/// <summary>
/// Clips tensor values to a maximum L2-norm.
/// </summary>
/// <remarks>
/// <para>
/// Given a tensor <paramref name="x"/>, and a maximum clip value <paramref name="clip_norm"/>, this operation normalizes
/// <paramref name="x"/> so that its L2-norm is less than or equal to <paramref name="clip_norm"/>, along the dimensions
/// given in <paramref name="axes"/>. Specifically, in the default case where all dimensions are used for calculation, if
/// the L2-norm of <paramref name="x"/> is already less than or equal to <paramref name="clip_norm"/>, then <paramref name="x"/>
/// is not modified. If the L2-norm is greater than <paramref name="clip_norm"/>, then this operation returns a tensor of
/// the same type and shape as <paramref name="x"/> with its values set to: <c>t* clip_norm / l2norm(t)</c></para>
/// </remarks>
/// <param name="x">The tensor.</param>
/// <param name="clip_norm">The minimum value to clip by. A 0 - D(scalar) tensor, or a tensor with the same shape as <paramref name="x"/>.</param>
/// <param name="axes">The minimum value to clip by. A 0 - D(scalar) tensor, or a tensor with the same shape as <paramref name="x"/>.</param>
/// <param name="operName">Operation name, optional.</param>
/// <returns>A clipped <see cref="TFOutput">tensor</see>.</returns>
public TFOutput ClipByNorm(TFOutput x, TFOutput clip_norm, TFOutput?axes = null, string operName = null)
{
// https://github.com/tensorflow/tensorflow/blob/r1.2/tensorflow/python/ops/clip_ops.py#L73
var scopeName = MakeName("ClipByNorm", operName);
using (var newScope = WithScope(scopeName)) {
// Calculate L2-norm, clip elements by ratio of clip_norm to L2-norm
var l2norm_inv = Rsqrt(ReduceSum(Mul(x, x), axes, keep_dims: true));
var intermediate = Mul(x, clip_norm);
var tclip = Identity(Mul(intermediate, Minimum(l2norm_inv, Div(Const(new TFTensor(1.0)), clip_norm), operName: operName)));
return(tclip);
}
}
/// <summary>
/// Clips tensor values to a maximum average L2-norm.
/// </summary>
/// <remarks>
/// Given a tensor <paramref name="x"/>, and a maximum clip value <paramref name="clip_norm"/>, this operation
/// normalizes <paramref name="x"/> so that its its average L2-norm is less than or equal to <paramref name="clip_norm"/>.
/// Specifically, if the average L2-norm is already less than or equal to <paramref name="clip_norm"/>, then <paramref name="x"/>
/// is not modified. If the average L2-norm is greater than <paramref name="clip_norm"/>, then this operation returns a tensor of the same
/// type and shape as <paramref name="x"/> with its values set to: <c>t* clip_norm / l2norm_avg(t)</c>. In this case,
/// the average L2-norm of the output tensor is <paramref name="clip_norm"/>.
/// </remarks>
/// <param name="x">The input tensor.</param>
/// <param name="clip_norm">A maximum clipping value.</param>
/// <param name="operName">Name of the oper.</param>
public TFOutput ClipByAverageNorm(TFOutput x, TFOutput clip_norm, string operName = null)
{
// https://github.com/tensorflow/tensorflow/blob/r1.2/tensorflow/python/ops/clip_ops.py#L251
var scopeName = MakeName("ClipByAverageNorm", operName);
using (var newScope = WithScope(scopeName)) {
// Calculate L2-norm per element, clip elements by ratio of clip_norm to
// L2-norm per element
TFOutput n_element = Cast(Size(x), TFDataType.Float);
TFOutput l2norm_inv = Rsqrt(ReduceSum(Mul(x, x), Range(Rank(x))));
TFOutput tclip = Identity(Mul(Mul(x, clip_norm), Minimum(Mul(l2norm_inv, n_element), Div(Const(new TFTensor(1.0)), clip_norm)), operName: operName));
return(tclip);
}
}
public void Should_EvaluateBitwiseAnd(int aValue, int bValue, int expected)
{
using (var graph = new TFGraph())
using (var session = new TFSession(graph))
{
TFOutput a = graph.Placeholder(TFDataType.Int32);
TFOutput b = graph.Placeholder(TFDataType.Int32);
TFOutput y = graph.BitwiseAnd(a, b);
TFTensor[] result = session.Run(new[] { a, b }, new TFTensor[] { aValue, bValue }, new[] { y });
Assert.Equal(expected, (int)result[0].GetValue());
}
}
public static TFGraph ImConvertGraph(TFTensor image, out TFOutput input, out TFOutput output)
{
var graph = new TFGraph();
input = graph.Placeholder(TFDataType.String);
const int W = 28;
const int H = 28;
