OpenMS  2.4.0
Public Types | Public Member Functions | Static Public Member Functions | Private Member Functions | Static Private Member Functions | Private Attributes | List of all members
SVMWrapper Class Reference

Serves as a wrapper for the libsvm. More...

#include <OpenMS/ANALYSIS/SVM/SVMWrapper.h>

Inheritance diagram for SVMWrapper:
ProgressLogger

Public Types

enum  SVM_parameter_type {
  SVM_TYPE, KERNEL_TYPE, DEGREE, C,
  NU, P, GAMMA, PROBABILITY,
  SIGMA, BORDER_LENGTH
}
 Parameters for the svm to be set from outside. More...
 
enum  SVM_kernel_type { OLIGO = 19, OLIGO_COMBINED }
 Kernel type. More...
 
- Public Types inherited from ProgressLogger
enum  LogType { CMD, GUI, NONE }
 Possible log types. More...
 

Public Member Functions

 SVMWrapper ()
 standard constructor More...
 
virtual ~SVMWrapper ()
 destructor More...
 
void setParameter (SVM_parameter_type type, Int value)
 You can set the parameters of the svm: More...
 
void setParameter (SVM_parameter_type type, double value)
 sets the double parameters of the svm More...
 
Int train (struct svm_problem *problem)
 trains the svm More...
 
Int train (SVMData &problem)
 trains the svm More...
 
void saveModel (std::string modelFilename) const
 saves the svm model More...
 
void loadModel (std::string modelFilename)
 loads the model More...
 
void predict (struct svm_problem *problem, std::vector< double > &predicted_labels)
 predicts the labels using the trained model More...
 
void predict (const SVMData &problem, std::vector< double > &results)
 predicts the labels using the trained model More...
 
Int getIntParameter (SVM_parameter_type type)
 You can get the actual int- parameters of the svm. More...
 
double getDoubleParameter (SVM_parameter_type type)
 You can get the actual double- parameters of the svm. More...
 
void predict (const std::vector< svm_node *> &vectors, std::vector< double > &predicted_rts)
 predicts the labels using the trained model More...
 
double performCrossValidation (svm_problem *problem_ul, const SVMData &problem_l, const bool is_labeled, const std::map< SVM_parameter_type, double > &start_values_map, const std::map< SVM_parameter_type, double > &step_sizes_map, const std::map< SVM_parameter_type, double > &end_values_map, Size number_of_partitions, Size number_of_runs, std::map< SVM_parameter_type, double > &best_parameters, bool additive_step_sizes=true, bool output=false, String performances_file_name="performances.txt", bool mcc_as_performance_measure=false)
 Performs a CV for the data given by 'problem'. More...
 
double getSVRProbability ()
 Returns the probability parameter sigma of the fitted Laplace model. More...
 
void getSignificanceBorders (svm_problem *data, std::pair< double, double > &borders, double confidence=0.95, Size number_of_runs=5, Size number_of_partitions=5, double step_size=0.01, Size max_iterations=1000000)
 calculates the significance borders of the error model and stores them in 'sigmas' More...
 
void getSignificanceBorders (const SVMData &data, std::pair< double, double > &sigmas, double confidence=0.95, Size number_of_runs=5, Size number_of_partitions=5, double step_size=0.01, Size max_iterations=1000000)
 calculates the significance borders of the error model and stores them in 'sigmas' More...
 
double getPValue (double sigma1, double sigma2, std::pair< double, double > point)
 calculates a p-value for a given data point using the model parameters More...
 
void getDecisionValues (svm_problem *data, std::vector< double > &decision_values)
 stores the prediction values for the encoded data in 'decision_values' More...
 
void scaleData (svm_problem *data, Int max_scale_value=-1)
 Scales the data such that every column is scaled to [-1, 1]. More...
 
svm_problem * computeKernelMatrix (svm_problem *problem1, svm_problem *problem2)
 computes the kernel matrix using the actual svm parameters and the given data More...
 
svm_problem * computeKernelMatrix (const SVMData &problem1, const SVMData &problem2)
 computes the kernel matrix using the actual svm parameters and the given data More...
 
void setTrainingSample (svm_problem *training_sample)
 This is used for being able to perform predictions with non libsvm standard kernels. More...
 
void setTrainingSample (SVMData &training_sample)
 This is used for being able to perform predictions with non libsvm standard kernels. More...
 
void getSVCProbabilities (struct svm_problem *problem, std::vector< double > &probabilities, std::vector< double > &prediction_labels)
 This function fills probabilities with the probability estimates for the first class. More...
 
void setWeights (const std::vector< Int > &weight_labels, const std::vector< double > &weights)
 Sets weights for the classes in C_SVC (see libsvm documentation for further details) More...
 
