This tool converts ONNX models to Apple Core ML format. To convert Core ML models to ONNX, use ONNXMLTools.
There's a comprehensive Tutorial showing how to convert PyTorch style transfer models through ONNX to Core ML models and run them in an iOS app.
To get the latest version of onnx-coreml from PyPI:
pip install --upgrade onnx-coreml
pip install --upgrade coremltools # onnx-coreml depends on the coremltools packageFor the latest changes please see the release notes.
To get the latest version from source (master branch of this repository), please see the installation section.
Please see the ONNX conversion section in the Neural network guide on how to use the converter.
There are a few notebook examples as well for reference.
def convert(model,
mode=None,
image_input_names=[],
preprocessing_args={},
image_output_names=[],
deprocessing_args={},
class_labels=None,
predicted_feature_name='classLabel',
add_custom_layers=False,
custom_conversion_functions={},
minimum_ios_deployment_target='13')__model__: ONNX model | str
An ONNX model with parameters loaded in onnx package or path to file
with models.
__mode__: str ('classifier', 'regressor' or None)
Mode of the converted coreml model:
'classifier', a NeuralNetworkClassifier spec will be constructed.
'regressor', a NeuralNetworkRegressor spec will be constructed.
__image_input_names__: list of strings
Name of the inputs to be defined as image type. Otherwise, by default all inputs are MultiArray type.
__preprocessing_args__: dict
Specify preprocessing parameters, that are be applied to all the image inputs specified through the "image_input_names" parameter.
'is_bgr', 'red_bias', 'green_bias', 'blue_bias', 'gray_bias',
'image_scale' keys with the same meaning as coremltools set_pre_processing_parameters: https://apple.github.io/coremltools/generated/coremltools.models.neural_network.builder.html#coremltools.models.neural_network.builder.NeuralNetworkBuilder.set_pre_processing_parameters
__image_output_names__: list of strings
Name of the outputs to be defined as image type. Otherwise, by default all outputs are MultiArray type.
__deprocessing_args__: dict
Same as 'preprocessing_args' but for the outputs.
__class_labels__: A string or list of strings.
As a string it represents the name of the file which contains
the classification labels (one per line).
As a list of strings it represents a list of categories that map
the index of the output of a neural network to labels in a classifier.
__predicted_feature_name__: str
Name of the output feature for the class labels exposed in the Core ML
model (applies to classifiers only). Defaults to 'classLabel'
__add_custom_layers__: bool
If True, then ['custom'](https://developer.apple.com/documentation/coreml/core_ml_api/integrating_custom_layers?language=objc) layers will be added to the model in place of unsupported onnx ops or for the ops
that have unsupported attributes.
Parameters for these custom layers should be filled manually by editing the mlmodel
or the 'custom_conversion_functions' argument can be used to do the same during the process of conversion
__custom_conversion_fuctions__: dict (str: function)
Specify custom function to be used for conversion for given op.
User can override existing conversion function and provide their own custom implementation to convert certain ops.
Dictionary key must be string specifying ONNX Op name or Op type and value must be a function implementation available in current context.
Example usage: {'Flatten': custom_flatten_converter, 'Exp': exp_converter}
`custom_flatten_converter()` and `exp_converter()` will be invoked instead of internal onnx-coreml conversion implementation for these two Ops;
Hence, User must provide implementation for functions specified in the dictionary. If user provides two separate functions for node name and node type, then custom function tied to node name will be used. As, function tied to node type is more generic than one tied to node name.
`custom_conversion_functions` option is different than `add_custom_layers`. Both options can be used in conjuction in which case, custom function will be invoked for provided ops and custom layer will be added for ops with no respective conversion function.
This option gives finer control to user. One use case could be to modify input attributes or certain graph properties before calling
existing onnx-coreml conversion function. Note that, It is custom conversion function's responsibility to add respective Core ML layer into builder(coreml tools's NeuralNetworkBuilder).
Examples: https://github.com/onnx/onnx-coreml/blob/master/tests/custom_layers_test.py#L43
__onnx_coreml_input_shape_map__: dict (str: List[int])
(Optional)
(only used if `minimum_ios_deployment_target` version is less than '13')
A dictionary with keys corresponding to the model input names. Values are a list of integers that specify
how the shape of the input is mapped to Core ML. Convention used for Core ML shapes is:
0: Sequence, 1: Batch, 2: channel, 3: height, 4: width.
For example, an input of rank 2 could be mapped as [3,4] (i.e. H,W) or [1,2] (i.e. B,C) etc.
__minimum_ios_deployment_target__: str
Target Deployment iOS version (default: '12')
Supported values: '11.2', '12', '13'
Core ML model produced by the converter will be compatible with the iOS version specified in this argument.
e.g. if `minimum_ios_deployment_target` = '12', the converter would only utilize Core ML features released till iOS12
(equivalently macOS 10.14, watchOS 5 etc).
Release notes:
* iOS 11 / Core ML 1: https://github.com/apple/coremltools/releases/tag/v0.8
* iOS 12 / Core ML 2: https://github.com/apple/coremltools/releases/tag/v2.0
* iOS 13 / Core ML 3: https://github.com/apple/coremltools/releases/tag/v3.0-beta
__model__: A Core ML model.
Also you can use command-line script for simplicity:
convert-onnx-to-coreml [OPTIONS] ONNX_MODEL
The command-line script currently doesn't support all options mentioned above. For more advanced use cases, you have to call the python function directly.
pip install -U onnx-coremlTo get the latest version of the converter, install from source by cloning the repository along with its submodules and running the install.sh script. That is,
git clone --recursive https://github.com/onnx/onnx-coreml.git
cd onnx-coreml
./install.shTo get the latest version of the converter, install from source by cloning the repository along with its submodules and running the install-develop.sh script. That is,
git clone --recursive https://github.com/onnx/onnx-coreml.git
cd onnx-coreml
./install-develop.shIn order to run unit tests, you need pytest.
pip install pytest
pip install pytest-covTo run all unit tests, navigate to the tests/ folder and run
pytestTo run a specific unit test, for instance the custom layer test, run
pytest -s custom_layers_test.py::CustomLayerTest::test_unsupported_ops_provide_functionsModels from https://github.com/onnx/models that have been tested to work with this converter:
- BVLC Alexnet
- BVLC Caffenet
- BVLC Googlenet
- BVLC reference_rcnn_ilsvrc13
- Densenet
- Emotion-FERPlus
- Inception V1
- Inception V2
- MNIST
- Resnet50
- Shufflenet
- SqueezeNet
- VGG
- ZFNet
You can find examples for converting a model through ONNX -> CoreML here
List of ONNX operators supported in Core ML 2.0 via the converter
List of ONNX operators supported in Core ML 3.0 via the converter
Some of the operators are partially compatible with Core ML, for example gemm with more than 1 non constant input is not supported in Core ML 2, or scale as an input for upsample layer is not supported in Core ML 3 etc.
For unsupported ops or unsupported attributes within supported ops, Core ML custom layers or custom functions can be used.
See the testing script tests/custom_layers_test.py on how to produce Core ML models with custom layers and custom functions.
Copyright © 2018 by Apple Inc., Facebook Inc., and Prisma Labs Inc.
Use of this source code is governed by the MIT License that can be found in the LICENSE.txt file.