|
PLearn 0.1
|
Computes the NLL, given a probability vector and the true class. More...
#include <NLLCostModule.h>


Public Member Functions | |
| NLLCostModule () | |
| Default constructor. | |
| virtual void | fprop (const Vec &input, const Vec &target, Vec &cost) const |
| given the input and target, compute the cost | |
| virtual void | fprop (const Mat &inputs, const Mat &targets, Mat &costs) const |
| batch version | |
| virtual void | fprop (const TVec< Mat * > &ports_value) |
| new version | |
| virtual void | bpropUpdate (const Vec &input, const Vec &target, real cost, Vec &input_gradient, bool accumulate=false) |
| Adapt based on the output gradient: this method should only be called just after a corresponding fprop. | |
| virtual void | bpropUpdate (const Mat &inputs, const Mat &targets, const Vec &costs, Mat &input_gradients, bool accumulate=false) |
| Overridden. | |
| virtual void | bpropUpdate (const Vec &input, const Vec &target, real cost) |
| Does nothing. | |
| virtual void | bpropAccUpdate (const TVec< Mat * > &ports_value, const TVec< Mat * > &ports_gradient) |
| New version of backpropagation. | |
| virtual void | bbpropUpdate (const Vec &input, const Vec &target, real cost, Vec &input_gradient, Vec &input_diag_hessian, bool accumulate=false) |
| Similar to bpropUpdate, but adapt based also on the estimation of the diagonal of the Hessian matrix, and propagates this back. | |
| virtual void | bbpropUpdate (const Vec &input, const Vec &target, real cost) |
| Does nothing. | |
| virtual void | setLearningRate (real dynamic_learning_rate) |
| Overridden to do nothing (in particular, no warning). | |
| virtual TVec< string > | costNames () |
| Indicates the name of the computed costs. | |
| virtual string | classname () const |
| virtual OptionList & | getOptionList () const |
| virtual OptionMap & | getOptionMap () const |
| virtual RemoteMethodMap & | getRemoteMethodMap () const |
| virtual NLLCostModule * | deepCopy (CopiesMap &copies) const |
| virtual void | build () |
| Post-constructor. | |
| 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 PPath & | declaringFile () |
Static Public Attributes | |
| static StaticInitializer | _static_initializer_ |
Static Protected Member Functions | |
| static void | declareOptions (OptionList &ol) |
| Declares the class options. | |
Private Types | |
| typedef CostModule | inherited |
Private Member Functions | |
| void | build_ () |
| This does the actual building. | |
Computes the NLL, given a probability vector and the true class.
If input is the probability vector, and target the index of the true class, this module computes cost = -log( input[target] ), and back-propagates the gradient and diagonal of Hessian.
Definition at line 53 of file NLLCostModule.h.
typedef CostModule PLearn::NLLCostModule::inherited [private] |
Reimplemented from PLearn::CostModule.
Definition at line 55 of file NLLCostModule.h.
| PLearn::NLLCostModule::NLLCostModule | ( | ) |
Default constructor.
Definition at line 53 of file NLLCostModule.cc.
{
output_size = 1;
target_size = 1;
}
| string PLearn::NLLCostModule::_classname_ | ( | ) | [static] |
Reimplemented from PLearn::CostModule.
Definition at line 51 of file NLLCostModule.cc.
| OptionList & PLearn::NLLCostModule::_getOptionList_ | ( | ) | [static] |
Reimplemented from PLearn::CostModule.
Definition at line 51 of file NLLCostModule.cc.
| RemoteMethodMap & PLearn::NLLCostModule::_getRemoteMethodMap_ | ( | ) | [static] |
Reimplemented from PLearn::CostModule.
Definition at line 51 of file NLLCostModule.cc.
Reimplemented from PLearn::CostModule.
Definition at line 51 of file NLLCostModule.cc.
| Object * PLearn::NLLCostModule::_new_instance_for_typemap_ | ( | ) | [static] |
Reimplemented from PLearn::CostModule.
Definition at line 51 of file NLLCostModule.cc.
| StaticInitializer NLLCostModule::_static_initializer_ & PLearn::NLLCostModule::_static_initialize_ | ( | ) | [static] |
Reimplemented from PLearn::CostModule.
Definition at line 51 of file NLLCostModule.cc.
| virtual void PLearn::NLLCostModule::bbpropUpdate | ( | const Vec & | input, |
| const Vec & | target, | ||
| real | cost | ||
| ) | [inline, virtual] |
Does nothing.
Reimplemented from PLearn::CostModule.
Definition at line 103 of file NLLCostModule.h.
{}
| void PLearn::NLLCostModule::bbpropUpdate | ( | const Vec & | input, |
| const Vec & | target, | ||
| real | cost, | ||
| Vec & | input_gradient, | ||
| Vec & | input_diag_hessian, | ||
| bool | accumulate = false |
||
| ) | [virtual] |
Similar to bpropUpdate, but adapt based also on the estimation of the diagonal of the Hessian matrix, and propagates this back.
Reimplemented from PLearn::CostModule.
Definition at line 297 of file NLLCostModule.cc.
References PLearn::TVec< T >::clear(), PLASSERT_MSG, PLearn::TVec< T >::resize(), and PLearn::TVec< T >::size().
{
if( accumulate )
{
PLASSERT_MSG( input_diag_hessian.size() == input_size,
"Cannot resize input_diag_hessian AND accumulate into it"
);
}
else
{
input_diag_hessian.resize( input_size );
input_diag_hessian.clear();
}
// input_diag_hessian[ i ] = 0 if i!=t
// input_diag_hessian[ t ] = 1/(x[t])^2
int the_target = (int) round( target[0] );
real input_t = input[ the_target ];
input_diag_hessian[ the_target ] += 1. / (input_t * input_t);
bpropUpdate( input, target, cost, input_gradient, accumulate );
}

