ONNXMLTools enables you to convert models from different machine learning toolkits into ONNX. Currently the following toolkits are supported:
- Keras (a wrapper of keras2onnx converter)
- Tensorflow (a wrapper of tf2onnx converter)
- scikit-learn (a wrapper of skl2onnx converter)
- Apple Core ML
- Spark ML (experimental)
You can install latest release of ONNXMLTools from PyPi:
pip install onnxmltools
or install from source:
pip install git+https://github.com/onnx/onnxmltools
If you choose to install
onnxmltools from its source code, you must set the environment variable
ONNX_ML=1 before installing the
This package relies on ONNX, NumPy, and ProtoBuf. If you are converting a model from scikit-learn, Core ML, Keras, LightGBM, SparkML, XGBoost, or LibSVM, you will need an environment with the respective package installed from the list below:
- Keras (version 2.0.8 or higher) with the corresponding Tensorflow version
- LightGBM (scikit-learn interface)
- XGBoost (scikit-learn interface)
ONNXMLTools has been tested with Python 2.7, 3.5, 3.6, and 3.7.
Note: some wrapped converters may not support python 2.x anymore.
If you want the converted ONNX model to be compatible with a certain ONNX version, please specify the target_opset parameter upon invoking the convert function. The following Keras model conversion example demonstrates this below. You can identify the mapping from ONNX Operator Sets (referred to as opsets) to ONNX releases in the versioning documentation.
Keras to ONNX Conversion
Next, we show an example of converting a Keras model into an ONNX model with
target_opset=7, which corresponds to ONNX release version 1.2.
import onnxmltools from keras.layers import Input, Dense, Add from keras.models import Model # N: batch size, C: sub-model input dimension, D: final model's input dimension N, C, D = 2, 3, 3 # Define a sub-model, it will become a part of our final model sub_input1 = Input(shape=(C,)) sub_mapped1 = Dense(D)(sub_input1) sub_model1 = Model(inputs=sub_input1, outputs=sub_mapped1) # Define another sub-model, it will become a part of our final model sub_input2 = Input(shape=(C,)) sub_mapped2 = Dense(D)(sub_input2) sub_model2 = Model(inputs=sub_input2, outputs=sub_mapped2) # Define a model built upon the previous two sub-models input1 = Input(shape=(D,)) input2 = Input(shape=(D,)) mapped1_2 = sub_model1(input1) mapped2_2 = sub_model2(input2) sub_sum = Add()([mapped1_2, mapped2_2]) keras_model = Model(inputs=[input1, input2], output=sub_sum) # Convert it! The target_opset parameter is optional. onnx_model = onnxmltools.convert_keras(keras_model, target_opset=7)
CoreML to ONNX Conversion
Here is a simple code snippet to convert a Core ML model into an ONNX model.
import onnxmltools import coremltools # Load a Core ML model coreml_model = coremltools.utils.load_spec('example.mlmodel') # Convert the Core ML model into ONNX onnx_model = onnxmltools.convert_coreml(coreml_model, 'Example Model') # Save as protobuf onnxmltools.utils.save_model(onnx_model, 'example.onnx')
Spark ML to ONNX Conversion (experimental)
Please refer to the following documents:
Testing model converters
onnxmltools converts models into the ONNX format which can be then used to compute predictions with the backend of your choice.
Checking the operator set version of your converted ONNX model
You can check the operator set of your converted ONNX model using Netron, a viewer for Neural Network models. Alternatively, you could identify your converted model's opset version through the following line of code.
opset_version = onnx_model.opset_import.version
If the result from checking your ONNX model's opset is smaller than the
target_opset number you specified in the onnxmltools.convert function, be assured that this is likely intended behavior. The ONNXMLTools converter works by converting each operator to the ONNX format individually and finding the corresponding opset version that it was most recently updated in. Once all of the operators are converted, the resultant ONNX model has the maximal opset version of all of its operators.
To illustrate this concretely, let's consider a model with two operators, Abs and Add. As of December 2018, Abs was most recently updated in opset 6, and Add was most recently updated in opset 7. Therefore, the converted ONNX model's opset will always be 7, even if you request
target_opset=8. The converter behavior was defined this way to ensure backwards compatibility.
Test all existing converters
There exists a way
to automatically check every converter with
This process requires the user to clone the onnxmltools repository.
The following command runs all unit tests and generates
dumps of models, inputs, expected outputs and converted models
python tests/main.py DUMP
It requires onnxruntime, numpy for most models, pandas for transforms related to text features, and scipy for sparse features. One test also requires keras to test a custom operator. That means sklearn or any machine learning library is requested.
Add a new converter
Once the converter is implemented, a unit test is added to confirm that it works. At the end of the unit test, function dump_data_and_model or any equivalent function must be called to dump the expected output and the converted model. Once these file are generated, a corresponding test must be added in tests_backend to compute the prediction with the runtime.