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

Deep Belief Net where the stacked RBMs also use a discriminative criteria. More...

#include <DiscriminativeDeepBeliefNet.h>

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

List of all members.

Public Member Functions

 DiscriminativeDeepBeliefNet ()
 Default constructor.
virtual int outputsize () const
 Returns the size of this learner's output, (which typically may depend on its inputsize(), targetsize() and set options).
virtual void forget ()
 (Re-)initializes the PLearner in its fresh state (that state may depend on the 'seed' option) and sets 'stage' back to 0 (this is the stage of a fresh learner!).
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 void computeCostsFromOutputs (const Vec &input, const Vec &output, const Vec &target, Vec &costs) const
 Computes the costs from already computed output.
virtual TVec< std::string > getTestCostNames () const
 Returns the names of the costs computed by computeCostsFromOutpus (and thus the test method).
virtual TVec< std::string > getTrainCostNames () const
 Returns the names of the objective costs that the train method computes and for which it updates the VecStatsCollector train_stats.
virtual void setTrainingSet (VMat training_set, bool call_forget=true)
 Declares the training set.
void updateNearestNeighbors ()
void greedyStep (const Vec &input, const Vec &target, int index, Vec train_costs, int stage, Vec dissimilar_example)
void fineTuningStep (const Vec &input, const Vec &target, Vec &train_costs)
void computeRepresentation (const Vec &input, Vec &representation, int layer) const
virtual string classname () const
virtual OptionListgetOptionList () const
virtual OptionMapgetOptionMap () const
virtual RemoteMethodMapgetRemoteMethodMap () const
virtual
DiscriminativeDeepBeliefNet
deepCopy (CopiesMap &copies) const
virtual void build ()
 Finish building the object; just call inherited::build followed by build_()
virtual void makeDeepCopyFromShallowCopy (CopiesMap &copies)
 Transforms a shallow copy into a deep copy.

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

real cd_learning_rate
 Contrastive divergence learning rate.
real cd_decrease_ct
 Contrastive divergence decrease constant.
real fine_tuning_learning_rate
 The learning rate used during the fine tuning gradient descent.
real fine_tuning_decrease_ct
 The decrease constant of the learning rate used during fine tuning gradient descent.
TVec< inttraining_schedule
 Number of examples to use during each phase of greedy pre-training.
TVec< PP< RBMLayer > > layers
 The layers of units in the network.
TVec< PP< RBMConnection > > connections
 The weights of the connections between the layers.
TVec< PP< RBMLayer > > unsupervised_layers
 Additional units for greedy unsupervised learning.
TVec< PP< RBMConnection > > unsupervised_connections
 Additional connections for greedy unsupervised learning.
int k_neighbors
 Number of good nearest neighbors to attract and bad nearest neighbors to repel.
int n_classes
 Number of classes.
real discriminative_criteria_weight
 Weight of the discriminative criteria.
real output_weights_l1_penalty_factor
 Output weights l1_penalty_factor.
real output_weights_l2_penalty_factor
 Output weights l2_penalty_factor.
bool compare_joint_in_discriminative_criteria
 Indication that the discriminative criteria should use the joint over the input and the hidden units, instead of the conditional over the hidden units given the input units.
bool do_not_use_generative_criteria
 Indication that the generative criteria should not be used during learning (does not work with compare_joint_in_discriminative_criteria = true).
int n_layers
 Number of layers.

Static Public Attributes

static StaticInitializer _static_initializer_

Static Protected Member Functions

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

Protected Attributes

bool nearest_neighbors_are_up_to_date
 Indication that nearest_neighbors_indices is up to date.
TVec< Vecactivations
 Stores the activations of the input and hidden layers (at the input of the layers)
TVec< Vecexpectations
 Stores the expectations of the input and hidden layers (at the output of the layers)
TVec< Vecactivation_gradients
 Stores the gradient of the cost wrt the activations of the input and hidden layers (at the input of the layers)
TVec< Vecexpectation_gradients
 Stores the gradient of the cost wrt the expectations of the input and hidden layers (at the output of the layers)
TVec< PP< RBMLayer > > greedy_layers
 Layers used for greedy learning.
TVec< PP< RBMConnection > > greedy_connections
 Connections used for greedy learning.
Vec dissimilar_example_representation
 Dissimilar example representation.
Vec input_representation
 Example representation.
Vec pos_down_val
 Positive down statistic.
Vec pos_up_val
 Positive up statistic.
Vec neg_down_val
 Negative down statistic.
Vec neg_up_val
 Negative up statistic.
Vec disc_pos_down_val1
 First discriminative positive down statistic.
Vec disc_pos_up_val1
 First discriminative positive up statistic.
Vec disc_pos_down_val2
 Second discriminative positive down statistic.
Vec disc_pos_up_val2
 Second discriminative positive up statistic.
Vec disc_neg_down_val
 Negative down statistic.
Vec disc_neg_up_val
 Negative up statistic.
Vec final_cost_input
 Input of cost function.
Vec final_cost_value
 Cost value.
Vec final_cost_gradient
 Cost gradient on output layer.
TVec< PP< ClassSubsetVMatrix > > other_class_datasets
 Datasets for each class.
TMat< intnearest_neighbors_indices
 Nearest neighbors for each training example.
TVec< intgreedy_stages
 Stages of the different greedy phases.
int currently_trained_layer
 Currently trained layer (1 means the first hidden layer, n_layers means the output layer)
PP< OnlineLearningModulefinal_module
 Output layer of neural net.
PP< CostModulefinal_cost
 Cost on output layer of neural net.

Private Types

typedef PLearner inherited

Private Member Functions

void build_ ()
 This does the actual building.
void build_layers_and_connections ()
void build_output_layer_and_cost ()
void setLearningRate (real the_learning_rate)

Detailed Description

Deep Belief Net where the stacked RBMs also use a discriminative criteria.

Definition at line 64 of file DiscriminativeDeepBeliefNet.h.


Member Typedef Documentation

Reimplemented from PLearn::PLearner.

Definition at line 66 of file DiscriminativeDeepBeliefNet.h.


Constructor & Destructor Documentation

PLearn::DiscriminativeDeepBeliefNet::DiscriminativeDeepBeliefNet ( )

Default constructor.

Definition at line 59 of file DiscriminativeDeepBeliefNet.cc.

References PLearn::PLearner::nstages, and PLearn::PLearner::random_gen.

                                                         :
    cd_learning_rate( 0. ),
    cd_decrease_ct( 0. ),
    fine_tuning_learning_rate( 0. ),
    fine_tuning_decrease_ct( 0. ),
    k_neighbors( 1 ),
    n_classes( -1 ),
    discriminative_criteria_weight( 0. ), 
    output_weights_l1_penalty_factor(0),
    output_weights_l2_penalty_factor(0),
    compare_joint_in_discriminative_criteria( false ),
    do_not_use_generative_criteria( false ),
//    cancel_normalization_terms( false ),
    n_layers( 0 ),
    nearest_neighbors_are_up_to_date( false ),
    currently_trained_layer( 0 )
{
    // random_gen will be initialized in PLearner::build_()
    random_gen = new PRandom();
    nstages = 0;
}

Member Function Documentation

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

Reimplemented from PLearn::PLearner.

Definition at line 57 of file DiscriminativeDeepBeliefNet.cc.

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

Reimplemented from PLearn::PLearner.

Definition at line 57 of file DiscriminativeDeepBeliefNet.cc.

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

Reimplemented from PLearn::PLearner.

Definition at line 57 of file DiscriminativeDeepBeliefNet.cc.