output = graph.Div(graph.ResizeBilinear(graph.ExpandDims(graph.Cast(
graph.DecodePng(contents: input, channels: 1), DstT: TFDataType.Float),
dim: graph.Const(0)), size: graph.Const(new int[] { W, H })
), y: graph.Const((float)255)
);
return(graph);
}
private static void ConstructGraphToNormalizeImage(out TFGraph graph, out TFOutput input, out TFOutput output, TFDataType destinationDataType = TFDataType.Float)
{
graph = new TFGraph();
input = graph.Placeholder(TFDataType.String);
output = graph.Cast(graph.Div(
x: graph.Sub(
x: graph.ResizeBilinear(
images: graph.ExpandDims(
input: graph.Cast(
graph.DecodeJpeg(contents: input, channels: 3), DstT: TFDataType.Float),
dim: graph.Const(0, "make_batch")),
size: graph.Const(new int[] { W, H }, "size")),
y: graph.Const(Mean, "mean")),
y: graph.Const(Scale, "scale")), destinationDataType);
}
public void Should_CalculateTanhGrad_Correctly()
{
using (TFGraph graph = new TFGraph())
using (TFSession session = new TFSession(graph))
{
TFOutput x = graph.Const(new TFTensor(0.7));
TFOutput y = graph.Tanh(x);
TFOutput dy = graph.Const(new TFTensor(new [] { 1.0 }));
TFOutput grad = graph.TanhGrad(y, dy);
TFTensor [] result = session.Run(new TFOutput [] { }, new TFTensor [] { }, new [] { grad });
double value = (double)result [0].GetValue();
Assert.Equal(0.634739589982459, value, 15);
}
}
/// <summary>
/// Variable node, with a starting initial value. Convenience that registers the init variable to a global queue.
/// </summary>
/// <param name="initialValue">Initial value.</param>
/// <param name="operName">Operation name, optional.</param>
/// <returns>The returning TFOutput returns the handle to the variable.</returns>
/// <remarks>
/// Variables need to be initialized before the main execution so you will typically want to
/// run the session on the variable.
///
/// The init sequence for the variable is stored in the graph, you must manually initialize
/// those by running the session on the global variables.
/// </remarks>
public TFOutput Variable(TFOutput initialValue, string operName = null)
{
var scopeName = MakeName("Variable", operName);
using (var newScope = WithScope(scopeName)) {
var type = initialValue.OutputType;
// This should be VariableV2, but requires Ref support in the C API.
var handle = VarHandleOp(type, new TFShape(GetShape(initialValue)));
using (var aScope = WithScope("Assign")) {
var init = AssignVariableOp(handle, initialValue);
AddInitVariable(init);
return(handle);
}
}
}
/// <summary>
/// Variable node, with a starting initial value.
/// </summary>
/// <param name="initialValue">Initial value.</param>
/// <param name="init">Returns the operation that initializes the value of the variable.</param>
/// <param name="value">Returns the value of the variable.</param>
/// <param name="operName">Operation name, optional.</param>
/// <returns>The returning TFOutput returns the handle to the variable.</returns>
/// <remarks>
/// Variables need to be initialized before the main execution so you will typically want to
/// run the session on the variable
/// </remarks>
public TFOutput Variable(TFOutput initialValue, out TFOperation init, out TFOutput value, string operName = null)
{
var scopeName = MakeName("Variable", operName);
using (var newScope = WithScope(scopeName)) {
var type = initialValue.OutputType;
var handle = VarHandleOp(type, new TFShape(GetShape(initialValue)));
using (var aScope = WithScope("Assign")) {
init = AssignVariableOp(handle, initialValue);
using (var rScope = WithScope("Read")) {
value = ReadVariableOp(handle, type);
return(handle);
}
}
}
}
private TFOutput AddActivation(TFGraph graph, TFOutput y, ActivationFunctionType fusedActivationFunction, Layer nonTrivialActivation)
{
if (fusedActivationFunction != ActivationFunctionType.Linear)
{
y = graph.AddActivation(y, fusedActivationFunction);
}
else if (nonTrivialActivation != null)
{
if (nonTrivialActivation is LeakyRelu leakyRelu)
{
y = graph.Maximum(y, graph.Mul(y, graph.Const(leakyRelu.Slope)));
}
}
return(y);
}
// The inception model takes as input the image described by a Tensor in a very
// specific normalized format (a particular image size, shape of the input tensor,
// normalized pixel values etc.).
