Problem of genderage.onnx converted to .tflite

[MMantis]

I want to use genderage.onnx to recognize the age and gender of photos on Android, so I used the onnx2tflite tool to convert genderage.onnx to genderage.tflite, and loaded it with TensorFlow Lite. It ran normally and output the results, but the results were not accurate at all. I want to know where I went wrong. Here is my kotlin code
<
val interpreter = Interpreter(FileUtil.loadMappedFile( applicationContext , "genderage.tflite"), options )
val inputImageProcessor =
ImageProcessor.Builder()
.add(ResizeOp(inputImageSize, inputImageSize, ResizeOp.ResizeMethod.BILINEAR))
.add(NormalizeOp(0f, 255f))
.build()
val tensorInputImage = TensorImage.fromBitmap(image)
val ageOutputArray = Array(1){ FloatArray(3) }
val processedImageBuffer = inputImageProcessor.process(tensorInputImage).buffer
interpreter?.run(
processedImageBuffer,
ageOutputArray
)
Log.d("test","predictAge out [0]:${ageOutputArray[0][0]} [1]:${ageOutputArray[0][1]} [2]:${ageOutputArray[0][2]}")
Log.d("test","age:${ageOutputArray[0][2] * 100}")

The out put all ways like :
| predictAge out [0]:0.36035335 [1]:-0.3603145 [2]:0.35510007
| age:35.510006
The three output values are always around 0.35