The ComputeGraph class provides a simple computation graph of operations used in a forward pass that are stored in an array on each call and then unwound with calls that calculate the gradients on the backward pass. More...

Inheritance diagram for MyCaffe.common.ComputeGraph< T >:

Public Member Functions
	ComputeGraph (CudaDnn< T > cuda, Log log, int nAxis, bool bNeedsBackprop=true, bool bClipGradients=false, bool bCheckNans=false, bool bAddDebug=false)
	The constructor. More...

void	Dispose ()
	Release all resources used. More...

void	DebugOp (params Blob< T >[] rgB)
	DebugOp operation places a debug stub in the backpropagation chain for debugging only. More...

Blob< T >	Round (Blob< T > b, int nDecimals=6)
	Round operation, rounds the values to the nearest specified decimal. More...

Blob< T >	PeekRow (Blob< T > btm, Blob< T > top, int ix)
	PeeKRow operation copies data and diffs from one row from within the bottom matrix and places it in the top vector. More...

void	PeekItem (Blob< T > btm, Blob< T > top, int ix)
	PeekItem operation copies a single item from the bottom to the top. More...

Blob< T >	CopyToRow (Blob< T > btm, Blob< T > top, int ix, bool bCopyDiff=false)
	CopyToRow operation copies the bottom vector into the top matrix. More...

void	CopyToCache (Blob< T > btm, Cache< T > cache)
	CopyToCache operation copies the blob into the cache. More...

Blob< T >	tanh (Blob< T > btm, Blob< T > top)
	'tanh' operation runs the tanh on each item in the btm and places the results in the top. More...

Blob< T >	sigmoid (Blob< T > btm, Blob< T > top)
	'sigmoid' operation runs the sigmoid on each item in the btm and places the results in the top. More...

Blob< T >	eltmul (Blob< T > btm1, Blob< T > btm2, Blob< T > top)
	'elthmul' operation mutliplies each element of the 'btm1' with the 'btm2' and places the results in 'top'. More...

Blob< T >	scalemul (Blob< T > btm1, Blob< T > btm2, Blob< T > top, int nIdx=0)
	'scalemul' operation mutliplies each element of the 'btm1' with the first item within 'btm2' and places the results in 'top'. More...

Blob< T >	mul (Blob< T > btm1, Blob< T > btm2, Blob< T > top, bool bAccumulateGrad=true)
	'mul' operation performs a blas gemm operation on the 'btm1' matrix with the 'btm2' matrix and places the results in 'top'. More...

Blob< T >	add (Blob< T > btm1, Blob< T > btm2, Blob< T > top, bool bAccumulateGrad=true)
	'elthmul' operation adds each element of the 'btm1' with the 'btm2' and places the results in 'top'. More...

Blob< T >	clear_grad (Blob< T > b)
	'clear_grad' operation only runs on the backward pass and zeros out the gradients on an input. More...

void	clear_grad (BlobCollection< T > rg)
	'clear_grad' operation only runs on the backward pass and zeros out the gradients of the inputs. More...

Blob< T >	softmax (Blob< T > btm, Blob< T > top)
	'softmax' operation runs the softmax on each item in the btm and places the results in the top. More...

void	BackwardOne (int nIdx)
	Runs a backward operation at a given index. More...

void	Backward (bool bClear=false)
	Runs a backward operation on all items starting from the last and running through the first. More...

void	Clear ()
	Clears all backward operations from the list. More...

Properties
Dictionary< string, Blob< T >[]>	Debug `[get]`
	Returns a dictionary of Blobs used during each operation, only filled when 'bAddDebug' = true in the constructor. More...

string	marker `[getset]`
	Get/set a string marker added to the debug information and used to indicate where in the code a given operation takes place. More...

bool	needs_backprop `[getset]`
	Get/set whether or not to back propagate. More...

int	axis `[get]`
	Returns the axis on which all operations are performed. More...

int	BackwardCount `[get]`
	Returns the backward operation count. More...

Detailed Description

The ComputeGraph class provides a simple computation graph of operations used in a forward pass that are stored in an array on each call and then unwound with calls that calculate the gradients on the backward pass.

This class is a re-write of the ComputeGraph originally created by Mohamed Ashmawy in the open-source project mashmawy/Seq2SeqLearn distributed under the MIT license.

Template Parameters

T	Specifies the base type.

Definition at line 21 of file ComputeGraph.cs.

Constructor & Destructor Documentation

◆ ComputeGraph()

MyCaffe.common.ComputeGraph< T >.ComputeGraph	(	CudaDnn< T >	cuda,
		Log	log,
		int	nAxis,
		bool	bNeedsBackprop = `true`,
		bool	bClipGradients = `false`,
		bool	bCheckNans = `false`,
		bool	bAddDebug = `false`
	)

The constructor.

