Tflite model giving different output in Android (ml vision) and Python

Question

I am using ML Vision api to create embeddings from a FaceNet model then comparing cosine distance between two embeddings. The output of Android version and Python differs a lot. Python version is performing way better than android one. What can be the issue? I am using FaceNet model in both.

I am using ML kit for inferencing https://firebase.google.com/docs/ml-kit/android/use-custom-models

I think may be it is caused by the way java reads images as the image array made in android differs with that of same image in python.

Answer 1

So I was stuck at this problem as I was following google documentation at ML vision docs Where the image was converted to float array before feeding it to classifier, It looks like this:

val bitmap = Bitmap.createScaledBitmap(yourInputImage, 224, 224, true)

val batchNum = 0
val input = Array(1) { Array(224) { Array(224) { FloatArray(3) } } }
for (x in 0..223) {
    for (y in 0..223) {
        val pixel = bitmap.getPixel(x, y)
        // Normalize channel values to [-1.0, 1.0]. This requirement varies by
        // model. For example, some models might require values to be normalized
        // to the range [0.0, 1.0] instead.
        input[batchNum][x][y][0] = (Color.red(pixel) - 127) / 255.0f
        input[batchNum][x][y][1] = (Color.green(pixel) - 127) / 255.0f
        input[batchNum][x][y][2] = (Color.blue(pixel) - 127) / 255.0f
    }
}

Then I analysed every step one by one found that the way pixels are fetched are wrong and were completely different from the way python does all of it.

Then I found this way of doing it from this source I changed that function with mine:

    private fun convertBitmapToByteBuffer(bitmap: Bitmap): ByteBuffer {
        val imgData = ByteBuffer.allocateDirect(4 * INPUT_SIZE * INPUT_SIZE * PIXEL_SIZE)
        imgData.order(ByteOrder.nativeOrder())
        val intValues = IntArray(INPUT_SIZE * INPUT_SIZE)


        imgData.rewind()
        bitmap.getPixels(intValues, 0, bitmap.width, 0, 0, bitmap.width, bitmap.height)
        // Convert the image to floating point.
        var pixel = 0
        for (i in 0 until INPUT_SIZE) {
            for (j in 0 until INPUT_SIZE) {
                val `val` = intValues[pixel++]
                imgData.putFloat(((`val`.shr(16) and 0xFF) - IMAGE_MEAN)/IMAGE_STD)
                imgData.putFloat(((`val`.shr(8) and 0xFF)- IMAGE_MEAN)/ IMAGE_STD)
                imgData.putFloat(((`val` and 0xFF) - IMAGE_MEAN)/IMAGE_STD)
            }
        }
        return imgData;
   }

And it worked!