- Public Member Functions inherited from ProgressLogger
 ProgressLogger ()
 Constructor. More...
 
 ~ProgressLogger ()
 Destructor. More...
 
 ProgressLogger (const ProgressLogger &other)
 Copy constructor. More...
 
ProgressLoggeroperator= (const ProgressLogger &other)
 Assignment Operator. More...
 
void setLogType (LogType type) const
 Sets the progress log that should be used. The default type is NONE! More...
 
LogType getLogType () const
 Returns the type of progress log being used. More...
 
void startProgress (SignedSize begin, SignedSize end, const String &label) const
 Initializes the progress display. More...
 
void setProgress (SignedSize value) const
 Sets the current progress. More...
 
void endProgress () const
 Ends the progress display. More...
 
void nextProgress () const
 increment progress by 1 (according to range begin-end) More...
 

Static Public Member Functions

static void createRandomPartitions (svm_problem *problem, Size number, std::vector< svm_problem *> &partitions)
 You can create 'number' equally sized random partitions. More...
 
static void createRandomPartitions (const SVMData &problem, Size number, std::vector< SVMData > &problems)
 You can create 'number' equally sized random partitions. More...
 
static svm_problem * mergePartitions (const std::vector< svm_problem *> &problems, Size except)
 You can merge partitions excluding the partition with index 'except'. More...
 
static void mergePartitions (const std::vector< SVMData > &problems, Size except, SVMData &merged_problem)
 You can merge partitions excluding the partition with index 'except'. More...
 
static void getLabels (svm_problem *problem, std::vector< double > &labels)
 Stores the stored labels of the encoded SVM data at 'labels'. More...
 
static double kernelOligo (const std::vector< std::pair< int, double > > &x, const std::vector< std::pair< int, double > > &y, const std::vector< double > &gauss_table, int max_distance=-1)
 returns the value of the oligo kernel for sequences 'x' and 'y' More...
 
static double kernelOligo (const svm_node *x, const svm_node *y, const std::vector< double > &gauss_table, double sigma_square=0, Size max_distance=50)
 calculates the oligo kernel value for the encoded sequences 'x' and 'y' More...
 
static void calculateGaussTable (Size border_length, double sigma, std::vector< double > &gauss_table)
 

Private Member Functions

bool nextGrid_ (const std::vector< double > &start_values, const std::vector< double > &step_sizes, const std::vector< double > &end_values, const bool additive_step_sizes, std::vector< double > &actual_values)
 find next grid search parameter combination More...
 
Size getNumberOfEnclosedPoints_ (double m1, double m2, const std::vector< std::pair< double, double > > &points)
 
void initParameters_ ()
 Initializes the svm with standard parameters. More...
 

Static Private Member Functions

static void printToVoid_ (const char *)
 This function is passed to lib svm for output control. More...
 

Private Attributes

svm_parameter * param_
 
svm_model * model_
 
double sigma_
 
std::vector< doublesigmas_
 
std::vector< doublegauss_table_
 
std::vector< std::vector< double > > gauss_tables_
 
Size kernel_type_
 
Size border_length_
 
svm_problem * training_set_
 
svm_problem * training_problem_
 
SVMData training_data_
 

Additional Inherited Members

- Static Protected Member Functions inherited from ProgressLogger
static String logTypeToFactoryName_ (LogType type)
 Return the name of the factory product used for this log type. More...
 
- Protected Attributes inherited from ProgressLogger
LogType type_
 
time_t last_invoke_
 
ProgressLoggerImplcurrent_logger_
 
- Static Protected Attributes inherited from ProgressLogger
static int recursion_depth_
 

Detailed Description

Serves as a wrapper for the libsvm.