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

Reimplemented from PLearn::PLearner.

Definition at line 57 of file DiscriminativeDeepBeliefNet.cc.

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

Reimplemented from PLearn::Object.

Definition at line 57 of file DiscriminativeDeepBeliefNet.cc.

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

Reimplemented from PLearn::PLearner.

Definition at line 57 of file DiscriminativeDeepBeliefNet.cc.

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

Finish building the object; just call inherited::build followed by build_()

Reimplemented from PLearn::PLearner.

Definition at line 470 of file DiscriminativeDeepBeliefNet.cc.

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

Here is the call graph for this function:

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

This does the actual building.

Reimplemented from PLearn::PLearner.

Definition at line 213 of file DiscriminativeDeepBeliefNet.cc.

References build_layers_and_connections(), build_output_layer_and_cost(), PLearn::TVec< T >::clear(), compare_joint_in_discriminative_criteria, currently_trained_layer, do_not_use_generative_criteria, PLearn::endl(), final_cost, final_module, greedy_stages, PLearn::PLearner::inputsize_, k_neighbors, layers, PLearn::TVec< T >::length(), n_classes, n_layers, PLERROR, PLearn::TVec< T >::resize(), PLearn::PLearner::stage, PLearn::PLearner::targetsize_, training_schedule, and PLearn::PLearner::weightsize_.

Referenced by build().

{
    // ### This method should do the real building of the object,
    // ### according to set 'options', in *any* situation.
    // ### Typical situations include:
    // ###  - Initial building of an object from a few user-specified options
    // ###  - Building of a "reloaded" object: i.e. from the complete set of
    // ###    all serialised options.
    // ###  - Updating or "re-building" of an object after a few "tuning"
    // ###    options have been modified.
    // ### You should assume that the parent class' build_() has already been
    // ### called.

    MODULE_LOG << "build_() called" << endl;

    if(inputsize_ > 0 && targetsize_ > 0)
    {
        // Initialize some learnt variables
        n_layers = layers.length();
        
        if( n_classes <= 0 )
            PLERROR("DiscriminativeDeepBeliefNet::build_() - \n"
                    "n_classes should be > 0.\n");

        if( k_neighbors <= 0 )
            PLERROR("DiscriminativeDeepBeliefNet::build_() - \n"
                    "k_neighbors should be > 0.\n");

        if( weightsize_ > 0 )
            PLERROR("DiscriminativeDeepBeliefNet::build_() - \n"
                    "usage of weighted samples (weight size > 0) is not\n"
                    "implemented yet.\n");

        if( training_schedule.length() != n_layers-1 )        
            PLERROR("DiscriminativeDeepBeliefNet::build_() - \n"
                    "training_schedule should have %d elements.\n",
                    n_layers-1);
        
        if( compare_joint_in_discriminative_criteria && do_not_use_generative_criteria)
            PLERROR("DiscriminativeDeepBeliefNet::build_() - \n"
                    "compare_joint_in_discriminative_criteria can't be used with\n"
                    "do_not_use_generative_criteria.\n");

//        if( (!compare_joint_in_discriminative_criteria || do_not_use_generative_criteria)
//            && cancel_normalization_terms )
//            PLERROR("DiscriminativeDeepBeliefNet::build_() - \n"
//                    "cancel_normalization_terms should be used with\n"
//                    "compare_joint_in_discriminative_criteria and \n"
//                    "do_not_use_generative_criteria without .\n");
            
        if(greedy_stages.length() == 0)
        {
            greedy_stages.resize(n_layers-1);
            greedy_stages.clear();
        }

        if(stage > 0)
            currently_trained_layer = n_layers;
        else
        {            
            currently_trained_layer = n_layers-1;
            while(currently_trained_layer>1
                  && greedy_stages[currently_trained_layer-1] <= 0)
                currently_trained_layer--;
        }

        build_layers_and_connections();

        if( !final_module || !final_cost )
            build_output_layer_and_cost();
    }
}

Here is the call graph for this function:

Here is the caller graph for this function:

void PLearn::DiscriminativeDeepBeliefNet::build_layers_and_connections ( ) [private]

Definition at line 334 of file DiscriminativeDeepBeliefNet.cc.

References activation_gradients, activations, connections, PLearn::endl(), expectation_gradients, expectations, greedy_connections, greedy_layers, i, PLearn::PLearner::inputsize_, layers, PLearn::TVec< T >::length(), n_layers, PLERROR, PLearn::PLearner::random_gen, PLearn::TVec< T >::resize(), PLearn::TVec< T >::size(), unsupervised_connections, and unsupervised_layers.

Referenced by build_().

{
    MODULE_LOG << "build_layers_and_connections() called" << endl;

    if( connections.length() != n_layers-1 )
        PLERROR("DiscriminativeDeepBeliefNet::build_layers_and_connections() - \n"
                "there should be %d connections.\n",
                n_layers-1);
     
    if(unsupervised_layers.length() != n_layers-1 
       && unsupervised_layers.length() != 0)
        PLERROR("DiscriminativeDeepBeliefNet::build_layers_and_connections() - \n"
                "there should be either 0 of %d unsupervised_layers.\n",
                n_layers-1);
        
    if(unsupervised_connections.length() != n_layers-1 
       && unsupervised_connections.length() != 0)
        PLERROR("DiscriminativeDeepBeliefNet::build_layers_and_connections() - \n"
                "there should be either 0 of %d unsupervised_connections.\n",
                n_layers-1);
        
    if(unsupervised_connections.length() != unsupervised_layers.length())
        PLERROR("DiscriminativeDeepBeliefNet::build_layers_and_connections() - \n"
                "there should be as many unsupervised_connections and "
                "unsupervised_layers.\n");
        

    if(layers[0]->size != inputsize_)
        PLERROR("DiscriminativeDeepBeliefNet::build_layers_and_connections() - \n"
                "layers[0] should have a size of %d.\n",
                inputsize_);
    

    activations.resize( n_layers );
    expectations.resize( n_layers );
    activation_gradients.resize( n_layers );
    expectation_gradients.resize( n_layers );

    greedy_layers.resize(n_layers-1);
    greedy_connections.resize(n_layers-1);
    for( int i=0 ; i<n_layers-1 ; i++ )
    {
        if( layers[i]->size != connections[i]->down_size )
            PLERROR("DiscriminativeDeepBeliefNet::build_layers_and_connections() "
                    "- \n"
                    "connections[%i] should have a down_size of %d.\n",
                    i, layers[i]->size);

        if( connections[i]->up_size != layers[i+1]->size )
            PLERROR("DiscriminativeDeepBeliefNet::build_layers_and_connections() "
                    "- \n"
                    "connections[%i] should have a up_size of %d.\n",
                    i, layers[i+1]->size);

        if(unsupervised_layers.length() != 0 &&
           unsupervised_connections.length() != 0 && 
           unsupervised_layers[i] && unsupervised_connections[i])
        {
            if( layers[i]->size != 
                unsupervised_connections[i]->down_size )
                PLERROR("DiscriminativeDeepBeliefNet::build_layers_and_connections() "
                        "- \n"
                        "connections[%i] should have a down_size of %d.\n",
                        i, unsupervised_layers[i]->size);
            
            if( unsupervised_connections[i]->up_size != 
                unsupervised_layers[i]->size )
                PLERROR("DiscriminativeDeepBeliefNet::build_layers_and_connections() "
                        "- \n"
                        "connections[%i] should have a up_size of %d.\n",
                        i, unsupervised_layers[i+1]->size);
            
            if( !(unsupervised_layers[i]->random_gen) )
            {
                unsupervised_layers[i]->random_gen = random_gen;
                unsupervised_layers[i]->forget();
            }
            
            if( !(unsupervised_connections[i]->random_gen) )
            {
                unsupervised_connections[i]->random_gen = random_gen;
                unsupervised_connections[i]->forget();
            }

