// Use for multiple dimension arrays
static IntPtr SetupMulti(TFDataType dt, long [] dims, Array data, long bytes)
{
var dataHandle = GCHandle.Alloc(data, GCHandleType.Pinned);
if (dims == null)
{
return(TF_NewTensor(dt, IntPtr.Zero, 0, dataHandle.AddrOfPinnedObject(), (UIntPtr)bytes, FreeTensorHandle, GCHandle.ToIntPtr(dataHandle)));
}
else
{
return(TF_NewTensor(dt, dims, dims.Length, dataHandle.AddrOfPinnedObject(), (UIntPtr)bytes, FreeTensorHandle, GCHandle.ToIntPtr(dataHandle)));
}
}
private static void FetchMultiDimensionalArray(Array target, TFDataType dt, IntPtr data, long[] shape)
{
var idx = new int[shape.Length];
for (int i = 0; i < shape.Length; i++)
{
if (shape[i] > Int32.MaxValue)
{
throw new ArgumentOutOfRangeException("Shape can not be longer than 32 bits");
}
}
Copy(target, dt, shape, idx, 0, ref data);
}
/// <summary>
/// Input() is used to instantiate a Keras tensor.
/// </summary>
///
/// <remarks>
/// A Keras tensor is a tensor object from the underlying backend (Theano or TensorFlow), which we
/// augment with certain attributes that allow us to build a Keras model just by knowing the inputs
/// and outputs of the model.
/// </remarks>
///
/// <param name="shape">A shape tuple (integer), including the batch size. For instance,
/// <c>batch_shape= (10, 32)</c> indicates that the expected input will be batches of 10 32-dimensional
/// vectors. <c>batch_shape= (None, 32)</c> indicates batches of an arbitrary number of 32-dimensional
/// vectors.</param>
/// <param name="batch_shape">The batch shape.</param>
/// <param name="name">An optional name string for the layer. Should be unique in a model (do not reuse
/// the same name twice). It will be autogenerated if it isn't provided.</param>
/// <param name="dtype">The data type expected by the input, as a string
/// (`float32`, `float64`, `int32`...)</param>
/// <param name="sparse">A boolean specifying whether the placeholder to be created is sparse.</param>
/// <param name="tensor">Optional existing tensor to wrap into the `Input` layer.
/// If set, the layer will not create a placeholder tensor.</param>
///
public static List <Tensor> Input(int?[] shape = null, int?[] batch_shape = null, string name = null,
TFDataType dtype = KerasSharp.Tools.DEFAULT_DTYPE, bool sparse = false, Tensor tensor = null)
{
if (batch_shape == null && tensor == null)
{
throw new ArgumentException("Please provide to Input either a `shape` or a `batch_shape` argument. Note that " +
"`shape` does not include the batch dimension.");
}
if (shape != null && batch_shape != null)
{
batch_shape = new int?[] { null }
}
private static (string[], int[], bool[], TFShape[], TFDataType[]) GetInputMetaData(TFGraph graph, string[] source, ISchema inputSchema)
{
var tfShapes = new TFShape[source.Length];
var tfTypes = new TFDataType[source.Length];
var colNames = new string[source.Length];
var inputColIndices = new int[source.Length];
var isInputVector = new bool[source.Length];
for (int i = 0; i < source.Length; i++)
{
colNames[i] = source[i];
if (!inputSchema.TryGetColumnIndex(colNames[i], out inputColIndices[i]))
{
throw Contracts.Except($"Column '{colNames[i]}' does not exist");
}
var tfoutput = new TFOutput(graph[colNames[i]]);
if (!TensorFlowUtils.IsTypeSupported(tfoutput.OutputType))
{
throw Contracts.Except($"Input type '{tfoutput.OutputType}' of input column '{colNames[i]}' is not supported in TensorFlow");
}
tfShapes[i] = graph.GetTensorShape(tfoutput);
var type = inputSchema.GetColumnType(inputColIndices[i]);
var shape = tfShapes[i].ToIntArray().Skip(tfShapes[i][0] == -1 ? BatchSize : 0);
if (type.AsVector.DimCount == 1)
{
int valCount = shape.Aggregate((x, y) => x * y);
if (type.ValueCount != valCount)
{
throw Contracts.Except($"Input shape mismatch: Input '{colNames[i]}' has shape {tfShapes[i].ToString()}, but input data is of length {valCount}.");
}
}
else if (shape.Select((dim, j) => dim != type.AsVector.GetDim(j)).Any(b => b))
{
throw Contracts.Except($"Input shape mismatch: Input '{colNames[i]}' has shape {tfShapes[i].ToString()}, but input data is {type.AsVector.ToString()}.");
}
isInputVector[i] = type.IsVector;
tfTypes[i] = tfoutput.OutputType;
var l = new long[tfShapes[i].NumDimensions];
for (int ishape = 0; ishape < tfShapes[i].NumDimensions; ishape++)
{
l[ishape] = tfShapes[i][ishape] == -1 ? BatchSize : tfShapes[i][ishape];
}
tfShapes[i] = new TFShape(l);
}
return(colNames, inputColIndices, isInputVector, tfShapes, tfTypes);
}
/// <summary>
/// Converts a <see cref="TFDataType"/> to a system type.
