Null check operator used on a null value in the method interpreter.run([input], [output]);

[peterbullmatti]

Im newbie to Tensorflow world， I learned tutorial and creat such method for text embedding, the print looks ok like this:

I/flutter (23098): input [10000, 10001, 0, 0, 0, 0, ... 
I/flutter (23098): output check [0.0, 0.0, 0.0, ...
I/flutter (23098): interpreter true
I/flutter (23098): inputShape: [1, 384]
I/flutter (23098): outputShape: [1, 384]
I/flutter (23098): input length: 384
I/flutter (23098): output length: 384

Here is my method:

 Future<List<double>> runModel(
      Interpreter interpreter, String inputText, List<int> inputShape) async {
    List<int> input = await stringToInputVector(inputText, inputShape[1]);

    List<int> outputShape = interpreter.getOutputTensor(0).shape;
    List<double> output =
        List<double>.generate(outputShape[0] * outputShape[1], (index) => 0.0);

    print('input $input');
    print('output check $output');
    print('interpreter ${interpreter.isAllocated}');
    print('inputShape: ${inputShape}');
    print('outputShape: ${outputShape}');
    print('input length: ${input.length}');
    print('output length: ${output.length}');
    
    interpreter.allocateTensors();
    interpreter.run([input], [output]);
    return output;
  }

  Future<String> extractRelevantInfo(String question, List<String> historyTexts,
      Interpreter modelInterpreter, List<int> inputShape) async {
    List<double> questionVector =
        await runModel(modelInterpreter, question, inputShape);
    double maxSimilarity = -1;
    String mostRelevantText;
    for (String historyText in historyTexts) {
      List<double> historyTextVector =
          await runModel(modelInterpreter, historyText, inputShape);
      double similarity =
          await computeCosineSimilarity(questionVector, historyTextVector);

      if (similarity > maxSimilarity) {
        maxSimilarity = similarity;
        mostRelevantText = historyText;
      }
    }

    return mostRelevantText;
  }

But now I met error:

E/flutter (23098): [ERROR:flutter/runtime/dart_vm_initializer.cc(41)] Unhandled Exception: Null check operator used on a null value
E/flutter (23098): #0      Interpreter.runForMultipleInputs (package:tflite_flutter/src/interpreter.dart:196:41)
E/flutter (23098): #1      Interpreter.run (package:tflite_flutter/src/interpreter.dart:157:5)
E/flutter (23098): #2      TensorflowAPI.runModel (package:accounting/api_services/tensorflow_api.dart:135:17)

I have no idea where is this error from, here is how I use this metho, the model I tried mobilebert_1_default.tflite and albert_lite_base_squadv1_1.tflite and lite-model_universal-sentence-encoder-qa-ondevice.tflite for text embedding, all of them have this error

await tensorflowAPI.extractRelevantInfo(
        userQuestion, historyStrings, interpreter, inputShape);

[JSYRD]

If your input is not a single tensor, try useinterpreter.runForMultipleInputs(input, output), rather than interpreter.run([input], [output]).
And also, if it's a single tensor, try useinterpreter.run(input, output) directly, and do not embrace them with [].

[Mroles]

I also encountered the same issue, but for object detection models. I am following the example in the tflite_flutter repo. After playing around a bit, I realized my outputs are null beyond index 0. Is there something I am missing?

Future<void> processImage() async {
    if (imagePath != null) {
      final imageData = File(imagePath!).readAsBytesSync();
      image = img.decodeImage(imageData);
      setState(() {});

      final imageInput = img.copyResize(
        image!,
        width: WIDTH,
        height: HEIGHT,
      );

      final imageMatrix = List.generate(
        imageInput.height,
        (y) => List.generate(
          imageInput.width,
          (x) {
            final pixel = imageInput.getPixel(x, y);
            return [img.getRed(pixel), img.getGreen(pixel), img.getBlue(pixel)];
          },
        ),
      );
      runInference(imageMatrix);
    }
  }

and here is the runInference method

  
  Future<void> runInference(
    List<List<List<num>>> imageMatrix,
  ) async {
    final input = [imageMatrix];

    //output shape [1,100, 4]
    final output = [
      List.generate(
        100,
        (index) => List<double>.filled(4, 0.0),
      )
    ];
    // Run inference
    interpreter.run(input, output);

    // Get first output tensor
    final result = output.first;

    setState(() {});

    interpreter.close();
  }
  

[kvanry]

@Mroles

I believe it's a bug with the code, I forked and fixed. The problem is it's trying to load a single buffer for each input even though there's multiple buffers for the output, so runForMultiple is expecting something like

Map<int, Buffer>

But Object detection is

Map<int, Map<int, Buffer>>

I forked and fixed for my purposes, but hope that helps. This repo is no longer maintained and the TF team at Google has a new repo, though the fix isn't there yet either...

https://github.com/tensorflow/flutter-tflite