Skip to content

TensorFlow Lite Micro for easy integration with Infineon ModusToolbox projects

License

Notifications You must be signed in to change notification settings

jeldriks/mtb-tflite-micro

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

9 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

TensorFlow Lite Micro (TFLM) Library for ModusToolbox

Overview

This library provides a pre-compiled static library for TensorFlow Lite Micro for easy integration with Infineon ModusToolbox projects. The official TFLM library by Infineon only supports Infineon PSoC 6; hence, this additional library was created to support other MCUs.

No modifications have been made to the TensorFlow Lite Micro codebase. The purpose of this library is solely to provide a head start with TensorFlow Lite Micro inside ModusToolbox. This pre-compiled library removes the complexity of building the TensorFlow Lite Micro in a Linux or macOS environment and integrating it into ModusToolbox.

Currently, static libraries are only provided for the GNU Arm Embedded Toolchain (default of the ModusToolbox build system).

Usage

Adding the Library

To use the library in ModusToolbox v3.x, add a manifest.loc file to the .modustoolbox folder in your home directory. Inside this file, paste the following URI:

https://github.com/jeldriks/mtb-tflite-micro/raw/main/mtb/manifests/mtb-super-manifest-supplement-tflm.xml

Quick Start

To add this library to a ModusToolbox project, install it from the Library Manager. Then add the following DEFINES, COMPONENTS, and CXXFLAGS to the application's Makefile:

  • DEFINES+= TF_LITE_STATIC_MEMORY

  • COMPONENTS+= TFLM_CMSIS_NN

    or with CMSIS NN disabled: COMPONENTS+= TFLM

  • CXXFLAGS+= -std=c++11

Please be aware that this library does not include any modifications/enhancements for Infineon microcontrollers or the ModusToolbox environment. Therefore, the usage might differ from Infineon code examples for machine learning applications.

The Machine Learning (ML) Configurator inside ModusToolbox is only available for PSoC 6. Therefore, model conversion and validation need to be performed outside of ModusToolbox. Refer to the official TensorFlow documentation for information on model conversion and quantization.

For more detailed instructions on using ModusToolbox, refer to ModusToolbox tools package user guide.

Application Code

Please refer to the TensorFlow documentation to learn how to run inference. A complete application example can be found in the Hello World; refer to hello_world_test.cc.

A platform-specific implementation needs to be added, which is not covered in the above examples: registering a debug log callback.

Building the Library

In case you want to update the library or build it for other targets:

  • Make sure you're using Linux or macOS to build the library. Tested with Ubuntu 22.04 and macOS 14.
  • Install required dependencies: A recent version of Make, Python, and ModusToolbox (only needed if the flag -m is used; see below). Python packages required by TFLM.
  • Navigate to the directory mtb.
  • For compatibility reasons, the build is fixed to a specific commit of the TFLM repo. To update the library to a newer version/commit, edit the variable TFLM_SHA in the below Bash script.
  • Run the Bash script build.sh with the following flags:
    • -a (string): architecture (find the available architectures in the Bash script). In case no architecture is provided, all available architectures will be built.
    • -c (bool): enable CMSIS NN kernels.
    • -m (bool): use ModusToolbox GCC instead of one downloaded on demand.

For example, to build the library for Cortex-M7 with FPU support, using CMSIS NN and ModusToolbox GCC, run the following command:

./build.sh -a cortex-m7+fp -c true -m true

To build for all architectures with CMSIS NN support:

./build.sh -c true