/// </summary>
/// <param name="type">The <see cref="TFDataType"/> to be converted.</param>
/// <returns>The system type corresponding to the given <paramref name="type"/>.</returns>
public static Type TypeFromTensorType(TFDataType type)
{
switch (type)
{
case TFDataType.Float:
return(typeof(float));
case TFDataType.Float_ref:
return(typeof(float));
case TFDataType.Double:
return(typeof(double));
case TFDataType.Int8:
return(typeof(sbyte));
case TFDataType.Int16:
return(typeof(short));
case TFDataType.Int32:
return(typeof(int));
case TFDataType.Int64:
return(typeof(long));
case TFDataType.UInt8:
return(typeof(byte));
case TFDataType.UInt16:
return(typeof(ushort));
case TFDataType.UInt32:
return(typeof(uint));
case TFDataType.UInt64:
return(typeof(ulong));
case TFDataType.String:
throw new NotSupportedException();
case TFDataType.Bool:
return(typeof(bool));
case TFDataType.Complex128:
return(typeof(Complex));
default:
return(null);
}
}
/// <summary>
/// Input() is used to instantiate a Keras tensor.
/// </summary>
///
/// <remarks>
/// A Keras tensor is a tensor object from the underlying backend (Theano or TensorFlow), which we
/// augment with certain attributes that allow us to build a Keras model just by knowing the inputs
/// and outputs of the model.
/// </remarks>
///
/// <param name="shape">A shape tuple (integer), including the batch size. For instance,
/// <c>batch_shape= (10, 32)</c> indicates that the expected input will be batches of 10 32-dimensional
/// vectors. <c>batch_shape= (None, 32)</c> indicates batches of an arbitrary number of 32-dimensional
/// vectors.</param>
/// <param name="batch_shape">The batch shape.</param>
/// <param name="name">An optional name string for the layer. Should be unique in a model (do not reuse
/// the same name twice). It will be autogenerated if it isn't provided.</param>
/// <param name="dtype">The data type expected by the input, as a string
/// (`float32`, `float64`, `int32`...)</param>
/// <param name="sparse">A boolean specifying whether the placeholder to be created is sparse.</param>
/// <param name="tensor">Optional existing tensor to wrap into the `Input` layer.
/// If set, the layer will not create a placeholder tensor.</param>
///
public static List <Tensor> Input(int?[] shape = null, int?[] batch_shape = null, string name = null,
TFDataType dtype = KerasSharp.Utils.DEFAULT_DTYPE, bool sparse = false, Tensor tensor = null)
{
// https://github.com/fchollet/keras/blob/f65a56fb65062c8d14d215c9f4b1015b97cc5bf3/keras/engine/topology.py#L1416
if (batch_shape == null && tensor == null)
{
throw new ArgumentException("Please provide to Input either a 'shape' or a 'batch_shape' argument. Note that " +
"'shape' does not include the batch dimension.");
}
if (shape != null && batch_shape != null)
{
batch_shape = new int?[] { null }
}
private TFTensor CreateTensorFromImage(byte[] contents, TFDataType dataType = TFDataType.Float)
{
// DecodeJpeg uses a scalar String-valued tensor as input.
