string PaddingInBytes(string value, PadType type, int byteCount)
{
Encoding enc = Encoding.GetEncoding("Shift_JIS");
if (byteCount < enc.GetByteCount(value))
{
// valueが既定のバイト数を超えている場合は、切り落とし
value = value.Substring(0, byteCount);
}
switch (type)
{
case PadType.Char:
// 文字列の場合 左寄せ＋空白埋め
return(value.PadRight(byteCount - (enc.GetByteCount(value) - value.Length)));
case PadType.Number:
// 数値の場合 右寄せ＋0埋め
return(value.PadLeft(byteCount, ' '));
default:
// 上記以外は全部空白
return(value.PadLeft(byteCount));
}
}
internal void HitByPad(PadType padType, Vector3 forceDir, Vector3 impulse)
{
if (isPlayer)
{
forceDir *= 1.1f;
}
if (padType == PadType.Persistant)
{
AddImpulse(GameSettings.instance.data.percistantPadJumpImpulse * forceDir);
}
else if (padType == PadType.Bottom)
{
AddImpulse(impulse);
}
else if (padType == PadType.Destructable)
{
AddImpulse(GameSettings.instance.data.destructablePadJumpImpulse * forceDir);
}
else if (padType == PadType.Finish)
{
anim.SetLayerWeight(1, 0);
anim.SetTrigger("victory");
}
}
/// <summary>
/// Padding the specified str.
/// </summary>
/// <returns>The result string.</returns>
/// <param name="str">String to be padded.</param>
/// <param name="length">Total Length.</param>
/// <param name="padStr">Padding string, default is " ".</param>
/// <param name="type">Pad Type,padleft ,padright or padboth.</param>
public static string Pad(this string str, int length, string padStr = null, PadType type = PadType.Left)
{
if (str.Length >= length)
{
return(str);
}
int needPading = length - str.Length;
int leftPadding = 0, rightPadding = 0;
switch (type)
{
case PadType.Left:
leftPadding = needPading;
break;
case PadType.Right:
rightPadding = needPading;
break;
case PadType.Both:
leftPadding = Convert.ToInt32(Math.Floor(needPading / 2f));
rightPadding = Convert.ToInt32(Math.Ceiling(needPading / 2f));
break;
}
padStr = padStr == null || padStr.Length <= 0 ? " " : padStr;
return(Repeat(padStr, leftPadding / padStr.Length + 1).Substring(0, leftPadding)
+ str
+ Repeat(padStr, rightPadding / padStr.Length + 1).Substring(0, rightPadding));
}
public static Conv2DInfo computePool2DInfo(int[] inShape, int[] filterSize,
int[] strides, PadType pad, roundingMode roundingMode,
ConvDataFormat dataFormat = ConvDataFormat.channelsLast, Nullable <int> padValue = null)
{
var filterHeight = filterSize[0];
var filterWidth = filterSize[1];
int[] filterShape;
if (dataFormat == ConvDataFormat.channelsLast)
{
filterShape = new int[] { filterHeight, filterWidth, inShape[3], inShape[3] };
}
else if (dataFormat == ConvDataFormat.channelsFirst)
{
filterShape = new int[] { filterHeight, filterWidth, inShape[1], inShape[1] };
}
else
{
throw new Exception("Unknown dataFormat");
}
var dilations = 1;
return(computeConv2DInfo(
inShape, filterShape, strides, new int[] { dilations }, pad, roundingMode, false,
dataFormat, padValue));
}
/// <summary>
/// Computes a 1D convolution over the input x.
/// </summary>
/// <param name="x">The input tensor, of Rank 3 or Rank 2, of shape
/// "[batch, width, inChannels]". If Rank 2, batch of 1 is assumed.</param>
/// <param name="filter">The filter, Rank 3, of shape
/// "[filterWidth, inDepth, outDepth]".</param>
/// <param name="stride">The number of entries by which the filter is moved right at
/// each step.</param>
/// <param name="pad">The type of padding algorithm.
/// - "same" and stride 1: output will be of same size as input,
/// regardless of filter size.
/// - "valid": output will be smaller than input if filter is larger
/// than 1x1.