            PP<RBMMixedLayer> greedy_layer = new RBMMixedLayer();
            greedy_layer->sub_layers.resize(2);
            greedy_layer->sub_layers[0] = layers[i+1];
            greedy_layer->sub_layers[1] = unsupervised_layers[i];
            greedy_layer->size = layers[i+1]->size + unsupervised_layers[i]->size;
            greedy_layer->build();

            PP<RBMMixedConnection> greedy_connection = new RBMMixedConnection();
            greedy_connection->sub_connections.resize(2,1);
            greedy_connection->sub_connections(0,0) = connections[i];
            greedy_connection->sub_connections(1,0) = unsupervised_connections[i];
            greedy_connection->build();
            
            greedy_layers[i] = greedy_layer;
            greedy_connections[i] = greedy_connection;
        }
        else
        {
            greedy_layers[i] = layers[i+1];
            greedy_connections[i] = connections[i];
        }

        if( !(layers[i]->random_gen) )
        {
            layers[i]->random_gen = random_gen;
            layers[i]->forget();
        }

        if( !(connections[i]->random_gen) )
        {
            connections[i]->random_gen = random_gen;
            connections[i]->forget();
        }

        activations[i].resize( layers[i]->size );
        expectations[i].resize( layers[i]->size );
        activation_gradients[i].resize( layers[i]->size );
        expectation_gradients[i].resize( layers[i]->size );
    }

    if( !(layers[n_layers-1]->random_gen) )
    {
        layers[n_layers-1]->random_gen = random_gen;
        layers[n_layers-1]->forget();
    }
    activations[n_layers-1].resize( layers[n_layers-1]->size );
    expectations[n_layers-1].resize( layers[n_layers-1]->size );
    activation_gradients[n_layers-1].resize( layers[n_layers-1]->size );
    expectation_gradients[n_layers-1].resize( layers[n_layers-1]->size );
}

Here is the call graph for this function:

Here is the caller graph for this function:

void PLearn::DiscriminativeDeepBeliefNet::build_output_layer_and_cost ( ) [private]

Definition at line 286 of file DiscriminativeDeepBeliefNet.cc.

References PLearn::GradNNetLayerModule::build(), PLearn::SoftmaxModule::build(), PLearn::ModuleStackModule::build(), PLearn::ClassErrorCostModule::build(), PLearn::CombiningCostsModule::build(), PLearn::NLLCostModule::build(), PLearn::class_error(), PLearn::CombiningCostsModule::cost_weights, final_cost, final_module, PLearn::OnlineLearningModule::input_size, PLearn::GradNNetLayerModule::L1_penalty_factor, PLearn::GradNNetLayerModule::L2_penalty_factor, layers, PLearn::ModuleStackModule::modules, n_classes, n_layers, PLearn::OnlineLearningModule::output_size, output_weights_l1_penalty_factor, output_weights_l2_penalty_factor, PLearn::OnlineLearningModule::random_gen, PLearn::PLearner::random_gen, PLearn::TVec< T >::resize(), PLearn::TVec< T >::size(), and PLearn::CombiningCostsModule::sub_costs.

Referenced by build_(), and forget().

{
    GradNNetLayerModule* gnl = new GradNNetLayerModule();
    gnl->input_size = layers[n_layers-1]->size;
    gnl->output_size = n_classes;
    gnl->L1_penalty_factor = output_weights_l1_penalty_factor;
    gnl->L2_penalty_factor = output_weights_l2_penalty_factor;
    gnl->random_gen = random_gen;
    gnl->build();

    SoftmaxModule* sm = new SoftmaxModule();
    sm->input_size = n_classes;
    sm->random_gen = random_gen;
    sm->build();

    ModuleStackModule* msm = new ModuleStackModule();
    msm->modules.resize(2);
    msm->modules[0] = gnl;
    msm->modules[1] = sm;
    msm->random_gen = random_gen;
    msm->build();
    final_module = msm;

    final_module->forget();

    NLLCostModule* nll = new NLLCostModule();
    nll->input_size = n_classes;
    nll->random_gen = random_gen;
    nll->build();
    
    ClassErrorCostModule* class_error = new ClassErrorCostModule();
    class_error->input_size = n_classes;
    class_error->random_gen = random_gen;
    class_error->build();

    CombiningCostsModule* comb_costs = new CombiningCostsModule();
    comb_costs->cost_weights.resize(2);
    comb_costs->cost_weights[0] = 1;
    comb_costs->cost_weights[1] = 0;
    comb_costs->sub_costs.resize(2);
    comb_costs->sub_costs[0] = nll;
    comb_costs->sub_costs[1] = class_error;
    comb_costs->build();

    final_cost = comb_costs;
    final_cost->forget();
}

Here is the call graph for this function:

Here is the caller graph for this function:

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

Reimplemented from PLearn::Object.

Definition at line 57 of file DiscriminativeDeepBeliefNet.cc.

Referenced by train().

Here is the caller graph for this function:

void PLearn::DiscriminativeDeepBeliefNet::computeCostsFromOutputs ( const Vec input,
const Vec output,
const Vec target,
Vec costs 
) const [virtual]

Computes the costs from already computed output.

Implements PLearn::PLearner.

Definition at line 976 of file DiscriminativeDeepBeliefNet.cc.

References currently_trained_layer, PLearn::TVec< T >::fill(), getTestCostNames(), PLearn::TVec< T >::last(), MISSING_VALUE, n_layers, and PLearn::TVec< T >::resize().

{
    //Assumes that computeOutput has been called

    costs.resize( getTestCostNames().length() );
    costs.fill( MISSING_VALUE );

    if( currently_trained_layer>n_layers-1 )
        if( ((int)round(output[0])) == ((int)round(target[0])) )
            costs.last() = 0;
        else
            costs.last() = 1;
}

Here is the call graph for this function:

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

Computes the output from the input.

Reimplemented from PLearn::PLearner.

Definition at line 960 of file DiscriminativeDeepBeliefNet.cc.

References PLearn::argmax(), computeRepresentation(), currently_trained_layer, final_cost_input, final_module, input_representation, and n_layers.

Here is the call graph for this function:

void PLearn::DiscriminativeDeepBeliefNet::computeRepresentation ( const Vec input,
Vec representation,
int  layer 
) const

Definition at line 938 of file DiscriminativeDeepBeliefNet.cc.

References activations, connections, expectations, i, layers, PLearn::TVec< T >::length(), and PLearn::TVec< T >::resize().

Referenced by computeOutput(), fineTuningStep(), and greedyStep().

{
    if(layer == 0)
    {
        representation.resize(input.length());
        expectations[0] << input;
        representation << input;
        return;
    }

    expectations[0] << input;
    for( int i=0 ; i<layer; i++ )
    {
        connections[i]->fprop( expectations[i], activations[i+1] );
        layers[i+1]->fprop(activations[i+1],expectations[i+1]);
    }
    representation.resize(expectations[layer].length());
    representation << expectations[layer];
}

Here is the call graph for this function:

Here is the caller graph for this function:

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

Declares the class options.

Reimplemented from PLearn::PLearner.

Definition at line 81 of file DiscriminativeDeepBeliefNet.cc.

