I’m starting learing about TinyML a.k.a “Machine Learning for Embedded Systems”, and I’m logging my journey. This post is about how to setup your toolchain, and deploy the “Hello World” on the ESP-EYE.

Btw, I used Linux Terminal throughout the post. But I think you can easily this everything under Windows and macOS.

• Install ESP-IDF Toolchain
• Download Tensorflow Micro
• Hello World
  • Compile Hello World
  • Run Hello World
• Workflow Inside A Docker Container
• Building more examples
• Conclusion

Install ESP-IDF Toolchain

Install IDF Toolchain unter Linux using docker. If you need to know how to install docker, read this.

docker pull espressif/idf

That’s all, but there is the option to use different versions. To download a specific docker image version or “tag”, specify it as espressif/idf:TAG, for example:

docker pull espressif/idf:release-v4.0

For more detailed info about ESP-IDF toolchain, you can read the ESP-IDF Programming Guide.
And if you want to dig more into IDF Docker you can follow this link.

Download Tensorflow Micro

clone the repository: https://github.com/espressif/tensorflow/

git clone https://github.com/espressif/tensorflow/

Hello World

Compile Hello World

Generate the “Hello World” example with the following command:

make -f tensorflow/lite/micro/tools/make/Makefile TARGET=esp generate_hello_world_esp_project

Go the the example project directory

cd tensorflow/lite/micro/tools/make/gen/esp_xtensa-esp32/prj/hello_world/esp-idf

Then build with idf.py

docker run --rm -v $PWD:/project -w /project espressif/idf idf.py build

Run Hello World

To flash your device (replace /dev/ttyUSB0 with the device serial port):

docker run --rm -v $PWD:/project -w /project --device=/dev/ttyUSB0 espressif/idf idf.py flash -p /dev/ttyUSB0

Workflow Inside A Docker Container

Sometimes is useful to work directly from Docker, for example for debugging purposes.

Let’s get inside the container:

docker run --rm -v $PWD:/project -w /project --device=/dev/ttyUSB0 -it espressif/idf

Now that you’re inside the container, you can run IDF commands directly:

idf.py build
idf.py flash -p /dev/ttyUSB0

If you want to see the debug output using the serial port. Note: you need use docker with --device=/dev/ttyUSB0 to be able to use monitor the serial port.

idf.py monitor

Use Ctrl+] to exit.

You could use combine the following commands:

idf.py --port /dev/ttyUSB0 flash monitor

You can find More info about IDF Monitor here.

Building more examples

You can more examples if you check the README files under tensorflow/lite/micro/examples/<EXAMPLE_NAME>/README.md. These are the following examples available at the moment:

$  tree -L 1 tensorflow/lite/micro/examples/
tensorflow/lite/micro/examples/
├── doorbell_camera
├── hello_world
├── image_recognition_experimental
├── magic_wand
├── micro_speech
├── network_tester
├── person_detection
└── person_detection_experimental

8 directories, 0 files

This file tensorflow/lite/micro/examples/hello_world/README.md, for instance, describes how to configure and build hello_world example.

Conclusion

TinyML is a new and growing field, for that reasing it’s still evolving but that means that there is a lot room for improving, and most important. TinyML enables a new set of real-world applications.

I hope you enjoy this introduction. Until next time ;)