Embedded AI Torch to TFLite on ESP32

We provide an example to train a Conv model with Pytorch and convert to Tensorflow Lite (Micro) and then deploy it on ESP32-S3-DevKitC-1-N8R8.

Qucik start

Train and convert model

You can play the notebook in Google Colab.

You may see model.h and data.h in your colab folder after running the notebook. Please put them in esp32/main.

Flash model to ESP32-S3

Please go to esp32 to get more details. You need to install ESP-IDF idf-v4.4.6.

How does it work?

There is a gap between deep models and resource-limited devices. From Kb-level microcontroller to powerful Edge devices with hardware accelerator, the on-board resource varies. It requires model compression and lightweight model designs.

We take the deployment of a simple CNN model on ESP32 as an example. We implement a tiny CNN model with Pytorch and train it on MNIST for epochs, and then convert it to ONNX and then take onnx2tf to perform conversion and full-integer quantization. The well-quantized model is exported to a .CC model file.

To deploy TFlite model to ESP32, we implement an inference program for ESP32-S3 and flash it into board with ESP-IDF.