tflite生成android,tensorflow lite - TFlite Android bytebuffer creation for inference

I have a customly trained mobilenetV2 model which accepts as input a 128x101x3 array of FLOAT32.

In Android (Java), when calling the tflite model inference, the float[x][y][z] input must be converted into a bytebuffer of size 4128101*3 (4 for the float size and the rest for the image size).

The problem is that I there are many ways to make the conversion and I cannot find which is the right one. I can think to add to the bytebuffer fist all the z for each x and y, or I can add all the y for each x and for each z.

For example, let's suppose for sake of simplicity that the 3rd dimension is just a repetition, i.e. [x][y][0] == [x][y][1] == [x][y][2]. Now I can create the bytebuffer like this:

ByteBuffer byteBuffer = ByteBuffer.allocateDirect(4 * 128 * 101 * 3);

byteBuffer.order(ByteOrder.nativeOrder());

for (int i=0; i

for(int j=0; j

byteBuffer.putFloat(myArray[i][j]); // z=0

byteBuffer.putFloat(myArray[i][j]); // z=1

byteBuffer.putFloat(myArray[i][j]); // z=2

}

byteBuffer.rewind();

Or I can create a bytebuffer like this:

for (int i=0; i

int [] inpShapeDim = {1, 1, myArray[0].length, 1};

TensorBuffer valInTnsrBuffer = TensorBuffer.createDynamic(imageDataType); // imageDataType = FLOAT32

valInTnsrBuffer.loadArray(myArray[i], inpShapeDim); //inpShapeDim=1x128x101x3

byteBuffer.put(valInTnsrBuffer.getBuffer());

}

int oneDeltaBufferPosition = byteBuffer.position();

for (int z=0; z<2; deltas++) {

for (int i = 0; i < oneDeltaBufferPosition; i++) {

byteBuffer.put(byteBuffer2.get(i));

}

byteBuffer.rewind();

Both of them are "valid" conversions, but then the inference doesn't work as expected, meaning that the recognition accuracy is not the same as in python.