singa Namespace Reference

Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. More...

Classes
struct	_Context
class	Accuracy
	Compute the accuray of the prediction, which is matched against the ground truth labels. More...
class	AdaGrad
class	Block
	Block represent a chunk of memory (on device or host). More...
class	Channel
	Channel for appending metrics or other information into files or screen. More...
class	ChannelManager
class	Constraint
	Apply constraints for parameters (gradient). More...
class	CppCPU
	Represent a CPU device which may have multiple threads/executors. More...
class	CSVDecoder
	Decode the string of csv formated data into data tensor (dtype is kFloat32) and optionally a label tensor (dtype is kInt). More...
class	CSVEncoder
	Convert values from tensors into a csv formated string. More...
class	Decoder
	The base decoder that converts a string into a set of tensors. More...
class	Device
	Allocate memory and execute Tensor operations. More...
class	DeviceMemPool
class	Encoder
	Base encoder class that convert a set of tensors into string for storage. More...
class	FeedForwardNet
	The feed-forward neural net. More...
class	ImageTransformer
class	Initializer
class	Layer
	The base layer class. More...
class	LocalUpdater
	LocalUpdater do gradient aggregation and update gradient calling the wrapped Optimizer on a specific device (i.e., CPU or GPU). More...
class	Loss
	The base loss class, which declares the APIs for computing the objective score (loss) for a pair of prediction (from the model) and the target (i.e. More...
class	Metric
	The base metric class, which declares the APIs for computing the performance evaluation metrics given the prediction of the model and the ground truth, i.e., the target. More...
class	MSE
	MSE is for mean squared error or squared euclidean distance. More...
class	Nesterov
class	Optimizer
	The base class for gradient descent algorithms used to update the model parameters in order to optimize the objective (loss) function. More...
class	Platform
	This class queries all available calculating devices on a given machine grouped according to manufacturer or device drivers. More...
class	Regularizer
	Apply regularization for parameters (gradient), e.g., L1 norm and L2 norm. More...
class	RMSProp
class	Scheduler
	Scheduling Tensor operations with dependency detection. More...
class	SGD
class	Snapshot
	The snapshot management. More...
class	SoftmaxCrossEntropy
	Softmax + cross entropy for multi-category classification. More...
class	Tensor
	A Tensor instance is a multi-dimensional array resident on a Device (default device is the host CPU). More...
class	Timer
	For benchmarking the time cost of operations. More...
class	Tokenizer
	Tokenize a string. More...
class	Transformer
	Base apply class that does data transformations in pre-processing stage. More...
class	Updater
	Basic Updater class just forward all the method function call to the wrapped Optimizer. More...
class	VirtualMemory
	Manage device memory pool including garbage collection, memory opt. More...

Typedefs
typedef struct singa::_Context	Context
typedef vector< size_t >	Shape
using	InitializerConf = FillerConf
	Base class for initializing parameter values.

