What’s New in Core ML, Part 2
What’s New in Core ML, Part 2
What’s New in Core ML, Part 2
How to Get a model
- https://github.com/tf-coreml/tf-coreml
- In collaboration with Google
- Support for custom layers
- Support for Quantized TensorFlow models
- https://github.com/onnx/onnx-coreml
- In collaboration with Facebook and Prisma
- https://github.com/apple/coremltools
- Support for latest Core ML
.mlmodel specification - Quantization utilities
- Flexible shape utilities
Quantization Utilities
Quantization conceptLinear Quantization ( 3 bit )Lookup Table Quantization ( 3 bit )- Decide on precision and algorithm
- Let Core ML Tools work its magic!
Execute JupyterLinear QuantizationOriginal vs Linear quantizedLookup Table QuantizationOriginal vs LoockUpTable quantizedSave (5 MB → 1.3 MB)- Metrics
- Accuracy
- Signal-to-noise ratio
- Visual inspection
- Model-specific metrics
Execute JupyterConvert failed! We need a custom layerMake a custom layerCoreML Graph VisualizationSaveYou have to implement a custom class
