I’ve created a CoreML model using Create ML tool and have been testing it on a physical device. As part of debugging the app I decided to hardcode the data to match a csv file that I had already ran in the Create ML preview feature.
9 times out of 10, the prediction label and prediction probability match, but 1 out 10 I get a prediction probability of NaN.

I read that initialising the input state with zeroes could impact it, which I’ve now done, but still can’t figure out what I randomly get NaNs.

Has anyone encountered probability issues with Core ML? My model is an activity classifier.

This is how I initialise the state:

func initializeState() -> MLMultiArray? {
        do {
            let state = try MLMultiArray(shape: [400], dataType: .double)
            for i in 0..<state.count {
                state[i] = 0.0
            }
            return state
        } catch {
            print("Error initializing state: (error)")
            return nil
        }
    }

This is how state is declared:

@State var currentState = try? MLMultiArray(
        shape: [(ModelConstants.hiddenInLength +
          ModelConstants.hiddenCellInLength) as NSNumber],
        dataType: MLMultiArrayDataType.double)

The input parameters for the model are declared like this:

let rotX = try? MLMultiArray(
            shape: [ModelConstants.predictionWindowSize] as [NSNumber],
            dataType: MLMultiArrayDataType.double)

and this is how I run the prediction model:

do {
            let prediction = try gymModel!.prediction(motionRotationRateX_rad_s_: rotX!,
                                                       motionRotationRateY_rad_s_: rotY!,
                                                       motionRotationRateZ_rad_s_: rotZ!,
                                                       motionUserAccelerationX_G_: accX!,
                                                       motionUserAccelerationY_G_: accY!,
                                                       motionUserAccelerationZ_G_: accZ!,
                                                       stateIn: currentState!)
            currentState = prediction.stateOut
            if let probability = prediction.labelProbability[prediction.label]{
                print("prediction is: (prediction.label) with (probability*100)")
                
                if probability.isNaN {
                    print("help")

                }
                
                return "(Date()). prediction is: (prediction.label) with (probability*100)"
            }
        } catch {
            print("Error running prediction: (error)")
        }

Have been going around in circles trying to figure out why I randomly get NaN. Any suggestions/ ideas on how to debug further would be massively appreciated
cheers