Functions
size_t	SizeOf (DataType t)
size_t	Product (const Shape &shape, int start=0, size_t len=0)
void	CheckDataTypeAndLang (const Tensor &in1, const Tensor &in2)
template<typename FromType , typename ToType >
ToType	TypeCast (const FromType &x)
Tensor	Boradcast (const Shape &shape)
Tensor	Reshape (const Tensor &in, const Shape &s)
	Reshape the given tensor and generate a new tensor; the total vol should match which shares the memory with in if possible.
Tensor	Resize (const Tensor &in, const Shape &s)
Tensor	Transpose (const Tensor &in)
	Reverse the shape vector.
Tensor	Broadcast (const Tensor &in, const Shape &shape)
	Return a view of the input tensor whose shape is broadcasted to be compitable with the given shape.
Tensor	Transpose (const Tensor &in, const vector< size_t > &axes)
	Change the axes.
void	CopyDataToFrom (Tensor *dst, const Tensor &src, const size_t num, const size_t dst_offset=0, const size_t src_offset=0)
	Copy 'num' elements of src to dst. More...
void	RepeatDataToFrom (bool broadcast_flag, const vector< size_t > &repeats, int axis, Tensor *dst, const Tensor &in, const size_t num)
Tensor	Abs (const Tensor &in)
Tensor	Exp (const Tensor &in)
Tensor	Log (const Tensor &in)
Tensor	ReLU (const Tensor &in)
Tensor	Sigmoid (const Tensor &in)
Tensor	Sign (const Tensor &in)
Tensor	Sqrt (const Tensor &in)
Tensor	Square (const Tensor &in)
Tensor	Tanh (const Tensor &in)
Tensor	Transform (const Tensor &in)
void	Abs (const Tensor &in, Tensor *out)
void	Exp (const Tensor &in, Tensor *out)
void	Log (const Tensor &in, Tensor *out)
void	ReLU (const Tensor &in, Tensor *out)
void	Sigmoid (const Tensor &in, Tensor *out)
void	Sign (const Tensor &in, Tensor *out)
void	Sqrt (const Tensor &in, Tensor *out)
void	Square (const Tensor &in, Tensor *out)
void	Tanh (const Tensor &in, Tensor *out)
void	Transform (const Tensor &in, Tensor *out)
template<typename SType >
Tensor	Pow (const Tensor &in, const SType x)
	Element-wise opeartion, out[i]=in[i]^x.
template<typename SType >
void	Pow (const Tensor &in, const SType x, Tensor *out)
	Element-wise opeartion, out[i]=in[i]^x.
Tensor	Pow (const Tensor &base, const Tensor &exp)
	Element-wise opeartion, out[i]=baes[i]^exp[i].
void	Pow (const Tensor &base, const Tensor &exp, Tensor *out)
	Element-wise opeartion, out[i]=baes[i]^exp[i].
template<typename SType >
Tensor	operator< (const Tensor &in, const SType x)
	Element-wise operation, out[i]= (in[i] < x) ? 1.f : 0.f.
template<typename SType >
void	LT (const Tensor &in, const SType x, Tensor *out)
Tensor	operator< (const Tensor &in1, const Tensor &in2)
	Element-wise operation, out[i]= (in1[i] < in2[i]) ? 1.f : 0.f.
void	LT (const Tensor &in1, const Tensor &in2, Tensor *out)
template<typename SType >
Tensor	operator<= (const Tensor &in, const SType x)
	Element-wise operation, out[i]= (in[i] <= x) ? 1.f : 0.f.
template<typename SType >
void	LE (const Tensor &in, const SType x, Tensor *out)
Tensor	operator<= (const Tensor &in1, const Tensor &in2)
	Element-wise operation, out[i]= (in1[i] <= in2[i]) ? 1.f : 0.f.
void	LE (const Tensor &in1, const Tensor &in2, Tensor *out)
template<typename SType >
Tensor	operator> (const Tensor &in, const SType x)
	Element-wise operation, out[i]= (in[i] > x) ? 1.f : 0.f.
template<typename SType >
void	GT (const Tensor &in, const SType x, Tensor *out)
Tensor	operator> (const Tensor &in1, const Tensor &in2)
	Element-wise operation, out[i]= (in1[i] > in2[i]) ? 1.f : 0.f.
void	GT (const Tensor &in1, const Tensor &in2, Tensor *out)
template<typename SType >
Tensor	operator>= (const Tensor &in, const SType x)
	Element-wise operation, out[i]= (in[i] >= x) ? 1.f : 0.f.
template<typename SType >
void	GE (const Tensor &in, const SType x, Tensor *out)
Tensor	operator>= (const Tensor &in1, const Tensor &in2)
	Element-wise operation, out[i]= (in1[i] >= in2[i]) ? 1.f : 0.f.
void	GE (const Tensor &in1, const Tensor &in2, Tensor *out)
Tensor	operator+ (const Tensor &lhs, const Tensor &rhs)
void	Add (const Tensor &lhs, const Tensor &rhs, Tensor *out)
