Similar output with completely different inputs in facenet machine studying mannequin in iOS utilizing tensorflow lite

I’m implementing face recognition and clustering utility in iOS, For that first I’ve detected faces from picture and cropped face photos utilizing MLKit, That works nicely.

Now I’m giving cropped face photos in face recognition machine studying fashions utilizing tensorflow lite to get vectors for face.

However On this second step, I get similar output for all face photos, To date I’ve tried three completely different variations of facenet machine studying fashions.

Beneath is the code to initialise machine studying mannequin:

init() throws {
        do {
            // Initialize an interpreter with the mannequin.
            interpreter = strive Interpreter(modelPath: modelPath)
            // Allocate reminiscence for the mannequin's enter `Tensor`s.
            strive interpreter.allocateTensors()
            
            // Get enter particulars
            let inputDetails = strive interpreter.enter(at: 0)
            inputWidth = inputDetails.form.dimensions[1]
            inputHeight = inputDetails.form.dimensions[2]

            // Print enter form
            print("Enter Form: (inputDetails.form), inputWidth: (inputWidth), inputHeight: (inputHeight)")
        } catch {
            throw error
        }
    }

Beneath is the code to course of cropped face photos:

func generateVectors(photos: [UIImage]) {
        do {
            for picture in photos {
                guard let resizedImage = picture.resized(to: CGSize(width: inputWidth, top: inputHeight)),
                      let inputData = resizedImage.normalizedData() else {
                    print("Did not course of picture")
                    proceed
                }
                
                // Feed information to mannequin
                strive interpreter.copy(inputData, toInputAt: 0)

                // Run inference
                strive interpreter.invoke()

                // Get output
                let outputTensor = strive interpreter.output(at: 0)
                // Convert uncooked Knowledge to Float32 array
                let floatArray = outputTensor.information.withUnsafeBytes {
                    Array(UnsafeBufferPointer(
                        begin: $0.baseAddress?.assumingMemoryBound(to: Float32.self),
                        rely: outputTensor.form.dimensions[1] // 128 values
                    ))
                }
                print("Characteristic Vector Output: (floatArray)")
            }
        } catch let error {
            print("Error working mannequin: (error)")
        }
        return output
    }

Beneath is the output vectors:
[-0.009220079, 0.038770936, 0.01804376, 0.0017266792, -0.06855814, 0.07214466, -0.06875174, -0.06868738, -0.09033966, -0.01930608, -0.0195967, 0.011521962, -0.008457049, 0.03948402, -0.002224167, -0.058808904, -0.030869633, -0.009765219, 0.000973374, 0.0037677472, -0.19366269, 0.08309283, -0.047775332, 0.016188627, 0.02456835, 0.01125063, -0.05468251, 0.08173297, 0.19266018, -0.059302196, -0.01258643, 0.28976113, 0.09724123, 0.002565295, -0.0486573, 0.10080463, -0.026181312, -0.014548875, 0.0027476046, -0.010945596, 0.009463895, 0.005252799, 0.019287268, -0.009135995, 0.008196506, -0.0612358, 0.009160928, 0.004444012, -0.010797426, 0.06427721, 0.0635613, -0.00705058, -0.23729105, -0.0038941011, -0.021929592, 0.0070061935, 0.007189498, 0.0051750448, -0.1103151, 0.034162432, 0.038976744, -0.13650523, -0.050892524, 0.04492468, -0.016929258, 0.017758325, -0.007822959, 0.019677926, 0.00964872, 0.0012666954, -0.044016927, -0.034723878, -0.08498321, 0.012788337, -0.041820418, 0.009875813, 0.003703687, -0.0010015147, 0.2554124, 0.046859697, -0.008996924, 0.028291812, -0.0138691915, 0.197871, -0.08906104, -0.003602797, -0.00471391, -0.009401287, 0.040403757, -0.14449984, 0.015253428, -0.0011967644, 0.012509383, -0.029001027, -0.017686743, -0.15465711, -0.010484967, 0.006724833, -0.0010327495, 0.004064801, 0.0033183543, -0.0161573, -0.008383467, -0.00051657186, -0.0041507203, 0.0070093162, -0.17482482, 0.0017419434, -0.005793525, 0.022245506, -0.1609532, 0.012124289, 0.012564862, 0.27895862, 0.013894456, 0.06850863, -0.003990949, -0.045771595, 0.11117715, 0.11178145, 0.15556717, -0.017600833, -0.15439847, -0.0024967825, 0.002401842, -0.0032675255, 0.0061577093, -0.001042465, -0.0018741376, -0.116546564, 0.014505089, 0.030642524, -0.00013455506, -0.00610297, 0.07161295, -0.014800259, -0.18568702, -0.010711309, 0.009508076, 0.012341375, -0.005046241, 0.0009768148, -0.0007504827, -0.057355925, -0.089514, 0.07412282, -0.020803858, -0.0035957384, 0.009036027, -0.01416573, -0.011037713, -0.1024432, -0.0001064163, -0.034000482, -0.0112722535, -0.002430556, 0.005671933, 0.009195163, -0.13230675, -0.0062144296, -0.07031675, -0.0019532659, -0.015630726, 0.0007332002, 0.005727988, 0.024316145, 0.02418903, -0.09510416, -0.006006736, -0.00040619334, 0.18963154, 0.032816295, -0.0025780753, -0.013093517, 0.02123824, 0.005748838, -0.089399874, -0.0964007, -0.0134110255, 0.010199679, -0.07981146, -0.03332103, 0.00030683028, -0.0036552714, 0.16836038, 0.02386042, -0.11319512, 0.10237385, 0.11490546, -0.06683481, -0.083391674, -0.013886787]

I’m not including a number of output right here, as a result of they’re similar and I do not need to add them pointless.

I doubt there’s some problem with processing cropped face photos, As I’ve tried three completely different machine studying fashions, and I’m going through similar problem with all of them.

One of many facenet mannequin I’m utilizing I took from this code:
https://github.com/njordsir/Clustering-faces-android

All the machine studying fashions are giving completely different outputs in contrast to one another, however all of them are giving similar output for each enter picture.

Thanks prematurely!