using (var tensor = TFTensor.CreateString(contents))
using (var graph = ConstructGraphToNormalizeImage(out TFOutput input, out TFOutput output, dataType))
using (var session = new TFSession(graph))
{
var normalized = session.Run(
inputs: new[] { input },
inputValues: new[] { tensor },
outputs: new[] { output });
return(normalized[0]);
}
}
internal static bool IsTypeSupported(TFDataType tfoutput)
{
switch (tfoutput)
{
case TFDataType.Float:
case TFDataType.Double:
case TFDataType.UInt8:
case TFDataType.UInt16:
case TFDataType.UInt32:
case TFDataType.UInt64:
return(true);
default:
return(false);
}
}
private static unsafe object FetchSimple(TFDataType dt, IntPtr data)
{
switch (dt)
{
case TFDataType.Float:
return(*(float *)data);
case TFDataType.Double:
return(*(double *)data);
case TFDataType.Int8:
return(*(sbyte *)data);
case TFDataType.Int16:
return(*(short *)data);
case TFDataType.Int32:
return(*(int *)data);
case TFDataType.Int64:
return(*(long *)data);
case TFDataType.UInt8:
return(*(byte *)data);
case TFDataType.UInt16:
return(*(ushort *)data);
case TFDataType.UInt32:
return(*(uint *)data);
case TFDataType.UInt64:
return(*(ulong *)data);
case TFDataType.String:
throw new NotImplementedException();
case TFDataType.Bool:
return(*(bool *)data);
case TFDataType.Complex128:
return(*(Complex *)data);
default:
return(null);
}
}
private static PrimitiveDataViewType Tf2MlNetTypeOrNull(TFDataType type)
{
switch (type)
{
case TFDataType.Float:
return(NumberDataViewType.Single);
case TFDataType.Float_ref:
return(NumberDataViewType.Single);
case TFDataType.Double:
return(NumberDataViewType.Double);
case TFDataType.UInt8:
return(NumberDataViewType.Byte);
case TFDataType.UInt16:
return(NumberDataViewType.UInt16);
case TFDataType.UInt32:
return(NumberDataViewType.UInt32);
case TFDataType.UInt64:
return(NumberDataViewType.UInt64);
case TFDataType.Int8:
return(NumberDataViewType.SByte);
case TFDataType.Int16:
return(NumberDataViewType.Int16);
case TFDataType.Int32:
return(NumberDataViewType.Int32);
case TFDataType.Int64:
return(NumberDataViewType.Int64);
case TFDataType.Bool:
return(BooleanDataViewType.Instance);
case TFDataType.String:
return(TextDataViewType.Instance);
default:
return(null);
}
}
/// <summary>
/// Specifies the ndim, dtype and shape of every input to a layer.
/// </summary>
///
/// <remarks>
/// Every layer should expose (if appropriate) an `input_spec` attribute:
/// a list of instances of InputSpec(one per input tensor).
/// A null entry in a shape is compatible with any dimension,
/// a null shape is compatible with any shape.
/// </remarks>
///
/// <param name="dtype">Expected datatype of the input.</param>
/// <param name="shape">Shape tuple, expected shape of the input
/// (may include null for unchecked axes).</param>
/// <param name="ndim">Integer, expected rank of the input.</param>
/// <param name="max_ndim">The maximum rank of the input.</param>
/// <param name="min_ndim">The minimum rank of the input.</param>
/// <param name="axes">Dictionary mapping integer axes to
/// a specific dimension value.</param>
public InputSpec(TFDataType dtype = Tools.DEFAULT_DTYPE, int[] shape = null, int?ndim = null,
int?max_ndim = null, int?min_ndim = null, Dictionary <int, int> axes = null)
{
this.dtype = dtype;
this.shape = shape;
if (shape != null)
{
this.ndim = shape.Length;
}
else
{
this.ndim = ndim;
}
this.max_ndim = max_ndim;
this.min_ndim = min_ndim;
this.axes = axes ?? new Dictionary <int, int>();
}
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);
}
// Convert the image in filename to a Tensor suitable as input to the Inception model.
public static TFTensor CreateTensorFromImage(byte[] image, TFDataType destinationDataType = TFDataType.Float)
{
// DecodeJpeg uses a scalar String-valued tensor as input.