//
// This function constructs a graph of TensorFlow operations which takes as
// input a JPEG-encoded string and returns a tensor suitable as input to the
// inception model.
//
private static TFGraph ConstructGraphToNormalizeImage(out TFOutput input, out TFOutput output, TFDataType destinationDataType = TFDataType.Float)
{
var graph = new TFGraph();
input = graph.Placeholder(TFDataType.String);
output = graph.Cast(
graph.ExpandDims(
input: graph.Cast(graph.DecodeJpeg(contents: input, channels: 3), DstT: TFDataType.Float),
dim: graph.Const(0, "make_batch")
)
, destinationDataType
);
return(graph);
}
public void Should_Stack(double[] x, double[] y, double[] z, int?axis, double[,] expected)
{
using (var graph = new TFGraph())
using (var session = new TFSession(graph))
{
var a = graph.Placeholder(TFDataType.Double, new TFShape(2));
var b = graph.Placeholder(TFDataType.Double, new TFShape(2));
var c = graph.Placeholder(TFDataType.Double, new TFShape(2));
TFOutput r = graph.Stack(new[] { a, b, c }, axis: axis);
TFTensor[] result = session.Run(new[] { a, b, c }, new TFTensor[] { x, y, z }, new[] { r });
double[,] actual = (double[, ])result[0].GetValue();
TestUtils.MatrixEqual(expected, actual, precision: 10);
}
}
public static TFOutput AddActivation(this TFGraph graph, TFOutput input, ActivationFunctionType activation)
{
switch (activation)
{
case ActivationFunctionType.Linear:
return(input);
case ActivationFunctionType.Relu:
return(graph.Relu(input));
case ActivationFunctionType.Relu6:
return(graph.Relu6(input));
default:
throw new NotSupportedException();
}
}
public void Should_ClipByValue(double[,] m, double min, double max, double[,] expected)
{
using (var graph = new TFGraph())
using (var session = new TFSession(graph))
{
var matrix = graph.Placeholder(TFDataType.Double);
var clip_min = graph.Placeholder(TFDataType.Double);
var clip_max = graph.Const(new TFTensor(max));
TFOutput y = graph.ClipByValue(matrix, clip_min, clip_max);
TFTensor[] result = session.Run(new[] { matrix, clip_min }, new TFTensor[] { m, min }, new[] { y });
double[,] actual = (double[, ])result[0].GetValue();
Assert.Equal(expected, actual);
}
}
// Convert the image in filename to a Tensor suitable as input to the Inception model.
static TFTensor CreateTensorFromImageFile(byte[] contents)
{
// DecodeJpeg uses a scalar String-valued tensor as input.
var inputTensor = TFTensor.CreateString(contents);
TFGraph graph = new TFGraph();
TFOutput input = graph.Placeholder(TFDataType.String);
TFOutput output = graph.DecodeJpeg(contents: input, channels: 3);
using (var session = new TFSession(graph))
{
var tensor = session.Run(
inputs: new[] { input },
inputValues: new[] { inputTensor },
outputs: new[] { output });
return(tensor[0]);
}
}
public ObjectDetector(string graphPath)
{
using (var graph = new TFGraph())
{
var model = File.ReadAllBytes(graphPath);
graph.Import(new TFBuffer(model));
this._session = new TFSession(graph);
this._detectionInputTensor = graph["image_tensor"][0];
this._detectionOutputTensors = new[]
{
graph["detection_boxes"][0],
graph["detection_scores"][0],
graph["detection_classes"][0],
// graph["num_detections"][0],
};
}
}
public void Train(List <TFTensor> inputs, List <TFTensor> expectedOutputs, int numEpochs)
{
for (int i = 0; i < numEpochs; i++)
{
int test = Random.Next(0, inputs.Count);
TFTensor expectedOutput = expectedOutputs[test];
TFTensor actualOutput = Think(inputs[test]);
TFTensor error;
using (TFGraph graph = new TFGraph())
{
TFOutput expected = graph.Const(expectedOutput);
TFOutput actual = graph.Const(actualOutput);
TFOutput errorOutput = graph.Sub(expected, actual);
TFSession session = new TFSession(graph);
error = session.GetRunner().Run(errorOutput);
}
}
}
// Returns range(0, rank(x)) if reduction_indices is null
TFOutput ReduceDims (TFOutput input, TFOutput? axis = null)
{
if (axis.HasValue)
return axis.Value;
// Fast path: avoid creating Rank and Range ops if ndims is known.
