Tflite regression model gives different output while input is the same

[am15h]

Tflite regression model gives different output while input is the same

[am15h]

Strange, the interpreter generally works on all tflite models, (not tested for custom ops till now). Please share your flutter doctor -v as well.

These kinds of issues are mostly a result of input/output format mismatch. I recommend using https://pub.dev/packages/tflite_flutter_helper, to eliminate such errors.

[MojtabaYEG]

Here is my flutter doctor -v:

[√] Flutter (Channel stable, v1.17.5, on Microsoft Windows, locale en-CA)
• Flutter version 1.17.5 at C:\Path
• Framework revision 8af6b2f038 (7 weeks ago), 2020-06-30 12:53:55 -0700
• Engine revision ee76268252
• Dart version 2.8.4

[√] Android toolchain - develop for Android devices (Android SDK version 29.0.3)
• Android SDK at C:\Path
• Platform android-29, build-tools 29.0.3
• Java binary at: C:\Path
• Java version OpenJDK Runtime Environment (build 1.8.0_212-release-1586-b04)
• All Android licenses accepted.

[√] Android Studio (version 3.6)
• Android Studio at C:\Path
• Flutter plugin version 47.1.1
• Dart plugin version 192.8052
• Java version OpenJDK Runtime Environment (build 1.8.0_212-release-1586-b04)

[√] Connected device (1 available)
• Android SDK built for x86 • emulator-5554 • android-x86 • Android 10 (API 29) (emulator)

• No issues found!

I also tried tflite_flutter_helper. here is my code:

var interpreterOptions = InterpreterOptions()..addDelegate(NnApiDelegate());
final interpreter = await Interpreter.fromAsset('path',options: interpreterOptions);
var imageBytes = (await rootBundle.load('path')).buffer;
imageLib.Image oriImage = imageLib.decodePng(imageBytes.asUint8List());
imageLib.Image copyImage = imageLib.copyCrop(oriImage, 60, 0, 60, 30);
TensorImage tensorImage = TensorImage.fromImage(copyImage);
TensorBuffer probabilityBuffer = TensorBuffer.createFixedSize([1,1], TfLiteType.float32);
interpreter.run(tensorImage.buffer, probabilityBuffer.buffer);

which doesn't work and gives me the following error:
[ERROR:flutter/lib/ui/ui_dart_state.cc(157)] Unhandled Exception: Bad state: failed precondition

the error referes to this line of code:
interpreter.run(tensorImage.buffer, probabilityBuffer.buffer);

The problem goes away when I use the ImageProcessorBuilder. Then my code is:

var interpreterOptions = InterpreterOptions()..addDelegate(NnApiDelegate());
final interpreter = await Interpreter.fromAsset('path',options: interpreterOptions);
var imageBytes = (await rootBundle.load('path')).buffer;
imageLib.Image oriImage = imageLib.decodePng(imageBytes.asUint8List());
imageLib.Image copyImage = imageLib.copyCrop(oriImage, 60, 0, 60, 30);
TensorImage tensorImage = TensorImage.fromImage(copyImage);
ImageProcessor imageProcessor = ImageProcessorBuilder().add(DequantizeOp(0, 1 / 255.0)).build();
tensorImage = imageProcessor.process(tensorImage);
TensorBuffer probabilityBuffer = TensorBuffer.createFixedSize([1,1], TfLiteType.float32);
interpreter.run(tensorImage.buffer, probabilityBuffer.buffer);

But all the input and output stuff will be incorrect. the input shape should be 13060*3 = 5400 while after using the imageProcessor it's 21600 ( which is 5400 * 4). also the ouput is [35, 204, 170, 74], and all of them are incorrect. Also outputs are integer instead of float. The output should be just 1 float number ( 60.5)

[MojtabaYEG]

I also changed the model from a regression to a classifier by making 118 output neurons instead of 1 neuron. I trained the new classifier CNN. The outputs of the classifier is also different from what I'm getting in python from h5 and tflite model.

@am15h Did you find a chance to make some test experiments? Is it possible to make an example that classifies MNIST with the package?

[MojtabaYEG]

Finally working with tflite_flutter_helper the problem is gone using the following code:

var interpreterOptions = InterpreterOptions()..addDelegate(NnApiDelegate());
final interpreter = await Interpreter.fromAsset('ModelPath',options: interpreterOptions);
var imageBytes = (await rootBundle.load('ImagePath')).buffer;
imageLib.Image oriImage = imageLib.decodePng(imageBytes.asUint8List());
imageLib.Image copyImage = imageLib.copyCrop(oriImage, 60, 0, 60, 30);
TensorImage tensorImage = TensorImage.fromImage(copyImage);
ImageProcessor imageProcessor = ImageProcessorBuilder().add(NormalizeOp(0, 255)).build();
tensorImage = imageProcessor.process(tensorImage);
TensorBuffer probabilityBuffer = TensorBuffer.createFixedSize(<int>[1,1], TfLiteType.float32);
interpreter.run(tensorImage.buffer, probabilityBuffer.buffer);
print(probabilityBuffer.buffer.asFloat32List());