var tensor = TFTensor.CreateString(image);
// Construct a graph to normalize the image
var(graph, input, output) = ConstructGraphToNormalizeImage(destinationDataType);
// Execute that graph to normalize this one image
using (var session = new TFSession(graph))
{
var normalized = session.Run(
inputs: new[] { input },
inputValues: new[] { tensor },
outputs: new[] { output });
return(normalized[0]);
}
}
// Use for single dimension arrays
internal static IntPtr SetupTensor(TFDataType dt, long[] dims, Array data, int start, int count, int size)
{
if (start < 0 || start > data.Length - count)
{
throw new ArgumentException("start + count > Array size");
}
var dataHandle = GCHandle.Alloc(data, GCHandleType.Pinned);
if (dims == null)
{
return(TF_NewTensor(dt, IntPtr.Zero, 0, dataHandle.AddrOfPinnedObject() + start * size, (UIntPtr)(count * size), FreeTensorHandleDelegate, GCHandle.ToIntPtr(dataHandle)));
}
else
{
return(TF_NewTensor(dt, dims, dims.Length, dataHandle.AddrOfPinnedObject() + start * size, (UIntPtr)(count * size), FreeTensorHandleDelegate, GCHandle.ToIntPtr(dataHandle)));
}
}
private static PrimitiveType Tf2MlNetTypeOrNull(TFDataType type)
{
switch (type)
{
case TFDataType.Float:
return(NumberType.R4);
case TFDataType.Float_ref:
return(NumberType.R4);
case TFDataType.Double:
return(NumberType.R8);
case TFDataType.UInt8:
return(NumberType.U1);
case TFDataType.UInt16:
return(NumberType.U2);
case TFDataType.UInt32:
return(NumberType.U4);
case TFDataType.UInt64:
return(NumberType.U8);
case TFDataType.Int8:
return(NumberType.I1);
case TFDataType.Int16:
return(NumberType.I2);
case TFDataType.Int32:
return(NumberType.I4);
case TFDataType.Int64:
return(NumberType.I8);
case TFDataType.Bool:
return(BoolType.Instance);
default:
return(null);
}
}
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);
}
// Convert the image in filename to a Tensor suitable as input to the Inception model.
public static TFTensor CreateTensorFromImageFile(string file, TFDataType destinationDataType = TFDataType.Float)
{
var contents = File.ReadAllBytes(file);
// DecodeJpeg uses a scalar String-valued tensor as input.
using (var tensor = TFTensor.CreateString(contents))
// Construct a graph to normalize the image
using (var graph = ConstructGraphToNormalizeImage(out TFOutput input, out TFOutput output, destinationDataType))
// Execute that graph to normalize this one image
using (var session = new TFSession(graph))
{
var normalized = session.Run(
inputs: new[] { input },
inputValues: new[] { tensor },
outputs: new[] { output });
Console.WriteLine(normalized.Length);
return(normalized[0]);
}
}
public static TFTensor CreateTensor(TFTensor tensor, TFDataType destinationDataType = TFDataType.Float)
{
TFGraph graph;
TFOutput input, output;
// Construct a graph to normalize the image
ConstructGraphToNormalizeImage(out graph, out input, out output, destinationDataType);
// Execute that graph to normalize this one image
using (var session = new TFSession(graph))
{
var normalized = session.Run(
inputs: new[] { input },
inputValues: new[] { tensor },
outputs: new[] { output });
return(normalized[0]);
}
}
internal static PrimitiveType Tf2MlNetType(TFDataType type)
{
switch (type)
{
case TFDataType.Float:
return(NumberType.R4);
case TFDataType.Double:
return(NumberType.R8);
case TFDataType.UInt32:
return(NumberType.U4);
case TFDataType.UInt64:
return(NumberType.U8);
default:
throw new NotSupportedException("TensorFlow type not supported.");
}
}
static Type TypeFromTensorType(TFDataType type)
{
switch (type)
{
case TFDataType.Float:
return(typeof(float));
case TFDataType.Double:
return(typeof(double));
case TFDataType.Int32:
return(typeof(int));
case TFDataType.UInt8:
return(typeof(byte));
case TFDataType.Int16:
return(typeof(short));
case TFDataType.Int8:
return(typeof(sbyte));
case TFDataType.String:
return(typeof(TFString));
case TFDataType.Int64:
return(typeof(long));
case TFDataType.Bool:
return(typeof(bool));
case TFDataType.UInt16:
return(typeof(ushort));
case TFDataType.Complex128:
return(typeof(Complex));
default:
return(null);
}
}
private unsafe TFTensor CreateTensorFromImageFileAlt(Mat inputFileName, TFDataType destinationDataType = TFDataType.Float)
{
Cv2.Resize(inputFileName, inputFileName, new OpenCvSharp.Size(64, 64)); // 일반 graph일 때
var array = new byte[inputFileName.Width * inputFileName.Height * inputFileName.ElemSize()]; // 임시 array 생성
Marshal.Copy(inputFileName.Data, array, 0, array.Length); // MS에서 제공하는 COPY Method
var matrix = new byte[1, inputFileName.Rows, inputFileName.Width, 3];
for (int i = 0; i < inputFileName.Rows; i++)
{
for (int j = 0; j < inputFileName.Cols; j++)
{
// 혹시라도 안되면 byte의 channel에 주목해보기!!! 난 bgr2rgb도 따로 안해서 둘다 시험해보기.