/// - For more info, see this guide:
/// [https://www.tensorflow.org/api_guides/python/nn#Convolution](
/// https://www.tensorflow.org/api_guides/python/nn#Convolution)
/// </param>
/// <param name="dilation">The dilation rate in which we sample input values in
/// atrous convolution. Defaults to "1". If it is greater than 1, then
/// stride must be "1".</param>
/// <param name="dimRoundingMode">The rounding mode used when computing output
/// dimensions if pad is a number. If none is provided, it will not round
/// and error if the output is of fractional size.</param>
/// <param name="padvalue">the value of pad if pad is number</param>
/// <returns></returns>
public static Tensor conv1d(this Tensor x, Tensor filter, int stride,
PadType pad, int dilation = 1,
roundingMode dimRoundingMode = roundingMode.none, Nullable <int> padvalue = null)
{
var input3D = x;
var reshapedTo3D = false;
if (x.Rank == 2)
{
reshapedTo3D = true;
input3D = x.as3D(1, x.Shape[0], x.Shape[1]);
}
var filter4D =
filter.as4D(1, filter.Shape[0], filter.Shape[1], filter.Shape[2]);
var input4D =
input3D.as4D(input3D.Shape[0], 1, input3D.Shape[1], input3D.Shape[2]);
int[] strides = new int[] { 1, stride };
int[] dilations = new int[] { 1, dilation };
var res = conv2d(
input4D, filter4D, strides, pad, dilations,
dimRoundingMode, padvalue);
if (reshapedTo3D)
{
return(res.as2D(res.Shape[2], res.Shape[3]));
}
return(res.as3D(res.Shape[0], res.Shape[2], res.Shape[3]));
}
public Padding(int count, PadType padType)
{
padTypes = Enumerable.Repeat(padType, count).ToArray();
splitPattern = DEFAULT_SPLIT_PATTERN;
columnSeparator = " ";
text = "";
sizes = new int[0];
}
/// <summary>
/// Computes the 2D max pooling of an image.
/// </summary>
/// <param name="x">The input tensor, of rank 4 or rank 3 of shape
/// `[batch, height, width, inChannels]`. If rank 3, batch of 1 is assumed</param>
/// <param name="filterSize">The filter size, a tuple `[filterHeight, filterWidth]`.</param>
/// <param name="strides">The strides of the pooling: `[strideHeight, strideWidth]`.</param>
/// <param name="pad"> The type of padding algorithm.
/// - `same` and stride 1: output will be of same size as input,
/// regardless of filter size.
/// - `valid`: output will be smaller than input if filter is larger
/// than 1x1.
/// - For more info, see this guide:
/// [https://www.tensorflow.org/api_guides/python/nn#Convolution](
/// https://www.tensorflow.org/api_guides/python/nn#Convolution)</param>
/// <param name="dimRoundingMode">The rounding mode used when computing output
/// dimensions if pad is a number. If none is provided, it will not round
/// and error if the output is of fractional size.</param>
/// <param name="padvalue"></param>
/// <returns></returns>
public static Tensor maxPool(this Tensor x, int[] filterSize, int[] strides, PadType pad,
roundingMode dimRoundingMode = roundingMode.none, Nullable <int> padvalue = null)
{
Tensor x4D = null;
var reshapedTo4D = false;
if (x.Rank == 3)
{
reshapedTo4D = true;
x4D = x.as4D(1, x.Shape[0], x.Shape[1], x.Shape[2]);
}
else
{
x4D = x as Tensor;
}
var convInfo = Util.computePool2DInfo(
x4D.Shape, filterSize, strides, pad, dimRoundingMode, ConvDataFormat.channelsLast, padvalue);
Func <Tensor, List <Tensor>, NamedGradientMap> grad = (Tensor dy, List <Tensor> s) =>
{
NamedGradientMap g = new NamedGradientMap();
g.gradient = new Dictionary <string, Func <Tensor> >();
g.gradient.Add("x", () =>
{
var y4D = s[0];
return(maxPoolBackprop(dy, x4D, y4D, filterSize, strides, pad, dimRoundingMode, padvalue));
});
return(g);
};
Engine e = ENV.engine;
ForwardFunc f = (IBackend bk, Func <Tensor, Tensor> saved) =>
{
return(saved(bk.maxPool(x4D, convInfo)));
};
var inputs = new Dictionary <string, Tensor>();
inputs.Add("x", x4D);
var res = e.runKernel(f, inputs, grad);
if (reshapedTo4D)
{
return(res.as3D(res.Shape[1], res.Shape[2], res.Shape[3]));
}
return(res);
}
static Tensor depthwiseConv2dDerInput(int[] xShape, Tensor dy, Tensor filter,
int[] strides, PadType pad, roundingMode dimRoundingMode, Nullable <int> padValue = null)
{
int[] xShape4D = xShape;
Tensor dy4D = null;
var reshapedTo4D = false;
if (dy.Rank == 3)
{
reshapedTo4D = true;
dy4D = dy.as4D(1, dy.Shape[0], dy.Shape[1], dy.Shape[2]);
xShape4D = new int[] { 1, xShape[0], xShape[1], xShape[2] };
}
else
{
dy4D = dy as Tensor;
}
var inDepth = xShape4D[3];
var outDepth = dy4D.Shape[3];
var dilations = 1;
var convInfo = Util.computeConv2DInfo(
xShape4D, filter.Shape, strides, new int[] { dilations, dilations },
pad, dimRoundingMode,
false, ConvDataFormat.channelsLast, padValue);
Engine e = ENV.engine;
ForwardFunc f = (IBackend bk, Func <Tensor, Tensor> saved) =>
{
return(bk.depthwiseConv2DDerInput(dy4D, filter, convInfo));
};
var inputs = new Dictionary <string, Tensor>();
inputs.Add("dy4D", dy4D);
var res = e.runKernel(f, inputs);
if (reshapedTo4D)
{
return(res.as3D(res.Shape[1], res.Shape[2], res.Shape[3]));
}
return(res);
}
/// <summary>
/// Computes the backprop of a max pool.