References PLearn::OptionBase::buildoption, cd_decrease_ct, cd_learning_rate, compare_joint_in_discriminative_criteria, connections, PLearn::declareOption(), PLearn::PLearner::declareOptions(), discriminative_criteria_weight, do_not_use_generative_criteria, final_cost, final_module, fine_tuning_decrease_ct, fine_tuning_learning_rate, greedy_stages, k_neighbors, layers, PLearn::OptionBase::learntoption, n_classes, n_layers, output_weights_l1_penalty_factor, output_weights_l2_penalty_factor, training_schedule, unsupervised_connections, and unsupervised_layers.

{
    declareOption(ol, "cd_learning_rate", 
                  &DiscriminativeDeepBeliefNet::cd_learning_rate,
                  OptionBase::buildoption,
                  "The learning rate used during the RBM "
                  "contrastive divergence training.\n");

    declareOption(ol, "cd_decrease_ct", 
                  &DiscriminativeDeepBeliefNet::cd_decrease_ct,
                  OptionBase::buildoption,
                  "The decrease constant of the learning rate used during "
                  "the RBMs contrastive\n"
                  "divergence training. When a hidden layer has finished "
                  "its training,\n"
                  "the learning rate is reset to it's initial value.\n");

    declareOption(ol, "fine_tuning_learning_rate", 
                  &DiscriminativeDeepBeliefNet::fine_tuning_learning_rate,
                  OptionBase::buildoption,
                  "The learning rate used during the fine tuning gradient descent.\n");

    declareOption(ol, "fine_tuning_decrease_ct", 
                  &DiscriminativeDeepBeliefNet::fine_tuning_decrease_ct,
                  OptionBase::buildoption,
                  "The decrease constant of the learning rate used during "
                  "fine tuning\n"
                  "gradient descent.\n");

    declareOption(ol, "training_schedule", 
                  &DiscriminativeDeepBeliefNet::training_schedule,
                  OptionBase::buildoption,
                  "Number of examples to use during each phase of greedy pre-training.\n"
                  "The number of fine-tunig steps is defined by nstages.\n"
        );

    declareOption(ol, "layers", &DiscriminativeDeepBeliefNet::layers,
                  OptionBase::buildoption,
                  "The layers of units in the network. The first element\n"
                  "of this vector should be the input layer and the\n"
                  "subsequent elements should be the hidden layers. The\n"
                  "output layer should not be included in layers.\n");

    declareOption(ol, "connections", &DiscriminativeDeepBeliefNet::connections,
                  OptionBase::buildoption,
                  "The weights of the connections between the layers.\n");

    declareOption(ol, "unsupervised_layers", 
                  &DiscriminativeDeepBeliefNet::unsupervised_layers,
                  OptionBase::buildoption,
                  "Additional units for greedy unsupervised learning.\n");

    declareOption(ol, "unsupervised_connections", 
                  &DiscriminativeDeepBeliefNet::unsupervised_connections,
                  OptionBase::buildoption,
                  "Additional connections for greedy unsupervised learning.\n");

    declareOption(ol, "k_neighbors", 
                  &DiscriminativeDeepBeliefNet::k_neighbors,
                  OptionBase::buildoption,
                  "Number of good nearest neighbors to attract and bad nearest "
                  "neighbors to repel.\n");

    declareOption(ol, "n_classes", 
                  &DiscriminativeDeepBeliefNet::n_classes,
                  OptionBase::buildoption,
                  "Number of classes.\n");

    declareOption(ol, "discriminative_criteria_weight", 
                  &DiscriminativeDeepBeliefNet::discriminative_criteria_weight,
                  OptionBase::buildoption,
                  "Weight of the discriminative criteria.\n");

    declareOption(ol, "output_weights_l1_penalty_factor", 
                  &DiscriminativeDeepBeliefNet::output_weights_l1_penalty_factor,
                  OptionBase::buildoption,
                  "Output weights l1_penalty_factor.\n");

    declareOption(ol, "output_weights_l2_penalty_factor", 
                  &DiscriminativeDeepBeliefNet::output_weights_l2_penalty_factor,
                  OptionBase::buildoption,
                  "Output weights l2_penalty_factor.\n");

    declareOption(ol, "compare_joint_in_discriminative_criteria", 
                  &DiscriminativeDeepBeliefNet::compare_joint_in_discriminative_criteria,
                  OptionBase::buildoption,
                  "Indication that the discriminative criteria should use the joint\n"
                  "over the input and the hidden units, instead of the conditional\n"
                  "over the hidden units given the input units.\n");

    declareOption(ol, "do_not_use_generative_criteria", 
                  &DiscriminativeDeepBeliefNet::do_not_use_generative_criteria,
                  OptionBase::buildoption,
                  "Indication that the generative criteria should not be used during learning\n"
                  "(does not work with compare_joint_in_discriminative_criteria = true).\n");

//    declareOption(ol, "cancel_normalization_terms", 
//                  &DiscriminativeDeepBeliefNet::cancel_normalization_terms,
//                  OptionBase::buildoption,
//                  "Indication that the discriminative and generative criteria should cancel\n"
//                  "their normalization terms. This is for the "
//                  "compare_joint_in_discriminative_criteria\n"
//                  "option, and this option ignores the value of discriminative_criteria_weight.\n");

    declareOption(ol, "greedy_stages", 
                  &DiscriminativeDeepBeliefNet::greedy_stages,
                  OptionBase::learntoption,
                  "Number of training samples seen in the different greedy "
                  "phases.\n"
        );

    declareOption(ol, "n_layers", &DiscriminativeDeepBeliefNet::n_layers,
                  OptionBase::learntoption,
                  "Number of layers.\n"
        );

    declareOption(ol, "final_module", 
                  &DiscriminativeDeepBeliefNet::final_module,
                  OptionBase::learntoption,
                  "Output layer of neural net.\n"
        );

    declareOption(ol, "final_cost", 
                  &DiscriminativeDeepBeliefNet::final_cost,
                  OptionBase::learntoption,
                  "Cost on output layer of neural net.\n"
        );

    // Now call the parent class' declareOptions
    inherited::declareOptions(ol);
}

Here is the call graph for this function:

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

Reimplemented from PLearn::PLearner.

Definition at line 197 of file DiscriminativeDeepBeliefNet.h.

:
    //#####  Not Options  #####################################################
DiscriminativeDeepBeliefNet * PLearn::DiscriminativeDeepBeliefNet::deepCopy ( CopiesMap copies) const [virtual]

Reimplemented from PLearn::PLearner.

Definition at line 57 of file DiscriminativeDeepBeliefNet.cc.

void PLearn::DiscriminativeDeepBeliefNet::fineTuningStep ( const Vec input,
const Vec target,
Vec train_costs 
)

Definition at line 902 of file DiscriminativeDeepBeliefNet.cc.

References activation_gradients, activations, computeRepresentation(), connections, expectation_gradients, expectations, final_cost, final_cost_gradient, final_cost_input, final_cost_value, final_module, i, input_representation, layers, and n_layers.

Referenced by train().

Here is the call graph for this function:

Here is the caller graph for this function:

void PLearn::DiscriminativeDeepBeliefNet::forget ( ) [virtual]

(Re-)initializes the PLearner in its fresh state (that state may depend on the 'seed' option) and sets 'stage' back to 0 (this is the stage of a fresh learner!).

(Re-)initialize the PLearner in its fresh state (that state may depend on the 'seed' option) and sets 'stage' back to 0 (this is the stage of a fresh learner!)