matrix[0, i, j, 0] = array[i * inputFileName.Step() + j * 3 + 0]; // 행-렬이다 여기서도. 즉 y,x이다.
matrix[0, i, j, 1] = array[i * inputFileName.Step() + j * 3 + 1];
matrix[0, i, j, 2] = array[i * inputFileName.Step() + j * 3 + 2]; // c++에서 복사해보고 제대로 복사되나 보고
}
}
UInt16[,,,] float_array = new UInt16[1, inputFileName.Rows, inputFileName.Cols, 3];
for (int i = 0; i < inputFileName.Rows; i++)
{
for (int j = 0; j < inputFileName.Cols; j++)
{ // 안된다면 byte에
// 혹시라도 안되면 byte의 channel에 주목해보기!!! 난 bgr2rgb도 따로 안해서 둘다 시험해보기.
float_array[0, i, j, 0] = (UInt16)(matrix[0, i, j, 0]);
float_array[0, i, j, 1] = (UInt16)(matrix[0, i, j, 1]);
float_array[0, i, j, 2] = (UInt16)(matrix[0, i, j, 2]); // c++에서 복사해보고 제대로 복사되나 보고
}
}
//Buffer.BlockCopy(array, 0, float_array, 0, matrix.Length);
TFTensor tensor = float_array; // 그대로 주게되면 제대로 들어가는지도 보고.
return(tensor);
}
public static TFTensor CreateTensorFromImageFile(byte[] contents, TFDataType destinationDataType = TFDataType.Float)
{
var tensor = TFTensor.CreateString(contents);
TFOutput input, output;
// Construct a graph to normalize the image
using (var graph = ConstructGraphToNormalizeImage(out input, out output, destinationDataType))
{
// Execute that graph to normalize this one image
using (var session = new TFSession(graph))
{
var normalized = session.Run(
inputs: new[] { input },
inputValues: new[] { tensor },
outputs: new[] { output });
return(normalized[0]);
}
}
}
/// <summary>
/// Creates a sequence of numbers.
/// </summary>
/// <remarks>
/// Creates a sequence of numbers that begins at `start` and extends by increments of `delta` up to but not including
/// `limit`. The dtype of the resulting tensor is inferred from the inputs unless it is provided explicitly.
/// </remarks>
/// <param name="start">A 0 - D `Tensor` (scalar).Acts as first entry in the range if `limit` is not None; otherwise, acts as range limit and first entry defaults to 0.</param>
/// <param name="limit">A 0 - D `Tensor` (scalar).Upper limit of sequence, exclusive. If None, defaults to the value of `start` while the first entry of the range defaults to 0.</param>
/// <param name="delta">A 0 - D `Tensor` (scalar).Number that increments `start`. Defaults to 1.</param>
/// <param name="dataType">The type of the elements of the resulting tensor.</param>
/// <param name="operName">A name for the operation.Defaults to "range".</param>
public TFOutput Range(TFOutput start, TFOutput?limit = null, TFOutput?delta = null, TFDataType?dataType = null, string operName = "range")
{
// https://github.com/tensorflow/tensorflow/blob/r1.2/tensorflow/python/ops/math_ops.py#L1156
if (limit == null)
{
limit = start;
start = Cast(Const(new TFTensor(0.0)), start.OutputType); // TODO: Maybe add dataType as convenience in Const?
