I met the same problem and I think I've found how to solve it.
First of all, the demo app is designed to use the quantized model, in which the input type and parameters' type are both 8-bit. But for the float version, the input type should be 32-bit. You can see from your error log that the expected bytes is exactly 4 times of the actual bytes!

So I think you need to modify the input type of the demo app to fit the 32-bit input.

Thank you for the reply. @pkurogerjs
However, The image data is already converted to floating point in the TFLiteDemo ImageClassifier.java file

"/** Writes Image data into a {@code ByteBuffer}. */
private void convertBitmapToByteBuffer(Bitmap bitmap) {
if (imgData == null) {
return;
}
imgData.rewind();
bitmap.getPixels(intValues, 0, bitmap.getWidth(), 0, 0, bitmap.getWidth(), bitmap.getHeight());
// Convert the image to floating point.
int pixel = 0;
long startTime = SystemClock.uptimeMillis();
for (int i = 0; i < DIM_IMG_SIZE_X; ++i) {
for (int j = 0; j < DIM_IMG_SIZE_Y; ++j) {
final int val = intValues[pixel++];
imgData.put((byte) ((val >> 16) & 0xFF));
imgData.put((byte) ((val >> 8) & 0xFF));
imgData.put((byte) (val & 0xFF));
}
}
"

Well, I don't agree with the comment in the code. It just extract the RGB value from a pixel into 8-bit bytes.
You can see in ImageClassifier.java that the line "tflite.run(imgData, labelProbArray);" takes a 8-bit format input and a 8-bit format output tensors. You need to change both of them to 32-bit float arrays.

I've successfully run the mobilenet_224 float version. the input is a 4-dim float array and the output put is a 1*num_of_label_types array.

I found this problem by reading the source code in the directory org.tenslow.lite.

Hope this will help you.

I think this exception is because your labelProbArray(line 77) is still a byte array. Change its type to float, please. The exception is thrown when calling something like input.copyTo(output) in Tensorflow Lite's java code, which requires that the type input and output should be the same.

Besides, it seems that for inception v3, the input image size should be 299*299.
I ran this model, the recognition results seems right, but the output probabilities are larger than 1.0. I'm confused about that

@pkurogerjs, Thankyou for the help. It works. I am initially more eager towards performance benchmarking between quantized vs float models.

Any difference between Slim and non-slim one? I have had also the output probabilities are larger than 1.0. and the recognition does work well. It does not return a right answer
@atrah22 Could you share your moditication?
I hvae changed
imgData = new float[DIM_BATCH_SIZE][DIM_IMG_SIZE_X][DIM_IMG_SIZE_Y][DIM_PIXEL_SIZE];
labelProbArray = new float[1][labelList.size()];

new AbstractMap.SimpleEntry<>(labelList.get(i), (labelProbArray[0][i])));

I was able to get it working with inception v3 non-slim 2015 by making @kidsung 's changes plus:

replacing the imgData.put's with:

 imgData[0][i][j][0] = (float) ((val >> 16) & 0xFF);
 imgData[0][i][j][1] = (float) ((val >> 8) & 0xFF);
 imgData[0][i][j][2] = (float) (val  & 0xFF);

commenting out

imgData.rewind();

and

imgData.order(ByteOrder.nativeOrder());

and modifying the declarations:

private float[][][][] imgData = null;
private float[][] labelProbArray = null;

Additionally, I received a tensor length error for the labels, the file had 1001 and (somewhere) it expected 1008, so I filled in 7 lines of foo1, foo2, etc.

Seems like the issue is resolved. Could you please close it if it is?

After modifying ImageClassifier.java , I can run float model from https://github.com/tensorflow/tensorflow/blob/master/tensorflow/contrib/lite/g3doc/models.md;
test the Inception model(Inception V3 Slim 2016)and Mobilenet model(Mobilenet 1.0 224 Float)
the inceptionv3_slim_2016.tflite model inference time is 2260ms,
the mobilenet_v1_1.0_224.tflite model inference time is 500ms and mobilenet_quant_v1_224.tflite is 65ms.
add:
private static final int IMAGE_MEAN = 128;
private static final float IMAGE_STD = 128.0f;
change:
private byte[][] labelProbArray = null;
.............................................................
private float[][] labelProbArray = null;

imgData = ByteBuffer.allocateDirect(
DIM_BATCH_SIZE * DIM_IMG_SIZE_X * DIM_IMG_SIZE_Y * DIM_PIXEL_SIZE);
................................................................
imgData = ByteBuffer.allocateDirect(
4*DIM_BATCH_SIZE * DIM_IMG_SIZE_X * DIM_IMG_SIZE_Y * DIM_PIXEL_SIZE);

labelProbArray = new byte[1][labelList.size()];
.............................................................
labelProbArray = new float[1][labelList.size()];

imgData.put((byte) ((val >> 16) & 0xFF));
imgData.put((byte) ((val >> 8) & 0xFF));
imgData.put((byte) (val & 0xFF));
.............................................................
imgData.putFloat((((val >> 16) & 0xFF)-IMAGE_MEAN)/IMAGE_STD);
imgData.putFloat((((val >> 8) & 0xFF)-IMAGE_MEAN)/IMAGE_STD);
imgData.putFloat(((val & 0xFF)-IMAGE_MEAN)/IMAGE_STD);

if use the inception model,modifying
static final int DIM_IMG_SIZE_X = 224;
static final int DIM_IMG_SIZE_Y = 224;
.............................................................
static final int DIM_IMG_SIZE_X = 299;
static final int DIM_IMG_SIZE_Y = 299;

@atrah22 Is there any result about quantized mode vs. float model ?