var shape = GetTensorShape (input);
if (shape.Length >= 0) {
// The python code distinguishes between tensor and sparsetensor
var array = new int [shape.Length];
for (int i = 0; i < array.Length; i++)
array [i] = i;
return this.Const (array, TFDataType.Int32);
}
return Range (Const (0), Const (shape.Length), Const (1));
}
/// <summary>
/// Computes the global norm of multiple tensors.
/// </summary>
/// <remarks>
/// <para>
/// Given a tuple or list of tensors <paramref name="tensors"/>, this operation returns the global norm of the elements in all tensors
/// in <paramref name="tensors"/>. The global norm is computed as: <c>global_norm = sqrt(sum([l2norm(t)**2 for t in t_list]))</c>. Any
/// entries in <paramref name="tensors"/> that are of type None are ignored.</para>
/// </remarks>
/// <param name="tensors">The input tensors.</param>
/// <param name="operName">Operation name, optional.</param>
/// <returns>A clipped <see cref="TFOutput">tensor</see>.</returns>
public TFOutput GlobalNorm(TFOutput [] tensors, string operName = null)
{
// https://github.com/tensorflow/tensorflow/blob/r1.2/tensorflow/python/ops/clip_ops.py#L122
var scopeName = MakeName("GlobalNorm", operName);
using (var newScope = WithScope(scopeName)) {
TFOutput [] half_squared_norms = new TFOutput [tensors.Length];
for (int i = 0; i < half_squared_norms.Length; i++)
{
half_squared_norms [i] = L2Loss(tensors [i]);
}
TFOutput half_squared_norm = ReduceSum(Stack(half_squared_norms));
TFOutput norm = Sqrt(Mul(half_squared_norm, Const(2.0)), operName: "global_norm");
return(norm);
}
}
/// <summary>
/// Create learning rate time decay operation.
/// </summary>
protected void CreateDecayOps(float decay, TFOutput initialLearningRate)
{
if (decay > 0)
{
var _decay = _graph.Const(decay, "Decay");
var one = _graph.Const(1f);
_updateOps.Add(_graph.AssignVariableOp(LearningRate,
_graph.Mul(initialLearningRate,
_graph.Div(one,
_graph.Add(one,
_graph.Mul(_decay,
_graph.Cast(Iterations.Read, _decay.OutputType)
)
)
)
)));
}
}
public void TestSession()
{
var status = new TFStatus();
using (var graph = new TFGraph()) {
var feed = Placeholder(graph, status);
var two = ScalarConst(2, graph, status);
var add = Add(feed, two, graph, status);
Assert(status);
// Create a session for this graph
using (var session = new TFSession(graph, status)) {
Assert(status);
// Run the graph
var inputs = new TFOutput [] {
new TFOutput(feed, 0)
};
var input_values = new TFTensor [] {
3
};
var outputs = new TFOutput [] {
new TFOutput(add, 0)
};
var output_values = new TFTensor [] {
3
};
var results = session.Run(runOptions: null,
inputs: inputs,
inputValues: input_values,
outputs: outputs,
targetOpers: null,
runMetadata: null,
status: status);
Assert(status);
var res = results [0];
Assert(res.TensorType == TFDataType.Int32);
Assert(res.NumDims == 0); // Scalar
Assert(res.TensorByteSize == (UIntPtr)4);
Assert(Marshal.ReadInt32(res.Data) == 3 + 2);
}
}
}
private static TFGraph ConstructGraphToNormalizeRawImage(out TFOutput input, out TFOutput output, int resizeWidth, int resizeHeight,
TFDataType destinationDataType = TFDataType.Float)
{
// Some constants specific to the pre-trained model at:
// https://storage.googleapis.com/download.tensorflow.org/models/inception5h.zip
//
// - The model was trained after with images scaled to 224x224 pixels.
