create new tag
view all tags

The TinyML examples written in C++

The TinyML examples in C++, ported to the ESP32 can be found at https://github.com/espressif/tflite-micro-esp-examples/tree/master/examples. There you also find the procedures to actually build the programs. In fact there are two methods of building:

  • use thee Espressif esp-idf build system
  • use the Arduino IDE
  • use the MicroPython interface to TensorFlow MicroPython Examples
The test programs allowing to test the model before deploying it to the target processor can be found in the TensorFlow Lite Micro repository.


For the Arduino SDK (Software Development Kit) you can install the TensorFlowLite_ESP32 library, which does not only contain the TensforFlowLite library but also the examples

  • hello_world
  • wake word detection
  • person detection
  • ... and even some older version of magic wand, which is not available elsewhere


Attention: When using Serial.print on the LoLin ESP32S3 mini, you must set the option USB CDC on boot to enabled! The option is found in the Tools menu. Oherwise you will not see any output on the Serial monitor.

The Hello World test programs

Three test programs are available:

  • hello_world_test.cc
  • evaluate.py
  • evaluate_test.py
These programs are built using the bazel build tool. In order tto build hello_world_test.cc, got to the base directory of TensorFlow Lite Micro directory and enter the commands. First the hello_world_test binary must be built from its source:


before the program can be run:


For a detailed analysis of the test program please set analysis cc test

The Hello World application with esp-idf

Let's start with the hello_world example. This example uses a pre-built model that has gone through the training process and it has been quantized to reduce its size such that it can be loaded into the small micro-controller memory. Building, training and quantization is described at https://github.com/tensorflow/tflite-micro/tree/main/tensorflow/lite/micro/examples/hello_world.

We must first tell the idf.py that we are using an ESP32S3 chip:

idf.py set-target esp32s3

After that we may have to clean previous builds:

idf.py fullclean

and finally we can build the hello_world program:

idf.py build

Once the program is built we can flash it and connect the serial monitor to see the results:

idf.py flash monitor

The program repeatedly runs the induction for 20 angle values between 0 and 2Π. These values can easily be captured by redirecting the output to a file. I then used an editor to prepare the data to be plotted with gnuplot.

idf.py flash monitor | tee results.txt

Finally I wrote a simple Python program calculating the correct 20 sine values, which allows to compare them to the results from TinyML.


Creating the model and training it

When working with Machine Learning, data play a primordal role because data are used to train the model and therefore influence the opaque algorithm within the model.

The model for the hello world example is created with the Python script train.py.

-- Uli Raich - 2023-09-02


Topic attachments
I Attachment History Action Size Date Who Comment
PNGpng hello_world_test_build.png r1 manage 52.2 K 2023-09-08 - 07:30 UliRaich  
PNGpng hello_world_test_run.png r1 manage 152.4 K 2023-09-08 - 07:30 UliRaich  
PNGpng results.png r1 manage 24.2 K 2023-09-02 - 14:28 UliRaich  
PNGpng tflite_examples_arduino.png r1 manage 178.7 K 2023-09-07 - 13:05 UliRaich  
Edit | Attach | Watch | Print version | History: r6 < r5 < r4 < r3 < r2 | Backlinks | Raw View | Raw edit | More topic actions
Topic revision: r6 - 2023-11-08 - UliRaich
This site is powered by the TWiki collaboration platform Powered by PerlCopyright © 2008-2024 by the contributing authors. All material on this collaboration platform is the property of the contributing authors.
Ideas, requests, problems regarding TWiki? Send feedback