| void PLearn::NLLCostModule::bpropAccUpdate | ( | const TVec< Mat * > & | ports_value, |
| const TVec< Mat * > & | ports_gradient | ||
| ) | [virtual] |
New version of backpropagation.
Reimplemented from PLearn::CostModule.
Definition at line 236 of file NLLCostModule.cc.
References PLearn::TMat< T >::isEmpty(), PLearn::TMat< T >::length(), PLearn::TVec< T >::length(), PLASSERT, PLERROR, PLearn::TMat< T >::resize(), and PLearn::TMat< T >::width().
{
PLASSERT( ports_value.length() == nPorts() );
PLASSERT( ports_gradient.length() == nPorts() );
Mat* prediction = ports_value[0];
Mat* target = ports_value[1];
#ifndef NDEBUG
Mat* cost = ports_value[2];
#endif
Mat* prediction_grad = ports_gradient[0];
Mat* target_grad = ports_gradient[1];
Mat* cost_grad = ports_gradient[2];
// If we have cost_grad and we want prediction_grad
if( prediction_grad && prediction_grad->isEmpty()
&& cost_grad && !cost_grad->isEmpty() )
{
PLASSERT( prediction );
PLASSERT( target );
PLASSERT( cost );
PLASSERT( !target_grad );
PLASSERT( prediction->width() == port_sizes(0,1) );
PLASSERT( target->width() == port_sizes(1,1) );
PLASSERT( cost->width() == port_sizes(2,1) );
PLASSERT( prediction_grad->width() == port_sizes(0,1) );
PLASSERT( cost_grad->width() == port_sizes(2,1) );
PLASSERT( cost_grad->width() == 1 );
int batch_size = prediction->length();
PLASSERT( target->length() == batch_size );
PLASSERT( cost->length() == batch_size );
PLASSERT( cost_grad->length() == batch_size );
prediction_grad->resize(batch_size, port_sizes(0,1));
for( int k=0; k<batch_size; k++ )
{
// input_gradient[ i ] = 0 if i != t,
// input_gradient[ t ] = -1/x[t]
int target_k = (int) round((*target)(k, 0));
(*prediction_grad)(k, target_k) -=
(*cost_grad)(k, 0) / (*prediction)(k, target_k);
}
}
else if( !prediction_grad && !target_grad && !cost_grad )
return;
else if( !cost_grad && prediction_grad && prediction_grad->isEmpty() )
PLERROR("In NLLCostModule::bpropAccUpdate - cost gradient is NULL,\n"
"cannot compute prediction gradient. Maybe you should set\n"
"\"propagate_gradient = 0\" on the incoming connection.\n");
else
PLERROR("In OnlineLearningModule::bpropAccUpdate - Port configuration "
"not implemented for class '%s'", classname().c_str());
checkProp(ports_value);
checkProp(ports_gradient);
}

