Browse Source

Created TensorConverter

tags/v0.12
Eli Belash 6 years ago
parent
commit
94c8c6b5ad
1 changed files with 285 additions and 0 deletions
  1. +285
    -0
      src/TensorFlowNET.Core/Tensors/TensorConverter.cs

+ 285
- 0
src/TensorFlowNET.Core/Tensors/TensorConverter.cs View File

@@ -0,0 +1,285 @@
using System;
using System.Threading.Tasks;
using NumSharp;
using NumSharp.Backends;
using NumSharp.Utilities;

namespace Tensorflow
{
/// <summary>
/// Provides various methods to conversion between types and <see cref="Tensor"/>.
/// </summary>
public static class TensorConverter
{
/// <summary>
/// Convert given <see cref="Array"/> to <see cref="Tensor"/>.
/// </summary>
/// <param name="nd">The ndarray to convert, can be regular, jagged or multi-dim array.</param>
/// <param name="astype">Convert <see cref="Array"/> to given <paramref name="astype"/> before inserting it into a <see cref="Tensor"/>.</param>
/// <exception cref="NotSupportedException"></exception>
public static Tensor ToTensor(NDArray nd, TF_DataType? astype = null)
{
return new Tensor(astype == null ? nd : nd.astype(astype.Value.as_numpy_typecode(), false));
}
/// <summary>
/// Convert given <see cref="NDArray"/> to <see cref="Tensor"/>.
/// </summary>
/// <param name="nd">The ndarray to convert.</param>
/// <param name="astype">Convert <see cref="Array"/> to given <paramref name="astype"/> before inserting it into a <see cref="Tensor"/>.</param>
/// <exception cref="NotSupportedException"></exception>
public static Tensor ToTensor(NDArray nd, NPTypeCode? astype = null)
{
return new Tensor(astype == null ? nd : nd.astype(astype.Value, false));
}
/// <summary>
/// Convert given <see cref="Array"/> to <see cref="Tensor"/>.
/// </summary>
/// <param name="array">The array to convert, can be regular, jagged or multi-dim array.</param>
/// <param name="astype">Convert <see cref="Array"/> to given <paramref name="astype"/> before inserting it into a <see cref="Tensor"/>.</param>
/// <exception cref="NotSupportedException"></exception>
public static Tensor ToTensor(Array array, TF_DataType? astype = null)
{
if (array == null) throw new ArgumentNullException(nameof(array));
var arrtype = array.ResolveElementType();

var astype_type = astype?.as_numpy_dtype() ?? arrtype;
if (astype_type == arrtype)
{
//no conversion required
if (astype == TF_DataType.TF_STRING)
{
throw new NotSupportedException(); //TODO! when string is fully implemented.
}

if (astype == TF_DataType.TF_INT8)
{
if (array.Rank != 1 || array.GetType().GetElementType()?.IsArray == true) //is multidim or jagged
array = Arrays.Flatten(array);

return new Tensor((sbyte[]) array);
}

//is multidim or jagged, if so - use NDArrays constructor as it records shape.
if (array.Rank != 1 || array.GetType().GetElementType().IsArray)
return new Tensor(new NDArray(array));

#if _REGEN
#region Compute
switch (arrtype)
{
%foreach supported_dtypes,supported_dtypes_lowercase%
case NPTypeCode.#1: return new Tensor((#2[])arr);
%
default:
throw new NotSupportedException();
}
#endregion
#else

#region Compute

switch (arrtype.GetTypeCode())
{
case NPTypeCode.Boolean: return new Tensor((bool[]) array);
case NPTypeCode.Byte: return new Tensor((byte[]) array);
case NPTypeCode.Int16: return new Tensor((short[]) array);
case NPTypeCode.UInt16: return new Tensor((ushort[]) array);
case NPTypeCode.Int32: return new Tensor((int[]) array);
case NPTypeCode.UInt32: return new Tensor((uint[]) array);
case NPTypeCode.Int64: return new Tensor((long[]) array);
case NPTypeCode.UInt64: return new Tensor((ulong[]) array);
case NPTypeCode.Char: return new Tensor((char[]) array);
case NPTypeCode.Double: return new Tensor((double[]) array);
case NPTypeCode.Single: return new Tensor((float[]) array);
default:
throw new NotSupportedException();
}

#endregion

#endif
} else
{
//conversion is required.
//by this point astype is not null.

//flatten if required
if (array.Rank != 1 || array.GetType().GetElementType()?.IsArray == true) //is multidim or jagged
array = Arrays.Flatten(array);

try
{
return ToTensor(
ArrayConvert.To(array, astype.Value.as_numpy_typecode()),
null
);
} catch (NotSupportedException)
{
//handle dtypes not supported by ArrayConvert
var ret = Array.CreateInstance(astype_type, array.LongLength);
Parallel.For(0, ret.LongLength, i => ret.SetValue(Convert.ChangeType(array.GetValue(i), astype_type), i));
return ToTensor(ret, null);
}
}
}

/// <summary>
/// Convert given <see cref="Array"/> to <see cref="Tensor"/>.
/// </summary>
/// <param name="constant">The constant scalar to convert</param>
/// <param name="astype">Convert <paramref name="constant"/> to given <paramref name="astype"/> before inserting it into a <see cref="Tensor"/>.</param>
/// <exception cref="NotSupportedException"></exception>
public static Tensor ToTensor<T>(T constant, TF_DataType? astype = null) where T : unmanaged
{
//was conversion requested?
if (astype == null)
{
//No conversion required
var constantType = typeof(T).as_dtype();
if (constantType == TF_DataType.TF_INT8)
return new Tensor((sbyte) (object) constant);

if (constantType == TF_DataType.TF_STRING)
return new Tensor((string) (object) constant);

