public async Task<IEnumerable<Prediction>> ClassifyAsync(byte[] bytes)
{
var mappedByteBuffer = GetModelAsMappedByteBuffer();
//var interpreter = new Xamarin.TensorFlow.Lite.Interpreter(mappedByteBuffer);
System.Console.WriteLine($"Running Tensorflow interpreter");
System.Console.WriteLine($"Tensorflow runtime version {TensorFlowLite.RuntimeVersion()}");
System.Console.WriteLine($"Tensorflow schema version {TensorFlowLite.SchemaVersion()}");
var interpreterOptions = new Interpreter.Options();
//TODO: Pass from UI?
var numThreads = 1;
interpreterOptions.SetNumThreads(numThreads);
//TODO: Look into use of GPU delegate vs NNAPI
// https://developer.android.com/ndk/guides/neuralnetworks
interpreterOptions.SetUseNNAPI(true);
interpreterOptions.SetAllowFp16PrecisionForFp32(true);
//var interpreter = new Interpreter(mappedByteBuffer);
var interpreter = new Interpreter(mappedByteBuffer, interpreterOptions);
var tensor = interpreter.GetInputTensor(0);
var shape = tensor.Shape();
var width = shape[1];
var height = shape[2];
var labels = await LoadLabelsAsync(LabelsFileName);
var byteBuffer = GetPhotoAsByteBuffer(bytes, width, height);
//var outputLocations = new float[1][] { new float[labels.Count] };
var outputLocations = new[] { new float[labels.Count] };
var outputs = Java.Lang.Object.FromArray(outputLocations);
interpreter.Run(byteBuffer, outputs);
var classificationResult = outputs.ToArray<float[]>();
var result = new List<Prediction>();
for (var i = 0; i < labels.Count; i++)
{
var label = labels[i];
result.Add(new Prediction(label, classificationResult[0][i]));
}
//TODO: Consider using this event or MediatR to return results to view model
//https://blog.duijzer.com/posts/mvvmcross_with_mediatr/
PredictionCompleted?.Invoke(this, new PredictionCompletedEventArgs(result));
return result;
}
void Start()
{
var options = new Interpreter.Options()
{
threads = 2,
gpuDelegate = null,
};
interpreter = new Interpreter(FileUtil.LoadFile(fileName), options);
interpreter.ResizeInputTensor(0, new int[] { 1, 28, 28, 1 });
interpreter.AllocateTensors();
inputBuffer = new ComputeBuffer(28 * 28, sizeof(float));
}
private void LoadTensorflowModel()
{
var assets = Android.App.Application.Context.Assets;
AssetFileDescriptor fileDescriptor = assets.OpenFd(MODEL_FILE);
FileInputStream inputStream = new FileInputStream(fileDescriptor.FileDescriptor);
FileChannel fileChannel = inputStream.Channel;
long startOffset = fileDescriptor.StartOffset;
long declaredLength = fileDescriptor.DeclaredLength;
_model = fileChannel.Map(FileChannel.MapMode.ReadOnly, startOffset, declaredLength);
Interpreter.Options options = new Interpreter.Options();
options.SetNumThreads(THREAD_NUM);
_interpreter = new Interpreter(_model, options);
}
void Start()
{
// recursively add children to list
componentList.Add(character);
AddDescendants(character.transform, componentList);
// options
var options = new Interpreter.Options()
{
threads = 2,
gpuDelegate = CreateGpuDelegate()// useGPU ? CreateGpuDelegate() : null,
};
// load decoder
string path = Path.Combine(Application.streamingAssetsPath, tfliteFileName);
Debug.Log(path);
interpreter = new Interpreter(FileUtil.LoadFile(path), options);
Debug.LogFormat(
"InputCount: {0}, OutputCount: {1}",
interpreter.GetInputTensorCount(), // always 1,1
interpreter.GetOutputTensorCount()
);
// add variable for input size
inputs = new float[2];
// load mean features from file
//path = Path.Combine(Application.streamingAssetsPath, vectorFileName);
//string[] numbers = File.ReadAllLines(path);
TextAsset meanfile = Resources.Load(vectorFileName) as TextAsset;
string[] numbers = meanfile.text.Split(new string[] { "\n" }, StringSplitOptions.None);