Tensor	operator- (const Tensor &lhs, const Tensor &rhs)
void	Sub (const Tensor &lhs, const Tensor &rhs, Tensor *out)
Tensor	operator* (const Tensor &lhs, const Tensor &rhs)
void	EltwiseMult (const Tensor &lhs, const Tensor &rhs, Tensor *out)
Tensor	operator/ (const Tensor &lhs, const Tensor &rhs)
void	Div (const Tensor &lhs, const Tensor &rhs, Tensor *out)
template<typename SType >
Tensor	operator+ (const Tensor &in, const SType x)
template<typename SType >
void	Add (const Tensor &in, const SType x, Tensor *out)
template<typename SType >
Tensor	operator- (const Tensor &in, const SType x)
template<typename SType >
void	Sub (const Tensor &in, const SType x, Tensor *out)
template<typename SType >
Tensor	operator* (const Tensor &in, const SType x)
template<typename SType >
void	EltwiseMult (const Tensor &in, const SType x, Tensor *out)
template<typename SType >
Tensor	operator/ (const Tensor &in, const SType x)
	For each element e of Tensor 'in', compute e / x.
template<typename SType >
void	Div (const Tensor &in, const SType x, Tensor *out)
	For each element e of Tensor 'in', compute e / x into out.
template<typename SType >
Tensor	Div (const SType x, const Tensor &in)
	For each element e of Tensor 'in', compute x/e.
template<typename SType >
void	Div (const SType x, const Tensor &in, Tensor *out)
	For each element e of Tensor 'in', compute x/e into 'out'.
template<typename SType = float>
SType	Sum (const Tensor &in)
Tensor	Average (const Tensor &in, const int axis)
	Average elements in the Tensor, currently only support vector and matrix. More...
void	AddColumn (const Tensor &v, Tensor *M)
	Add column 'v' with each column of matrix M.
template<typename SType >
void	AddColumn (const SType alpha, const SType beta, const Tensor &v, Tensor *out)
	For each column 'c' of matrix out, do c=alpha*v + beta*c.
void	AddRow (const Tensor &v, Tensor *out)
	Add row 'v' with each row of matrix M; write results into 'out'.
template<typename SType >
void	AddRow (const SType alpha, const SType beta, const Tensor &v, Tensor *M)
	For each row 'r' of matrix out, do r=alpha*v + beta*r.
void	DivColumn (const Tensor &v, Tensor *M)
	Divide column 'v' by each column of matrix M; write results into 'out'.
void	DivRow (const Tensor &v, Tensor *M)
	Divide row 'v' by each row of matrix M; write results into 'out'.
void	MultColumn (const Tensor &v, Tensor *M)
	Multiply column 'v' and each column of matrix M; write results into 'out'.
void	MultRow (const Tensor &v, Tensor *M)
	Multiply row 'v' with each row of matrix M; write results into 'out'.
Tensor	SoftMax (const Tensor &in)
	Do softmax for each row. 'in' could be a 1-d or 2-d Tensor.
Tensor	RowMax (const Tensor &in)
void	SoftMax (const Tensor &in, Tensor *out)
	Do softmax for each row. 'in' could be a 1-d or 2-d Tensor.
void	SubColumn (const Tensor &v, Tensor *M)
	Sub column 'v' by each column of matrix M.
void	SubRow (const Tensor &v, Tensor *M)
	Sub row 'v' by each row of matrix M; write results into 'out'.
void	SumColumns (const Tensor &M, Tensor *out)
	Sum all columns of matrix M into a single column as 'out'.
void	SumRows (const Tensor &M, Tensor *out)
	Sum all rows of matrix M into a single row as 'out'.
Tensor	Sum (const Tensor &in, const int axis)
	Sum elements in the Tensor, currently only support vector and matrix. More...
template<typename SType >
void	Bernoulli (const SType p, Tensor *out)
	For each element x set x = 1 if random() < p; otherwise x = 1.
template<typename SType >
void	Gaussian (const SType mean, const SType std, Tensor *out)
	Fill in Tensor 't' following Gaussian distribution.
template<typename SType >
void	Uniform (const SType low, const SType high, Tensor *out)
	Fill in Tensor 't' following uniform distribution.
template<typename SType >
void	Axpy (SType alpha, const Tensor &in, Tensor *out)
	out = alpha*in + out
Tensor	Mult (const Tensor &A, const Tensor &B)
	Do matrix vector multipication or matrix matrix multiplication depdending on the Tensor shape. More...
