Onnx Coreml Versions Save

• Miscellaneous Bug fixes

This release includes

• Generic remove unused layer pass to eliminate layers whose output is not being used
• Transforming Conv -> Crop -> BN to Conv -> BN -> Crop to allow Conv-BN fusion
• Includes bug fixes related to expand_dims being generated internally and error messages

Breaking change: converter argument target_ios has been renamed to minimum_ios_deployment_target, as that is more accurate description of what it represents (details here)

This release introduces support for more layers and operators which can be found here. This release adds support for new layers introduces in Core ML 3.

• Added the argument target_ios to choose the Core ML spec version that is produced by the converter. target_ios='13' will enable the converter to use all the new layers added in Core ML 3.
• Added custom_conversion_function option where users can provide their own custom conversion function. Please check example here.
• Supports new model specification version 4 with target iOS 13 or later.
• Find examples for converting PyTorch models into Core ML format here.

Any questions or concerns related to this release can be submitted as an issue and will be review by the team. All comments are welcomed and will be used to improve the existing documentation.

• Added the argument target_ios to choose the Core ML spec version that is produced by the converter. target_ios = '13' will enable the converter to use all the new layers added in Core ML 3.
• (diff with 1.0b2) use rank inference implemented in neural network builder API of coremltools 3.0b6. This helps in converting more models, which previously produced errors.
• (diff with 1.0b2) added support for ops GRU, ROIAlign, TopK.

Known Issues

• Model with Upsample may require using custom_conversion_function, through which value of the scale parameter can be provided. This is due to the fact that CoreML upsample layer supports only statically known scale factors, and in certain cases the ONNX graph has dynamic scale inputs, even though the source program (e.g. pytorch code) uses static scales.

• Added Tensor shape mapping between ONNX and CoreML to improve conversion

• Miscellaneous Bug fixes