}
if (delta == null)
{
delta = Cast(Const(new TFTensor(1.0)), start.OutputType);
}
using (var newScope = WithScope(MakeName("Range", operName)))
{
// infer dtype if not explicitly provided
if (dataType == null)
{
var dtype_hierarchy = new[] { TFDataType.Int32, TFDataType.Int64, TFDataType.Float, TFDataType.Double };
if (!dtype_hierarchy.Contains(start.OutputType) ||
!dtype_hierarchy.Contains(limit.Value.OutputType) ||
!dtype_hierarchy.Contains(delta.Value.OutputType))
{
throw new ArgumentException("Unexpected type");
}
TFDataType[] dtypes = new[] { start.OutputType, limit.Value.OutputType, delta.Value.OutputType };
int imax = dtypes.Select(x => Array.IndexOf(dtype_hierarchy, x)).Max();
TFDataType inferred_dtype = dtype_hierarchy[imax];
start = Cast(start, inferred_dtype);
limit = Cast(limit.Value, inferred_dtype);
delta = Cast(delta.Value, inferred_dtype);
}
return(Range(start, limit.Value, delta.Value, operName: operName));
}
}
/// <summary>
/// Create a constant tensor based on a shape
/// Used by Zeros and Ones
/// </summary>
/// <param name="value">Value for tensor</param>
/// <param name="tfshape">Shape of the tensor</param>
/// <param name="dtype">Optional Type of the Zero value. Default: Double</param>
/// <param name="operName">Operation name, optional.</param>
/// <returns></returns>
/// see https://github.com/tensorflow/tensorflow/blob/r1.1/tensorflow/python/framework/constant_op.py
public TFOutput Constant(object value, TFShape tfshape, TFDataType dtype = TFDataType.Double, string operName = null)
{
if (tfshape.NumDimensions <= 0)
{
TFTensor tensor = TFTensor.Create1DTensor(dtype, value);
return(Const(tensor, tensor.TensorType, operName));
}
//convert the .net type to relevant tensorflow type
object dtvalue = TFTensor.FetchSimple(dtype, value);
var shape = tfshape.ToArray();
var idx = new int [shape.Length];
for (int i = 0; i < shape.Length; i++)
{
if (shape [i] > Int32.MaxValue)
{
throw new ArgumentOutOfRangeException("Shape can not be longer than 32 bits");
}
}
Array data = null;
if (tfshape.IsLongArray)
{
data = Array.CreateInstance(dtvalue.GetType(), tfshape.ToArray());
}
else
{
data = Array.CreateInstance(dtvalue.GetType(), tfshape.ToIntArray());
}
TFTensor.Set(data, dtype, shape, idx, 0, value);
TFTensor tensor_value = new TFTensor(data);
return(Const(tensor_value, tensor_value.TensorType, operName));
}
// Convert the image in filename to a Tensor suitable as input to the Inception model.
public static TFTensor CreateTensorFromImageFile(string file, TFDataType destinationDataType = TFDataType.Float)
{
contents = File.ReadAllBytes(file);
System.Buffer.BlockCopy(contents, 99712, crop, 0, 336 * 336);
// DecodeJpeg uses a scalar String-valued tensor as input.
var tensor = TFTensor.CreateString(contents);
TFOutput input, output;
// Construct a graph to normalize the image
using (var graph = ConstructGraphToNormalizeImage(out input, out output, destinationDataType)){
// Execute that graph to normalize this one image
using (var session = new TFSession(graph)) {
var normalized = session.Run(
inputs: new [] { input },
inputValues: new [] { tensor },
outputs: new [] { output });
return(normalized [0]);
}
}
}
// Additional pointers for using TensorFlow & CustomVision together
// Python: https://github.com/tensorflow/tensorflow/blob/master/tensorflow/examples/label_image/label_image.py
// C++: https://github.com/tensorflow/tensorflow/blob/master/tensorflow/examples/label_image/main.cc
// Java: https://github.com/Azure-Samples/cognitive-services-android-customvision-sample/blob/master/app/src/main/java/demo/tensorflow/org/customvision_sample/MSCognitiveServicesClassifier.java
private static (TFGraph, TFOutput, TFOutput) ConstructGraphToNormalizeImage(TFDataType destinationDataType = TFDataType.Float)
{
const int W = 227;
const int H = 227;
//const float Scale = 1;
// Depending on your CustomVision.ai Domain - set appropriate Mean Values (RGB)
// https://github.com/Azure-Samples/cognitive-services-android-customvision-sample for RGB values (in BGR order)
//var bgrValues = new TFTensor(new float[] { 104.0f, 117.0f, 123.0f }); // General (Compact) & Landmark (Compact)