This class can be used for svm predictions. You can either perform classification or regression and choose certain kernel functions and additional parameters. Furthermore the models can be saved and loaded and we support also a new kernel function that was specially designed for learning with small sequences of different lengths.

Member Enumeration Documentation

◆ SVM_kernel_type

Kernel type.

Enumerator
OLIGO 
OLIGO_COMBINED 

◆ SVM_parameter_type

Parameters for the svm to be set from outside.

This type is used to specify the kind of parameter that is to be set or retrieved by the set/getParameter methods.

Enumerator
SVM_TYPE 

the svm type cab be NU_SVR or EPSILON_SVR

KERNEL_TYPE 

the kernel type

DEGREE 

the degree for the polynomial- kernel

the C parameter of the svm

NU 

the nu parameter for nu-SVR

the epsilon parameter for epsilon-SVR

GAMMA 

the gamma parameter of the POLY, RBF and SIGMOID kernel

PROBABILITY 
SIGMA 
BORDER_LENGTH 

Constructor & Destructor Documentation

◆ SVMWrapper()

standard constructor

◆ ~SVMWrapper()

virtual ~SVMWrapper ( )
virtual

destructor

Member Function Documentation

◆ calculateGaussTable()

static void calculateGaussTable ( Size  border_length,
double  sigma,
std::vector< double > &  gauss_table 
)
static

◆ computeKernelMatrix() [1/2]

svm_problem* computeKernelMatrix ( svm_problem *  problem1,
svm_problem *  problem2 
)

computes the kernel matrix using the actual svm parameters and the given data

This function can be used to compute a kernel matrix. 'problem1' and 'problem2' are used together wit the oligo kernel function (could be extended if you want to use your own kernel functions).

◆ computeKernelMatrix() [2/2]

svm_problem* computeKernelMatrix ( const SVMData problem1,
const SVMData problem2 
)

computes the kernel matrix using the actual svm parameters and the given data

This function can be used to compute a kernel matrix. 'problem1' and 'problem2' are used together wit the oligo kernel function (could be extended if you want to use your own kernel functions).

◆ createRandomPartitions() [1/2]

static void createRandomPartitions ( svm_problem *  problem,
Size  number,
std::vector< svm_problem *> &  partitions 
)
static

You can create 'number' equally sized random partitions.

This function creates 'number' equally sized random partitions and stores them in 'partitions'.

◆ createRandomPartitions() [2/2]

static void createRandomPartitions ( const SVMData problem,
Size  number,
std::vector< SVMData > &  problems 
)
static

You can create 'number' equally sized random partitions.

This function creates 'number' equally sized random partitions and stores them in 'partitions'.

◆ getDecisionValues()

void getDecisionValues ( svm_problem *  data,
std::vector< double > &  decision_values 
)

stores the prediction values for the encoded data in 'decision_values'

This function can be used to get the prediction values of the data if a model is already trained by the train() method. For regression the result is the same as for the method predict. For classification this function returns the distance from the separating hyperplane. For multiclass classification the decision_values vector will be empty.

◆ getDoubleParameter()

double getDoubleParameter ( SVM_parameter_type  type)

You can get the actual double- parameters of the svm.

Parameter types
C the C parameter of the svm
P the P parameter of the svm (sets the epsilon in epsilon-svr)
NU the nu parameter in nu-SVR
GAMMA for POLY, RBF and SIGMOID
Parameters
typeThe parameter that should be returned.

◆ getIntParameter()

Int getIntParameter ( SVM_parameter_type  type)

You can get the actual int- parameters of the svm.

Parameter types
KERNEL_TYPE LINEAR for the linear kernel
RBF for the rbf kernel
POLY for the polynomial kernel
SIGMOID for the sigmoid kernel
DEGREE the degree for the polynomial- kernel and the locality- improved kernel
SVM_TYPE he SVM type of the svm: can be NU_SVR or EPSILON_SVR
Parameters
typeThe parameter that should be returned.

◆ getLabels()

static void getLabels ( svm_problem *  problem,
std::vector< double > &  labels 
)
static

Stores the stored labels of the encoded SVM data at 'labels'.

◆ getNumberOfEnclosedPoints_()

Size getNumberOfEnclosedPoints_ ( double  m1,
double  m2,
const std::vector< std::pair< double, double > > &  points 
)
private

◆ getPValue()

double getPValue ( double  sigma1,
double  sigma2,
std::pair< double, double point 
)

calculates a p-value for a given data point using the model parameters

Uses the model parameters to calculate the p-value for 'point' which has the data entries: measured, predicted retention time.

