1

I'd like to convert the hosted models TensorFlow-Lite hosted models mainly the mobilenets into ONNX format. So I'd like to try the quantized version of those hosted models and run them with onnx-runtime.

What would be the right procedure for converting those models to be consumed by onnx-runtime?

user3428154
  • 1,174
  • 5
  • 18
  • 38

2 Answers2

0

I'm pretty new to ONNX, but according to their official tutorial page, there isn't a TFLite to ONNX converter.

You could still use the TensorFlow -> ONNX conversion path. When you download one of the TFLite hosted models, you'll get an archive, which contains the original TensorFlow frozen graph used for conversion (xxx_frozen.pb), as well as the converted .tflite. You could take the frozen graph format and feed it into the TensorFlow-ONNX converter as instructed below:

https://github.com/onnx/tensorflow-onnx#getting-started

yyoon
  • 3,555
  • 3
  • 21
  • 17
0

There is a tflite to onnx converter. I have tried this and it has worked for me. I'm not sure about the tensorflow-lite hosted models, but you can give it a try. In my case I used this converter for the models developed by me and it has worked fine.

import tflite2onnx

tflite_path = "path/to/the/tflitemodel"
onnx_path = "path/where/you/want/to/save/your/model" #modelname.onnx
tflite2onnx.convert(tflite_path,onnx_path)
Hitesh Kumar
  • 193
  • 3
  • 11