vmean = new float[componentList.Count * 4 + 3];
for (int i = 0; i < numbers.Length; i++)
{
if (float.TryParse(numbers[i], out float parsedValue))
{
vmean[i] = parsedValue;
}
}
}
//bool _isStoppedDrag = true;
//bool _hasPrinted = false;
public void Start()
{
shouldClearText = true;
var options = new Interpreter.Options()
{
threads = 2,
gpuDelegate = null,
};
interpreter = new Interpreter(FileUtil.LoadFile(fileName), options);
interpreter.ResizeInputTensor(0, new int[] { 1, 28, 28, 1 });
interpreter.AllocateTensors();
inputBuffer = new ComputeBuffer(28 * 28, sizeof(float));
//_hasPrinted = false;
//GameObject inputImage = GameObject.Find("Input Image");
//LineDrawer lineDrawer = inputImage.GetComponent<LineDrawer>();
//_isStoppedDrag = lineDrawer.isStoppedDrag;
}
void Start()
{
Debug.LogFormat("TensorFlow Lite Verion: {0}", Interpreter.GetVersion());
var options = new Interpreter.Options()
{
threads = 2,
};
interpreter = new Interpreter(model.bytes, options);
int inputCount = interpreter.GetInputTensorCount();
int outputCount = interpreter.GetOutputTensorCount();
for (int i = 0; i < inputCount; i++)
{
Debug.LogFormat("Input {0}: {1}", i, interpreter.GetInputTensorInfo(i));
}
for (int i = 0; i < inputCount; i++)
{
Debug.LogFormat("Output {0}: {1}", i, interpreter.GetOutputTensorInfo(i));
}
}
public async Task <PredictionResult> PredictAsync(Stream stream)
{
//https://www.tensorflow.org/lite/models/image_classification/android
//Code https://github.com/tensorflow/examples/blob/master/lite/examples/image_classification/android/app/src/main/java/org/tensorflow/lite/examples/classification/tflite/Classifier.java
//Kotlin example https://medium.com/@teresa.wu/tensorflow-image-recognition-on-android-with-kotlin-cee8d977ae9
// - code at https://github.com/teresawu/random
//var tf = new Tensor();
_labels = _labels ?? (_labels = await LoadLabels(LabelsFileName));
//_model = _model ?? (_model = await LoadModel(ModelFileName));
//_modelFile = _modelFile ?? (_modelFile = await LoadModelFile(ModelFileName));
_modelBuffer = _modelBuffer ?? (_modelBuffer = await LoadModelFileBuffer(ModelFileName));
//var device = Android.App.Application.Context.Dev
//Xamarin.T.TensorFlow.Lite.
//var x = new TensorFlowInferenceInterface()
//TODO: Look into GPU - GpuHelperDelegate g;
//TODO: Pass from UI?
var numThreads = 1;
var tfOptions = new Interpreter.Options();
tfOptions.SetNumThreads(numThreads);
//tfOptions.SetUseNNAPI(true);
tfOptions.SetAllowFp16PrecisionForFp32(true);
//TODO: Look into use of GPU delegate vs NNAPI
// https://developer.android.com/ndk/guides/neuralnetworks
//https://stackoverflow.com/questions/30716027/c-sharp-equivalent-of-java-memory-mapping-methods
//ByteBuffer.FromArray()
//var buff = new ByteBuffer()
var interpreter = new Interpreter(_modelBuffer, tfOptions);
//var imgData = new int[ImageWidth * ImageHeight];
//bitmap.GetPixels(argbPixelArray, 0, bitmap.Width, 0, 0, bitmap.Width, bitmap.Height);
var bitmap = CreateBitmap(stream);
var imgData = ConvertBitmapToByteBuffer(bitmap);
//var normalizedPixelComponents = new float[argbPixelArray.Length * 3];
//var arr = new float[1, _labels.Count];
//var arr1 = new float[_labels.Count];
//var labelProbArray = new float[_labels.Count];
//var objectArray = new object[1];
var labelProbArray = new float[1][] {
new float[_labels.Count]
};
//https://forums.xamarin.com/discussion/1930/creating-arrays-of-custom-java-objects-in-jni
//var labelProbArray = Java.Lang.Object.FromArray<IJavaObject>();
//var javaProbArray = Java.Lang.Object.FromArray(labelArray);
//var javaObjectArray = Java.Lang.Object.FromArray(objectArray);
var javaLabelProbArray = Java.Lang.Object.FromArray(labelProbArray);
