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);
}
internal static IEnumerable <object> EnumerateOverAxis(this NumSharp.NDArray array, int axis = 0)
{
var prefix = Enumerable.Repeat(All, axis).ToArray();
foreach (var idx in Enumerable.Range(0, array.Shape[axis]))
{
Slice[] slice = prefix.Concat(new Slice[] { idx, Ellipsis }).ToArray();
yield return(array[slice]);
}
}
public static byte[] PackImg(IRHeader header, NumSharp.NDArray img, int quality = 95, string img_fmt = ".jpg")
{
int[] encodeParams = null;
OpenCvSharp.ImageEncodingParam imageEncodingParam = new ImageEncodingParam(ImwriteFlags.JpegQuality, quality);
if (img_fmt.ToLower() == ".jpg" || img_fmt.ToLower() == ".jpeg")
{
encodeParams = new int[] { }
}
;
Cv2.ImEncode(img_fmt, new Mat(img.GetMemPtr()), out var buf, imageEncodingParam);
return(Pack(header, buf));
}
}
public void UpdateGraph(Series s, NumSharp.NDArray x, NumSharp.NDArray y)
{
s.Points.Clear();
for (int i = 0; i < 1000; i++)
{
var _x = x[i].GetSingle();
var _y = y[i].GetSingle();
if (i % 1 == 0)
{
//Console.WriteLine($"{i}:({_x},{_y}),");
s.Points.AddXY(_x, _y);
}
}
}
public static Tensor tensor(NumSharp.NDArray data, dtype?dtype = null, device?device = null, bool?requires_grad = null, bool?pin_memory = null) => PyTorch.Instance.tensor(data, dtype: dtype, device: device, requires_grad: requires_grad, pin_memory: pin_memory);
private NDArray AminImpl <T>(int?axis = null) where T : struct
{
var res = new NDArray(dtype);
if (axis == null)
{
var npArr = this.Storage.GetData <int>();
int min = npArr[0];
for (int i = 0; i < npArr.Length; i++)
{
min = Math.Min(min, npArr[i]);
}
res.Storage = res.TensorEngine.GetStorage(dtype);
res.Storage.Allocate(new Shape(1));
res.Storage.ReplaceData(new int[1] {
min
});
}
else
{
if (axis < 0 || axis >= this.ndim)
{
throw new Exception("Invalid input: axis");
}
int[] resShapes = new int[this.shape.Length - 1];
int index = 0; //index for result shape set
//axis departs the shape into three parts: prev, cur and post. They are all product of shapes
int prev = 1;
int cur = 1;
int post = 1;
int size = 1; //total number of the elements for result
//Calculate new Shape
for (int i = 0; i < this.shape.Length; i++)
{
if (i == axis)
{
cur = this.shape[i];
}
else
{
resShapes[index++] = this.shape[i];
size *= this.shape[i];
if (i < axis)
{
prev *= this.shape[i];
}
else
{
post *= this.shape[i];
}
}
}
//Fill in data
index = 0; //index for result data set
int sameSetOffset = this.Storage.Shape.Strides[axis.Value];
int increments = cur * post;
switch (typeof(T).Name)
{
case "Int32":
{
var resData = new int[size]; //res.Data = new double[size];
var npArr = Data <int>();
int start = 0;
int min = 0;
for (int i = 0; i < this.size; i += increments)
{
for (int j = i; j < i + post; j++)
{
start = j;
min = npArr[start];
for (int k = 0; k < cur; k++)
{
min = Math.Min(min, npArr[start]);
start += sameSetOffset;
}
resData[index++] = min;
}
}
res.Storage.Allocate(new Shape(resShapes));
res.Storage.ReplaceData(resData);
}
break;
case "Single":
{
var resData = new float[size]; //res.Data = new double[size];
var npArr = Data <float>();
int start = 0;
float min = 0;
for (int i = 0; i < this.size; i += increments)
{
for (int j = i; j < i + post; j++)
{
start = j;
min = npArr[start];
for (int k = 0; k < cur; k++)
{
min = Math.Min(min, npArr[start]);
start += sameSetOffset;
}
resData[index++] = min;
}
}
res.Storage.Allocate(new Shape(resShapes));
res.Storage.ReplaceData(resData);
}
break;
}
}
return(res);
}
/// <summary>
/// Join a sequence of arrays along an existing axis.
/// </summary>
/// <param name="axis">The axis along which the arrays will be joined. If axis is None, arrays are flattened before use. Default is 0.</param>
/// <param name="arrays">The arrays must have the same shape, except in the dimension corresponding to axis (the first, by default).</param>
/// <returns>The concatenated array.</returns>
/// <remarks>https://docs.scipy.org/doc/numpy/reference/generated/numpy.concatenate.html</remarks>
public static NDArray concatenate(NDArray[] arrays, int axis = 0)
{
//What we do is we have the axis which is the only dimension that is allowed to be different
//We need to perform a check if the dimensions actually match.
//After we have the axis ax=1 where shape is (3,ax,3) - ax is the only dimension that can vary.
//So if we got input of (3,5,3) and (3,1,3), we create a return shape of (3,6,3).
//We perform the assignment by iterating a slice: (:,n,:) on src and dst where dst while n of dst grows as we iterate all arrays.
if (arrays == null)
{
throw new ArgumentNullException(nameof(arrays));
}
if (arrays.Length == 0)
{
throw new ArgumentException("Value cannot be an empty collection.", nameof(arrays));
}
if (arrays.Length == 1)
{
return(arrays[0]);
}
var first = arrays[0];
var firstShape = (int[])first.shape.Clone();
while (axis < 0)
{
axis = first.ndim + axis; //translate negative axis
}
int i, j;
int axisSize = 0; //accumulated shape[axis] size for return shape.
NPTypeCode retType = first.GetTypeCode;
foreach (var src in arrays)
{
//accumulate the concatenated axis
var shape = src.shape;
axisSize += shape[axis];
if (ReferenceEquals(src, first))
{
continue;
}
var srcType = src.GetTypeCode;
//resolve what the return type should be and should we perform casting.
if (first.GetTypeCode != srcType)
{
if (srcType.CompareTo(retType) == 1)
{
retType = srcType;
}
}
if (shape.Length != first.ndim)
{
throw new IncorrectShapeException("all the input arrays must have same number of dimensions.");
}
//verify the shapes are equal
for (j = 0; j < shape.Length; j++)
{
if (axis == j)
{
continue;
}
if (shape[j] != firstShape[j])
{
throw new IncorrectShapeException("all the input array dimensions except for the concatenation axis must match exactly.");
}
}
}
//prepare return shape
firstShape[axis] = axisSize;
var retShape = new Shape(firstShape);
var dst = new NDArray(retType, retShape);
var accessorDst = new Slice[retShape.NDim];
var accessorSrc = new Slice[retShape.NDim];
for (i = 0; i < accessorDst.Length; i++)
{
accessorSrc[i] = accessorDst[i] = Slice.All;
}
accessorSrc[axis] = Slice.Index(0);
accessorDst[axis] = Slice.Index(0);
foreach (var src in arrays)
{
var len = src.shape[axis];
for (i = 0; i < len; i++)
{
var writeTo = dst[accessorDst];
var writeFrom = src[accessorSrc];
MultiIterator.Assign(writeTo.Storage, writeFrom.Storage);
accessorSrc[axis]++;
accessorDst[axis]++; //increment every step
}
accessorSrc[axis] = Slice.Index(0); //reset src
}
return(dst);
}
