Exporting a specific function from a TFLite model to ONNX, instead of just the "default" one

[josephrocca]

Describe the feature request

Currently it seems that if a TFLite file contains more than one function, it's only possible to convert one of them. Seems like it just converts the first function and ignores the rest, and there's no way to choose a different one as far as I can see. Here's an example tflite file that I'm working with here. If you open it in netron.app you can see it has two graphs - "encode" and "decode". When I export it, it only exports the "decode" graph. Here's a notebook showing the exporting process: https://colab.research.google.com/gist/josephrocca/3f3d940091c3a6e105155e4b77226487/notebook.ipynb

This request is related to this one, except that issue is asking for the ability to export multiple functions in a TF model to an ONNX file that has multiple functions. I don't need mutliple functions in my ONNX file, but I need to be able to export all the functions (separately, or together) into one or more ONNX files. Currently I can only get one of them, and there's no way to choose which one I get.

Describe scenario use case

Google recently released Lyra V2, but they only released the TFLite files, and it's uncertain whether they'll be able to release it in other model formats. So I'm trying to do my best to work with the models that they've given us. The end goal is to get Lyra V2 working on the web with ORT Web.

[dramaticlama]

Hi Joseph,

I understand that you want a real solution for this. Meanwhile if you are looking for a workaround I believe that it should work by manipulating the tflite flatbuffer from python:

import sys
sys.path.append('<path to tensorflow git>/tensorflow/lite/tools')
import flatbuffer_utils as fbu

model = fbu.read_model('<path to lyra git>/lyra/model_coeffs/quantizer.tflite')
# Only keep the encoder.
model.subgraphs = [model.subgraphs[1]]
fbu.write_model(model, '<path to lyra git>/lyra/model_coeffs/quantizer_encoder.tflite')

I hope that this workaround works for you and that it will be temporary until someone can fix it !

[josephrocca]

@dramaticlama Thank you! This is a super helpful workaround in the meantime. I was assuming that this sort of thing would be more complicated due to some shared parameters or something like that.

Here's a working Colab using dramaticlama's code for anyone else that runs into this issue: https://colab.research.google.com/gist/josephrocca/320fd77316bd09cef52c3fc8cb657a14/notebook.ipynb