| void PLearn::NLLCostModule::bpropUpdate | ( | const Mat & | inputs, |
| const Mat & | targets, | ||
| const Vec & | costs, | ||
| Mat & | input_gradients, | ||
| bool | accumulate = false |
||
| ) | [virtual] |
Overridden.
Reimplemented from PLearn::CostModule.
Definition at line 210 of file NLLCostModule.cc.
References PLearn::TMat< T >::clear(), i, PLearn::TMat< T >::length(), PLASSERT, PLASSERT_MSG, PLearn::TMat< T >::resize(), and PLearn::TMat< T >::width().
{
PLASSERT( inputs.width() == input_size );
PLASSERT( targets.width() == target_size );
if( accumulate )
{
PLASSERT_MSG( input_gradients.width() == input_size &&
input_gradients.length() == inputs.length(),
"Cannot resize input_gradients and accumulate into it" );
}
else
{
input_gradients.resize(inputs.length(), input_size );
input_gradients.clear();
}
// input_gradient[ i ] = 0 if i != t,
// input_gradient[ t ] = -1/x[t]
for (int i = 0; i < inputs.length(); i++) {
int the_target = (int) round( targets(i, 0) );
input_gradients(i, the_target) -= 1. / inputs(i, the_target);
}
}

| void PLearn::NLLCostModule::bpropUpdate | ( | const Vec & | input, |
| const Vec & | target, | ||
| real | cost, | ||
| Vec & | input_gradient, | ||
| bool | accumulate = false |
||
| ) | [virtual] |
Adapt based on the output gradient: this method should only be called just after a corresponding fprop.
Reimplemented from PLearn::CostModule.
Definition at line 187 of file NLLCostModule.cc.
References PLearn::TVec< T >::clear(), PLASSERT, PLASSERT_MSG, PLearn::TVec< T >::resize(), and PLearn::TVec< T >::size().
{
PLASSERT( input.size() == input_size );
PLASSERT( target.size() == target_size );
if( accumulate )
{
PLASSERT_MSG( input_gradient.size() == input_size,
"Cannot resize input_gradient AND accumulate into it" );
}
else
{
input_gradient.resize( input_size );
input_gradient.clear();
}
int the_target = (int) round( target[0] );
// input_gradient[ i ] = 0 if i != t,
// input_gradient[ t ] = -1/x[t]
input_gradient[ the_target ] -= 1. / input[ the_target ];
}

| virtual void PLearn::NLLCostModule::bpropUpdate | ( | const Vec & | input, |
| const Vec & | target, | ||
| real | cost | ||
| ) | [inline, virtual] |
Does nothing.
Reimplemented from PLearn::CostModule.
Definition at line 89 of file NLLCostModule.h.
{}
| void PLearn::NLLCostModule::build | ( | ) | [virtual] |
Post-constructor.
The normal implementation should call simply inherited::build(), then this class's build_(). This method should be callable again at later times, after modifying some option fields to change the "architecture" of the object.
Reimplemented from PLearn::CostModule.
Definition at line 74 of file NLLCostModule.cc.
Referenced by PLearn::TopDownAsymetricDeepNetwork::build_output_layer_and_cost(), PLearn::StackedFocusedAutoassociatorsNet::build_output_layer_and_cost(), and PLearn::DiscriminativeDeepBeliefNet::build_output_layer_and_cost().
{
inherited::build();
build_();
}