A typical forget() method should do the following:

  • call inherited::forget() to initialize its random number generator with the 'seed' option
  • initialize the learner's parameters, using this random generator
  • stage = 0

Reimplemented from PLearn::PLearner.

Definition at line 528 of file DiscriminativeDeepBeliefNet.cc.

References build_output_layer_and_cost(), PLearn::TVec< T >::clear(), connections, PLearn::PLearner::forget(), greedy_stages, i, layers, PLearn::TVec< T >::length(), n_layers, PLearn::PLearner::stage, unsupervised_connections, and unsupervised_layers.

{

    inherited::forget();

    for( int i=0 ; i<n_layers ; i++ )
        layers[i]->forget();
    
    for( int i=0 ; i<n_layers-1 ; i++ )
        connections[i]->forget();
    
    if(unsupervised_layers.length() != 0)
        for( int i=0 ; i<n_layers-1 ; i++ )
            unsupervised_layers[i]->forget();
    
    if(unsupervised_connections.length() != 0)
        for( int i=0 ; i<n_layers-1 ; i++ )
            unsupervised_connections[i]->forget();
    
    build_output_layer_and_cost();

    stage = 0;
    greedy_stages.clear();
}

Here is the call graph for this function:

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

Reimplemented from PLearn::Object.

Definition at line 57 of file DiscriminativeDeepBeliefNet.cc.

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

Reimplemented from PLearn::Object.

Definition at line 57 of file DiscriminativeDeepBeliefNet.cc.

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

Reimplemented from PLearn::Object.

Definition at line 57 of file DiscriminativeDeepBeliefNet.cc.

TVec< string > PLearn::DiscriminativeDeepBeliefNet::getTestCostNames ( ) const [virtual]

Returns the names of the costs computed by computeCostsFromOutpus (and thus the test method).

Implements PLearn::PLearner.

Definition at line 991 of file DiscriminativeDeepBeliefNet.cc.

References PLearn::TVec< T >::append().

Referenced by computeCostsFromOutputs(), and getTrainCostNames().

{
    // Return the names of the costs computed by computeCostsFromOutputs
    // (these may or may not be exactly the same as what's returned by
    // getTrainCostNames).

    TVec<string> cost_names(0);

    cost_names.append( "class_error" );

    return cost_names;
}

Here is the call graph for this function:

Here is the caller graph for this function:

TVec< string > PLearn::DiscriminativeDeepBeliefNet::getTrainCostNames ( ) const [virtual]

Returns the names of the objective costs that the train method computes and for which it updates the VecStatsCollector train_stats.

Implements PLearn::PLearner.

Definition at line 1004 of file DiscriminativeDeepBeliefNet.cc.

References getTestCostNames().

Referenced by train().

{
    return getTestCostNames();
}

Here is the call graph for this function:

Here is the caller graph for this function:

void PLearn::DiscriminativeDeepBeliefNet::greedyStep ( const Vec input,
const Vec target,
int  index,
Vec  train_costs,
int  stage,
Vec  dissimilar_example 
)

Definition at line 708 of file DiscriminativeDeepBeliefNet.cc.

References cd_decrease_ct, cd_learning_rate, compare_joint_in_discriminative_criteria, computeRepresentation(), connections, disc_neg_down_val, disc_neg_up_val, disc_pos_down_val1, disc_pos_down_val2, disc_pos_up_val1, disc_pos_up_val2, discriminative_criteria_weight, dissimilar_example_representation, do_not_use_generative_criteria, greedy_connections, greedy_layers, input_representation, layers, n_layers, neg_down_val, neg_up_val, PLASSERT, pos_down_val, pos_up_val, and PLearn::sample().

Referenced by train().

{
    PLASSERT( index < n_layers );
    real lr;

    // Get dissimilar example representation
    computeRepresentation(dissimilar_example, dissimilar_example_representation, 
                          index);

    // Get example representation
    computeRepresentation(input, input_representation, 
                          index);

    if( !do_not_use_generative_criteria )
    {
        // CD generative learning stats
        
        // Positive phase
        greedy_connections[index]->setAsDownInput( input_representation );
        greedy_layers[index]->getAllActivations( greedy_connections[index] );
        greedy_layers[index]->computeExpectation();
        greedy_layers[index]->generateSample();
        
        pos_down_val << input_representation;
        pos_up_val << greedy_layers[index]->expectation;
        
        if( !compare_joint_in_discriminative_criteria )
        {
            disc_pos_down_val1 << input_representation;
            disc_pos_up_val1 << layers[index+1]->expectation;
        }
        
//        if( !cancel_normalization_terms )
//        {
        // Negative phase
        greedy_connections[index]->setAsUpInput( greedy_layers[index]->sample );    
        layers[index]->getAllActivations( greedy_connections[index] );
        layers[index]->computeExpectation();
        layers[index]->generateSample();
        
        greedy_connections[index]->setAsDownInput( layers[index]->sample );
        greedy_layers[index]->getAllActivations( greedy_connections[index] );
        greedy_layers[index]->computeExpectation();
        
        neg_down_val << layers[index]->sample;
        neg_up_val << greedy_layers[index]->expectation;
//      }
    }
    else if( !compare_joint_in_discriminative_criteria )
    {
        
        connections[index]->setAsDownInput( input_representation );
        layers[index+1]->getAllActivations( connections[index] );
        layers[index+1]->computeExpectation();
        
        disc_pos_down_val1 << input_representation;
        disc_pos_up_val1 << layers[index+1]->expectation;
    }

    // CD discriminative criteria stats

    if( !compare_joint_in_discriminative_criteria )
    {
        // Positive phase
        connections[index]->setAsDownInput( dissimilar_example_representation );
        layers[index+1]->getAllActivations( connections[index] );
        layers[index+1]->computeExpectation();
        
        disc_pos_down_val2 << dissimilar_example_representation;
        disc_pos_up_val2 << layers[index+1]->expectation;
    }

    // Negative phase
    disc_neg_down_val << input_representation;
    disc_neg_down_val += dissimilar_example_representation;
    disc_neg_down_val /= 2;
    connections[index]->setAsDownInput( disc_neg_down_val );
    layers[index+1]->getAllActivations( connections[index] );
    layers[index+1]->computeExpectation();

    disc_neg_up_val << layers[index+1]->expectation;

    if( compare_joint_in_discriminative_criteria )
        //&& !cancel_normalization_terms)
    {
        layers[index+1]->generateSample();
        connections[index]->setAsUpInput( layers[index+1]->sample );
        layers[index]->getAllActivations( connections[index] );
        layers[index]->computeExpectation();
        layers[index]->generateSample();

        connections[index]->setAsDownInput( layers[index]->sample );
        layers[index+1]->getAllActivations( connections[index] );
        layers[index+1]->computeExpectation();

        disc_pos_down_val1 << layers[index]->sample;
        disc_pos_up_val1 << layers[index+1]->expectation;
    }

