PLearn 0.1
|
NLL (and derivatives thereof) between the target and input. More...
#include <NLLErrModule.h>
Public Member Functions | |
NLLErrModule () | |
Default constructor. | |
virtual int | getTarget (const Vec &input) const |
virtual void | fprop (const Vec &input, Vec &output) const |
given the input, compute the output (possibly resize it appropriately) SOON TO BE DEPRECATED, USE fprop(const TVec<Mat*>& ports_value) | |
virtual void | bpropUpdate (const Vec &input, const Vec &output, const Vec &output_gradient) |
SOON TO BE DEPRECATED, USE bpropAccUpdate(const TVec<Mat*>& ports_value, const TVec<Mat*>& ports_gradient) Adapt based on the output gradient: this method should only be called just after a corresponding fprop; it should be called with the same arguments as fprop for the first two arguments (and output should not have been modified since then). | |
virtual void | bpropUpdate (const Vec &input, const Vec &output, Vec &input_gradient, const Vec &output_gradient) |
virtual void | bbpropUpdate (const Vec &input, const Vec &output, const Vec &output_gradient, const Vec &output_diag_hessian) |
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 &output, Vec &input_gradient, const Vec &output_gradient, Vec &input_diag_hessian, const Vec &output_diag_hessian) |
virtual void | forget () |
reset the parameters to the state they would be BEFORE starting training. | |
virtual string | classname () const |
virtual OptionList & | getOptionList () const |
virtual OptionMap & | getOptionMap () const |
virtual RemoteMethodMap & | getRemoteMethodMap () const |
virtual NLLErrModule * | 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. | |
Protected Attributes | |
int | target_size |
Private Types | |
typedef OnlineLearningModule | inherited |
Private Member Functions | |
void | build_ () |
This does the actual building. | |
Private Attributes | |
Vec | fp_sm |
NLL (and derivatives thereof) between the target and input.
This class computes the Negative Log-Likelihood of the input, given the desired 'target'. Also propagates gradient and diagonal of Hessian backwards.
Definition at line 64 of file NLLErrModule.h.
typedef OnlineLearningModule PLearn::NLLErrModule::inherited [private] |
Reimplemented from PLearn::OnlineLearningModule.
Definition at line 66 of file NLLErrModule.h.
PLearn::NLLErrModule::NLLErrModule | ( | ) |
Default constructor.
Definition at line 58 of file NLLErrModule.cc.
References PLearn::OnlineLearningModule::output_size.
: target_size(1) /* ### Initialize all fields to their default value */ { output_size = 1; }
string PLearn::NLLErrModule::_classname_ | ( | ) | [static] |
Reimplemented from PLearn::OnlineLearningModule.
Definition at line 56 of file NLLErrModule.cc.
OptionList & PLearn::NLLErrModule::_getOptionList_ | ( | ) | [static] |
Reimplemented from PLearn::OnlineLearningModule.
Definition at line 56 of file NLLErrModule.cc.
RemoteMethodMap & PLearn::NLLErrModule::_getRemoteMethodMap_ | ( | ) | [static] |
Reimplemented from PLearn::OnlineLearningModule.
Definition at line 56 of file NLLErrModule.cc.
Reimplemented from PLearn::OnlineLearningModule.
Definition at line 56 of file NLLErrModule.cc.
Object * PLearn::NLLErrModule::_new_instance_for_typemap_ | ( | ) | [static] |
Reimplemented from PLearn::Object.
Definition at line 56 of file NLLErrModule.cc.
StaticInitializer NLLErrModule::_static_initializer_ & PLearn::NLLErrModule::_static_initialize_ | ( | ) | [static] |
Reimplemented from PLearn::OnlineLearningModule.
Definition at line 56 of file NLLErrModule.cc.
void PLearn::NLLErrModule::bbpropUpdate | ( | const Vec & | input, |
const Vec & | output, | ||
const Vec & | output_gradient, | ||
const Vec & | output_diag_hessian | ||
) | [virtual] |
Similar to bpropUpdate, but adapt based also on the estimation of the diagonal of the Hessian matrix, and propagates this back.
If these methods are defined, you can use them INSTEAD of bpropUpdate(...) THE DEFAULT IMPLEMENTATION PROVIDED HERE JUST CALLS bbpropUpdate(input, output, input_gradient, output_gradient, in_hess, out_hess) AND IGNORES INPUT HESSIAN AND INPUT GRADIENT
Reimplemented from PLearn::OnlineLearningModule.
