DataBlock

Full name Microsoft.VisualBasic.MachineLearning.CNN.data.DataBlock Assembly Microsoft.VisualBasic.DeepLearning Members 47

Holding all the data handled by the network. So a layer will receive this class and return a similar block as a output that will be used by the next layer in the chain.

@author Daniel Persson (mailto.woden@gmail.com)

01 Syntax

Microsoft.VisualBasic.MachineLearning.CNN.data.DataBlock

02 Methods

Name	Overloads	Summary
SetValues	1	把整块数据写入 DataBlock.w，并声明值数据的设备端缓存已失效
SetGradients	1	把整块数据写入 DataBlock.dw，并声明梯度数据的设备端缓存已失效
MarkValueModified	1	声明值数据的设备端缓存已失效（就地改写 DataBlock.w 之后必须调用）
MarkGradientModified	1	声明梯度数据的设备端缓存已失效（就地改写 DataBlock.dw 之后必须调用）
MarkHostModified	1	对已创建的视图逐个推进版本号（未创建的视图本来就没有显存副本，无需处理）
.ctor	4	Creates an empty block, used by the deserializer.
ToString	1	Returns a short diagnostic description of the block.
addImageData	3	Prepares the input by copying the pixels into the weight vector and normalizing them into [-0.5, 0.5].
getWeight	2	Gets the weight at the given three dimensional coordinate.
setWeight	2	get DataBlock.w
addWeight	1	Adds a value to the weight at the given three dimensional coordinate.
getGradient	2	Gets the gradient at the given three dimensional coordinate.
setGradient	3	Replaces the whole gradient vector with the given values.
addGradient	2	Adds a value to the gradient at the given flat index.
subGradient	1	Subtracts a value from the gradient at the given flat index.
mulGradient	1	val * DataBlock.Gradients
cloneAndZero	1	Creates a new block with the same shape and trace but with all weights zeroed.
clone	1	make a copy of DataBlock.w or DataBlock.Weights
clearGradient	1	just set DataBlock.dw vector to zero
addFrom	1	Copies the weights of another block into this one.
addFromScaled	1	Copies the weights of another block into this one, scaled by a constant factor.

03 Properties

Name	Overloads	Summary
SX	1	Gets the width of the block.
SY	1	Gets the height of the block.
Depth	1	Gets the number of channels (depth) of the block.
trace	1	Optional diagnostic label that identifies where this block came from.
Weights	1	the multiple class classify probability weight (or the prediction result)
Gradients	1	Gets the gradient vector that corresponds one to one with DataBlock.Weights.
Value	1	值数据的张量视图（零拷贝，与 DataBlock.Weights 共享同一份底层存储）。形状为 (SY, SX, Depth)，正好对应索引公式 (SXy + x)Depth + depth。
Grad	1	梯度数据的张量视图（零拷贝，与 DataBlock.Gradients 共享同一份底层存储）。
TensorShape4D	1	把本块当作单样本图像时的四维形状 (N=1, H=SY, W=SX, C=Depth)
Value4D	1	值数据的四维视图 (1, SY, SX, Depth)：供卷积/池化算子使用
Grad4D	1	梯度数据的四维视图 (1, SY, SX, Depth)：作为卷积/池化的 gradOutput
Value2D	1	值数据的二维列向量视图 (n, 1)：供矩阵乘使用
Grad2D	1	梯度数据的二维列向量视图 (n, 1)：作为矩阵乘的 gradOutput

04 Fields

Name	Overloads	Summary
w	1	backend of DataBlock.Weights
dw	1	backend of DataBlock.Gradients

05 Members

method SetValues #

SetValues(Double())

把整块数据写入 DataBlock.w，并声明值数据的设备端缓存已失效

method SetGradients #

SetGradients(Double())

把整块数据写入 DataBlock.dw，并声明梯度数据的设备端缓存已失效

method MarkValueModified #

MarkValueModified

声明值数据的设备端缓存已失效（就地改写 DataBlock.w 之后必须调用）

Remarks

需要把本块持有的全部视图都标记一遍：三维的 DataBlock.Value、四维的 DataBlock.Value4D、二维的 DataBlock.Value2D。它们各自持有独立的版本号，少标一个就会出现 "某个算子读到旧值"的静默错误。由于视图都由本块缓存，这里已经不需要再调用全局的 Tensor.InvalidateAllDeviceCaches，失效粒度从"所有张量"收敛到"本块"。

method MarkGradientModified #

MarkGradientModified

声明梯度数据的设备端缓存已失效（就地改写 DataBlock.dw 之后必须调用）

method MarkHostModified #

MarkHostModified(Tensor)

