Skip to content

Latest commit

 

History

History
94 lines (66 loc) · 3.13 KB

README.en.md

File metadata and controls

94 lines (66 loc) · 3.13 KB
Raw

tflite4zero_env
a C++ Environment for Building Tensorflow Lite Projects on Raspberry Pi Zero (armv6)

zh en

General information

Version: TensorFlow 2.4.1

Encode: UTF-8

GPU/XNNPACK acceleration not supported, yet.

Update

20210401：

  1. Refactored the way the project was built. Now every project has its own Makefile, which is placed in the folder of the corresponding project.
    The syntax for building a project is still ./build_project.sh <project name>

  2. Modified the Makefile of the sample project. Now each project can add multiple source file directories in the SRC_DIRS variable of Makefile.

Instructions for use

Create and build the project

  1. Give the project build script ./build_project.sh execute permission:

    chmod +x ./build_project.sh

  2. Create your tf-lite project in the ./project directory. You can refer to existing sample projects.
    The name of the project folder is your project name.

  3. Write the project code and Makefile.

  4. Execute the build script:

    ./build_project.sh <project name>

# Please do not use the source ./build_project.sh or. ./build_project.sh command, the exit command in the script will cause the current ssh window to exit

  5. The generated project will be in the ./build directory.

Run the sample project

This is an example of running the MobileNetV3 target recognition network model.

  1. Build the sample project: 
    ./build_project.sh label_image_tf1.14
  2. Copy the network model, label, data set and test picture to the project build directory: 
    cp -r ./project/label_image_tf1.14/data ./build/label_image_tf1.14/
  3. Switch to the directory where the test picture is located: 
    cd ./build/label_image_tf1.14/data
  4. Run the MobileNetV3 network model and perform target recognition on dogs.bmp: 
    ../bin/label_image_tf1.14 -m ./mobnet_v3_coco_official.tflite -l ./labelmap.txt -i dogs.bmp -o 1 -t 1

Note: The image file input to the model must be in bmp format, and can only be a file name without path. The output picture is in the current directory, such as out_dogs.bmp.

If cross compile is needed

Cross compile toolchain (use toolchain 6.5.0 in tar.gz):

https://github.com/rvagg/rpi-newer-crosstools/archive/eb68350c5c8ec1663b7fe52c742ac4271e3217c5.tar.gz

CC/CXX/AR variables need to be set in Makefile.

/your/path/to/rpi-newer-crosstools/x64-gcc-6.5.0/arm-rpi-linux-gnueabihf/bin/arm-rpi-linux-gnueabihf-gcc 
 
/your/path/to/rpi-newer-crosstools/x64-gcc-6.5.0/arm-rpi-linux-gnueabihf/bin/arm-rpi-linux-gnueabihf-g++ 
 
/your/path/to/rpi-newer-crosstools/x64-gcc-6.5.0/arm-rpi-linux-gnueabihf/bin/arm-rpi-linux-gnueabihf-ar

More info about cross-compiling tensorflowlite library on rpi zero (already compiled in our environment):

https://blog.csdn.net/weixin_41973774/article/details/114807080 (in Chinese)