How to apply Sigmoid on MLMultiarray elements?

Question

for a torch tensor in python I can use
y = x[i].sigmoid()

how do I replicate this using MLMultiArray in Swift ?
I can get the array by VNCoreMLFeatureValueObservation.featureValue.multiArrayValue
now I want to apply sigmoid on this

the sigmoid function I'm using is:

func sigmoid(z: Double) -> Double {
        return 1.0 / (1.0 + exp(-z))
    }

Answer 1

Typed this from memory, so there may be mistakes, but basically you grab a pointer to the memory and then loop through it:

let ptr = multiArray.dataPointer.assumingMemoryBound(to: Double.self)
for i in 0..<multiArray.count {
  ptr[i] = sigmoid(ptr[i])
}

However, it's much simpler to add the sigmoid operation to the Core ML model itself and let Core ML worry about it. You can do this by adding the sigmoid operation to the original model before you convert it, or afterwards (see my book Core ML Survival Guide for instructions on how to do this).

Answer 2

I didn't test this function, but I think that it can achieve what you want

public func sigmoid_multiarray(inputs: [MLMultiArray], outputs: [MLMultiArray]) {
for i in 0..<inputs.count {
    let input=inputs[i]
    let output=outputs[i]
    
    let count = input.count
    let iptr = UnsafeMutablePointer<Float>(OpaquePointer(input.dataPointer))
    let optr = UnsafeMutablePointer<Float>(OpaquePointer(output.dataPointer))
    
    // output = -input
    vDSP_vneg(iptr, 1, optr, 1, vDSP_Length(count))

    // output = exp(-input)
    var countAsInt32 = Int32(count)
    vvexpf(optr, optr, &countAsInt32)

    // output = 1 + exp(-input)
    var one: Float = 1
    vDSP_vsadd(optr, 1, &one, optr, 1, vDSP_Length(count))
    
    // output = 1 / 1 + exp(-input)
    var one_ar = [Float] (repeating: 1, count: input.count)
    vvdivf(optr, &one_ar, optr, &countAsInt32)
}

}