/// </summary>
/// <param name="dy">The dy error, of rank 4 or rank 3 of shape
/// [batchSize, height, width, channels]. If rank 3, batch of 1 is assumed.</param>
/// <param name="input">The original input image, of rank 4, of shape
/// [batchSize, height, width, channels].</param>
/// <param name="output ">The original output image, of rank 4, of shape
/// [batchSize, outHeight, outWidth, channels].</param>
/// <param name="filterSize">The filter size, a tuple [filterHeight, filterWidth].</param>
/// <param name="strides">The strides of the pooling: [strideHeight, strideWidth].</param>
/// <param name="pad">A string from: 'same', 'valid'. The type of padding algorithm used in the forward prop of the op.</param>
/// <param name="dimRoundingMode">A string from: 'ceil', 'round', 'floor'. The
/// rounding mode used when computing output dimensions if pad is a
/// number. If none is provided, it will not round and error if the output
/// is of fractional size.</param>
/// <param name="padvalue"></param>
/// <returns></returns>
private static Tensor maxPoolBackprop(Tensor dy, Tensor input, Tensor output, int[] filterSize,
int[] strides, PadType pad, roundingMode dimRoundingMode, int?padvalue)
{
var convInfo = Util.computePool2DInfo(
input.Shape, filterSize, strides, pad, dimRoundingMode, ConvDataFormat.channelsLast, padvalue);
Engine e = ENV.engine;
ForwardFunc f = (IBackend bk, Func <Tensor, Tensor> saved) =>
{
return(bk.maxPoolBackprop(dy, input, output, convInfo));
};
var inputs = new Dictionary <string, Tensor>();
inputs.Add("dy", dy);
inputs.Add("input", input);
var res = e.runKernel(f, inputs);
return(res);
}
/// <summary>
/// Computes the backprop of an avg pool.
/// </summary>
/// <param name="dy">The dy error, of rank 4 or rank 3 of shape
/// [batchSize, height, width, channels]. If rank 3, batch of 1 is assumed</param>
/// <param name="input">The input image, of rank 4 or rank 3 of shape
/// [batchSize, height, width, channels]. If rank 3, batch of 1 is assumed</param>
/// <param name="filterSize">The filter size, a tuple [filterHeight, filterWidth].</param>
/// <param name="strides">The strides of the pooling: [strideHeight, strideWidth].</param>
/// <param name="pad"> A string from: 'same', 'valid'. The type of padding algorithm used in the forward prop of the op.</param>
/// <param name="padvalue"></param>
/// <returns></returns>
private static Tensor avgPoolBackprop(Tensor dy, Tensor input, int[] filterSize,
int[] strides, PadType pad, int?padvalue)
{
Tensor input4D = null;
Tensor dy4D = null;
var reshapedTo4D = false;
if (input.Rank == 3)
{
reshapedTo4D = true;
input4D = input.as4D(1, input.Shape[0], input.Shape[1], input.Shape[2]);
dy4D = dy.as4D(1, dy.Shape[0], dy.Shape[1], dy.Shape[2]);
}
else
{
input4D = input as Tensor;
dy4D = dy as Tensor;
}
var convInfo = Util.computePool2DInfo(
input4D.Shape, filterSize, strides, pad, roundingMode.none, ConvDataFormat.channelsLast, padvalue);
Engine e = ENV.engine;
ForwardFunc f = (IBackend bk, Func <Tensor, Tensor> saved) =>
{
return(bk.avgPoolBackprop(dy4D, input4D, convInfo));
};
var inputs = new Dictionary <string, Tensor>();
inputs.Add("dy4D", dy4D);
inputs.Add("input4D", input4D);
var res = e.runKernel(f, inputs);
if (reshapedTo4D)
{
return(res.as3D(res.Shape[1], res.Shape[2], res.Shape[3]));
}
return(res);
}