void	Mult (const Tensor &A, const Tensor &B, Tensor *C)
	Do matrix vector multipication or matrix matrix multiplication depdending on the Tensor shape. More...
template<typename SType >
void	Mult (const SType alpha, const Tensor &A, const Tensor &B, const SType beta, Tensor *C)
	Do matrix vector multipication or matrix matrix multiplication depdending on the Tensor shape. More...
void	ComputeCrossEntropy (const Tensor &p, const Tensor &t, Tensor *loss)
	Compute the cross entropy loss given the prediction probability 'p' and the target (ground truth) labels 't'. More...
void	SoftmaxCrossEntropyBwd (const Tensor &t, Tensor *p)
	Compute the dx, given prediction probability 'p' (p=softmax(x)) and the target (ground truth) labels 't'. More...
Tensor	CrossEntropyFwd (const Tensor &p, const Tensor &t)
	To be called by pysinga autograd operations; swig ignores the const qualifier http://www.swig.org/Doc3.0/SWIGPlus.html#SWIGPlus_const.
Tensor	SoftmaxCrossEntropyBwd (const Tensor &p, const Tensor &t)
Tensor	CopyRows (const Tensor &in, const size_t start, const size_t end)
	Return a tensor consisting of rows ([start, end)) from 'in'. More...
Tensor	SliceRows (const Tensor &in, const size_t start, const size_t end)
	Alias of CopyRows.
Tensor	SliceOn (const Tensor &in, const size_t start, const size_t end, int axis)
	Slice the input tensor along the give axis to generate a new tensor.
Tensor	CopyColumns (const Tensor &in, const size_t start, const size_t end)
	Return a tensor consisting of columns ([start, end)) from 'in'. More...
Tensor	SliceColumns (const Tensor &in, const size_t start, const size_t end)
	Alias of CopyColumns.
Tensor	ConcatenateRows (const vector< Tensor > &in)
	Return a tensor which is vertically stacked from tensors in 'in'. More...
Tensor	ConcatOn (const std::vector< Tensor > &in, int axis)
	Return a tensor concatenated of the input tensors along the give axis.
Tensor	ConcatRows (const vector< Tensor > &in)
	Alias name for function ConcatenateRows.
Tensor	ConcatenateColumns (const vector< Tensor > &in)
	Return a tensor which is horizontally stacked from tensors in 'in'. More...
Tensor	ConcatColumns (const vector< Tensor > &in)
	Alias name for function ConcatenateColumns.
Tensor	crop (Tensor &input, const size_t crop_height, const size_t crop_width, const size_t crop_h_offset, const size_t crop_w_offset, const string &image_dim_order)
Tensor	mirror (Tensor &input, const bool horizontal_mirror, const bool vertical_mirror, const string &image_dim_order)
std::shared_ptr< Initializer >	CreateInitializer (const InitializerConf &conf)
	TODO(wangwei) create the initializers from factory like that for Layer.
std::shared_ptr< Layer >	CreateLayer (const std::string type)
const std::vector< std::string >	GetRegisteredLayers ()
std::shared_ptr< Constraint >	CreateConstraint (std::string type)
std::shared_ptr< Regularizer >	CreateRegularizer (std::string type)
std::shared_ptr< Optimizer >	CreateOptimizer (const string &type)
void	InitChannel (const char *argv)
	Initial function for global usage of channel. More...
void	SetChannelDirectory (const char *path)
	Set the directory name for persisting channel content.
Channel *	GetChannel (const std::string &channel_name)
	Get the channel instance.
void	InitLogging (const char *argv)
	Global functions for both glog and built-in log.
void	LogToStderr ()
	Make it so that all log messages go only to stderr.
void	SetStderrLogging (int severity)
	Make it so that all log messages of at least a particular severity are logged to stderr (in addtion to logging to the usual log files)
void	SetLogDestination (int severity, const char *path)
	Set the file name for logging (and disable logging to stderr)
mkldnn::memory::data_type	GetMKLDNNDataType (DataType dtype)
bool	icasecmp (const string &l, const string &r)
string	ToLowerCase (const string &input)
int	ArgPos (int argc, char **arglist, const char *arg)
template<typename T >
std::string	VecToStr (const std::vector< T > &in)