//var bgrValues = new TFTensor(0f); // Retail (Compact)
var graph = new TFGraph();
var input = graph.Placeholder(TFDataType.String);
var caster = graph.Cast(graph.DecodeJpeg(contents: input, channels: 3), DstT: TFDataType.Float);
var dims_expander = graph.ExpandDims(caster, graph.Const(0, "batch"));
var resized = graph.ResizeBilinear(dims_expander, graph.Const(new int[] { H, W }, "size"));
//var resized_mean = graph.Sub(resized, graph.Const(bgrValues, "mean"));
//var normalised = graph.Div(resized_mean, graph.Const(Scale));
var output = resized;
return(graph, input, output);
}
// 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 (TFGraph graph, TFOutput input, TFOutput output) PreprocessNeuralNetwork(int width, int height)
{
// Custom model has been processed using Keras
// which uses its own algorithm for normalizing images
// RGB values are normalized between -1 and 1
// image shape can be customized in keras,
// so width and height are converted to method arguments
const TFDataType destinationDataType = TFDataType.Float;
TFGraph graph = new TFGraph();
TFOutput input = graph.Placeholder(TFDataType.String);
var decodeJpeg = graph.DecodeJpeg(contents: input, channels: 3);
var castToFloat = graph.Cast(decodeJpeg, DstT: TFDataType.Float);
var expandDims = graph.ExpandDims(castToFloat, dim: graph.Const(0, "make_batch")); // shape: [1, height, width, channels]
var resize = graph.ResizeBilinear(expandDims, size: graph.Const(new int[] { width, height }, "size"));
var divide = graph.Div(resize, y: graph.Const(255f, destinationDataType, "divide"));
var substract = graph.Sub(divide, y: graph.Const(.5f, destinationDataType, "sub"));
var multiply = graph.Mul(substract, y: graph.Const(2f, destinationDataType, "multiply"));
var castToDestinationDataType = graph.Cast(multiply, destinationDataType);
return(graph, input, castToDestinationDataType);
}
public static TFTensor CreateTensorFromMemoryJpg(Byte[] jpeg, TFDataType destinationDataType = TFDataType.Float)
{
// DecodeJpeg uses a scalar String-valued tensor as input.
var tensor = TFTensor.CreateString(jpeg);
TFOutput input;
TFOutput output;
// Construct a graph to normalize the image
using (var graph = ConstructGraphToNormalizeImage2(out input, out output, destinationDataType))
{
// Execute that graph to normalize this one image
using (var session = new TFSession(graph))
{
var normalized = session.Run(
inputs: new[] { input },
inputValues: new[] { tensor },
outputs: new[] { output });
return(normalized[0]);
}
}
}
private static TFTensor CreateTensorFromBitmap(Image bitmap, TFDataType destinationDataType = TFDataType.Float)
{
//var contents = File.ReadAllBytes(file);
var contents = ImageUtils.ImageToByte(bitmap);
// DecodeJpeg uses a scalar String-valued tensor as input.
var tensor = TFTensor.CreateString(contents);
// Construct a graph to normalize the image
using (var graph = ConstructGraphToNormalizeImage(out TFOutput input, out TFOutput output, destinationDataType))
{
// Execute that graph to normalize this one image
using (var session = new TFSession(graph))
{
var normalized = session.Run(
inputs: new[] { input },
inputValues: new[] { tensor },
outputs: new[] { output });
return(normalized[0]);
}
}
}
public static TFTensor CreateTensorFromBGR(Byte[] rgb, TFDataType destinationDataType = TFDataType.Float)
{
// DecodeJpeg uses a scalar String-valued tensor as input.
//var tensor = TFTensor.CreateString(bgr);
/*
* byte[] floatValues = new byte[640 * 480 * 3];
* int j = 0;
* for (int i = 0; i < rgb.Length/4; i++)
* {
* floatValues[j * 3 + 0] = (rgb[i * 4 + 0]) ;
* floatValues[j * 3 + 1] = (rgb[i * 4 + 1]) ;
* floatValues[j * 3 + 2] = (rgb[i * 4 + 2]) ;
* j++;
* }*/
TFShape shape = new TFShape(480, 640, 3);
var tensor = TFTensor.FromBuffer(shape, rgb, 0, rgb.Length);
TFOutput input;
TFOutput output;
// Construct a graph to normalize the image
using (var graph = ConstructGraphToNormalizeImage(out input, out output, destinationDataType))
{
// Execute that graph to normalize this one image
using (var session = new TFSession(graph))
{
var normalized = session.Run(
inputs: new[] { input },
inputValues: new[] { tensor },
outputs: new[] { output });
return(normalized[0]);
}
}
}