static Tensor depthwiseConv2dDerFilter(this Tensor x, Tensor dy, int[] filterShape, int[] strides,
PadType pad, roundingMode dimRoundingMode = roundingMode.none, Nullable <int> padValue = null)
{
Tensor x4D = null;
if (x.Rank == 3)
{
x4D = x.as4D(1, x.Shape[0], x.Shape[1], x.Shape[2]);
}
else
{
x4D = x as Tensor;
}
Tensor dy4D = null;
if (dy.Rank == 3)
{
dy4D = dy.as4D(1, dy.Shape[0], dy.Shape[1], dy.Shape[2]);
}
else
{
dy4D = dy as Tensor;
}
var dilations = 1;
var convInfo = Util.computeConv2DInfo(
x4D.Shape, filterShape, strides, new int[] { dilations, dilations }, pad, dimRoundingMode, false, ConvDataFormat.channelsLast, padValue);
Engine e = ENV.engine;
ForwardFunc f = (IBackend bk, Func <Tensor, Tensor> saved) =>
{
return(bk.depthwiseConv2DDerFilter(x4D, dy4D, convInfo));
};
var inputs = new Dictionary <string, Tensor>();
inputs.Add("x4D", x4D);
inputs.Add("dy4D", dy4D);
return(e.runKernel(f, inputs));
}
string PaddingInBytes(string value, PadType type, int byteCount)
{
Encoding enc = Encoding.GetEncoding("Shift_JIS");
if (byteCount < enc.GetByteCount(value))
{
// valueが既定のバイト数を超えている場合は、切り落とし
value = value.Substring(0, byteCount);
}
switch (type)
{
case PadType.Char:
// 文字列の場合 左寄せ＋空白埋め
return value.PadRight(byteCount - (enc.GetByteCount(value) - value.Length));
case PadType.Number:
// 数値の場合 右寄せ＋0埋め
return value.PadLeft(byteCount, ' ');
default:
// 上記以外は全部空白
return value.PadLeft(byteCount);
}
}
/// <summary>Returns a new instance of the appropriate class of an object based on its object ID.</summary> /// <param name="objectID">The object ID of the new object.</param> public static GeneralObject GetNewObjectInstance(PadType objectID) => GetNewObjectInstance((int)objectID);
/// <summary>Creates a new instance of the <seealso cref="ObjectIDAttribute"/> attribute.</summary>
/// <param name="objectID">The object ID of the <seealso cref="Pad"/>.</param>
public ObjectIDAttribute(PadType objectID) : this((int)objectID)
{
}
public static string Pad(string text, PadType type, int maximumLength, char paddingCharacter = ' ') => type switch
{
/// <summary>
/// 2-D convolution with separable filters.
///
/// Performs a depthwise convolution that acts separately on channels followed
/// by a pointwise convolution that mixes channels. Note that this is
/// separability between dimensions [1, 2] and 3, not spatial separability
/// between dimensions 1 and 2.
///
/// See
/// [https://www.tensorflow.org/api_docs/python/tf/nn/separable_conv2d](
/// https://www.tensorflow.org/api_docs/python/tf/nn/separable_conv2d)
/// for more details.
/// </summary>
/// <param name="x">The input tensor, of rank 4 or rank 3, of shape
/// `[batch, height, width, inChannels]`. If rank 3, batch of 1 is
/// assumed.</param>
/// <param name="depthwiseFilter">The depthwise filter tensor, rank 4, of shape
/// `[filterHeight, filterWidth, inChannels, channelMultiplier]`. This is
/// the filter used in the first step.</param>
/// <param name="pointwiseFilter">The pointwise filter tensor, rank 4, of shape
/// `[1, 1, inChannels * channelMultiplier, outChannels]`. This is
/// the filter used in the second step.</param>
/// <param name="strides">The strides of the convolution: `[strideHeight,
/// strideWidth]`. If strides is a single number, then `strideHeight ==
/// strideWidth`.</param>
/// <param name="pad"> The type of padding algorithm.
