private static NDArray _broadcast_to <T>(NDArray nd, Shape shape, bool subok, bool rdonly)
{
T[,] table = new T[shape.Dimensions[0], shape.Dimensions[1]];
if (nd.shape[0] == 1)
{// (1,2,3)
for (int i = 0; i < shape.Dimensions[0]; i++)
{
for (int j = 0; j < shape.Dimensions[1]; j++)
{
table[i, j] = nd.Data <T>(0, j);
}
}
}
else if (nd.shape[1] == 1)
{
for (int i = 0; i < shape.Dimensions[0]; i++)
{
for (int j = 0; j < shape.Dimensions[1]; j++)
{
table[i, j] = nd.Data <T>(i, 0);
}
}
}
return(np.array <T>(table));
}
public Tensor tensor(NumSharp.NDArray data, dtype?dtype = null, device?device = null, bool?requires_grad = null, bool?pin_memory = null)
{
// note: this implementation works only for device CPU
// todo: implement for GPU
var type = data.dtype.ToDtype();
if (dtype != null && type != dtype)
{
throw new NotImplementedException("Type of the array is different from specified dtype. Data conversion is not supported (yet)");
}
var tensor = torch.empty((Shape)data.shape, dtype: type, device: device,
requires_grad: requires_grad, pin_memory: pin_memory);
var storage = tensor.PyObject.storage();
long ptr = storage.data_ptr();
switch (type)
{
case Torch.dtype.UInt8: Marshal.Copy(data.Data <byte>(), 0, new IntPtr(ptr), data.len); break;
case Torch.dtype.Int32: Marshal.Copy(data.Data <int>(), 0, new IntPtr(ptr), data.len); break;
case Torch.dtype.Int64: Marshal.Copy(data.Data <long>(), 0, new IntPtr(ptr), data.len); break;
case Torch.dtype.Float32: Marshal.Copy(data.Data <float>(), 0, new IntPtr(ptr), data.len); break;
case Torch.dtype.Float64: Marshal.Copy(data.Data <double>(), 0, new IntPtr(ptr), data.len); break;
}
return(tensor);
}
/// <summary>
/// Draw samples from a uniform distribution.
/// Samples are uniformly distributed over the half-open interval [low, high) (includes low, but excludes high). In other words, any value within the given interval is equally likely to be drawn by uniform.
/// </summary>
/// <param name="low">Lower boundary of the output interval. All values generated will be greater than or equal to low. The default value is 0.</param>
/// <param name="high">Upper boundary of the output interval. All values generated will be less than high. The default value is 1.0.</param>
/// <param name="size">Output shape. If the given shape is, e.g., m, n, k, then m * n * k samples are drawn. If size is None (default), a single value is returned if low and high are both scalars. </param>
/// <returns>NDArray with values of type <see cref="double"/></returns>
public NDArray uniform(double low, double high, params int[] size)
{
if (size == null || size.Length == 0) //return scalar
{
var ret = new NDArray <double>(new Shape(1));
var data = new double[] { low + randomizer.NextDouble() * (high - low) };
ret.ReplaceData(data);
return(ret);
}
var result = new NDArray <double>(size);
double[] resultArray = result.Data <double>();
//parallelism is prohibited to make sure the result come out presistant
double diff = high - low;
for (int i = 0; i < result.size; ++i)
{
resultArray[i] = low + randomizer.NextDouble() * diff;
}
result.ReplaceData(resultArray); //incase of a view //todo! incase of a view?
return(result);
}
public static NDArray prod(NDArray nd, int axis = -1, Type dtype = null)
{
NDArray result = null;
if (nd.size == 0)
{
return(1);
}
if (axis == -1)
{
switch (nd.dtype.Name)
{
case "Int32":
{
int prod = 1;
for (int i = 0; i < nd.size; i++)
{
prod *= nd.Data <int>(i);
}
result = prod;
}
break;
case "Int64":
{
long prod = 1;
for (int i = 0; i < nd.size; i++)
{
prod *= nd.Data <long>(i);
}
result = prod;
}
break;
}
}
else
{
throw new NotImplementedException($"np.prod axis {axis}");
}
return(result);
}
/// <summary>
/// Random values in a given shape.
/// Create an array of the given shape and populate it with random samples from a uniform distribution over [0, 1).