    // RBM updates
    if( !do_not_use_generative_criteria )
        //&& !cancel_normalization_terms )
    {
        lr = cd_learning_rate/(1 + cd_decrease_ct 
                               * this_stage); 
        
        layers[index]->setLearningRate( lr );
        greedy_connections[index]->setLearningRate( lr );
        greedy_layers[index]->setLearningRate( lr );
        
        layers[index]->update( pos_down_val, neg_down_val );
        greedy_connections[index]->update( pos_down_val, pos_up_val,
                                           neg_down_val, neg_up_val );
        greedy_layers[index]->update( pos_up_val, neg_up_val );
    }
    
    if( //cancel_normalization_terms || 
        discriminative_criteria_weight != 0 )
    {
        lr = discriminative_criteria_weight * 
            cd_learning_rate/(1 + cd_decrease_ct 
                              * this_stage); 
        
        if( !compare_joint_in_discriminative_criteria )
        {
            layers[index]->setLearningRate( lr );
            connections[index]->setLearningRate( lr );
            layers[index+1]->setLearningRate( lr );
            
            layers[index]->accumulatePosStats( disc_pos_down_val1 );
            layers[index]->accumulatePosStats( disc_pos_down_val2 );
            layers[index]->accumulateNegStats( disc_neg_down_val );
            layers[index]->update();
            
            connections[index]->accumulatePosStats( disc_pos_down_val1,
                                                    disc_pos_up_val1 );
            connections[index]->accumulatePosStats( disc_pos_down_val2,
                                                    disc_pos_up_val2 );
            connections[index]->accumulateNegStats( disc_neg_down_val,
                                                    disc_neg_up_val );
            connections[index]->update();
            
            layers[index+1]->accumulatePosStats( disc_pos_up_val1 );
            layers[index+1]->accumulatePosStats( disc_pos_up_val2 );
            layers[index+1]->accumulateNegStats( disc_neg_up_val );
            layers[index+1]->update();
        }
        else //if( !cancel_normalization_terms )
        {
            layers[index]->setLearningRate( lr );
            connections[index]->setLearningRate( lr );
            layers[index+1]->setLearningRate( lr );
            
            layers[index]->accumulatePosStats( disc_pos_down_val1 );
            layers[index]->accumulateNegStats( disc_neg_down_val );
            layers[index]->update();
            
            connections[index]->accumulatePosStats( disc_pos_down_val1,
                                                    disc_pos_up_val1 );
            connections[index]->accumulateNegStats( disc_neg_down_val,
                                                    disc_neg_up_val );
            connections[index]->update();
            
            layers[index+1]->accumulatePosStats( disc_pos_up_val1 );
            layers[index+1]->accumulateNegStats( disc_neg_up_val );
            layers[index+1]->update();
        }
//        else
//        {
//            lr = cd_learning_rate/(1 + cd_decrease_ct 
//                                   * this_stage); 
//            layers[index]->setLearningRate( lr );
//            connections[index]->setLearningRate( lr );
//            layers[index+1]->setLearningRate( lr );
//            
//            layers[index]->accumulatePosStats( pos_down_val );
//            layers[index]->accumulateNegStats( disc_neg_down_val );
//            layers[index]->update();
//            
//            connections[index]->accumulatePosStats( pos_down_val,
//                                                    pos_up_val );
//            connections[index]->accumulateNegStats( disc_neg_down_val,
//                                                    disc_neg_up_val );
//            connections[index]->update();
//            
//            layers[index+1]->accumulatePosStats( pos_up_val );
//            layers[index+1]->accumulateNegStats( disc_neg_up_val );
//            layers[index+1]->update();
//        }
    }
}

Here is the call graph for this function:

Here is the caller graph for this function:

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

Transforms a shallow copy into a deep copy.

Reimplemented from PLearn::PLearner.

Definition at line 477 of file DiscriminativeDeepBeliefNet.cc.

References activation_gradients, activations, connections, PLearn::deepCopyField(), disc_neg_down_val, disc_neg_up_val, disc_pos_down_val1, disc_pos_down_val2, disc_pos_up_val1, disc_pos_up_val2, dissimilar_example_representation, expectation_gradients, expectations, final_cost, final_cost_gradient, final_cost_input, final_cost_value, final_module, greedy_connections, greedy_layers, greedy_stages, input_representation, layers, PLearn::PLearner::makeDeepCopyFromShallowCopy(), nearest_neighbors_indices, neg_down_val, neg_up_val, other_class_datasets, pos_down_val, pos_up_val, training_schedule, unsupervised_connections, and unsupervised_layers.

Here is the call graph for this function:

int PLearn::DiscriminativeDeepBeliefNet::outputsize ( ) const [virtual]

Returns the size of this learner's output, (which typically may depend on its inputsize(), targetsize() and set options).

Implements PLearn::PLearner.

Definition at line 520 of file DiscriminativeDeepBeliefNet.cc.

References currently_trained_layer, layers, n_layers, and PLearn::TVec< T >::size().

{
    if( currently_trained_layer>n_layers-1 )
        return 1;
    else
        return layers[currently_trained_layer]->size;
}

Here is the call graph for this function:

void PLearn::DiscriminativeDeepBeliefNet::setLearningRate ( real  the_learning_rate) [private]

Definition at line 1060 of file DiscriminativeDeepBeliefNet.cc.

References final_cost, final_module, greedy_connections, greedy_layers, i, layers, and n_layers.

Referenced by train().

{
    layers[0]->setLearningRate( the_learning_rate );
    for( int i=0 ; i<n_layers-1 ; i++ )
    {
        greedy_layers[i]->setLearningRate( the_learning_rate );
        greedy_connections[i]->setLearningRate( the_learning_rate );
    }

    final_module->setLearningRate( the_learning_rate );
    final_cost->setLearningRate( the_learning_rate );
}

Here is the caller graph for this function:

void PLearn::DiscriminativeDeepBeliefNet::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 1009 of file DiscriminativeDeepBeliefNet.cc.

References nearest_neighbors_are_up_to_date, and PLearn::PLearner::setTrainingSet().

{
    inherited::setTrainingSet(training_set,call_forget);
    nearest_neighbors_are_up_to_date = false;
}

Here is the call graph for this function:

void PLearn::DiscriminativeDeepBeliefNet::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.

Implements PLearn::PLearner.

Definition at line 562 of file DiscriminativeDeepBeliefNet.cc.

References classname(), currently_trained_layer, PLearn::TVec< T >::data(), disc_neg_down_val, disc_neg_up_val, disc_pos_down_val1, disc_pos_down_val2, disc_pos_up_val1, disc_pos_up_val2, dissimilar_example_representation, PLearn::endl(), PLearn::fast_exact_is_equal(), PLearn::TVec< T >::fill(), final_cost_gradient, final_cost_input, final_cost_value, fine_tuning_decrease_ct, fine_tuning_learning_rate, fineTuningStep(), PLearn::VMat::getExample(), getTrainCostNames(), greedy_layers, greedy_stages, greedyStep(), i, input_representation, PLearn::PLearner::inputsize(), k_neighbors, layers, PLearn::VMat::length(), PLearn::TVec< T >::length(), MISSING_VALUE, n_classes, n_layers, nearest_neighbors_indices, neg_down_val, neg_up_val, PLearn::PLearner::nstages, other_class_datasets, PLERROR, pos_down_val, pos_up_val, PLearn::PLearner::random_gen, PLearn::PLearner::report_progress, PLearn::TVec< T >::resize(), PLearn::sample(), setLearningRate(), PLearn::PLearner::stage, PLearn::TVec< T >::subVec(), PLearn::PLearner::targetsize(), PLearn::tostring(), PLearn::PLearner::train_set, PLearn::PLearner::train_stats, training_schedule, and updateNearestNeighbors().