/// - `same` and stride 1: output will be of same size as input,
/// regardless of filter size.
/// - `valid`: output will be smaller than input if filter is larger
/// than 1x1.
/// - For more info, see this guide:
/// [https://www.tensorflow.org/api_guides/python/nn#Convolution](
/// https://www.tensorflow.org/api_guides/python/nn#Convolution)</param>
/// <param name="dilation">The dilation rates: `[dilationHeight, dilationWidth]`
/// in which we sample input values across the height and width dimensions
/// in atrous convolution. Defaults to `[1, 1]`. If `rate` is a single
/// number, then `dilationHeight == dilationWidth`. If it is greater than
/// 1, then all values of `strides` must be 1.</param>
/// <param name="padvalue">the value of pad if pad is number</param>
/// <returns></returns>
public static Tensor separableConv2d(this Tensor x, Tensor depthwiseFilter, Tensor pointwiseFilter, int[] strides, PadType pad
, int[] dilation, Nullable <int> padvalue = null)
{
var x4D = x;
var reshapedTo4D = false;
if (x.Rank == 3)
{
reshapedTo4D = true;
x4D = x.as4D(1, x.Shape[0], x.Shape[1], x.Shape[2]);
}
Util.assert(
x4D.Rank == 4,
"Error in separableConv2d: input must be Rank 4, but got " +
"Rank ${x4D.Rank}.");
Util.assert(
depthwiseFilter.Rank == 4,
"Error in separableConv2d: depthwise filter must be Rank 4, but got " +
"Rank ${depthwiseFilter.Rank}.");
Util.assert(
pointwiseFilter.Rank == 4,
"Error in separableConv2d: pointwise filter must be Rank 4, but got " +
"Rank ${depthwiseFilter.Rank}.");
Util.assert(
pointwiseFilter.Shape[0] == 1,
"Error in separableConv2d: the first dimension of pointwise filter " +
" must be 1, but got ${pointwiseFilter.shape[0]}.");
Util.assert(
pointwiseFilter.Shape[1] == 1,
"Error in separableConv2d: the second dimension of pointwise filter " +
" must be 1, but got ${pointwiseFilter.shape[1]}.");
var inChannels = depthwiseFilter.Shape[2];
var channelMultiplier = depthwiseFilter.Shape[3];
Util.assert(
pointwiseFilter.Shape[2] == inChannels * channelMultiplier,
"Error in separableConv2d: the third dimension of pointwise filter " +
"must be ${inChannels * channelMultiplier}, " +
"but got ${pointwiseFilter.shape[2]}.");
var depthwise = depthwiseConv2d(
x4D, depthwiseFilter, strides, pad, dilation, roundingMode.none, padvalue);
var pointwiseStride = 1;
var res = conv2d(
depthwise, pointwiseFilter,
new int[] { pointwiseStride, pointwiseStride }, PadType.valid);
if (reshapedTo4D)
{
return(res.as3D(res.Shape[1], res.Shape[2], res.Shape[3]));
}
return(res);
}
public static Conv2DInfo computeConv2DInfo(int[] inShape, int[] filterShape,
int[] strides, int[] dilations, PadType pad,
roundingMode roundingMode = roundingMode.none, bool depthwise = false,
ConvDataFormat dataFormat = ConvDataFormat.channelsLast, Nullable <int> padValue = null)
{
var batchSize = -1;
var inHeight = -1;
var inWidth = -1;
var inChannels = -1;
if (dataFormat == ConvDataFormat.channelsLast)
{
batchSize = inShape[0];
inHeight = inShape[1];
inWidth = inShape[2];
inChannels = inShape[3];
}
else
{
batchSize = inShape[0];
inChannels = inShape[1];
inHeight = inShape[2];
inWidth = inShape[3];
}
var filterHeight = filterShape[0];
var filterWidth = filterShape[1];
var filterChannels = filterShape[3];
var strideHeight = strides[0];
var strideWidth = strides[1];
var dilationHeight = dilations[0];
int dilationWidth = 0;
if (dilations.Length > 1)
{
dilationWidth = dilations[1];
}
else
{
dilationWidth = dilations[0];
}
var effectiveFilterHeight = getEffectiveFilterSize(filterHeight, dilationHeight);
var effectiveFilterWidth =
getEffectiveFilterSize(filterWidth, dilationWidth);
var d = getPadAndOutInfo(
pad, inHeight, inWidth, strideHeight, strideWidth, effectiveFilterHeight,
effectiveFilterWidth, roundingMode, padValue);
var padInfo = d.Item1;
var outHeight = d.Item2;
var outWidth = d.Item3;
var outChannels = depthwise ? filterChannels * inChannels : filterChannels;
int[] outShape = null;
if (dataFormat == ConvDataFormat.channelsFirst)
{
outShape = new int[] { batchSize, outChannels, outHeight, outWidth };
}
else if (dataFormat == ConvDataFormat.channelsLast)
{
outShape = new int[] { batchSize, outHeight, outWidth, outChannels };
}
return(new Conv2DInfo()
{
batchSize = batchSize,
dataFormat = dataFormat,
inHeight = inHeight,
inWidth = inWidth,
inChannels = inChannels,
outHeight = outHeight,
outWidth = outWidth,
outChannels = outChannels,
padInfo = padInfo,
strideHeight = strideHeight,
strideWidth = strideWidth,
filterHeight = filterHeight,
filterWidth = filterWidth,
dilationHeight = dilationHeight,
dilationWidth = dilationWidth,
inShape = inShape,
outShape = outShape,
filterShape = filterShape
});
}
/// <summary>
/// Computes the transposed 2D convolution of an image, also known as a
/// deconvolution.
