How to force the opset version on ONNX to allow quantization?

Question

I'm making some test converting ML models from several frameworks to ONNX, but I'm consistently getting warnings and errors related with the opset, E.G. the bellow code is for a lgmb model:

import numpy as np
import lightgbm as lgb
import timeit
import onnxruntime as ort
from onnxmltools.convert import convert_lightgbm
from onnxconverter_common.data_types import FloatTensorType

# Create some random data for binary classification
max_depth = 8
num_classes = 2
n_estimators = 1000
n_features = 3000
n_fit = 100
n_pred= 100000
X = np.random.rand(n_fit, n_features)
X = np.array(X, dtype=np.float32)
y = np.random.randint(num_classes, size=n_fit)
test_data = np.random.rand(n_pred, n_features).astype('float32')

model = lgb.LGBMClassifier(n_estimators=n_estimators, max_depth=max_depth, pred_early_stop=False)
model.fit(X, y)

# Use ONNXMLTOOLS to convert the model to ONNXML
input_types = [("input", FloatTensorType([n_pred, n_features]))] # Define the inputs for the ONNX
onnx_ml_model = convert_lightgbm(model, initial_types=input_types)

At the last line, I get the following error:

RuntimeError                              Traceback (most recent call last)
<ipython-input-13-11c5c9ef194b> in <module>
      1 # Use ONNXMLTOOLS to convert the model to ONNXML
      2 input_types = [("input", FloatTensorType([n_pred, n_features]))] # Define the inputs for the ONNX
----> 3 onnx_ml_model = convert_lightgbm(model, initial_types=input_types)
.............................
RuntimeError: target_opset 15 is higher than the number of the installed onnx package or the converter support (13).

The onnx version I have installed 1.10.2 which is the most recent one, the same for the onnxmltools the version I have is 1.10.0 which is also the most recent one.

To solve that I can use the parameter target_opset in the function convert_lightgbm, e.g.

onnx_ml_model = convert_lightgbm(model, initial_types=input_types,target_opset=13)

For that parameter I get the following message/warning:

The maximum opset needed by this model is only 9.

I get the same message for target_opset between 13 and 9, for target_opset bellow 9 I don´t get any message.

To continue I leave the model with target_opset = 13:

onnx_ml_model = convert_lightgbm(model, initial_types=input_types,target_opset=13)

And run the bellow code to quantize the model:

import onnx
from onnxruntime.quantization import quantize_dynamic, QuantType,quantize_qat

model_path = "ONNX_edge_deployment/src/APIs/YOLO_ONNX/lgbm.onnx"
model_quant = 'ONNX_edge_deployment/src/APIs/YOLO_ONNX/lgbm_quant.onnx'

onnx.save(onnx_ml_model, model_path)    
quantized_model = quantize_qat(new_model_path, model_quant)

But at the last line, I get the following error:

InferenceError                            Traceback (most recent call last)
..............
~\Anaconda3\lib\site-packages\onnx\shape_inference.py in infer_shapes(model, check_type, strict_mode, data_prop)
     40     if isinstance(model, (ModelProto, binary_type)):
     41         model_str = model if isinstance(model, binary_type) else model.SerializeToString()
---> 42         inferred_model_str = C.infer_shapes(model_str, check_type, strict_mode, data_prop)
     43         return onnx.load_from_string(inferred_model_str)
     44     elif isinstance(model, string_types):

InferenceError: [ShapeInferenceError] (op_type:ZipMap, node name: ZipMap): [ShapeInferenceError] type case unsupported for symbolic shape inference. inferred=5

Also there seems to be something odd related with the version of the model, following this tutorial of opsets at http://onnx.ai/, in particular this part:

domains = onnx_ml_model.opset_import
for dom in domains:
    print("domain: %r, version: %r" % (dom.domain, dom.version))

I get:

domain: '', version: 9
domain: 'ai.onnx.ml', version: 1

But in the tutorial they get:

domain: '', version: 15
domain: 'ai.onnx.ml', version: 2

How is that in a older tutorial they get newer version on both cases? 9 vs 15 and 1 vs 2