// - The colors, represented as R, G, B in 1-byte each were converted to
// float using (value - Mean)/Scale.
const float Mean = 117;
const float Scale = 1;
//var graph = new TFGraph();
//input = graph.Placeholder(TFDataType.UInt8);
//output = graph.Cast(graph.Div(
// x: graph.Sub(
// x: graph.ResizeBilinear(
// images: graph.ExpandDims(
// input: graph.Cast(input, DstT: TFDataType.Float),
// dim: graph.Const(0, "make_batch")),
// size: graph.Const(new int[] { resizeWidth, resizeHeight }, "size")),
// y: graph.Const(Mean, "mean")),
// y: graph.Const(Scale, "scale")), destinationDataType);
//return graph;
var graph = new TFGraph();
input = graph.Placeholder(TFDataType.UInt8);
output = graph.Cast(
graph.ExpandDims(
input: graph.Cast(input, DstT: TFDataType.Float),
dim: graph.Const(0, "make_batch")
)
, destinationDataType
);
return(graph);
}
public void MNISTTwoHiddenLayerNetworkGPUTest()
{
// Parameters
var learningRate = 0.1f;
var epochs = 5;
var numGPUs = 4;
var mnist = new Mnist();
mnist.ReadDataSets("/tmp");
int batchSize = 400;
int numBatches = mnist.TrainImages.Length / batchSize;
using (var graph = new TFGraph())
{
var X = graph.Placeholder(TFDataType.Float, new TFShape(-1, 784));
var Y = graph.Placeholder(TFDataType.Float, new TFShape(-1, 10));
var Xs = graph.Split(graph.Const(0), X, numGPUs);
var Ys = graph.Split(graph.Const(0), Y, numGPUs);
var sgd = new SGD(graph, learningRate, 0.9f);
TFOutput[] costs = new TFOutput[numGPUs];
TFOutput[] accuracys = new TFOutput[numGPUs];
var variablesAndGradients = new Dictionary <Variable, List <TFOutput> >();
for (int i = 0; i < numGPUs; i++)
{
using (var device = graph.WithDevice("/GPU:" + i))
{
(costs[i], _, accuracys[i]) = CreateNetwork(graph, Xs[i], Ys[i], variablesAndGradients);
foreach (var gv in sgd.ComputeGradient(costs[i], colocateGradientsWithOps: true))
{
if (!variablesAndGradients.ContainsKey(gv.variable))
{
variablesAndGradients[gv.variable] = new List <TFOutput>();
}
variablesAndGradients[gv.variable].Add(gv.gradient);
}
}
}
var cost = graph.ReduceMean(graph.Stack(costs));
var accuracy = graph.ReduceMean(graph.Stack(accuracys));
var gradientsAndVariables = new (TFOutput gradient, Variable variable)[variablesAndGradients.Count];
static void Main(string[] args)
{
using (var session = new TFSession())
{
TFGraph graph = session.Graph;
TFOutput a = graph.Const(2);
TFOutput b = graph.Const(3);
Console.WriteLine("a=2, b=3");
// 상수 더하기
TFTensor addProcess = session.GetRunner().Run(graph.Add(a, b));
string Result = addProcess.GetValue().ToString();
Console.WriteLine("a+b={0}", Result);
// 상수 곱하기
TFTensor multiProcess = session.GetRunner().Run(graph.Mul(a, b));
Result = multiProcess.GetValue().ToString();
Console.ReadLine();
}
}
/// <summary>
/// Computes dropout.
/// </summary>
/// <param name="x">A tensor.</param>
/// <param name="keep_prob">A scalar Tensor with the same type as x. The probability that each element is kept.</param>
/// <param name="noise_shape">A 1-D Tensor of type int32, representing the shape for randomly generated keep/drop flags.</param>
/// <param name="seed">Integer seed used for the random distribution, using the TensorFlow SetRandomSeed .</param>
/// <param name="operName">Operation name, optional.</param>
/// <remarks>
/// With probability keep_prob, outputs the input element scaled up by 1 / keep_prob,
/// otherwise outputs 0. The scaling is so that the expected sum is unchanged.
