PLearn 0.1
Public Member Functions | Static Public Member Functions | Public Attributes | Static Public Attributes | Protected Member Functions | Static Protected Member Functions | Protected Attributes | Private Types | Private Member Functions
PLearn::PruningLinearRegressor Class Reference

The first sentence should be a BRIEF DESCRIPTION of what the class does. More...

#include <PruningLinearRegressor.h>

Inheritance diagram for PLearn::PruningLinearRegressor:
Inheritance graph
[legend]
Collaboration diagram for PLearn::PruningLinearRegressor:
Collaboration graph
[legend]

List of all members.

Public Member Functions

 PruningLinearRegressor ()
 Default constructor.
virtual void train ()
 The role of the train method is to bring the learner up to stage==nstages, updating the train_stats collector with training costs measured on-line in the process.
virtual void computeOutput (const Vec &input, Vec &output) const
 Computes the output from the input.
virtual string classname () const
virtual OptionListgetOptionList () const
virtual OptionMapgetOptionMap () const
virtual RemoteMethodMapgetRemoteMethodMap () const
virtual PruningLinearRegressordeepCopy (CopiesMap &copies) const
virtual void build ()
 simply calls inherited::build() then build_()
virtual void makeDeepCopyFromShallowCopy (CopiesMap &copies)
 Transforms a shallow copy into a deep copy.
virtual void setTrainingSet (VMat training_set, bool call_forget=true)
 Declares the training set.

Static Public Member Functions

static string _classname_ ()
static OptionList_getOptionList_ ()
static RemoteMethodMap_getRemoteMethodMap_ ()
static Object_new_instance_for_typemap_ ()
static bool _isa_ (const Object *o)
static void _static_initialize_ ()
static const PPathdeclaringFile ()

Public Attributes

string pruning_method
 ### declare public option fields (such as build options) here
real min_t_ratio
int max_number
real max_fraction

Static Public Attributes

static StaticInitializer _static_initializer_

Protected Member Functions

void computeTRatio ()
 Utility function to compute the t-ratio for the estimator b (regression coefficients)
void newDatasetIndices ()
 Find the dataset indices corresponding to coefficients not pruned.

Static Protected Member Functions

static void declareOptions (OptionList &ol)
 Declares the class options.

Protected Attributes

Vec t_ratio
 t-ratio statistics for the estimator b (regression coefficients) Saved as a learned option to allow computing statistical significance of the weights when the model is reloaded and used in test mode.
TVec< intinput_indices
 Indices of inputs kept for regression.

Private Types

typedef LinearRegressor inherited

Private Member Functions

void build_ ()
 This does the actual building.

Detailed Description

The first sentence should be a BRIEF DESCRIPTION of what the class does.

Place the rest of the class programmer documentation here. Doxygen supports Javadoc-style comments. See http://www.doxygen.org/manual.html

Todo:
Write class to-do's here if there are any.
Deprecated:
Write deprecated stuff here if there is any. Indicate what else should be used instead.

Definition at line 57 of file PruningLinearRegressor.h.


Member Typedef Documentation

Reimplemented from PLearn::LinearRegressor.

Definition at line 59 of file PruningLinearRegressor.h.


Constructor & Destructor Documentation

PLearn::PruningLinearRegressor::PruningLinearRegressor ( )

Default constructor.

Definition at line 55 of file PruningLinearRegressor.cc.

    : pruning_method("max_number"),
      min_t_ratio(0.05),
      max_number(50),
      max_fraction(0.5)
{ }

Member Function Documentation

string PLearn::PruningLinearRegressor::_classname_ ( ) [static]

Reimplemented from PLearn::LinearRegressor.

Definition at line 53 of file PruningLinearRegressor.cc.

OptionList & PLearn::PruningLinearRegressor::_getOptionList_ ( ) [static]

Reimplemented from PLearn::LinearRegressor.

Definition at line 53 of file PruningLinearRegressor.cc.

RemoteMethodMap & PLearn::PruningLinearRegressor::_getRemoteMethodMap_ ( ) [static]

Reimplemented from PLearn::LinearRegressor.

Definition at line 53 of file PruningLinearRegressor.cc.

bool PLearn::PruningLinearRegressor::_isa_ ( const Object o) [static]

Reimplemented from PLearn::LinearRegressor.

Definition at line 53 of file PruningLinearRegressor.cc.

Object * PLearn::PruningLinearRegressor::_new_instance_for_typemap_ ( ) [static]

Reimplemented from PLearn::LinearRegressor.

Definition at line 53 of file PruningLinearRegressor.cc.