{
    MODULE_LOG << "train() called " << endl;
    MODULE_LOG << "  training_schedule = " << training_schedule << endl;

    Vec input( inputsize() );
    Vec dissimilar_example( inputsize() );
    Vec target( targetsize() );
    Vec target2( targetsize() );
    real weight; // unused
    real weight2; // unused
    
    TVec<string> train_cost_names = getTrainCostNames() ;
    Vec train_costs( train_cost_names.length() );
    train_costs.fill(MISSING_VALUE) ;

    int nsamples = train_set->length();
    int sample;

    PP<ProgressBar> pb;

    // clear stats of previous epoch
    train_stats->forget();

    int init_stage;

    /***** initial greedy training *****/
    for( int i=0 ; i<n_layers-1 ; i++ )
    {
        updateNearestNeighbors();
            
        MODULE_LOG << "Training connection weights between layers " << i
            << " and " << i+1 << endl;

        int end_stage = training_schedule[i];
        int* this_stage = greedy_stages.subVec(i,1).data();
        init_stage = *this_stage;

        MODULE_LOG << "  stage = " << *this_stage << endl;
        MODULE_LOG << "  end_stage = " << end_stage << endl;

        if( report_progress && *this_stage < end_stage )
            pb = new ProgressBar( "Training layer "+tostring(i)
                                  +" of "+classname(),
                                  end_stage - init_stage );

        train_costs.fill(MISSING_VALUE);
 
        dissimilar_example_representation.resize(layers[i]->size);
        input_representation.resize(layers[i]->size);

        pos_down_val.resize(layers[i]->size);
        pos_up_val.resize(greedy_layers[i]->size);
        neg_down_val.resize(layers[i]->size);
        neg_up_val.resize(greedy_layers[i]->size);

        disc_pos_down_val1.resize(layers[i]->size);
        disc_pos_up_val1.resize(layers[i+1]->size);
        disc_pos_down_val2.resize(layers[i]->size);
        disc_pos_up_val2.resize(layers[i+1]->size);
        disc_neg_down_val.resize(layers[i]->size);
        disc_neg_up_val.resize(layers[i+1]->size);

        for( ; *this_stage<end_stage ; (*this_stage)++ )
        {
            sample = *this_stage % nsamples;
            train_set->getExample(sample, input, target, weight);

            // Find dissimilar example
            int dissim_index = nearest_neighbors_indices(
                sample,random_gen->uniform_multinomial_sample(k_neighbors));
            
            other_class_datasets[(int)round(target[0])]->getExample(dissim_index,
                                                                    dissimilar_example, 
                                                                    target2, weight2);
            
            if(((int)round(target[0])) == ((int)round(target2[0])))
                PLERROR("DiscriminativeDeepBeliefNet::train(): dissimilar"
                        " example is from same class!");

            greedyStep( input, target, i, train_costs, *this_stage,
                        dissimilar_example);
            train_stats->update( train_costs );

            if( pb )
                pb->update( *this_stage - init_stage + 1 );
        }
    }

    /***** fine-tuning by gradient descent *****/
    if( stage < nstages )
    {

        MODULE_LOG << "Fine-tuning all parameters, by gradient descent" << endl;
        MODULE_LOG << "  stage = " << stage << endl;
        MODULE_LOG << "  nstages = " << nstages << endl;
        MODULE_LOG << "  fine_tuning_learning_rate = " << 
            fine_tuning_learning_rate << endl;

        init_stage = stage;
        if( report_progress && stage < nstages )
            pb = new ProgressBar( "Fine-tuning parameters of all layers of "
                                  + classname(),
                                  nstages - init_stage );

        setLearningRate( fine_tuning_learning_rate );
        train_costs.fill(MISSING_VALUE);

        final_cost_input.resize(n_classes);
        final_cost_value.resize(2); // Should be resized anyways
        final_cost_gradient.resize(n_classes);

        for( ; stage<nstages ; stage++ )
        {
            sample = stage % nsamples;
            if( !fast_exact_is_equal( fine_tuning_decrease_ct, 0. ) )
                setLearningRate( fine_tuning_learning_rate
                                 / (1. + fine_tuning_decrease_ct * stage ) );

            train_set->getExample( sample, input, target, weight );

            fineTuningStep( input, target, train_costs );
            train_stats->update( train_costs );

            if( pb )
                pb->update( stage - init_stage + 1 );
        }

    }
    
    train_stats->finalize();
    MODULE_LOG << "  train costs = " << train_stats->getMean() << endl;


    // Update currently_trained_layer
    if(stage > 0)
        currently_trained_layer = n_layers;
    else
    {            
        currently_trained_layer = n_layers-1;
        while(currently_trained_layer>1 
              && greedy_stages[currently_trained_layer-1] <= 0)
            currently_trained_layer--;
    }
}

Here is the call graph for this function:

void PLearn::DiscriminativeDeepBeliefNet::updateNearestNeighbors ( )

Definition at line 1015 of file DiscriminativeDeepBeliefNet.cc.

References PLearn::TVec< T >::append(), PLearn::computeNearestNeighbors(), PLearn::endl(), PLearn::VMat::getExample(), i, PLearn::PLearner::inputsize(), k_neighbors, PLearn::VMat::length(), n_classes, nearest_neighbors_are_up_to_date, nearest_neighbors_indices, other_class_datasets, PLearn::TMat< T >::resize(), PLearn::TVec< T >::resize(), PLearn::PLearner::targetsize(), and PLearn::PLearner::train_set.

Referenced by train().

{
    if( !nearest_neighbors_are_up_to_date )
    {
        MODULE_LOG << "Computing nearest neighbors" << endl;

        Vec input( inputsize() );
        Vec target( targetsize() );
        real weight; // unused
        
        other_class_datasets.resize(n_classes);
        for(int k=0; k<n_classes; k++)
        {
            other_class_datasets[k] = new ClassSubsetVMatrix();
            other_class_datasets[k]->classes.resize(0);
            for(int l=0; l<n_classes; l++)
                if( l != k )
                    other_class_datasets[k]->classes.append(l);
            other_class_datasets[k]->source = train_set;
            other_class_datasets[k]->build();
        }
        
        
        // Find training nearest neighbors
        input.resize(train_set->inputsize());
        target.resize(train_set->targetsize());
        nearest_neighbors_indices.resize(train_set->length(), k_neighbors);
        TVec<int> nearest_neighbors_indices_row;
        for(int i=0; i<train_set.length(); i++)
        {
            train_set->getExample(i,input,target,weight);
            nearest_neighbors_indices_row = nearest_neighbors_indices(i);
            computeNearestNeighbors(
                new GetInputVMatrix((VMatrix *)
                                    other_class_datasets[(int)round(target[0])]),
                input,
                nearest_neighbors_indices_row,
                -1);
        }
    }
    
    nearest_neighbors_are_up_to_date = true;
}

Here is the call graph for this function:

Here is the caller graph for this function:


Member Data Documentation

Reimplemented from PLearn::PLearner.

Definition at line 197 of file DiscriminativeDeepBeliefNet.h.

Stores the gradient of the cost wrt the activations of the input and hidden layers (at the input of the layers)

Definition at line 223 of file DiscriminativeDeepBeliefNet.h.

Referenced by build_layers_and_connections(), fineTuningStep(), and makeDeepCopyFromShallowCopy().

Stores the activations of the input and hidden layers (at the input of the layers)

Definition at line 214 of file DiscriminativeDeepBeliefNet.h.

Referenced by build_layers_and_connections(), computeRepresentation(), fineTuningStep(), and makeDeepCopyFromShallowCopy().

Contrastive divergence decrease constant.

Definition at line 75 of file DiscriminativeDeepBeliefNet.h.

Referenced by declareOptions(), and greedyStep().

Contrastive divergence learning rate.

Definition at line 72 of file DiscriminativeDeepBeliefNet.h.

Referenced by declareOptions(), and greedyStep().

Indication that the discriminative criteria should use the joint over the input and the hidden units, instead of the conditional over the hidden units given the input units.