/// </summary>
public NDArray rand(Shape shape)
{
NDArray ndArray = new NDArray(typeof(double), shape);
double[] numArray = ndArray.Data <double>();
for (int index = 0; index < ndArray.size; ++index)
{
numArray[index] = randomizer.NextDouble();
}
ndArray.SetData <double[]>(numArray);
return(ndArray);
}
private NDArray setValue1D <T>(NDArray indexes)
{
var buf = Data <T>();
var idx = indexes.Data <int>();
var values = new T[indexes.size];
Parallel.For(0, indexes.size, (row) =>
{
values[row] = buf[idx[row]];
});
return(new NDArray(values, indexes.size));
}
public static NDArray repeat(NDArray nd, int repeats, int axis = -1)
{
int size = nd.size * repeats;
// scalar
switch (nd.dtype.Name)
{
case "Int32":
{
var nd2 = new NDArray(new int[size], new Shape(size));
var data = nd.Data <int>();
for (int i = 0; i < nd.size; i++)
{
for (int j = 0; j < repeats; j++)
{
nd2.itemset(i * repeats + j, data[i]);
}
}
return(nd2);
}
case "Boolean":
{
var nd2 = new NDArray(new bool[size], new Shape(size));
var data = nd.Data <bool>();
for (int i = 0; i < nd.size; i++)
{
for (int j = 0; j < repeats; j++)
{
nd2.itemset(i * repeats + j, data[i]);
}
}
return(nd2);
}
}
throw new NotImplementedException("np.repeat");
}
private NDArray setValue4D <T>(NDArray indexes)
{
var nd = new NDArray(dtype, new Shape(indexes.size, shape[1], shape[2], shape[3]));
/*switch (Type.GetTypeCode(dtype))
* {
* case TypeCode.Byte:
* {
* var buf = Data<byte>().AsSpan();
*
* var idx = indexes.Data<int>().AsSpan();
*
* var shapes = nd.shape.AsSpan();
* var data = nd.Data<byte>().AsSpan();
*
* Parallel.For(0, nd.shape[0], (item) =>
* {
* for (int row = 0; row < shapes[1]; row++)
* for (int col = 0; col < shapes[2]; col++)
* for (int channel = 0; channel < shapes[3]; channel++)
* data[Storage.Shape.GetIndexInShape(item, row, col, channel)] = buf[Storage.Shape.GetIndexInShape(idx[item], row, col, channel)];
* });
* }
* break;
* }*/
{
var buf = Data <T>();
var idx = indexes.Data <int>();
var data = nd.Data <T>();
Parallel.For(0, nd.shape[0], (item) =>
{
for (int row = 0; row < nd.shape[1]; row++)
{
for (int col = 0; col < nd.shape[2]; col++)
{
for (int channel = 0; channel < nd.shape[3]; channel++)
{
data[Storage.Shape.GetIndexInShape(item, row, col, channel)] = buf[Storage.Shape.GetIndexInShape(idx[item], row, col, channel)];
}
}
}
});
}
return(nd);
}
private NDArray setValue2D <T>(NDArray indexes)
{
var buf = Data <T>();
var idx = indexes.Data <int>();
var nd = new NDArray(dtype, new Shape(indexes.size, shape[1]));
Parallel.For(0, nd.shape[0], (row) =>
{
for (int col = 0; col < nd.shape[1]; col++)
{
nd.SetData(buf[Storage.Shape.GetIndexInShape(idx[row], col)], row, col);
}
});
return(nd);
}
private NDArray setValue2D <T>(NDArray indexes)
{
var buf = Data <T>();
var idx = indexes.Data <int>();
var selectedValues = new NDArray(this.dtype, new Shape(indexes.size, shape[1]));
Parallel.ForEach(Enumerable.Range(0, selectedValues.shape[0]), (row) =>
{
for (int col = 0; col < selectedValues.shape[1]; col++)
{
selectedValues[row, col] = buf[Storage.Shape.GetIndexInShape(idx[row], col)];
}
});
return(selectedValues);
}
private NDArray setValue3D <T>(NDArray indexes)
{
var buf = Data <T>();
var selectedValues = new NDArray(dtype, new Shape(indexes.size, shape[1], shape[2]));
var idx = indexes.Data <int>();
Parallel.ForEach(Enumerable.Range(0, selectedValues.shape[0]), (item) =>
{
for (int row = 0; row < selectedValues.shape[1]; row++)
{
for (int col = 0; col < selectedValues.shape[2]; col++)
{
selectedValues.SetData(buf[Storage.Shape.GetIndexInShape(idx[item], row, col)], item, row, col);
}
}
});
return(selectedValues);
}
/// <summary>
/// Draw random samples from a normal (Gaussian) distribution.