//var labelProbArray = Java.Lang.Object.To(labelArray);
var javaLabelProbArray1 = Java.Lang.Object.FromArray(labelProbArray[0]);
//var javaLabelProbArray1 = new
var javaProbArrayContainer = new Java.Lang.Object[] { javaLabelProbArray1 };
var floatArrayArray = new Java.Lang.Object(JNIEnv.NewArray(labelProbArray), JniHandleOwnership.TransferLocalRef);
//var array = JNIEnv.NewArray(arr);
//Java.Lang.Object[] values = arr
// .Select(b => new Java.Lang.Object(b.ToLocalJniHandle(), JniHandleOwnership.TransferLocalRef))
// .ToArray();
//var intPtrFloatArray = JNIEnv.NewArray<float>(labelProbArray);
//var labelProbArrayContainer = new Java.Lang.Object[] { javaLabelProbArray };
//values2[0] = labelProbArray;
//Array<ByteArray>
try
{
// Run the inference
//Try RunForMultipleInputsOutputs
// ?? GetOutputTensor
// - see https://github.com/tensorflow/tensorflow/issues/25751
// or https://devblogs.microsoft.com/xamarin/android-apps-tensorflow/
interpreter.Run(imgData, floatArrayArray);
}
catch (Exception e)
{
System.Console.WriteLine(e);
throw;
}
//var outputArrayFirstDimension = javaProbArrayContainer[0];
//var outputs = outputArrayFirstDimension.ToArray<float[][]>();
//https://stackoverflow.com/questions/17305522/mono-for-android-binding-jagged-array
//float[][] payload = (float[][])JNIEnv.GetArray(floatArrayArray, JniHandleOwnership.DoNotTransfer, typeof(float[]));
//float[][] payload = JNIEnv.GetArray<float[]>(javaLabelProbArray);
float[][] outputs = JNIEnv.GetArray <float[]>(floatArrayArray.Handle);
for (int i = 0; i < outputs[0].Length; i++)
{
if (!float.IsNaN(outputs[0][i]))
{
}
if (outputs[0][i] > 0)
{
}
}
/*
* using (var byteArrayArray = new Java.Lang.Object(JNIEnv.NewArray(data), JniHandleOwnership.TransferLocalRef))
* {
* Console.WriteLine("# jonp [[b: {0}", JNIEnv.GetClassNameFromInstance(byteArrayArray.Handle));
* byte[][] data2 = JNIEnv.GetArray<byte[]>(byteArrayArray.Handle);
*/
//https://stackoverflow.com/questions/6594250/type-cast-from-java-lang-object-to-native-clr-type-in-monodroid
var obj = javaLabelProbArray;
var propertyInfo = obj.GetType().GetProperty("Instance");
var result = propertyInfo == null
? null
: propertyInfo.GetValue(obj, null) as float[][];
//var outputs = javaLabelProbArray.ToArray<float[][]>();
// Find the best classifications.
var recognitions = new List <Recognition>(_labels.Count);
for (var i = 0; i < _labels.Count; i++)
{
recognitions.Add(new Recognition
(
i.ToString(),
_labels[i],
labelProbArray[0][i],
null
));
}
// Sort high-to-low via confidence
//Array.Sort(results, (x, y) => y.Confidence.CompareTo(x.Confidence));
var orderedRecognitions = recognitions.OrderByDescending(x => x.Confidence).ToList();
foreach (var recognition in orderedRecognitions)
{
System.Console.WriteLine($"Result {recognition.Title} with confidence {recognition.Confidence}");
}
/*
* var pq =
* new PriorityQueue<Recognition>(
* 3,
* new Comparator<Recognition>() {
* @Override
* public int compare(Recognition lhs, Recognition rhs)
* {
* // Intentionally reversed to put high confidence at the head of the queue.
* return Float.compare(rhs.getConfidence(), lhs.getConfidence());
* }
* });
* for (int i = 0; i < labels.size(); ++i) {
* pq.add(
* new Recognition(
* "" + i,
* labels.size() > i? labels.get(i) : "unknown",
* getNormalizedProbability(i),
* null));
* }
* final ArrayList<Recognition> recognitions = new ArrayList<Recognition>();
* int recognitionsSize = Math.min(pq.size(), MAX_RESULTS);
* for (int i = 0; i<recognitionsSize; ++i) {
* recognitions.add(pq.poll());
* }
*/
return
(await Task.FromResult(
new PredictionResult
{
Success = true,
Answer = $"{orderedRecognitions.First().Title} with confidence {orderedRecognitions.First().Confidence:0.##}"
}));
}