/// </remarks>
public TFOutput Dropout(TFOutput x, double keep_prob, TFShape noise_shape = null, int?seed = null, string operName = null)
{
if (keep_prob < 0 || keep_prob >= 1)
{
throw new ArgumentOutOfRangeException("keep_prob must be a scalar tensor or a float in the range (0, 1], got " + keep_prob);
}
if (keep_prob == 1)
{
return(x);
}
var scopeName = MakeName("dropout", operName);
using (var newScope = WithScope(scopeName)) {
var tkeep_prob = Const(keep_prob);
return(Dropout(x, tkeep_prob, noise_shape, seed, operName));
}
}
// Returns range(0, rank(x)) if reduction_indices is null
TFOutput ReduceDims(TFOutput input, TFOutput?axis = null)
{
if (axis.HasValue)
{
return(axis.Value);
}
// Fast path: avoid creating Rank and Range ops if ndims is known.
var shape = GetTensorShape(input);
if (shape.IsFullySpecified)
{
// The python code distinguishes between tensor and sparsetensor
return(this.Const(shape.ToIntArray(), TFDataType.Int32));
}
// Otherwise, we rely on Range and Rank to do the right thing at run-time.
return(Range(Const(0), Rank(input), Const(1)));
}
private static TFGraph ConstructGraphToNormalizeImage(int channel, out TFOutput input, out TFOutput output)
{
var graph = new TFGraph();
input = graph.Placeholder(TFDataType.String);
output =
//[1 * W * H * RGB / 255]
graph.ExpandDims(
//[W * H * RGB / 255]
graph.Div(
//[W * H * RGB]
graph.Transpose(
//[H * W * RGB]
graph.Cast(graph.DecodeBmp(contents: input, channels: channel), DstT: TFDataType.Float)
, graph.Const(new int[] { 1, 0, 2 }))
, y: graph.Const(255f))
, dim: graph.Const(0));
return(graph);
}
public LinearLayer(TFGraph graph, int inSize, int outSize)
{
var wShape = new TFShape(inSize, outSize);
W = graph.VariableV2(wShape, TFDataType.Float);
TFOutput initFactor = graph.Sqrt(graph.Div(graph.Const(2F), graph.Const((float)inSize)));
TFOutput initialW = graph.Mul(graph.Cast(graph.RandomNormal(wShape, 0, 1), TFDataType.Float), initFactor);
InitW = graph.Assign(W, initialW);
var bShape = new TFShape(outSize);
b = graph.VariableV2(bShape, TFDataType.Float);
TFOutput initialB = graph.Zeros(bShape, TFDataType.Float);
InitB = graph.Assign(b, initialB);
}
public void Apply(TFGraph graph, float learningRate = 0.01F, float beta1 = 0.9F, float beta2 = 0.999F, float eps = 1e-8F)
{
TFOutput tfLearningRate = graph.Const(learningRate);
TFOutput tfBeta1 = graph.Const(beta1);
TFOutput tfBeta2 = graph.Const(beta2);
TFOutput tfEps = graph.Const(eps);
TFOutput tfBeta1Power = graph.Const(beta1);
TFOutput tfBeta2Power = graph.Const(beta2);
for (int i = 0; i < _variables.Count; i++)
{
Operations.Add(graph.ApplyAdam(_variables[i], _m[i], _v[i], tfBeta1Power, tfBeta2Power, tfLearningRate, tfBeta1, tfBeta2, tfEps, _gradients[i], null, true).Operation);
}
}
public IModelResult Eval(IFeatures feature)
{
var inputValuesTensors = feature.GetFeaturesTensors();
TFOutput output = new TFOutput();
TFOutput input = new TFOutput();
lock (_graph)
{
using (var session = new TFSession(Graph))
{
var result = session.Run(
inputs: new[] { input },
inputValues: inputValuesTensors.ToArray(),
outputs: new[] { output });
return(ResultConvertor.Convert(result));
}
}
}
public void tf_random_normal_test()
{
using (var graph = new TFGraph())
using (var session = new TFSession(graph))
{
TFOutput y = graph.RandomNormal(new TFShape(1000, 100), mean: 42, stddev: 0.4, seed: 1337);
TFTensor[] result = session.Run(new TFOutput[] { }, new TFTensor[] { }, new[] { y });
object output = result[0].GetValue();
double[,] actual = (double[, ])output;
Assert.AreEqual(1000, actual.Rows());
Assert.AreEqual(100, actual.Columns());
double actualMean = actual.Mean();
double expectedMean = 42;
Assert.AreEqual(expectedMean, actualMean, 1e-2);
}
}