| void PLearn::NLLCostModule::build_ | ( | ) | [private] |
This does the actual building.
Reimplemented from PLearn::CostModule.
Definition at line 69 of file NLLCostModule.cc.
{
}
| string PLearn::NLLCostModule::classname | ( | ) | const [virtual] |
Reimplemented from PLearn::CostModule.
Definition at line 51 of file NLLCostModule.cc.
| TVec< string > PLearn::NLLCostModule::costNames | ( | ) | [virtual] |
Indicates the name of the computed costs.
Reimplemented from PLearn::CostModule.
Definition at line 323 of file NLLCostModule.cc.
| void PLearn::NLLCostModule::declareOptions | ( | OptionList & | ol | ) | [static, protected] |
Declares the class options.
Reimplemented from PLearn::CostModule.
Definition at line 59 of file NLLCostModule.cc.
{
// declareOption(ol, "myoption", &NLLCostModule::myoption,
// OptionBase::buildoption,
// "Help text describing this option");
// Now call the parent class' declareOptions
inherited::declareOptions(ol);
}
| static const PPath& PLearn::NLLCostModule::declaringFile | ( | ) | [inline, static] |
Reimplemented from PLearn::CostModule.
Definition at line 117 of file NLLCostModule.h.
:
//##### Protected Member Functions ######################################
| NLLCostModule * PLearn::NLLCostModule::deepCopy | ( | CopiesMap & | copies | ) | const [virtual] |
Reimplemented from PLearn::CostModule.
Definition at line 51 of file NLLCostModule.cc.
| void PLearn::NLLCostModule::fprop | ( | const Mat & | inputs, |
| const Mat & | targets, | ||
| Mat & | costs | ||
| ) | const [virtual] |
batch version
Reimplemented from PLearn::CostModule.
Definition at line 110 of file NLLCostModule.cc.
References PLearn::TMat< T >::length(), pl_log, PLASSERT, PLearn::TMat< T >::resize(), and PLearn::TMat< T >::width().
{
PLASSERT( inputs.width() == input_size );
PLASSERT( targets.width() == target_size );
int batch_size = inputs.length();
PLASSERT( inputs.length() == batch_size );
PLASSERT( targets.length() == batch_size );
costs.resize(batch_size, output_size);
for( int k=0; k<batch_size; k++ )
{
int target_k = (int) round( targets(k, 0) );
costs(k, 0) = -pl_log( inputs(k, target_k) );
}
}

| void PLearn::NLLCostModule::fprop | ( | const Vec & | input, |
| const Vec & | target, | ||
| Vec & | cost | ||
| ) | const [virtual] |
given the input and target, compute the cost
Reimplemented from PLearn::CostModule.
Definition at line 90 of file NLLCostModule.cc.
References PLearn::TVec< T >::hasMissing(), PLearn::is_equal(), PLearn::min(), MISSING_VALUE, pl_log, PLASSERT, PLASSERT_MSG, PLearn::TVec< T >::resize(), PLearn::TVec< T >::size(), and PLearn::sum().
{
PLASSERT( input.size() == input_size );
PLASSERT( target.size() == target_size );
cost.resize( output_size );
if( input.hasMissing() )
cost[0] = MISSING_VALUE;
else
{
PLASSERT_MSG( min(input) >= 0.,
"Elements of \"input\" should be positive" );
PLASSERT_MSG( is_equal( sum(input), 1. ),
"Elements of \"input\" should sum to 1" );
int the_target = (int) round( target[0] );
cost[0] = -pl_log( input[ the_target ] );
}
}

