(including Beta 4 changes)
Neural Networks:
- Additional neural network layers: linear option in upsampling, reduction with axis mode, slice, spaces to depth, depth to spaces
- Separate preprocessing (channel bias and mean image subtraction) for multiple input images is now fully supported
API:
- New API to optionally constrain computation to use the CPU only: see MLModel.prediction(from:options:)
- New API to compile models on device: see MLModel.compileModel(at: )
- Additional API to query for details about the model inputs and outputs: see Additional Feature Constraints in MLFeatureDescription documentation
Model Validation:
- The model compiler (and Xcode viewer) now requires validated models with all type information filled in.
- You may see errors in Xcode or from the compiler with statements such as:
"validator error: Description of multiarray feature '__feature_vector__' has invalid or unspecified dataType. It must be specified as DOUBLE, FLOAT32 or INT32"
- If you see this or a similar validation error you must update your model. You can do this by either
- Re-converting the model from its source using the latest coremltools 0.5.1
- Try running the following on the macOS 10.13 command line:
xcrun coremlcompiler repair <path/to/invalid.mlmodel>
Xcode:
- Generated code now includes availability macros
- Known Bug: Availability for macOS/OSX is incorrect (13.0 instead of 10.13). This will be fixed in an upcoming release
Fixes:
- Removed memory leaks when a model outputs an image.
- Removed memory leaks in SVM models.
- Fixes to allow the use of pipeline classifiers where the class labels were integers.
- Fixes to input and output layer selection in neural networks.
- Float32 input and output MLMultiArrays for Neural Networks now fully supported
Coremltools - 0.5.1
- The new permanent home for the documentation is https://apple.github.io/coremltools/. The old documentation links on PyPI have been removed.
- Keras 2 converter now supports dilated convolution, depth-wise and separable convolution.
- Keras 1.2 converter now supports atrous convolution.
- Image outputs now supported in the Python bindings for model prediction.
- Fixes to the 'add' operator in the neural network builder.