CoreMLPredictiveModel

open class CoreMLPredictiveModel : PredictiveModel

ENTERPRISE EDITION ONLY

CoreMLPredictiveModel is a Core ML based implementation of the PredictiveModel protocol. Basically the CoreMLPredictiveModel used a Core ML model to return prediction results.

CoreMLPredictiveModel automatically converts between Couchbase Lite data and Core ML data when calling into the MLModel object to return prediction results. All Core ML data types including Int64, Double, String, Dictionary, MultiArray, Image, and Sequence are supported.

When the MLObject has a single input and the input type is Image, CoreMLPredictiveModel will use Vision framework via the VNCoreMLModel to process the input image and call into the MLModel object. The CoreMLPredictiveModel supports all VNObservation types including VNClassificationObservation, VNCoreMLFeatureValueObservation, and VNPixelBufferObservation as mentioned in https://developer.apple.com/documentation/vision/vncoremlrequest. However there is a compatibility limitation when the VNCoreMLModel returns VNCoreMLFeatureValueObservation or VNPixelBufferObservation results that the MLModel must return a single output, otherwise the observation outputs cannot be mapped to the MLModel outputs. When the VNCoreMLModel cannot be used to result the prediction result, CoreMLPredictiveModel will fall back to use the MLModel instead.

When converting blob data to VNPixelBuffer for an input image, only ARGB pixel format is currently supported. However this limitation is applied only when the VNCoreMLModel cannot be used.