◆ getSignificanceBorders() [1/2]

void getSignificanceBorders ( svm_problem *  data,
std::pair< double, double > &  borders,
double  confidence = 0.95,
Size  number_of_runs = 5,
Size  number_of_partitions = 5,
double  step_size = 0.01,
Size  max_iterations = 1000000 
)

calculates the significance borders of the error model and stores them in 'sigmas'

◆ getSignificanceBorders() [2/2]

void getSignificanceBorders ( const SVMData data,
std::pair< double, double > &  sigmas,
double  confidence = 0.95,
Size  number_of_runs = 5,
Size  number_of_partitions = 5,
double  step_size = 0.01,
Size  max_iterations = 1000000 
)

calculates the significance borders of the error model and stores them in 'sigmas'

◆ getSVCProbabilities()

void getSVCProbabilities ( struct svm_problem *  problem,
std::vector< double > &  probabilities,
std::vector< double > &  prediction_labels 
)

This function fills probabilities with the probability estimates for the first class.

The libSVM function svm_predict_probability is called to get probability estimates for the positive class. Since this is only used for binary classification it is sufficient for every test example to report the probability of the test example belonging to the positive class. Probability estimates have to be turned on during training (svm.setParameter(PROBABILITY, 1)), otherwise this method will fill the 'probabilities' vector with -1s.

◆ getSVRProbability()

double getSVRProbability ( )

Returns the probability parameter sigma of the fitted Laplace model.

The libsvm is used to fit a Laplace model to the prediction values by performing an internal cv using the training set if setParameter(PROBABILITY, 1) was invoked before using train. Look for your libsvm documentation for more details. The model parameter sigma is returned by this method. If no model was fitted during training zero is returned.

◆ initParameters_()

void initParameters_ ( )
private

Initializes the svm with standard parameters.

◆ kernelOligo() [1/2]

static double kernelOligo ( const std::vector< std::pair< int, double > > &  x,
const std::vector< std::pair< int, double > > &  y,
const std::vector< double > &  gauss_table,
int  max_distance = -1 
)
static

returns the value of the oligo kernel for sequences 'x' and 'y'

This function computes the kernel value of the oligo kernel, which was introduced by Meinicke et al. in 2004. 'x' and 'y' are encoded by encodeOligo and 'gauss_table' has to be constructed by calculateGaussTable.

'max_distance' can be used to speed up the computation even further by restricting the maximum distance between a k_mer at position i in sequence 'x' and a k_mer at position j in sequence 'y'. If i - j > 'max_distance' the value is not added to the kernel value. This approximation is switched off by default (max_distance < 0).

◆ kernelOligo() [2/2]

static double kernelOligo ( const svm_node *  x,
const svm_node *  y,
const std::vector< double > &  gauss_table,
double  sigma_square = 0,
Size  max_distance = 50 
)
static

calculates the oligo kernel value for the encoded sequences 'x' and 'y'

This kernel function calculates the oligo kernel value [Meinicke 04] for the sequences 'x' and 'y' that had been encoded by the encodeOligoBorder... function of the LibSVMEncoder class.

◆ loadModel()

void loadModel ( std::string  modelFilename)

loads the model

The svm- model is loaded. After this, the svm is ready for prediction.

Parameters
modelFilenameThe name of the model file that should be loaded.

◆ mergePartitions() [1/2]

static svm_problem* mergePartitions ( const std::vector< svm_problem *> &  problems,
Size  except 
)
static

You can merge partitions excluding the partition with index 'except'.

◆ mergePartitions() [2/2]

static void mergePartitions ( const std::vector< SVMData > &  problems,
Size  except,
SVMData merged_problem 
)
static

You can merge partitions excluding the partition with index 'except'.

◆ nextGrid_()

bool nextGrid_ ( const std::vector< double > &  start_values,
const std::vector< double > &  step_sizes,
const std::vector< double > &  end_values,
const bool  additive_step_sizes,
std::vector< double > &  actual_values 
)
private

find next grid search parameter combination

The current grid cell is given in actual_values. The result is returned in actual_values.

