Golioth TensorFlow Lite Model Update Example

Overview

This application downloads a TensorFlow Lite model from the Golioth servers and uses it to recognize spoken words using an i2s microphone.

• Compile and load the application on an m5stack CoreS3 device
• Upload the included TensorFlow models as Golioth artifacts
• Roll out a release with your desired artifact; the device will no be able to recognize simple words (eg: yes, no)
• Release the other artifact to load a new model that recognizes additional words (stop, go)

Use this application as a reference when adding multi-artifact download to update trained models on your machine-learning applications.

Local Setup

git submodule add https://github.com/golioth/golioth-firmware-sdk.git submodules/golioth-firmware-sdk
cd example-tensorflow-model-update
git submodule update --init --recursive

Building the Application and Assign Credentials

Build for m5stack CoreS3

idf.py set-target esp32s3
idf.py build
idf.py flash monitor

Provisioning

esp32> settings set wifi/ssid <my-wifi-ssid>
esp32> settings set wifi/psk <my-wifi-psk>
esp32> settings set golioth/psk-id <my-psk-id@my-project>
esp32> settings set golioth/psk <my-psk>
esp32> kernel reboot cold

TensorFlow Lite Models

TensorFlow Lite models are found in the models directory.

1. Upload these models to Golioth as artifacts using the name model
2. Create and rollout a release that includes a model artifact
3. The most recent release that includes a model artifact will be used by the device as the currently selected TensorFlow Lite model.

Available Models

• model.bin_header_yn: Trained to recognize yes, and no
• model.bin_header_ynsg: Trained to recognize yes, no, stop, and go

Model Formatting

Models may be trained by following the tflite-micro Micro Speech Training guide.

The models/model.tflite binary file output during the training process is given a custom header that indicates the text labels used during the training. Here is an example of adding a header to the training data:

echo "GLTHBEGIN;silence;unknown;yes;no;GLTHEND" > model.bin_header_yn
cat models/model.tflite >> model.bin_header_yn