createFloat32Tensor2DFromInts method

Pointer<Float> createFloat32Tensor2DFromInts(Pointer<Pointer<OrtValue>> inputTensorPointer, { required Pointer<OrtMemoryInfo> memoryInfo, required List<List<int>> values, })

Implementation

Pointer<Float> createFloat32Tensor2DFromInts(
  Pointer<Pointer<OrtValue>> inputTensorPointer, {
  required Pointer<OrtMemoryInfo> memoryInfo,
  required List<List<int>> values,
}) {
  final flatArray = values.expand((i) => i).toList();
  final inputTensorNative = calloc<Float>(flatArray.length);
  for (var i = 0; i < flatArray.length; i += 1) {
    inputTensorNative[i] = flatArray[i].toDouble();
    // Extremely useful for debugging failures, allows comparison of Magika
    // example code's input array to our input array.
    //
    // Magika can be thought of as a model that takes 1536 bytes and returns
    // a 113-length vector of floats.
    //
    // The bytes have to become floats, have leading and trailing whitespace
    // trimmed. These requirements are non-obvious and only were identified
    // through failing tests. Moreover, tests are very sensitive due to the
    // nature of the model and the size of the test files. ex. trimming one
    // whitespace character in html.htm led to it being detected as
    // javascript.
    // print(
    //     'inputTensorNative[$i] => ${flatArray[i]} => ${inputTensorNative[i]}');
  }
  final inputShape = calloc<Int64>(2);
  inputShape[0] = 1;
  inputShape[1] = flatArray.length;
  final ptrVoid = inputTensorNative.cast<Void>();
  final status = createTensorWithDataAsOrtValue(
    inputTensorPointer,
    memoryInfo: memoryInfo,
    inputData: ptrVoid,
    inputDataLengthInBytes: flatArray.length * sizeOf<Float>(),
    inputShape: inputShape,
    inputShapeLengthInBytes: 2,
    onnxTensorElementDataType:
        ONNXTensorElementDataType.ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT.value,
  );
  if (status.isError) {
    final error =
        'Code: ${getErrorCodeMessage(status)}\n'
        'Message: ${consumeErrorMessage(status)}';
    calloc.free(inputTensorNative);
    calloc.free(inputShape);
    throw Exception(error);
  }
  calloc.free(inputShape);
  return inputTensorNative;
}