◆ performCrossValidation()

double performCrossValidation ( svm_problem *  problem_ul,
const SVMData problem_l,
const bool  is_labeled,
const std::map< SVM_parameter_type, double > &  start_values_map,
const std::map< SVM_parameter_type, double > &  step_sizes_map,
const std::map< SVM_parameter_type, double > &  end_values_map,
Size  number_of_partitions,
Size  number_of_runs,
std::map< SVM_parameter_type, double > &  best_parameters,
bool  additive_step_sizes = true,
bool  output = false,
String  performances_file_name = "performances.txt",
bool  mcc_as_performance_measure = false 
)

Performs a CV for the data given by 'problem'.

◆ predict() [1/3]

void predict ( struct svm_problem *  problem,
std::vector< double > &  predicted_labels 
)

predicts the labels using the trained model

The prediction process is started and the results are stored in 'predicted_labels'.

◆ predict() [2/3]

void predict ( const SVMData problem,
std::vector< double > &  results 
)

predicts the labels using the trained model

The prediction process is started and the results are stored in 'predicted_labels'.

◆ predict() [3/3]

void predict ( const std::vector< svm_node *> &  vectors,
std::vector< double > &  predicted_rts 
)

predicts the labels using the trained model

The prediction process is started and the results are stored in 'predicted_rts'.

◆ printToVoid_()

static void printToVoid_ ( const char *  )
staticprivate

This function is passed to lib svm for output control.

The intention is to discard the output, as we don't need it.

◆ saveModel()

void saveModel ( std::string  modelFilename) const

saves the svm model

The model of the trained svm is saved into 'modelFilename'. Throws an exception if the model cannot be saved.

Exceptions
Exception::UnableToCreateFile
Parameters
modelFilenameThe file name where the model will be saved.

◆ scaleData()

void scaleData ( svm_problem *  data,
Int  max_scale_value = -1 
)

Scales the data such that every column is scaled to [-1, 1].

Scales the x[][].value values of the svm_problem* structure. If the second parameter is omitted, the data is scaled to [-1, 1]. Otherwise the data is scaled to [0, max_scale_value]

◆ setParameter() [1/2]

void setParameter ( SVM_parameter_type  type,
Int  value 
)

You can set the parameters of the svm:

Parameter types
KERNEL_TYPE LINEAR for the linear kernel
RBF for the rbf kernel
POLY for the polynomial kernel
SIGMOID for the sigmoid kernel
DEGREE the degree for the polynomial- kernel and the locality- improved kernel
C the C parameter of the svm
Parameters
typeThe type of parameter to set.
valueThe new value for parameter type.

◆ setParameter() [2/2]

void setParameter ( SVM_parameter_type  type,
double  value 
)

sets the double parameters of the svm

Parameters
typeThe type of parameter to set.
valueThe new value for parameter type.

◆ setTrainingSample() [1/2]

void setTrainingSample ( svm_problem *  training_sample)

This is used for being able to perform predictions with non libsvm standard kernels.

◆ setTrainingSample() [2/2]

void setTrainingSample ( SVMData training_sample)

This is used for being able to perform predictions with non libsvm standard kernels.

◆ setWeights()

void setWeights ( const std::vector< Int > &  weight_labels,
const std::vector< double > &  weights 
)

Sets weights for the classes in C_SVC (see libsvm documentation for further details)

◆ train() [1/2]

Int train ( struct svm_problem *  problem)

trains the svm

The svm is trained with the data stored in the 'svm_problem' structure.

◆ train() [2/2]

Int train ( SVMData problem)

trains the svm

The svm is trained with the data stored in the 'SVMData' structure.

Member Data Documentation

◆ border_length_

Size border_length_
private

◆ gauss_table_

std::vector<double> gauss_table_
private

◆ gauss_tables_

std::vector<std::vector<double> > gauss_tables_
private

◆ kernel_type_

Size kernel_type_
private

◆ model_

svm_model* model_
private

◆ param_

svm_parameter* param_
private

◆ sigma_

double sigma_
private

◆ sigmas_

std::vector<double> sigmas_
private

◆ training_data_

SVMData training_data_
private

◆ training_problem_

svm_problem* training_problem_
private

◆ training_set_

svm_problem* training_set_
private