Definition at line 119 of file DiscriminativeDeepBeliefNet.h.

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

The weights of the connections between the layers.

Definition at line 92 of file DiscriminativeDeepBeliefNet.h.

Referenced by build_layers_and_connections(), computeRepresentation(), declareOptions(), fineTuningStep(), forget(), greedyStep(), and makeDeepCopyFromShallowCopy().

Currently trained layer (1 means the first hidden layer, n_layers means the output layer)

Definition at line 282 of file DiscriminativeDeepBeliefNet.h.

Referenced by build_(), computeCostsFromOutputs(), computeOutput(), outputsize(), and train().

Negative down statistic.

Definition at line 260 of file DiscriminativeDeepBeliefNet.h.

Referenced by greedyStep(), makeDeepCopyFromShallowCopy(), and train().

Negative up statistic.

Definition at line 262 of file DiscriminativeDeepBeliefNet.h.

Referenced by greedyStep(), makeDeepCopyFromShallowCopy(), and train().

First discriminative positive down statistic.

Definition at line 252 of file DiscriminativeDeepBeliefNet.h.

Referenced by greedyStep(), makeDeepCopyFromShallowCopy(), and train().

Second discriminative positive down statistic.

Definition at line 256 of file DiscriminativeDeepBeliefNet.h.

Referenced by greedyStep(), makeDeepCopyFromShallowCopy(), and train().

First discriminative positive up statistic.

Definition at line 254 of file DiscriminativeDeepBeliefNet.h.

Referenced by greedyStep(), makeDeepCopyFromShallowCopy(), and train().

Second discriminative positive up statistic.

Definition at line 258 of file DiscriminativeDeepBeliefNet.h.

Referenced by greedyStep(), makeDeepCopyFromShallowCopy(), and train().

Weight of the discriminative criteria.

Definition at line 108 of file DiscriminativeDeepBeliefNet.h.

Referenced by declareOptions(), and greedyStep().

Dissimilar example representation.

Definition at line 237 of file DiscriminativeDeepBeliefNet.h.

Referenced by greedyStep(), makeDeepCopyFromShallowCopy(), and train().

Indication that the generative criteria should not be used during learning (does not work with compare_joint_in_discriminative_criteria = true).

Definition at line 123 of file DiscriminativeDeepBeliefNet.h.

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

Stores the gradient of the cost wrt the expectations of the input and hidden layers (at the output of the layers)

Definition at line 228 of file DiscriminativeDeepBeliefNet.h.

Referenced by build_layers_and_connections(), fineTuningStep(), and makeDeepCopyFromShallowCopy().

Stores the expectations of the input and hidden layers (at the output of the layers)

Definition at line 218 of file DiscriminativeDeepBeliefNet.h.

Referenced by build_layers_and_connections(), computeRepresentation(), fineTuningStep(), and makeDeepCopyFromShallowCopy().

Cost gradient on output layer.

Definition at line 269 of file DiscriminativeDeepBeliefNet.h.

Referenced by fineTuningStep(), makeDeepCopyFromShallowCopy(), and train().

Input of cost function.

Definition at line 265 of file DiscriminativeDeepBeliefNet.h.

Referenced by computeOutput(), fineTuningStep(), makeDeepCopyFromShallowCopy(), and train().

Cost value.

Definition at line 267 of file DiscriminativeDeepBeliefNet.h.

Referenced by fineTuningStep(), makeDeepCopyFromShallowCopy(), and train().

The decrease constant of the learning rate used during fine tuning gradient descent.

Definition at line 82 of file DiscriminativeDeepBeliefNet.h.

Referenced by declareOptions(), and train().

The learning rate used during the fine tuning gradient descent.

Definition at line 78 of file DiscriminativeDeepBeliefNet.h.

Referenced by declareOptions(), and train().

Connections used for greedy learning.

Definition at line 234 of file DiscriminativeDeepBeliefNet.h.

Referenced by build_layers_and_connections(), greedyStep(), makeDeepCopyFromShallowCopy(), and setLearningRate().

Layers used for greedy learning.

Definition at line 231 of file DiscriminativeDeepBeliefNet.h.

Referenced by build_layers_and_connections(), greedyStep(), makeDeepCopyFromShallowCopy(), setLearningRate(), and train().

Stages of the different greedy phases.

Definition at line 278 of file DiscriminativeDeepBeliefNet.h.

Referenced by build_(), declareOptions(), forget(), makeDeepCopyFromShallowCopy(), and train().

Example representation.

Definition at line 240 of file DiscriminativeDeepBeliefNet.h.

Referenced by computeOutput(), fineTuningStep(), greedyStep(), makeDeepCopyFromShallowCopy(), and train().

Number of good nearest neighbors to attract and bad nearest neighbors to repel.

Definition at line 102 of file DiscriminativeDeepBeliefNet.h.

Referenced by build_(), declareOptions(), train(), and updateNearestNeighbors().

Number of classes.

Definition at line 105 of file DiscriminativeDeepBeliefNet.h.

Referenced by build_(), build_output_layer_and_cost(), declareOptions(), train(), and updateNearestNeighbors().

Indication that nearest_neighbors_indices is up to date.

Definition at line 210 of file DiscriminativeDeepBeliefNet.h.

Referenced by setTrainingSet(), and updateNearestNeighbors().

Nearest neighbors for each training example.

Definition at line 275 of file DiscriminativeDeepBeliefNet.h.

Referenced by makeDeepCopyFromShallowCopy(), train(), and updateNearestNeighbors().

Negative down statistic.

Definition at line 247 of file DiscriminativeDeepBeliefNet.h.

Referenced by greedyStep(), makeDeepCopyFromShallowCopy(), and train().

Negative up statistic.

Definition at line 249 of file DiscriminativeDeepBeliefNet.h.

Referenced by greedyStep(), makeDeepCopyFromShallowCopy(), and train().

Datasets for each class.

Definition at line 272 of file DiscriminativeDeepBeliefNet.h.

Referenced by makeDeepCopyFromShallowCopy(), train(), and updateNearestNeighbors().

Output weights l1_penalty_factor.

Definition at line 111 of file DiscriminativeDeepBeliefNet.h.

Referenced by build_output_layer_and_cost(), and declareOptions().

Output weights l2_penalty_factor.

Definition at line 114 of file DiscriminativeDeepBeliefNet.h.

Referenced by build_output_layer_and_cost(), and declareOptions().

Positive down statistic.

Definition at line 243 of file DiscriminativeDeepBeliefNet.h.

Referenced by greedyStep(), makeDeepCopyFromShallowCopy(), and train().

Positive up statistic.

Definition at line 245 of file DiscriminativeDeepBeliefNet.h.

Referenced by greedyStep(), makeDeepCopyFromShallowCopy(), and train().

Number of examples to use during each phase of greedy pre-training.

The number of fine-tunig steps is defined by nstages.

Definition at line 86 of file DiscriminativeDeepBeliefNet.h.

Referenced by build_(), declareOptions(), makeDeepCopyFromShallowCopy(), and train().

Additional connections for greedy unsupervised learning.

Definition at line 98 of file DiscriminativeDeepBeliefNet.h.

Referenced by build_layers_and_connections(), declareOptions(), forget(), and makeDeepCopyFromShallowCopy().

Additional units for greedy unsupervised learning.

Definition at line 95 of file DiscriminativeDeepBeliefNet.h.

Referenced by build_layers_and_connections(), declareOptions(), forget(), and makeDeepCopyFromShallowCopy().


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