/// </summary>
/// <param name="x">The input image, of Rank 4 or Rank 3, of shape
/// "[batch, height, width, inDepth]". If Rank 3, batch of 1 is assumed.</param>
/// <param name="filter">The filter, Rank 4, of shape
/// "[filterHeight, filterWidth, outDepth, inDepth]".
/// "inDepth" must match "inDepth" in "x".</param>
/// <param name="outputShape">outputShape Output shape, of Rank 4 or Rank 3:
/// "[batch, height, width, outDepth]". If Rank 3, batch of 1 is assumed.</param>
/// <param name="strides">The strides of the original convolution:
/// "[strideHeight, strideWidth]".</param>
/// <param name="pad">The type of padding algorithm used in the non-transpose version
/// of the op.</param>
/// <param name="dimRoundingMode">The rounding mode used when computing output
/// dimensions if pad is a number. If none is provided, it will not round
/// and error if the output is of fractional size.</param>
/// <param name="padvalue">the value of pad if pad is number</param>
/// <returns></returns>
public static Tensor conv2dTranspose(this Tensor x, Tensor filter, int[] outputShape, int[] strides, PadType pad, roundingMode dimRoundingMode = roundingMode.none,
Nullable <int> padvalue = null)
{
return(conv2dDerInput(
outputShape, x, filter, strides, pad, dimRoundingMode, padvalue));
}
/// <summary>
/// Fill the string with the new length.
/// </summary>
/// <param name="length">The new string length. If the value is less than the original length of the string, no action is taken.</param>
/// <param name="str">The string to be filled.</param>
/// <param name="padStr">A string to be used for padding. The default is blank.</param>
/// <param name="type">
/// Fill in which side of the string.
/// <para><see cref="PadType.Both"/>Fill both sides of the string. If not even, get extra padding on the right side.</para>
/// <para><see cref="PadType.Left"/>Fill the left side of the string.</para>
/// <para><see cref="PadType.Right"/>Fill the right side of the string.</para>
/// </param>
/// <returns>Returns filled string.</returns>
public static string Pad(int length, string str = null, string padStr = null, PadType type = PadType.Right)
{
str = str ?? string.Empty;
var needlePadding = length - str.Length;
if (needlePadding <= 0)
{
return(str);
}
int rightPadding;
var leftPadding = rightPadding = 0;
if (type == PadType.Both)
{
leftPadding = needlePadding >> 1;
rightPadding = (needlePadding >> 1) + (needlePadding % 2 == 0 ? 0 : 1);
}
else if (type == PadType.Right)
{
rightPadding = needlePadding;
}
else
{
leftPadding = needlePadding;
}
padStr = padStr ?? Space;
padStr = padStr.Length <= 0 ? Space : padStr;
var leftPadCount = (leftPadding / padStr.Length) + (leftPadding % padStr.Length == 0 ? 0 : 1);
var rightPadCount = (rightPadding / padStr.Length) + (rightPadding % padStr.Length == 0 ? 0 : 1);
return(Repeat(padStr, leftPadCount).Substring(0, leftPadding) + str +
Repeat(padStr, rightPadCount).Substring(0, rightPadding));
}
public MetaData(PadType padType)
{
MaxLength = 0;
PadType = padType;
}
public static Tuple <PadInfo, int, int> getPadAndOutInfo(PadType pad, int inHeight, int inWidth,