对已创建的视图逐个推进版本号（未创建的视图本来就没有显存副本，无需处理）

method .ctor #

#ctor

Creates an empty block, used by the deserializer.

method .ctor overload 4 #

#ctor(Int32, Int32, Int32)

Creates a block of the given size with randomly initialized weights.

Parameters

Name	Type	Description
`sx`	`Int32`	Width.
`sy`	`Int32`	Height.
`depth`	`Int32`	Number of channels.

method .ctor overload 2 #

#ctor(Dimension, Int32, Double)

Creates a block from a Dimension and a weight initialization value.

Parameters

Name	Type	Description
`dims`	Dimension	The block dimensions.
`depth`	`Int32`	Number of channels.
`c`	`Double`	The constant used to fill the weights, or `-1` for random initialization.

method .ctor overload 3 #

#ctor(Int32, Int32, Int32, Double)

Creates a block of the given size.

Parameters

Name	Type	Description
`sx`	`Int32`	Width.
`sy`	`Int32`	Height.
`depth`	`Int32`	Number of channels.
`c`	`Double`	Constant used to initialize the weight vector; when it equals `-1` the weights are randomly initialized instead.

method ToString #

ToString

Returns a short diagnostic description of the block.

Returns

A text that reports the shape, the trace label and the first few weight values.

method addImageData #

addImageData(Byte(), Byte)

Prepares the input by copying the pixels into the weight vector and normalizing them into [-0.5, 0.5].

Parameters

Name	Type	Description
`imgData`	`Byte()`	The raw byte pixel data.
`maxvalue`	`Byte`	The maximum possible channel value, normally 255.

method addImageData overload 2 #

addImageData(Double(), Double)

Prepares the input by copying the pixels into the weight vector and normalizing them into [-0.5, 0.5].

Parameters

Name	Type	Description
`imgData`	`Double()`	The raw pixel data.
`maxvalue`	`Double`	The maximum possible channel value.

method addImageData overload 3 #

addImageData(Int32(), Int32)

Prepares the input by copying the pixels into the weight vector and normalizing them into [-0.5, 0.5].

Parameters

Name	Type	Description
`imgData`	`Int32()`	The raw integer pixel data.
`maxvalue`	`Int32`	The maximum possible channel value.

method getWeight overload 2 #

getWeight(Int32)

set DataBlock.w

Parameters

Name	Type	Description
`ix`	`Int32`	-

method getWeight #

getWeight(Int32, Int32, Int32)

Gets the weight at the given three dimensional coordinate.

Parameters

Name	Type	Description
`x`	`Int32`	The column index.
`y`	`Int32`	The row index.
`depth`	`Int32`	The channel index.

Returns

The weight value.

method setWeight #

setWeight(Int32, Double)

get DataBlock.w

Parameters

Name	Type	Description
`ix`	`Int32`	-
`val`	`Double`	-

method setWeight overload 2 #

setWeight(Int32, Int32, Int32, Double)

Sets the weight at the given three dimensional coordinate.

Parameters

Name	Type	Description
`x`	`Int32`	The column index.
`y`	`Int32`	The row index.
`depth`	`Int32`	The channel index.
`val`	`Double`	The value to store.

method addWeight #

addWeight(Int32, Int32, Int32, Double)

Adds a value to the weight at the given three dimensional coordinate.

Parameters

Name	Type	Description
`x`	`Int32`	The column index.
`y`	`Int32`	The row index.
`depth`	`Int32`	The channel index.
`val`	`Double`	The value to add.

method getGradient #

getGradient(Int32, Int32, Int32)

Gets the gradient at the given three dimensional coordinate.

Parameters

Name	Type	Description
`x`	`Int32`	The column index.
`y`	`Int32`	The row index.
`depth`	`Int32`	The channel index.