Variables
std::shared_ptr< Device >	defaultDevice
	a singleton CppDevice as the host for all devices.
const size_t	kDataWidth []
	hardcode the width of types defined in DataType More...
const int	INFO = 0
const int	WARNING = 1
const int	ERROR = 2
const int	FATAL = 3
const int	NUM_SEVERITIES = 4

Detailed Description

Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements.

See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at

http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.

Function Documentation

Average()

Tensor singa::Average	(	const Tensor &	in,
		const int	axis
	)

Average elements in the Tensor, currently only support vector and matrix.

if 'axis' is 0, average all rows into a single row if 'axis' is 1, average all columns into a single column TODO(wangwei) support arbitrary Tensor like numpy.average

ComputeCrossEntropy()

void singa::ComputeCrossEntropy	(	const Tensor &	p,
		const Tensor &	t,
		Tensor *	loss
	)

Compute the cross entropy loss given the prediction probability 'p' and the target (ground truth) labels 't'.

'p' could be either a 1-d vector for a single instance or a 2-d matrix for a batch of instances. t[i] could be the ground truth label index or a label weighted array of the i-th instance. For example, if there are 3 candidate labels for each instance, t[i] could be 2 or [0, 0, 1]. If one instance could have multiple labels, then t[i] could be [1, 0, 1]. The loss is computed into p.

ConcatenateColumns()

Tensor singa::ConcatenateColumns

(

const vector< Tensor > &

in

)

Return a tensor which is horizontally stacked from tensors in 'in'.

Each tensor in 'in' is a 2D tensor. Values are copied, no memory sharing.

ConcatenateRows()

Tensor singa::ConcatenateRows

(

const vector< Tensor > &

in

)

Return a tensor which is vertically stacked from tensors in 'in'.

Each tensor in 'in' is a 2D tensor. Values are copied, no memory sharing.

CopyColumns()

Tensor singa::CopyColumns	(	const Tensor &	in,
		const size_t	start,
		const size_t	end
	)

Return a tensor consisting of columns ([start, end)) from 'in'.

It copies the values from 'in'. 'in' is a 2D Tensor.

CopyDataToFrom()

void singa::CopyDataToFrom	(	Tensor *	dst,
		const Tensor &	src,
		const size_t	num,
		const size_t	dst_offset = 0,
		const size_t	src_offset = 0
	)

Copy 'num' elements of src to dst.

The first 'src_offset' ('dst_offset') elements will be skipped.

CopyRows()

Tensor singa::CopyRows	(	const Tensor &	in,
		const size_t	start,
		const size_t	end
	)

Return a tensor consisting of rows ([start, end)) from 'in'.

It copies the values from 'in'. 'in' ia a 2D Tensor.

InitChannel()

void singa::InitChannel

(

const char *

argv

)

Initial function for global usage of channel.

'argv' is for future use.

Mult() [1/3]

Tensor singa::Mult	(	const Tensor &	A,
		const Tensor &	B
	)

Do matrix vector multipication or matrix matrix multiplication depdending on the Tensor shape.

result = A * B

Mult() [2/3]

void singa::Mult	(	const Tensor &	A,
		const Tensor &	B,
		Tensor *	C
	)

Do matrix vector multipication or matrix matrix multiplication depdending on the Tensor shape.

C = A * B

Mult() [3/3]

template<typename SType >

void singa::Mult	(	const SType	alpha,
		const Tensor &	A,
		const Tensor &	B,
		const SType	beta,
		Tensor *	C
	)

Do matrix vector multipication or matrix matrix multiplication depdending on the Tensor shape.

out = alpha lhs * rhs + beta * out

SoftmaxCrossEntropyBwd()

void singa::SoftmaxCrossEntropyBwd	(	const Tensor &	t,
		Tensor *	p
	)

Compute the dx, given prediction probability 'p' (p=softmax(x)) and the target (ground truth) labels 't'.

'p' and 't' are either 1-d vector or 2-d matrix. 'grad' has the same shape as 'p'. dx is computed into p.

Sum()

Tensor singa::Sum	(	const Tensor &	in,
		const int	axis
	)

Sum elements in the Tensor, currently only support vector and matrix.

if 'axis' is 0, sum all rows into a single row if 'axis' is 1, sum all columns into a single column TODO(wangwei) support arbitrary Tensor like numpy.sum

Variable Documentation

kDataWidth

const size_t singa::kDataWidth[]

Initial value:

= {sizeof(float),  sizeof(float) / 2,
                             sizeof(int),    sizeof(char),
                             sizeof(double), sizeof(unsigned char)
                            }

hardcode the width of types defined in DataType