int strideHeight, int strideWidth, int filterHeight, int filterWidth,
roundingMode roundingMode = roundingMode.none, Nullable <int> padValue = null)
{
var padInfo = new PadInfo();
var outHeight = 0;
var outWidth = 0;
if (pad == PadType.number)
{
padInfo.bottom = padValue.Value;
padInfo.left = padValue.Value;
padInfo.right = padValue.Value;
padInfo.top = padValue.Value;
var outShape = computeOutputShape3D(
new int[] { inHeight, inWidth, 1 }, filterHeight, 1, strideHeight, padValue,
roundingMode);
outHeight = outShape[0];
outWidth = outShape[1];
padInfo.alongh = padValue.Value;
padInfo.alongw = padValue.Value;
}
else if (pad == PadType.valid)
{
padInfo.bottom = 0;
padInfo.left = 0;
padInfo.right = 0;
padInfo.top = 0;
outHeight = (int)Math.Ceiling((inHeight - filterHeight + 1d) / strideHeight);
outWidth = (int)Math.Ceiling((inWidth - filterWidth + 1d) / strideWidth);
padInfo.alongh = 0;
padInfo.alongw = 0;
}
else if (pad == PadType.same)
{
outHeight = (int)Math.Ceiling(inHeight / (float)strideHeight);
outWidth = (int)Math.Ceiling(inWidth / (float)strideWidth);
var padAlongHeight =
(outHeight - 1) * strideHeight + filterHeight - inHeight;
var padAlongWidth = (outWidth - 1) * strideWidth +
filterWidth - inWidth;
var top = (int)Math.Floor(padAlongHeight / 2f);
var bottom = (int)padAlongHeight - top;
var left = (int)Math.Floor(padAlongWidth / 2f);
var right = (int)padAlongWidth - left;
padInfo.bottom = bottom;
padInfo.left = left;
padInfo.right = right;
padInfo.top = top;
padInfo.alongh = padAlongHeight;
padInfo.alongw = padAlongWidth;
}
else
{
throw new Exception("Unknown padding parameter");
}
return(new Tuple <PadInfo, int, int>(padInfo, outHeight, outWidth));
}
/// <summary>
/// Depthwise 2D convolution.
/// * Given a 4D "input" array and a "filter" array of shape
/// "[filterHeight, filterWidth, inChannels, channelMultiplier]" containing
/// "inChannels" convolutional filters of depth 1, this op applies a
/// different filter to each input channel (expanding from 1 channel to
/// "channelMultiplier" channels for each), then concatenates the results
/// together. The output has "inChannels * channelMultiplier" channels.
///
/// See
/// [https://www.tensorflow.org/api_docs/python/tf/nn/depthwise_conv2d](
/// https://www.tensorflow.org/api_docs/python/tf/nn/depthwise_conv2d)
/// for more details.
/// </summary>
/// <param name="input">The input tensor, of Rank 4 or Rank 3, of shape
/// "[batch, height, width, inChannels]". If Rank 3, batch of 1 is assumed.</param>
/// <param name="filter">The filter tensor, Rank 4, of shape
/// "[filterHeight, filterWidth, inChannels, channelMultiplier]"</param>
/// <param name="strides">The strides of the convolution: "[strideHeight,
/// strideWidth]". If strides is a single number, then "strideHeight ==
/// strideWidth".</param>
/// <param name="pad">The type of padding algorithm.
/// - "same" and stride 1: output will be of same size as input,
/// regardless of filter size.
/// - "valid": output will be smaller than input if filter is larger
/// than 1x1.
