY_PATH=./ ./main
# you may get "Segmentation fault (core dumped)" without sudo

# GPU
cmake .. -DINFERENCE_HELPER_ENABLE_TFLITE_DELEGATE_EDGETPU=off -DINFERENCE_HELPER_ENABLE_TFLITE_DELEGATE_GPU=on  -DINFERENCE_HELPER_ENABLE_TFLITE_DELEGATE_XNNPACK=off
# you may need `sudo apt install ocl-icd-opencl-dev` or `sudo apt install libgles2-mesa-dev`

# XNNPACK
cmake .. -DINFERENCE_HELPER_ENABLE_TFLITE_DELEGATE_EDGETPU=off -DINFERENCE_HELPER_ENABLE_TFLITE_DELEGATE_GPU=off -DINFERENCE_HELPER_ENABLE_TFLITE_DELEGATE_XNNPACK=on

# NNAPI (Note: You use Android for NNAPI. Therefore, you will modify CMakeLists.txt in Android Studio rather than the following command)
cmake .. -DINFERENCE_HELPER_ENABLE_TFLITE_DELEGATE_EDGETPU=off -DINFERENCE_HELPER_ENABLE_TFLITE_DELEGATE_GPU=off -DINFERENCE_HELPER_ENABLE_TFLITE_DELEGATE_XNNPACK=off -DINFERENCE_HELPER_ENABLE_TFLITE_DELEGATE_NNAPI=on

You also need to select framework when calling InferenceHelper::create .

EdgeTPU

• Install the following library
  • Linux: https://github.com/google-coral/libedgetpu/releases/download/release-grouper/edgetpu_runtime_20210726.zip
  • Windows: https://github.com/google-coral/libedgetpu/releases/download/release-frogfish/edgetpu_runtime_20210119.zip
    • the latest version doesn't work. If you have already installed edgetpu_runtime_20210726.zip , uninstall it. Also uninstall UsbDk Runtime Libraries from Windows. Then install edgetpu_runtime_20210119.zip
    • it may be better to delete C:\Windows\System32\edgetpu.dll to ensure the program uses our pre-built library
• You may need something like the following commands to run the app

  cp libedgetpu.so.1.0 libedgetpu.so.1
  sudo LD_LIBRARY_PATH=./ ./main
  
  # You may also need the followings beforehand so that you can run X app with sudo via SSH
  touch ~/.Xauthority
  xauth generate :0 . trusted 
  xauth nlist $DISPLAY | sudo xauth nmerge -

NNAPI

By default, NNAPI will select the most appropreate accelerator for the model. You can specify which accelerator to use by yourself. Modify the following code in InferenceHelperTensorflowLite.cpp

// options.accelerator_name = "qti-default";
// options.accelerator_name = "qti-dsp";
// options.accelerator_name = "qti-gpu";

License

• play_with_tflite
• https://github.com/iwatake2222/play_with_tflite
• Copyright 2020 iwatake2222
• Licensed under the Apache License, Version 2.0

Acknowledgements

• This project utilizes OSS (Open Source Software)
  • NOTICE.md
• This project utilizes models from other projects:
  • Please find model_information.md in resource.zip