Linux	Windows

ONNXMLTools enables you to convert models from different machine learning toolkits into ONNX. Currently the following toolkits are supported:

• Apple Core ML
• scikit-learn (subset of models convertible to ONNX)
• Keras
• LightGBM (through its scikit-learn interface)

To convert Tensorflow models to ONNX, see tensorflow-onnx.

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 onnx package.

This package relies on ONNX, NumPy, and ProtoBuf. If you are converting a model from scikit-learn, Core ML, Keras, or LightGBM, you will need an environment with the respective package installed from the list below:

1. scikit-learn
2. CoreMLTools
3. Keras (version 2.0.8 or higher) with the corresponding Tensorflow version
4. LightGBM (scikit-learn interface)

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.

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 text
onnxmltools.utils.save_text(onnx_model, 'example.json')

# Save as protobuf
onnxmltools.utils.save_model(onnx_model, 'example.onnx')

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) 

onnxmltools converts models into the ONNX format which can be then used to compute predictions with the backend of your choice.

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[0].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, do not be alarmed. 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.

Documentation for the ONNX Model format and more examples for converting models from different frameworks can be found in the ONNX tutorials repository.

There exists a way to automatically check every converter with onnxruntime or onnxruntime-gpu. 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 in folder TESTDUMP.

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.

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.

MIT License