/// </summary>
/// <param name="loc">Mean of the distribution</param>
/// <param name="scale">Standard deviation of the distribution</param>
/// <param name="dims"></param>
/// <returns></returns>
public NDArray normal(double loc, double scale, params int[] dims)
{
var array = new NDArray(typeof(double), new Shape(dims));
double[] arr = array.Data <double>();
for (int i = 0; i < array.size; i++)
{
double u1 = 1.0 - randomizer.NextDouble(); //uniform(0,1] random doubles
double u2 = 1.0 - randomizer.NextDouble();
double randStdNormal = Math.Sqrt(-2.0 * Math.Log(u1)) *
Math.Sin(2.0 * Math.PI * u2); //random normal(0,1)
double randNormal = loc + scale * randStdNormal; //random normal(mean,stdDev^2)
arr[i] = randNormal;
}
array.ReplaceData(arr);
return(array);
}
private NDArray setValue <T>(NDArray indexes)
{
Shape newShape = new int[] { indexes.size }.Concat(shape.Skip(1)).ToArray();
var buf = Data <T>();
var idx = indexes.Data <int>();
var array = new T[newShape.Size];
var indice = Shape.GetShape(newShape.Dimensions, axis: 0);
var length = Shape.GetSize(indice);
for (var row = 0; row < newShape[0]; row++)
{
var d = buf.AsSpan(idx[row] * length, length);
d.CopyTo(array.AsSpan(row * length));
}
var nd = new NDArray(array, newShape);
return(nd);
}
public NDArray Permutation(int max)
{
var random = new Random();
int[] orders = new int[max];
var np = new NDArray(typeof(int)).arange(max);
int[] npData = np.Data <int>();
for (int i = 0; i < max; i++)
{
var pos = random.Next(0, max);
var zero = npData[0];
npData[0] = npData[pos];
npData[pos] = zero;
}
return(np);
}
/// <summary>
/// Least Square method
///
/// Determines NDArray X which reduces least square error of Linear System A * X = B.
/// This NDArray is equal to A.
/// </summary>
/// <param name="nDArrayB">Result NDArray B</param>
/// <param name="rcon"></param>
/// <returns>NArray X</returns>
public NDArray lstqr(NDArray nDArrayB, double rcon = 0.0001)
{
var A = (double[])Data <double>();
var b = (double[])nDArrayB.Data <double>();
int m = this.shape[0];
int n = this.shape[1];
int nrhs = (nDArrayB.ndim > 1) ? nDArrayB.shape[1] : 1;
int lda = m;
int ldb = m;
int rank = 0;
double[] work = new double[1];
int lwork = -1;
int info = 0;
double[] singVal = new double[m];
LAPACK.dgelss_(ref m, ref n, ref nrhs, A, ref lda, b, ref ldb, singVal, ref rcon, ref rank, work, ref lwork, ref info);
lwork = (int)work[0];
work = new double[lwork];
LAPACK.dgelss_(ref m, ref n, ref nrhs, A, ref lda, b, ref ldb, singVal, ref rcon, ref rank, work, ref lwork, ref info);
double[] sln = new double[n * nrhs];
for (int idx = 0; idx < sln.Length; idx++)
{
sln[idx] = b[m * (idx % nrhs) + idx / nrhs];
}
var slnArr = new NDArray(typeof(double), new Shape(n, nrhs));
slnArr.Storage.ReplaceData(sln);
return(slnArr);
}
public NDArray this[NDArray <bool> booleanArray]
{
get
{
if (!Enumerable.SequenceEqual(shape, booleanArray.shape))
{
throw new IncorrectShapeException();
}
var boolDotNetArray = booleanArray.Data <bool>();
switch (dtype.Name)
{
case "Int32":
{
var nd = new List <int>();
for (int idx = 0; idx < boolDotNetArray.Length; idx++)
{
if (boolDotNetArray[idx])
{
nd.Add(Data <int>(booleanArray.Storage.Shape.GetDimIndexOutShape(idx)));
}
}
return(new NDArray(nd.ToArray(), nd.Count));
}
case "Double":
{
var nd = new List <double>();
for (int idx = 0; idx < boolDotNetArray.Length; idx++)
{
if (boolDotNetArray[idx])
{
nd.Add(Data <double>(booleanArray.Storage.Shape.GetDimIndexOutShape(idx)));
}
}
return(new NDArray(nd.ToArray(), nd.Count));
}
}
throw new NotImplementedException("");
}
set
{
if (!Enumerable.SequenceEqual(shape, booleanArray.shape))
{
throw new IncorrectShapeException();
}
object scalarObj = value.Storage.GetData().GetValue(0);
bool[] boolDotNetArray = booleanArray.Storage.GetData() as bool[];
int elementsAmount = booleanArray.size;
for (int idx = 0; idx < elementsAmount; idx++)
{
if (boolDotNetArray[idx])
{
int[] indexes = booleanArray.Storage.Shape.GetDimIndexOutShape(idx);
Array.SetValue(scalarObj, Storage.Shape.GetIndexInShape(indexes));
}
}
}
}