Parameters

cuda	Specifies the connection to CUDA.
log	Specifies the general output log.
nAxis	Specifies the axis under which to perform all of the actions in the graph.
bNeedsBackprop	Optionally, specifies whether or not to back propagate (default = true).
bClipGradients	Optionally, specifies whether or not to clip the gradients (default = false, Debug only).
bCheckNans	Optionally, specifies whether or not to check for nans (default = false, Debug only).
bAddDebug	Optionally, specifies to add debug information (default = false, Debug only).

Definition at line 45 of file ComputeGraph.cs.

Member Function Documentation

◆ add()

Blob< T > MyCaffe.common.ComputeGraph< T >.add	(	Blob< T >	btm1,
		Blob< T >	btm2,
		Blob< T >	top,
		bool	bAccumulateGrad = `true`
	)

'elthmul' operation adds each element of the 'btm1' with the 'btm2' and places the results in 'top'.

Parameters

btm1	Specifies the first input.
btm2	Specifies the second input.
top	Specifies the output.
bAccumulateGrad	Optionally, specifies whether or not to acumulate the gradients (default = true).

Returns: The 'top' Blob is returned.

Definition at line 538 of file ComputeGraph.cs.

◆ Backward()

void MyCaffe.common.ComputeGraph< T >.Backward ( bool bClear = false )

Runs a backward operation on all items starting from the last and running through the first.

Parameters

bClear Optionally, specifies to clear the list of operations upon completing (default = false).

Definition at line 684 of file ComputeGraph.cs.

◆ BackwardOne()

void MyCaffe.common.ComputeGraph< T >.BackwardOne ( int nIdx )

Runs a backward operation at a given index.

Parameters

nIdx	Specifies the index of the backward operation to run.

Definition at line 675 of file ComputeGraph.cs.

◆ Clear()

void MyCaffe.common.ComputeGraph< T >.Clear ( )

Clears all backward operations from the list.

Definition at line 698 of file ComputeGraph.cs.

◆ clear_grad() [1/2]

Blob< T > MyCaffe.common.ComputeGraph< T >.clear_grad ( Blob< T > b )

'clear_grad' operation only runs on the backward pass and zeros out the gradients on an input.

Parameters

b	Specifies the input.

Returns: The input Blob is returned.

Definition at line 576 of file ComputeGraph.cs.

◆ clear_grad() [2/2]

void MyCaffe.common.ComputeGraph< T >.clear_grad ( BlobCollection< T > rg )

'clear_grad' operation only runs on the backward pass and zeros out the gradients of the inputs.

Parameters

rg	Specifies an array of inputs.

Definition at line 594 of file ComputeGraph.cs.

◆ CopyToCache()

void MyCaffe.common.ComputeGraph< T >.CopyToCache	(	Blob< T >	btm,
		Cache< T >	cache
	)

CopyToCache operation copies the blob into the cache.

Parameters

btm	Specifies the input vector to copy.
cache	Specifies the cache.

Definition at line 322 of file ComputeGraph.cs.

◆ CopyToRow()

Blob< T > MyCaffe.common.ComputeGraph< T >.CopyToRow	(	Blob< T >	btm,
		Blob< T >	top,
		int	ix,
		bool	bCopyDiff = `false`
	)

CopyToRow operation copies the bottom vector into the top matrix.

Parameters

btm	Specifies the input vector.
top	Specifies the output matrix.
ix	Specifies the row into which the data is copied within the top.
bCopyDiff	Optionally, specifies to copy the diff and the data.

Returns: The top blob is returned.

Definition at line 290 of file ComputeGraph.cs.

◆ DebugOp()

void MyCaffe.common.ComputeGraph< T >.DebugOp ( params Blob< T >[] rgB )

DebugOp operation places a debug stub in the backpropagation chain for debugging only.

Definition at line 186 of file ComputeGraph.cs.

◆ Dispose()

void MyCaffe.common.ComputeGraph< T >.Dispose ( )

Release all resources used.

Definition at line 60 of file ComputeGraph.cs.

◆ eltmul()

Blob< T > MyCaffe.common.ComputeGraph< T >.eltmul	(	Blob< T >	btm1,
		Blob< T >	btm2,
		Blob< T >	top
	)

'elthmul' operation mutliplies each element of the 'btm1' with the 'btm2' and places the results in 'top'.

Parameters

btm1	Specifies the first input.
btm2	Specifies the second input.
top	Specifies the output.

Returns: The 'top' Blob is returned.

Definition at line 412 of file ComputeGraph.cs.

◆ mul()

Blob< T > MyCaffe.common.ComputeGraph< T >.mul	(	Blob< T >	btm1,
		Blob< T >	btm2,
		Blob< T >	top,
		bool	bAccumulateGrad = `true`
	)

'mul' operation performs a blas gemm operation on the 'btm1' matrix with the 'btm2' matrix and places the results in 'top'.

Parameters

btm1	Specifies the first input.
btm2	Specifies the second input.
top	Specifies the output.
bAccumulateGrad	Optionally, specifies to accumulate the gradient (default = true).

Returns: The 'top' Blob is returned.

Definition at line 494 of file ComputeGraph.cs.

