public Tensor CreateVariable(float[] data, long[] shape, string name = "")
{
var arr = new CNTK.NDArrayView(BackendUtil.CastShapeInt(shape), data, DeviceManager.Current);
var v = new CNTK.Variable(BackendUtil.CastShapeInt(shape), VariableKind.Parameter, CNTK.DataType.Float, arr, false, new AxisVector(), false, name, name);
return(Out(v));
}
private Tensor CreateVariable(int[] data, long[] shape, string name = "")
{
shape = BackendUtil.Row2ColMajor(shape);
var arr = new CNTK.NDArrayView(BackendUtil.CastShapeInt(shape), Array.ConvertAll(data, x => (float)x), DeviceManager.Current);
var v = new CNTK.Variable(BackendUtil.CastShapeInt(shape), VariableKind.Input, CNTK.DataType.Float, arr, false, new AxisVector(), false, name, name);
return(Out(v));
}
/// <summary>
/// Detect what's in the given image.
/// </summary>
/// <param name="imagePath">The path of the image to check.</param>
/// <returns>A tuple of probabilities indicating if the image contains a dog or a cat.</returns>
private (double cat, double dog) DetectScene(string imagePath)
{
// load the neural network
if (network == null)
{
var path = @"..\..\..\..\..\models\catdogdetector.model";
network = Function.Load(path, NetUtil.CurrentDevice);
}
// grab the image
var imageData = new float[150 * 150 * 3];
using (var mat = Cv.Cv2.ImRead(imagePath))
{
using (var mat2 = new Cv.Mat(150, 150, mat.Type()))
{
Cv.Cv2.Resize(mat, mat2, new OpenCvSharp.Size(150, 150));
imageData = StyleTransfer.FlattenByChannel(mat2, new float[] { 0, 0, 0 });
}
}
// set up a tensor to hold the image data
var imageValue = new CNTK.NDArrayView(
new int[] { 150, 150, 3 },
imageData,
NetUtil.CurrentDevice);
// set up input and output dictionaries
var inputs = new Dictionary <CNTK.Variable, CNTK.Value>()
{
{ network.Arguments[0], new CNTK.Value(imageValue) }
};
var outputs = new Dictionary <CNTK.Variable, CNTK.Value>()
{
{ network.Outputs[0], null }
};
// run the neural network
network.Evaluate(inputs, outputs, NetUtil.CurrentDevice);
// return the result
var key = network.Outputs[0];
var result = outputs[key].GetDenseData <float>(key);
return(result[0][0], result[0][1]);
}
// The value represents a n-dimensional tensor with 2 dynamic axes: sequence and batch
// It assumes that only the highest 2 axes are dynamic, and all the other axes are static.
public static void CopyTo <T>(this Value value, Variable variable, List <List <T> > data)
{
if ((value.GetDataType() == DataType.Float) && (!typeof(T).Equals(typeof(float))) ||
(value.GetDataType() == DataType.Double) && (!typeof(T).Equals(typeof(double))))
{
throw new ArgumentException("The value type does not match the list type.");
}
// Todo: how to check whether the dynamic axes are the highest 2 axes in the shape.
if (variable.DynamicAxes().Count != 2)
{
throw new ArgumentException("The variable should have 2 dynamic axes.");
}
var variableShape = variable.Shape;
var valueShape = value.Shape;
if (variableShape != value.Shape.SubShape(0, valueShape.Rank - 2))
{
throw new ArgumentException("The variable and value does not have same shape.");
}
// Todo: transform sparse to dense
// Currently only for dense
if ((value.GetStorageFormat() != StorageFormat.Dense))
{
throw new ArgumentException("The value is not in denst format.");
}
var outputNDArrayView = value.Data();
var outputShape = outputNDArrayView.Shape();
var outputShapeRank = outputShape.Rank;
var numOfElementsInSample = variableShape.TotalSize;
var numOfSamplesInSequence = outputShape.GetDimensionSize(outputShapeRank - 2);
var numOfSequences = outputShape.GetDimensionSize(outputShapeRank - 1);
// Copy the data from the output buffer.
// Todo: directly access the data in output buffer?