Definition at line 198 of file NLLErrModule.cc.
References bpropUpdate(), PLearn::OnlineLearningModule::output_size, PLWARNING, and PLearn::TVec< T >::size().
{ int odh_size = output_diag_hessian.size(); if( odh_size != output_size ) { PLWARNING("NLLErrModule::bbpropUpdate:" " 'output_diag_hessian.size()'\n" " should be equal to 'output_size' (%i != %i)\n", odh_size, output_size); } bpropUpdate( input, output, output_gradient ); }
void PLearn::NLLErrModule::bbpropUpdate | ( | const Vec & | input, |
const Vec & | output, | ||
Vec & | input_gradient, | ||
const Vec & | output_gradient, | ||
Vec & | input_diag_hessian, | ||
const Vec & | output_diag_hessian | ||
) | [virtual] |
Definition at line 215 of file NLLErrModule.cc.
References bpropUpdate(), PLearn::TVec< T >::clear(), PLearn::OnlineLearningModule::estimate_simpler_diag_hessian, getTarget(), i, PLearn::OnlineLearningModule::input_size, PLearn::OnlineLearningModule::output_size, PLERROR, PLWARNING, PLearn::TVec< T >::resize(), PLearn::TVec< T >::size(), PLearn::softmax(), and PLearn::TVec< T >::subVec().
{ int odh_size = output_diag_hessian.size(); int target = getTarget( input ); bool is_final_cost = false; // if yes, output_diag_hessian is 0 // size check // others size checks will be done in bpropUpdate() if( odh_size == 0 ) { PLWARNING("NLLErrModule::bbpropUpdate: you are not providing" " output_diag_hessian.\n" "Assuming this is the final cost," " and output_diag_hessian=0.\n"); is_final_cost = true; } else if( odh_size != output_size ) { PLERROR("NLLErrModule::bbpropUpdate:" " 'output_diag_hessian.size()'\n" " should be equal to 'output_size' (%i != %i)\n", odh_size, output_size); } bpropUpdate( input, output, input_gradient, output_gradient ); Vec input_ = input.subVec( 0, input_size ); input_diag_hessian.resize( input_size ); Vec softmax_in = softmax( input_ ); // computation of term dC/dy d²y/dx², // skipped if estimate_simpler_diag_hessian, unless it is final cost if( estimate_simpler_diag_hessian && !is_final_cost ) { input_diag_hessian.clear(); } else { for( int i=0 ; i<input_size ; i++ ) { real sm_i = softmax_in[i]; input_diag_hessian[i] = sm_i*( 1-sm_i); } if( !is_final_cost ) input_diag_hessian *= output_gradient[0]; } // computation of term d²C/dy² (dy/dx)², // skipped if it is final cost, because then d²C/dy² == d²C/dC² == 0 if( !is_final_cost ) { Vec fprime = softmax_in; fprime[target] -= 1; fprime *= fprime; input_diag_hessian += output_diag_hessian[0] * fprime; } }
void PLearn::NLLErrModule::bpropUpdate | ( | const Vec & | input, |
const Vec & | output, | ||
const Vec & | output_gradient | ||
) | [virtual] |
SOON TO BE DEPRECATED, USE bpropAccUpdate(const TVec<Mat*>& ports_value, const TVec<Mat*>& ports_gradient) Adapt based on the output gradient: this method should only be called just after a corresponding fprop; it should be called with the same arguments as fprop for the first two arguments (and output should not have been modified since then).
Since sub-classes are supposed to learn ONLINE, the object is 'ready-to-be-used' just after any bpropUpdate. N.B. The DEFAULT IMPLEMENTATION JUST CALLS bpropUpdate(input, output, input_gradient, output_gradient) AND IGNORES INPUT GRADIENT.
Reimplemented from PLearn::OnlineLearningModule.
Definition at line 128 of file NLLErrModule.cc.
References getTarget(), PLearn::OnlineLearningModule::output_size, PLWARNING, and PLearn::TVec< T >::size().
Referenced by bbpropUpdate().