◆ PeekItem()

void MyCaffe.common.ComputeGraph< T >.PeekItem	(	Blob< T >	btm,
		Blob< T >	top,
		int	ix
	)

PeekItem operation copies a single item from the bottom to the top.

Parameters

btm	Specifies the input vector.
top	Specifies the output matrix.
ix	Specifies the row into which the data is copied within the top.

Definition at line 264 of file ComputeGraph.cs.

◆ PeekRow()

Blob< T > MyCaffe.common.ComputeGraph< T >.PeekRow	(	Blob< T >	btm,
		Blob< T >	top,
		int	ix
	)

PeeKRow operation copies data and diffs from one row from within the bottom matrix and places it in the top vector.

Parameters

btm	Specifies the input matrix.
top	Specifies the output vector.
ix	Specifies the row within the input matrix to copy.

Returns: The top blob is returned.

Definition at line 232 of file ComputeGraph.cs.

◆ Round()

Blob< T > MyCaffe.common.ComputeGraph< T >.Round	(	Blob< T >	b,
		int	nDecimals = `6`
	)

Round operation, rounds the values to the nearest specified decimal.

Parameters

b	Specifies the blob to round.
nDecimals	Optionally, specifies the decimals (defautl = 6).

Returns: Returns the blob.

Definition at line 209 of file ComputeGraph.cs.

◆ scalemul()

Blob< T > MyCaffe.common.ComputeGraph< T >.scalemul	(	Blob< T >	btm1,
		Blob< T >	btm2,
		Blob< T >	top,
		int	nIdx = `0`
	)

'scalemul' operation mutliplies each element of the 'btm1' with the first item within 'btm2' and places the results in 'top'.

Parameters

btm1	Specifies the first input.
btm2	Specifies the second input.
top	Specifies the output.
nIdx	Specifies the index of btm2 to use for scaling.

Returns: The 'top' Blob is returned.

Definition at line 451 of file ComputeGraph.cs.

◆ sigmoid()

Blob< T > MyCaffe.common.ComputeGraph< T >.sigmoid	(	Blob< T >	btm,
		Blob< T >	top
	)

'sigmoid' operation runs the sigmoid on each item in the btm and places the results in the top.

Parameters

btm	Specifies the input data.
top	Specifies the output data.

Returns: The top blob is returned.

Definition at line 377 of file ComputeGraph.cs.

◆ softmax()

Blob< T > MyCaffe.common.ComputeGraph< T >.softmax	(	Blob< T >	btm,
		Blob< T >	top
	)

'softmax' operation runs the softmax on each item in the btm and places the results in the top.

Parameters

btm	Specifies the input data.
top	Specifies the output data.

Returns: The top blob is returned.

Definition at line 608 of file ComputeGraph.cs.

◆ tanh()

Blob< T > MyCaffe.common.ComputeGraph< T >.tanh	(	Blob< T >	btm,
		Blob< T >	top
	)

'tanh' operation runs the tanh on each item in the btm and places the results in the top.

Parameters

btm	Specifies the input data.
top	Specifies the output data.

Returns: The top blob is returned.

Definition at line 343 of file ComputeGraph.cs.

Property Documentation

◆ axis

int MyCaffe.common.ComputeGraph< T >.axis

get

Returns the axis on which all operations are performed.

Definition at line 104 of file ComputeGraph.cs.

◆ BackwardCount

int MyCaffe.common.ComputeGraph< T >.BackwardCount

get

Returns the backward operation count.

Definition at line 666 of file ComputeGraph.cs.

◆ Debug

Dictionary<string, Blob<T>[]> MyCaffe.common.ComputeGraph< T >.Debug

get

Returns a dictionary of Blobs used during each operation, only filled when 'bAddDebug' = true in the constructor.

Definition at line 78 of file ComputeGraph.cs.

◆ marker

string MyCaffe.common.ComputeGraph< T >.marker

getset

Get/set a string marker added to the debug information and used to indicate where in the code a given operation takes place.

Definition at line 86 of file ComputeGraph.cs.

◆ needs_backprop

bool MyCaffe.common.ComputeGraph< T >.needs_backprop

getset

Get/set whether or not to back propagate.

Definition at line 95 of file ComputeGraph.cs.

The documentation for this class was generated from the following file:

C:/Data/Data/SS_Projects/Intelligence/GitHub/MyCaffe/MyCaffe/common/ComputeGraph.cs

Public Member Functions

Properties

Detailed Description

Constructor & Destructor Documentation

◆ ComputeGraph()

Member Function Documentation

◆ add()

◆ Backward()

◆ BackwardOne()

◆ Clear()

◆ clear_grad() [1/2]

◆ clear_grad() [2/2]

◆ CopyToCache()

◆ CopyToRow()

◆ DebugOp()

◆ Dispose()

◆ eltmul()

◆ mul()

◆ PeekItem()

◆ PeekRow()

◆ Round()

◆ scalemul()

◆ sigmoid()

◆ softmax()

◆ tanh()

Property Documentation

◆ axis

◆ BackwardCount

◆ Debug

◆ marker

◆ needs_backprop