StaticInitializer PruningLinearRegressor::_static_initializer_ & PLearn::PruningLinearRegressor::_static_initialize_ ( ) [static]

Reimplemented from PLearn::LinearRegressor.

Definition at line 53 of file PruningLinearRegressor.cc.

void PLearn::PruningLinearRegressor::build ( ) [virtual]

simply calls inherited::build() then build_()

Reimplemented from PLearn::LinearRegressor.

Definition at line 121 of file PruningLinearRegressor.cc.

References PLearn::LinearRegressor::build(), and build_().

Here is the call graph for this function:

void PLearn::PruningLinearRegressor::build_ ( ) [private]

This does the actual building.

Reimplemented from PLearn::LinearRegressor.

Definition at line 100 of file PruningLinearRegressor.cc.

References max_number, min_t_ratio, PLERROR, and pruning_method.

Referenced by build().

{
    if (pruning_method == "max_number")
    {
        if (max_number < 1)
            PLERROR("\"max_number\" should be strictly positive");
    }
    else if (pruning_method == "max_fraction")
    {
        if (max_fraction <= 0.0  ||  max_fraction >= 1.0)
            PLERROR("\"max_fraction\" should be in range ]0,1[");
    }
    else if (pruning_method == "min_t_ratio")
    {
        if (min_t_ratio <= 0.0)
            PLERROR("\"min_t_ratio\" should be strictly positive");
    }
    else
        PLERROR("Pruning method \"%s\" not supported", pruning_method.c_str());
}

Here is the caller graph for this function:

string PLearn::PruningLinearRegressor::classname ( ) const [virtual]

Reimplemented from PLearn::LinearRegressor.

Definition at line 53 of file PruningLinearRegressor.cc.

void PLearn::PruningLinearRegressor::computeOutput ( const Vec input,
Vec output 
) const [virtual]

Computes the output from the input.

Reimplemented from PLearn::LinearRegressor.

Definition at line 165 of file PruningLinearRegressor.cc.

References PLearn::LinearRegressor::computeOutput(), input_indices, PLearn::TVec< T >::length(), and PLearn::selectElements().

{
    Vec actual_input(input_indices.length());
    selectElements(input, input_indices, actual_input);
    inherited::computeOutput(actual_input, output);
}

Here is the call graph for this function:

void PLearn::PruningLinearRegressor::computeTRatio ( ) [protected]

Utility function to compute the t-ratio for the estimator b (regression coefficients)

Definition at line 172 of file PruningLinearRegressor.cc.

References PLearn::abs(), PLearn::TMat< T >::copy(), i, PLearn::TMat< T >::length(), PLearn::TVec< T >::length(), PLearn::matInvert(), PLASSERT, PLearn::LinearRegressor::resid_variance, PLearn::TVec< T >::resize(), PLearn::sqrt(), t_ratio, PLearn::LinearRegressor::weights, PLearn::TMat< T >::width(), and PLearn::LinearRegressor::XtX.

Referenced by newDatasetIndices().

{
    // We wish to compute the t-ratio of the estimator coefficients.
    // For that purpose, we use the following formula:
    //
    //    t = |b| / sigma_b
    //
    // where b is the coefficients vector and sigma_b is the stderr matrix
    // of the estimator of b.  The latter is computed as:
    //
    //     sigma_b = s^2 * inverse(X'X)
    //
    // where s^2 is the residual variance (estimated using the
    // LinearRegressor::computeResidualsVariance method)
    // and X is the matrix of regressors..

    const int ninputs = weights.length();
    Mat sigma_b(ninputs, ninputs);
    t_ratio.resize(ninputs);

    // We compute the estimator
    PLASSERT(resid_variance.length() == 1);
    PLASSERT(weights.width() == 1);
    real residual_variance = resid_variance[0];
 
    Mat XtX_copy = XtX.copy();  // matInvert overwrite the input matrix
    Mat XtX_inverse(XtX.length(), XtX.width());
    matInvert(XtX_copy, XtX_inverse);
    for (int i=0; i<ninputs; i++)
    {
        real sigma_b = sqrt(residual_variance*XtX_inverse(i,i));
        t_ratio[i] = abs(weights(i,0)) / sigma_b;
    }
}

Here is the call graph for this function:

Here is the caller graph for this function:

void PLearn::PruningLinearRegressor::declareOptions ( OptionList ol) [static, protected]

Declares the class options.

Reimplemented from PLearn::LinearRegressor.

Definition at line 62 of file PruningLinearRegressor.cc.

