Convert TF model in OpenCV model

[marbre2]

Hello,
I have trained a TF model with a jupyter notebook on Google colab, inspired by the jupyter notebook "StarDist_2D_ZeroCostDL4Mic.ipynb"
I would like to use this StarDist trained model directly in QuPath.
For this purpose, I need to convert the output saved_model.pb as OpenCV model. I am able to perform it afterwards, but I'd rather do it in the same jupyter notebook using tf2onnx package, and tf2onnx.convert.from_function().
However, I have issue following issue:

---

AttributeError Traceback (most recent call last)
in
4 output_path = model.name + ".pb"
5
----> 6 model_proto, _ = tf2onnx.convert.from_function(model, input_signature=spec, opset=10, output_path=output_path)
7 output_names = [n.name for n in model_proto.graph.output]

/usr/local/lib/python3.7/dist-packages/tf2onnx/convert.py in from_function(function, input_signature, opset, custom_ops, custom_op_handlers, custom_rewriter, inputs_as_nchw, outputs_as_nchw, extra_opset, shape_override, target, large_model, output_path)
547 raise ValueError("from_function requires input_signature")
548
--> 549 concrete_func = function.get_concrete_function(*input_signature)
550
551 input_names = [input_tensor.name for input_tensor in concrete_func.inputs

AttributeError: 'StarDist2D' object has no attribute 'get_concrete_function'

Could you please give me some advice about how to make it work?

Thank you for your time and help,

Best regards

[marbre2]

If anyone ever wonder, here is how I solved the issue - for jupyter notebook (code is not optimized).
The script sample (for Jupyter notebook) below performs
-> saving TF trained model into a ONNX model compatible with StarDist plugin for QuPath.

# SAVING AND EXPORT DL MODEL FOR STARDIST - QUPATH extension
## Install all packages and cuda libraries
!pip install tensorflow
!pip install -U tf2onnx
!pip install 'stardist[bioimageio]'
!pip install -q stardist # contains tools to operate STARDIST.
!pip install -q gputools # improves STARDIST performances
#!pip install edt # improves STARDIST performances
!pip install -q onnxruntime

## Loading libraries
import tf2onnx
from stardist.models import StarDist2D
import os
import tensorflow as tf
import numpy as np
import onnxruntime as rt
import shutil

## Select trained TF model directory
QC_model_folder = "/content/drive/MyDrive/TP-Stardist/Stardist/models/"  #@param {type:"string"}
#Here we define the loaded model name and path
QC_model_name = os.path.basename(QC_model_folder)
QC_model_path = os.path.dirname(QC_model_folder)
full_QC_model_path = QC_model_path+'/'+QC_model_name+'/'

if QC_model_name=="":
    print('!! WARNING: The chosen model does not exist !!')
    print('Please make sure you provide a valid model path (without "/" at the end of the string) before proceeding further.')
    quit()

    if os.path.exists(full_QC_model_path):
        print("The "+QC_model_name+" network will be evaluated")
    else:  
        print('!! WARNING: The chosen model does not exist !!')
        print('Please make sure you provide a valid model path (without "/" at the end of the string) before proceeding further.')

## Loading model
model = StarDist2D(None, name=QC_model_name, basedir=QC_model_path)

## Saving model for Qupath - Stardist extension

saved_model_path = os.path.join(QC_model_path, model.name)
model_path_zip = os.path.join(saved_model_path, "TF_SavedModel.zip")
print(model_path_zip)

output_model_tf = os.path.join(saved_model_path, 'TF_SavedModel')
print(output_model_tf)

shutil.unpack_archive(model_path_zip, output_model_tf)
output_model_onnx = os.path.join(saved_model_path, 'ONNX_SavedModel')
os.makedirs(output_model_onnx, exist_ok=True)
print(output_model_onnx)

!python -m tf2onnx.convert --opset 10 --saved-model {output_model_tf} --output_frozen_graph {os.path.join(output_model_onnx, "saved_model_opencv.pb")}