#if _REGEN
#region Compute
switch (InfoOf<T>.NPTypeCode)
{
%foreach supported_dtypes,supported_dtypes_lowercase%
case NPTypeCode.#1: return new Tensor((#2)(object)constant);
%
default:
throw new NotSupportedException();
}
#endregion
#else

#region Compute

switch (InfoOf<T>.NPTypeCode)
{
case NPTypeCode.Boolean: return new Tensor((bool) (object) constant);
case NPTypeCode.Byte: return new Tensor((byte) (object) constant);
case NPTypeCode.Int16: return new Tensor((short) (object) constant);
case NPTypeCode.UInt16: return new Tensor((ushort) (object) constant);
case NPTypeCode.Int32: return new Tensor((int) (object) constant);
case NPTypeCode.UInt32: return new Tensor((uint) (object) constant);
case NPTypeCode.Int64: return new Tensor((long) (object) constant);
case NPTypeCode.UInt64: return new Tensor((ulong) (object) constant);
case NPTypeCode.Char: return new Tensor(Converts.ToByte(constant));
case NPTypeCode.Double: return new Tensor((double) (object) constant);
case NPTypeCode.Single: return new Tensor((float) (object) constant);
default:
throw new NotSupportedException();
}

#endregion
#endif
}

//conversion required

if (astype == TF_DataType.TF_INT8)
return new Tensor(Converts.ToSByte(constant));

if (astype == TF_DataType.TF_STRING)
return new Tensor(Converts.ToString(constant));

var astype_np = astype?.as_numpy_typecode();

#if _REGEN
#region Compute
switch (astype_np)
{
%foreach supported_dtypes,supported_dtypes_lowercase%
case NPTypeCode.#1: return new Tensor(Converts.To#1(constant));
%
default:
throw new NotSupportedException();
}
#endregion
#else

#region Compute
switch (astype_np)
{
case NPTypeCode.Boolean: return new Tensor(Converts.ToBoolean(constant));
case NPTypeCode.Byte: return new Tensor(Converts.ToByte(constant));
case NPTypeCode.Int16: return new Tensor(Converts.ToInt16(constant));
case NPTypeCode.UInt16: return new Tensor(Converts.ToUInt16(constant));
case NPTypeCode.Int32: return new Tensor(Converts.ToInt32(constant));
case NPTypeCode.UInt32: return new Tensor(Converts.ToUInt32(constant));
case NPTypeCode.Int64: return new Tensor(Converts.ToInt64(constant));
case NPTypeCode.UInt64: return new Tensor(Converts.ToUInt64(constant));
case NPTypeCode.Char: return new Tensor(Converts.ToByte(constant));
case NPTypeCode.Double: return new Tensor(Converts.ToDouble(constant));
case NPTypeCode.Single: return new Tensor(Converts.ToSingle(constant));
default:
throw new NotSupportedException();
}
#endregion
#endif
}

/// <summary>
/// Convert given <see cref="Array"/> to <see cref="Tensor"/>.
/// </summary>
/// <param name="constant">The constant scalar to convert</param>
/// <param name="astype">Convert <paramref name="constant"/> to given <paramref name="astype"/> before inserting it into a <see cref="Tensor"/>.</param>
/// <exception cref="NotSupportedException"></exception>
public static Tensor ToTensor(string constant, TF_DataType? astype = null)
{
switch (astype)
{
//was conversion requested?
case null:
case TF_DataType.TF_STRING:
return new Tensor(constant);
//conversion required
case TF_DataType.TF_INT8:
return new Tensor(Converts.ToSByte(constant));
default:
{
var astype_np = astype?.as_numpy_typecode();

#if _REGEN
#region Compute
switch (astype_np)
{
%foreach supported_dtypes,supported_dtypes_lowercase%
case NPTypeCode.#1: return new Tensor(Converts.To#1(constant));
%
default:
throw new NotSupportedException();
}
#endregion
#else

#region Compute
switch (astype_np)
{
case NPTypeCode.Boolean: return new Tensor(Converts.ToBoolean(constant));
case NPTypeCode.Byte: return new Tensor(Converts.ToByte(constant));
case NPTypeCode.Int16: return new Tensor(Converts.ToInt16(constant));
case NPTypeCode.UInt16: return new Tensor(Converts.ToUInt16(constant));
case NPTypeCode.Int32: return new Tensor(Converts.ToInt32(constant));
case NPTypeCode.UInt32: return new Tensor(Converts.ToUInt32(constant));
case NPTypeCode.Int64: return new Tensor(Converts.ToInt64(constant));
case NPTypeCode.UInt64: return new Tensor(Converts.ToUInt64(constant));
case NPTypeCode.Char: return new Tensor(Converts.ToByte(constant));
case NPTypeCode.Double: return new Tensor(Converts.ToDouble(constant));
case NPTypeCode.Single: return new Tensor(Converts.ToSingle(constant));
default:
throw new NotSupportedException();
}
#endregion
#endif
}
}
}

}
}

Loading…
Cancel
Save