Returns

The gradient value.

method getGradient overload 2 #

getGradient(Int32)

get DataBlock.Gradients

Remarks

get element value from DataBlock.dw

Parameters

Name	Type	Description
`ix`	`Int32`	-

method setGradient overload 3 #

setGradient(Int32, Int32, Int32, Double)

Sets the gradient at the given three dimensional coordinate.

Parameters

Name	Type	Description
`x`	`Int32`	The column index.
`y`	`Int32`	The row index.
`depth`	`Int32`	The channel index.
`val`	`Double`	The value to store.

method setGradient overload 2 #

setGradient(Int32, Double)

set element value to DataBlock.Gradients

Remarks

set value to DataBlock.dw

Parameters

Name	Type	Description
`ix`	`Int32`	-
`val`	`Double`	-

method setGradient #

setGradient(Double())

Replaces the whole gradient vector with the given values.

Parameters

Name	Type	Description
`val`	`Double()`	The new gradient values.

method addGradient overload 2 #

addGradient(Int32, Int32, Int32, Double)

Adds a value to the gradient at the given three dimensional coordinate.

Parameters

Name	Type	Description
`x`	`Int32`	The column index.
`y`	`Int32`	The row index.
`depth`	`Int32`	The channel index.
`val`	`Double`	The value to add.

method addGradient #

addGradient(Int32, Double)

Adds a value to the gradient at the given flat index.

Parameters

Name	Type	Description
`ix`	`Int32`	The flat index into the gradient vector.
`val`	`Double`	The value to add.

method subGradient #

subGradient(Int32, Double)

Subtracts a value from the gradient at the given flat index.

Parameters

Name	Type	Description
`ix`	`Int32`	The flat index into the gradient vector.
`val`	`Double`	The value to subtract.

method mulGradient #

mulGradient(Double)

val * DataBlock.Gradients

Parameters

Name	Type	Description
`val`	`Double`	-

method cloneAndZero #

cloneAndZero

Creates a new block with the same shape and trace but with all weights zeroed.

Returns

A zero filled copy of this block.

method clone #

clone

make a copy of DataBlock.w or DataBlock.Weights

method clearGradient #

clearGradient

just set DataBlock.dw vector to zero

method addFrom #

addFrom(DataBlock)

Copies the weights of another block into this one.

Parameters

Name	Type	Description
`db`	DataBlock	The source block; it must have the same length.

method addFromScaled #

addFromScaled(DataBlock, Double)

Copies the weights of another block into this one, scaled by a constant factor.

Parameters

Name	Type	Description
`db`	DataBlock	The source block; it must have the same length.
`a`	`Double`	The scale factor applied to every source weight.

property SX #

Gets the width of the block.

property SY #

Gets the height of the block.

property Depth #

Depth

Gets the number of channels (depth) of the block.

property trace #

trace

Optional diagnostic label that identifies where this block came from.

property Weights #

Weights

the multiple class classify probability weight (or the prediction result)

Remarks

get DataBlock.w

property Gradients #

Gradients

Gets the gradient vector that corresponds one to one with DataBlock.Weights.

Remarks

Returns the backing field DataBlock.dw.

property Value #

Value

值数据的张量视图（零拷贝，与 DataBlock.Weights 共享同一份底层存储）。形状为 (SY, SX, Depth)，正好对应索引公式 (SXy + x)Depth + depth。

property Grad #

Grad

梯度数据的张量视图（零拷贝，与 DataBlock.Gradients 共享同一份底层存储）。

property TensorShape4D #

TensorShape4D

把本块当作单样本图像时的四维形状 (N=1, H=SY, W=SX, C=Depth)

property Value4D #

Value4D

值数据的四维视图 (1, SY, SX, Depth)：供卷积/池化算子使用

property Grad4D #

Grad4D

梯度数据的四维视图 (1, SY, SX, Depth)：作为卷积/池化的 gradOutput

property Value2D #

Value2D

值数据的二维列向量视图 (n, 1)：供矩阵乘使用

property Grad2D #

Grad2D

梯度数据的二维列向量视图 (n, 1)：作为矩阵乘的 gradOutput

field w #

backend of DataBlock.Weights

field dw #

backend of DataBlock.Gradients