Pytorch OCR models for deploying to ESP32?

I wanted to use Pytorch to target ESP32, because I don’t like TensorFlow.

The ESP32 docs (GitHub - espressif/esp-dl: Espressif deep-learning library for AIoT applications) say that Pytorch models converted to ONNX can be ran on the ESP32 as long as the esp-dl supports all the required operators.

However, what OCR models should I look at? I want the best accuracy and ideally a model that doesn’t use operators outside of those supported by esp-dl.