/// - For more info, see this guide:
/// [https://www.tensorflow.org/api_guides/python/nn#Convolution](
/// https://www.tensorflow.org/api_guides/python/nn#Convolution)</param>
/// <param name="dilations">The dilation rates: "[dilationHeight, dilationWidth]"
/// in which we sample input values across the height and width dimensions
/// in atrous convolution. Defaults to "[1, 1]". If "rate" is a single
/// number, then "dilationHeight == dilationWidth". If it is greater than
/// 1, then all values of "strides" must be 1.</param>
/// <param name="dimRoundingMode">The rounding mode used when computing output
/// dimensions if pad is a number. If none is provided, it will not round
/// and error if the output is of fractional size.</param>
/// <param name="padvalue">the value of pad if pad is number</param>
/// <returns></returns>
public static Tensor depthwiseConv2d(this Tensor input, Tensor filter,
int[] strides, PadType pad, int[] dilations = null,
roundingMode dimRoundingMode = roundingMode.none, Nullable <int> padvalue = null)
{
if (dilations == null)
{
dilations = new int[] { 1, 1 };
}
Tensor input4D = null;
var reshapedTo4D = false;
if (input.Rank == 3)
{
reshapedTo4D = true;
input4D = input.as4D(1, input.Shape[0], input.Shape[1], input.Shape[2]);
}
else
{
input4D = input as Tensor;
}
var convInfo = Util.computeConv2DInfo(
input4D.Shape, filter.Shape, strides, dilations, pad, dimRoundingMode,
true /* depthwise */, ConvDataFormat.channelsLast, padvalue);
Func <Tensor, List <Tensor>, NamedGradientMap> grad = (Tensor dy,
List <Tensor> s) =>
{
NamedGradientMap g = new NamedGradientMap();
g.gradient = new Dictionary <string, Func <Tensor> >();
g.gradient.Add("input4D", () =>
{
var resg = depthwiseConv2dDerInput(input4D.Shape, dy, filter, strides, pad, dimRoundingMode, padvalue);
if (reshapedTo4D)
{
resg = resg.as3D(resg.Shape[1], resg.Shape[2], resg.Shape[3]);
}
return(resg);
});
g.gradient.Add("filter", () =>
{
return(depthwiseConv2dDerFilter(input4D, dy, filter.Shape, strides, pad, dimRoundingMode, padvalue));
});
return(g);
};
Engine e = ENV.engine;
ForwardFunc f = (IBackend bk, Func <Tensor, Tensor> saved) =>
{
return(bk.depthwiseConv2D(input4D, filter, convInfo));
};
var inputs = new Dictionary <string, Tensor>();
inputs.Add("input4D", input4D);
inputs.Add("filter", filter);
var res = e.runKernel(f, inputs, grad);
if (reshapedTo4D)
{
return(res.as3D(res.Shape[1], res.Shape[2], res.Shape[3]));
}
return(res);
}
public static int[,] Filter_int(int[,] arr, int[,] filter, PadType padType)
{
int width = arr.GetLength(1);
int height = arr.GetLength(0);
int[,] temp;
PadMyArray <int> padArr = new PadMyArray <int>();
int[,] result = new int[height, width];
int[,] toConv = new int[filter.GetLength(0), filter.GetLength(1)];
if (arr.Length < filter.Length)
{
Console.WriteLine("Cannot filter image, less than filter window. Returned array with zeros. Method: filter_int");
return(result);
}
int padsizeR, padsizeC = 0;
if (filter.GetLength(0) % 2 == 0)
{
padsizeR = filter.GetLength(0) / 2;
}
else
{
padsizeR = (filter.GetLength(0) - 1) / 2;
}
if (filter.GetLength(1) % 2 == 0)
{
padsizeC = filter.GetLength(1) / 2;
}
else
{
padsizeC = (filter.GetLength(1) - 1) / 2;
}
temp = padArr.PadArray(arr, padsizeR, padsizeC, padType, Direction.both);
//filtering
//obtain part of image array and convolution with filter window
try
{
for (int i = 1; i <= height; i++)
{
for (int j = 1; j <= width; j++)
{
for (int m = 0; m < filter.GetLength(0); m++)
{
for (int n = 0; n < filter.GetLength(1); n++)
{
toConv[m, n] = temp[i + m - 1, j + n - 1];
}
}
//such size coz array mult by elements, and we take part same size as filter window
int[,] convolution = new int[filter.GetLength(0), filter.GetLength(1)];
convolution = toConv.ArrayMultElements(filter);
//get elemt after filter
if (convolution.Cast <int>().Sum() < 0)
{
result[i - 1, j - 1] = 0;
}
else if (convolution.Cast <int>().Sum() > 255)
{
result[i - 1, j - 1] = 255;
}
else
{
result[i - 1, j - 1] = convolution.Cast <int>().Sum();
}
//result[i - 1, j - 1] = convolution.Cast<int>().Sum(); //get elemt after filter
}
}
}
catch (Exception e)
{
Console.WriteLine("Problem in filter, most likely OutOfRangeException, here message:\n" +
e.Message);
}
return(result);
}
public void AddColumn(string text, PadType padType)
{
overriddenPadTypes[getKey()] = padType;
columns.Add(text);
}