new version
Reimplemented from PLearn::CostModule.
Definition at line 129 of file NLLCostModule.cc.
References i, PLearn::is_equal(), PLearn::is_missing(), PLearn::TMat< T >::isEmpty(), PLearn::TMat< T >::length(), PLearn::TVec< T >::length(), PLearn::min(), MISSING_VALUE, pl_log, PLASSERT, PLASSERT_MSG, PLCHECK_MSG, PLERROR, PLearn::TMat< T >::resize(), PLearn::sum(), and PLearn::TMat< T >::width().
{
PLASSERT( ports_value.length() == nPorts() );
Mat* prediction = ports_value[0];
Mat* target = ports_value[1];
Mat* cost = ports_value[2];
// If we have prediction and target, and we want cost
if( prediction && !prediction->isEmpty()
&& target && !target->isEmpty()
&& cost && cost->isEmpty() )
{
PLASSERT( prediction->width() == port_sizes(0, 1) );
PLASSERT( target->width() == port_sizes(1, 1) );
int batch_size = prediction->length();
PLASSERT( target->length() == batch_size );
cost->resize(batch_size, port_sizes(2, 1));
for( int i=0; i<batch_size; i++ )
{
if( (*prediction)(i).hasMissing() || is_missing((*target)(i,0)) )
(*cost)(i,0) = MISSING_VALUE;
else
{
#ifdef BOUNDCHECK
PLASSERT_MSG( min((*prediction)(i)) >= 0.,
"Elements of \"prediction\" should be positive" );
// Ensure the distribution probabilities sum to 1. We relax a
// bit the default tolerance as probabilities using
// exponentials could suffer numerical imprecisions.
if (!is_equal( sum((*prediction)(i)), 1., 1., 1e-5, 1e-5 ))
PLERROR("In NLLCostModule::fprop - Elements of"
" \"prediction\" should sum to 1"
" (found a sum = %f at row %d)",
sum((*prediction)(i)), i);
#endif
int target_i = (int) round( (*target)(i,0) );
PLASSERT( is_equal( (*target)(i, 0), target_i ) );
(*cost)(i,0) = -pl_log( (*prediction)(i, target_i) );
}
}
}
else if( !prediction && !target && !cost )
return;
else
PLCHECK_MSG( false, "Unknown port configuration" );
checkProp(ports_value);
}

| OptionList & PLearn::NLLCostModule::getOptionList | ( | ) | const [virtual] |
Reimplemented from PLearn::CostModule.
Definition at line 51 of file NLLCostModule.cc.
| OptionMap & PLearn::NLLCostModule::getOptionMap | ( | ) | const [virtual] |
Reimplemented from PLearn::CostModule.
Definition at line 51 of file NLLCostModule.cc.
| RemoteMethodMap & PLearn::NLLCostModule::getRemoteMethodMap | ( | ) | const [virtual] |
Reimplemented from PLearn::CostModule.
Definition at line 51 of file NLLCostModule.cc.
| void PLearn::NLLCostModule::makeDeepCopyFromShallowCopy | ( | CopiesMap & | copies | ) | [virtual] |
Transforms a shallow copy into a deep copy.
Reimplemented from PLearn::CostModule.
Definition at line 81 of file NLLCostModule.cc.
{
inherited::makeDeepCopyFromShallowCopy(copies);
}
| virtual void PLearn::NLLCostModule::setLearningRate | ( | real | dynamic_learning_rate | ) | [inline, virtual] |
Overridden to do nothing (in particular, no warning).
Reimplemented from PLearn::OnlineLearningModule.
Definition at line 107 of file NLLCostModule.h.
{}
Reimplemented from PLearn::CostModule.
Definition at line 117 of file NLLCostModule.h.
1.7.4