Package apple.coreml.protocol

Interface MLCustomLayer

  • public interface MLCustomLayer
    Protocol for specifying a custom layer implementation.

    • Method Detail

      • encodeToCommandBufferInputsOutputsError

        default boolean encodeToCommandBufferInputsOutputsError​(MTLCommandBuffer commandBuffer,
                                                        NSArray<?> inputs,
                                                        NSArray<?> outputs,
                                                        org.moe.natj.general.ptr.Ptr<NSError> error)
        Optional selector for encoding a shader implementing the custom layer to a command buffer. The execution of the buffer will occur inside Core ML. Providing this function does not guarantee GPU evaluation for each prediction. If not provided, the execution of this layer will always be on the CPU. Should return 'YES' on sucessfully encoding, or 'NO' and set an error if the encoding fails.

      • evaluateOnCPUWithInputsOutputsError

        boolean evaluateOnCPUWithInputsOutputsError​(NSArray<? extends MLMultiArray> inputs,
                                            NSArray<? extends MLMultiArray> outputs,
                                            org.moe.natj.general.ptr.Ptr<NSError> error)
        For the given inputs, populate the provide output memory. This will be called for each evaluation performed on the CPU. The inputs and outputs will have the shapes of the most recent call to outputShapesForInputShapes. The memory for both input and output arrays are preallocated, and should not be copied or moved. The implementation should not alter the inputs. Should return 'YES' on sucess, or 'NO' and set the error on failure.

      • initWithParameterDictionaryError

        java.lang.Object initWithParameterDictionaryError​(NSDictionary<java.lang.String,​?> parameters,
                                                  org.moe.natj.general.ptr.Ptr<NSError> error)
        Initialize the custom layer implementation. The dictionary contains the contents of the 'parameters' map from the model specification. This function is called once on model load. We expect the implementation to return 'nil' and set an error in the event of failure initializing the object.

      • outputShapesForInputShapesError

        NSArray<? extends NSArray<? extends NSNumber>> outputShapesForInputShapesError​(NSArray<? extends NSArray<? extends NSNumber>> inputShapes,
                                                                               org.moe.natj.general.ptr.Ptr<NSError> error)
        For the given input shapes, the implementation needs to return the output shape produced by this layer. This will be called at least once at load time, and will be called any time the size of the inputs changes in a call to predict. Consumes and returns 5D arrays of shapes in the usual Core ML convention - (Sequence, Batch, Channel Height, Width). See the Core ML neural network protobuf specification for more details about how layers use these dimensions. This will get called at load and run time. In the event of an error the implementation should return 'nil' and set an error.

      • setWeightDataError

        boolean setWeightDataError​(NSArray<? extends NSData> weights,
                           org.moe.natj.general.ptr.Ptr<NSError> error)
        The data encoded in the 'weights' field of the model specification will be loaded and made available to the implementation here, with one entry of the array for each entry in the 'repeated' weights field in the Core ML neural network specification . This is called when the model is loaded, but is a separate call from the initialization. The pointer to weights should be stored, but modifying or copying its contents can significantly increase an app's memory footprint. The implementation should return 'YES' on success, or return 'NO' and set an error in the event of a failure.