// Todo: need to map DataBuffer() to C#
NDArrayView cpuOutputNDArrayView;
uint numOfOutputData = outputNDArrayView.Shape().TotalSize;
// Todo: consider mask.
Debug.Assert(numOfElementsInSample * numOfSamplesInSequence * numOfSequences == numOfOutputData);
T[] outputData = new T[numOfOutputData];
if (value.GetDataType() == DataType.Float)
{
cpuOutputNDArrayView = new NDArrayView(outputNDArrayView.Shape(), outputData as float[], numOfOutputData, DeviceDescriptor.CPUDevice);
}
else if (value.GetDataType() == DataType.Double)
{
cpuOutputNDArrayView = new NDArrayView(outputNDArrayView.Shape(), outputData as double[], numOfOutputData, DeviceDescriptor.CPUDevice);
}
else
{
throw new ArgumentException("The data type " + value.GetDataType().ToString() + " is not supported. Only float or double is supported by CNTK.");
}
cpuOutputNDArrayView.CopyFrom(outputNDArrayView);
for (int seqIndex = 0, dataIndex = 0; seqIndex < numOfSequences; seqIndex++)
{
var seqData = new List <T>();
// Todo: consider mask
// Todo: make it more efficient.
for (int i = 0; i < numOfElementsInSample * numOfSamplesInSequence; i++)
{
seqData.Add(outputData[dataIndex++]);
}
data.Add(seqData);
}
}
// Copy Value to List<List<T>> for dense input
// Todo: could this be a extension to Value class??
public void CopyValueTo <T>(string varName, Value value, List <List <T> > sequences)
{
// Todo: deal with GPUDevice.
if (value.Device != DeviceDescriptor.CPUDevice)
{
throw new InvalidOperationException("Currently only CPU device is supported.");
}
if ((value.GetDataType() == DataType.Float) && (!typeof(T).Equals(typeof(float))) ||
(value.GetDataType() == DataType.Double) && (!typeof(T).Equals(typeof(double))))
{
throw new InvalidDataException("The value type does not match the list type.");
}
// Todo: transform sparse to dense
// Currently only for dense
if ((value.GetStorageFormat() != StorageFormat.Dense))
{
throw new InvalidDataException("The value is not in denst format.");
}
var variable = getVariableByName(varName);
var outputNDArrayView = value.Data();
var outputShape = outputNDArrayView.Shape();
var varRank = variable.Shape.Rank;
var valueRank = outputNDArrayView.Shape().Rank;
Debug.Assert(varRank + 2 == valueRank);
var numOfElementsInSample = variable.Shape.TotalSize;
var numOfSamplesInSequence = outputShape.GetDimensionSize(varRank);
var numOfSequences = outputShape.GetDimensionSize(varRank + 1);
//var outputShape = outputVar.Shape().AppendShape(new NDShape(dynamicAxisShape));
// Copy the data from the output buffer.
// Todo: directly access the data in output buffer?
// Todo: need to map DataBuffer() to C#
NDArrayView cpuOutputNDArrayView;
uint numOfOutputData = outputNDArrayView.Shape().TotalSize;
Debug.Assert(numOfElementsInSample * numOfSamplesInSequence * numOfSequences == numOfOutputData);
T[] outputData = new T[numOfOutputData];
if (value.GetDataType() == DataType.Float)
{
cpuOutputNDArrayView = new NDArrayView(outputNDArrayView.Shape(), outputData as float[], numOfOutputData, DeviceDescriptor.CPUDevice);
}
else if (value.GetDataType() == DataType.Double)
{
cpuOutputNDArrayView = new NDArrayView(outputNDArrayView.Shape(), outputData as double[], numOfOutputData, DeviceDescriptor.CPUDevice);
}
else
{
throw new InvalidDataException("The data type " + value.GetDataType().ToString() + " is not supported. Only float or double is supported by CNTK.");
}
cpuOutputNDArrayView.CopyFrom(outputNDArrayView);
for (int seqIndex = 0, dataIndex = 0; seqIndex < numOfSequences; seqIndex++)
{
var seqData = new List <T>();
// Todo: make it more efficient.
for (int i = 0; i < numOfElementsInSample * numOfSamplesInSequence; i++)
{
seqData.Add(outputData[dataIndex++]);
}
sequences.Add(seqData);
}
}