References PLearn::OptionBase::buildoption, PLearn::declareOption(), PLearn::LinearRegressor::declareOptions(), input_indices, PLearn::OptionBase::learntoption, max_fraction, max_number, min_t_ratio, pruning_method, and t_ratio.

{
    //#####  Build Options  ####################################################

    declareOption(ol, "pruning_method", &PruningLinearRegressor::pruning_method,
                  OptionBase::buildoption,
                  "The pruning method:\n"
                  " - \"max_number\"    = keep only the weights with the k-best t-ratio\n"
                  " - \"max_fraction\"  = same as \"max_number\", but using a fraction rather than a hard threshold\n"
                  " - \"min_t_ratio\"   = keep only the weights with t-ratio > min_t_ratio");

    declareOption(ol, "min_t_ratio", &PruningLinearRegressor::min_t_ratio,
                  OptionBase::buildoption,
                  "Minimum t-ratio for not pruning a coefficient");

    declareOption(ol, "max_number", &PruningLinearRegressor::max_number,
                  OptionBase::buildoption,
                  "Maximum number of coefficients (the default)");

    declareOption(ol, "max_fraction", &PruningLinearRegressor::max_fraction,
                  OptionBase::buildoption,
                  "Maximum fraction (in [0,1]) of coefficients");

    //#####  Learnt Options  ###################################################

    declareOption(ol, "t_ratio", &PruningLinearRegressor::t_ratio,
                  OptionBase::learntoption,
                  "t-ratio statistics for the estimator b (regression coefficients)\n"
                  "Saved as a learned option to allow computing statistical significance\n"
                  "of the coefficients when the model is reloaded and used in test mode.");

    declareOption(ol, "input_indices", &PruningLinearRegressor::input_indices,
                  OptionBase::learntoption,
                  "Indices of inputs kept for regression");

    inherited::declareOptions(ol);
}

Here is the call graph for this function:

static const PPath& PLearn::PruningLinearRegressor::declaringFile ( ) [inline, static]

Reimplemented from PLearn::LinearRegressor.

Definition at line 92 of file PruningLinearRegressor.h.

:
    //#####  Protected Options  ###############################################
PruningLinearRegressor * PLearn::PruningLinearRegressor::deepCopy ( CopiesMap copies) const [virtual]

Reimplemented from PLearn::LinearRegressor.

Definition at line 53 of file PruningLinearRegressor.cc.

OptionList & PLearn::PruningLinearRegressor::getOptionList ( ) const [virtual]

Reimplemented from PLearn::LinearRegressor.

Definition at line 53 of file PruningLinearRegressor.cc.

OptionMap & PLearn::PruningLinearRegressor::getOptionMap ( ) const [virtual]

Reimplemented from PLearn::LinearRegressor.

Definition at line 53 of file PruningLinearRegressor.cc.

RemoteMethodMap & PLearn::PruningLinearRegressor::getRemoteMethodMap ( ) const [virtual]

Reimplemented from PLearn::LinearRegressor.

Definition at line 53 of file PruningLinearRegressor.cc.

void PLearn::PruningLinearRegressor::makeDeepCopyFromShallowCopy ( CopiesMap copies) [virtual]

Transforms a shallow copy into a deep copy.

Reimplemented from PLearn::LinearRegressor.

Definition at line 127 of file PruningLinearRegressor.cc.

References PLearn::deepCopyField(), input_indices, PLearn::LinearRegressor::makeDeepCopyFromShallowCopy(), and t_ratio.

Here is the call graph for this function:

void PLearn::PruningLinearRegressor::newDatasetIndices ( ) [protected]

Find the dataset indices corresponding to coefficients not pruned.

Definition at line 207 of file PruningLinearRegressor.cc.

References PLearn::TVec< T >::append(), computeTRatio(), PLearn::TVec< T >::copy(), i, PLearn::LinearRegressor::include_bias, input_indices, PLearn::TMat< T >::length(), PLearn::TVec< T >::length(), max_fraction, max_number, PLearn::min(), min_t_ratio, PLASSERT, pruning_method, PLearn::TVec< T >::resize(), PLearn::sortElements(), t_ratio, and PLearn::LinearRegressor::weights.

Referenced by train().

{
    // Compute first the t-ratios
    computeTRatio();

    // Sort all t-ratios
    int nb_weights = weights.length();
    PLASSERT(nb_weights == t_ratio.length());
    Vec t_ratio_sort = t_ratio.copy();
    sortElements(t_ratio_sort, true);

    // Find the t-ratio threshold
    real t_ratio_threshold = 0.0;
    if (pruning_method == "max_number")
    {
        int keep_n_weights = min(max_number, nb_weights);
        // Add one coefficient if max is not yet reached
        if (include_bias  &&  keep_n_weights<nb_weights)
            ++keep_n_weights;
        t_ratio_threshold = t_ratio_sort[keep_n_weights-1];
    }
    else if (pruning_method == "max_fraction")
    {
        int keep_n_weights = max_fraction*nb_weights;
        t_ratio_threshold = t_ratio_sort[keep_n_weights-1];
    }
    else if (pruning_method == "min_t_ratio")
    {
        t_ratio_threshold = min_t_ratio;
    }

    // Find kept (not pruned) coefficient indices
    input_indices.resize(0);
    int offset = include_bias ? 1 : 0;
    for (int i=0; i<nb_weights; i++)
    {
        if (t_ratio[i] >= t_ratio_threshold)
        {
            int data_index = i - offset;
            if (data_index >= 0)  // = -1 for bias
                input_indices.append(data_index);
        }
    }
}

Here is the call graph for this function:

Here is the caller graph for this function:

void PLearn::PruningLinearRegressor::setTrainingSet ( VMat  training_set,
bool  call_forget = true 
) [virtual]

Declares the training set.

Then calls build() and forget() if necessary. Also sets this learner's inputsize_ targetsize_ weightsize_ from those of the training_set. Note: You shouldn't have to override this in subclasses, except in maybe to forward the call to an underlying learner.

Reimplemented from PLearn::PLearner.

Definition at line 134 of file PruningLinearRegressor.cc.

References PLERROR, PLearn::PLearner::setTrainingSet(), and PLearn::PLearner::targetsize().

Referenced by train().

{
    inherited::setTrainingSet(training_set, call_forget);
    if (targetsize() > 1)
        PLERROR("PruningLinearRegressor works only with single target problems");
}

Here is the call graph for this function:

Here is the caller graph for this function:

void PLearn::PruningLinearRegressor::train ( ) [virtual]

The role of the train method is to bring the learner up to stage==nstages, updating the train_stats collector with training costs measured on-line in the process.

Reimplemented from PLearn::LinearRegressor.

Definition at line 141 of file PruningLinearRegressor.cc.

References PLearn::TVec< T >::append(), PLearn::VMat::columns(), PLearn::TVec< T >::copy(), input_indices, PLearn::PLearner::inputsize(), PLearn::TVec< T >::length(), newDatasetIndices(), setTrainingSet(), PLearn::LinearRegressor::train(), PLearn::PLearner::train_set, and PLearn::PLearner::weightsize().

{
    // train with all coefficients
    inherited::train();

    // find the dataset indices corresponding to coefficients not pruned
    newDatasetIndices();

    // Add target and (possibly) weight indices
    TVec<int> all_indices = input_indices.copy();
    all_indices.append(inputsize());        // target index = inputsize
    if (weightsize())
        all_indices.append(inputsize()+1);  // weight index = target index + 1

    // Build new training set
    VMat new_trainset = train_set.columns(all_indices);
    int new_inputsize = input_indices.length();
    new_trainset->defineSizes(new_inputsize, 1, weightsize());

    // Train with this new training set
    setTrainingSet(new_trainset);
    inherited::train();
}

Here is the call graph for this function:


Member Data Documentation

Reimplemented from PLearn::LinearRegressor.

Definition at line 92 of file PruningLinearRegressor.h.

Indices of inputs kept for regression.

Definition at line 113 of file PruningLinearRegressor.h.

Referenced by computeOutput(), declareOptions(), makeDeepCopyFromShallowCopy(), newDatasetIndices(), and train().

Definition at line 68 of file PruningLinearRegressor.h.

Referenced by declareOptions(), and newDatasetIndices().

Definition at line 67 of file PruningLinearRegressor.h.

Referenced by build_(), declareOptions(), and newDatasetIndices().

Definition at line 66 of file PruningLinearRegressor.h.

Referenced by build_(), declareOptions(), and newDatasetIndices().

### declare public option fields (such as build options) here

Definition at line 65 of file PruningLinearRegressor.h.

Referenced by build_(), declareOptions(), and newDatasetIndices().

t-ratio statistics for the estimator b (regression coefficients) Saved as a learned option to allow computing statistical significance of the weights when the model is reloaded and used in test mode.

Definition at line 110 of file PruningLinearRegressor.h.

Referenced by computeTRatio(), declareOptions(), makeDeepCopyFromShallowCopy(), and newDatasetIndices().


The documentation for this class was generated from the following files:
 All Classes Namespaces Files Functions Variables Typedefs Enumerations Enumerator Friends Defines