{ int out_size = output.size(); int og_size = output_gradient.size(); // for size check getTarget( input ); // size check if( out_size != output_size ) { PLWARNING("NLLErrModule::bpropUpdate: output.size()' should be\n" " equal to 'output_size' (%i != %i)\n", out_size, output_size); } if( og_size != output_size ) { PLWARNING("NLLErrModule::bpropUpdate: 'output_gradient.size()'\n" " should be equal to 'output_size' (%i != %i)\n", og_size, output_size); } }
void PLearn::NLLErrModule::bpropUpdate | ( | const Vec & | input, |
const Vec & | output, | ||
Vec & | input_gradient, | ||
const Vec & | output_gradient | ||
) | [virtual] |
Definition at line 154 of file NLLErrModule.cc.
References fp_sm, getTarget(), PLearn::OnlineLearningModule::input_size, PLearn::OnlineLearningModule::output_size, PLERROR, PLearn::TVec< T >::resize(), and PLearn::TVec< T >::size().
{ int out_size = output.size(); int og_size = output_gradient.size(); int target = getTarget( input ); bool is_final_cost = false; // if yes, output_gradient is 1 // size check if( out_size != output_size ) { PLERROR("NLLErrModule::bpropUpdate: output.size()' should be\n" " equal to 'output_size' (%i != %i)\n", out_size, output_size); } if( og_size == 0 ) { /* PLWARNING("NLLErrModule::bpropUpdate: you are not providing" "output_gradient.\n" "Assuming this is the final cost, and output_gradient=1.\n"); */ is_final_cost = true; } else if( og_size != output_size ) { PLERROR("NLLErrModule::bpropUpdate: 'output_gradient.size()'\n" " should be equal to 'output_size' (%i != %i)\n", og_size, output_size); } // input_gradient[i] = output_gradient*( softmax(input)[i] ) if i!=target // input_gradient[target] = output_gradient*( softmax(input)[target] ) input_gradient.resize( input_size ); input_gradient << fp_sm; input_gradient[target] -= 1; if( !is_final_cost ) input_gradient *= output_gradient[0]; }
void PLearn::NLLErrModule::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::OnlineLearningModule.
Definition at line 291 of file NLLErrModule.cc.
References PLearn::OnlineLearningModule::build(), and build_().
{ inherited::build(); build_(); }
void PLearn::NLLErrModule::build_ | ( | ) | [private] |
This does the actual building.
Reimplemented from PLearn::OnlineLearningModule.
Definition at line 308 of file NLLErrModule.cc.
References fp_sm, PLearn::OnlineLearningModule::input_size, PLearn::OnlineLearningModule::output_size, PLWARNING, PLearn::TVec< T >::resize(), and target_size.
Referenced by build().
{ if( input_size < 0 ) { PLWARNING("NLLErrModule::build_: 'input_size' is < 0.\n" "You should set it to a positive integer.\n" "Defaulting to '1' (like a sigmoid function ?)\n"); input_size = 1; } if( output_size != 1 && output_size != 2 ) { PLWARNING("NLLErrModule::build_: 'output_size' (%i) should be 1.\n" "Setting 'output_size' to 1.\n", output_size); output_size = 1; } target_size = 1; fp_sm.resize(input_size); }
string PLearn::NLLErrModule::classname | ( | ) | const [virtual] |
Reimplemented from PLearn::Object.
Definition at line 56 of file NLLErrModule.cc.
void PLearn::NLLErrModule::declareOptions | ( | OptionList & | ol | ) | [static, protected] |
Declares the class options.
Reimplemented from PLearn::OnlineLearningModule.
Definition at line 303 of file NLLErrModule.cc.
References PLearn::OnlineLearningModule::declareOptions().
{ inherited::declareOptions(ol); }
static const PPath& PLearn::NLLErrModule::declaringFile | ( | ) | [inline, static] |
Reimplemented from PLearn::OnlineLearningModule.
Definition at line 106 of file NLLErrModule.h.
:
//##### Protected Options ###############################################
NLLErrModule * PLearn::NLLErrModule::deepCopy | ( | CopiesMap & | copies | ) | const [virtual] |
Reimplemented from PLearn::OnlineLearningModule.
Definition at line 56 of file NLLErrModule.cc.
void PLearn::NLLErrModule::forget | ( | ) | [virtual] |
reset the parameters to the state they would be BEFORE starting training.
Note that this method is necessarily called from build().
Implements PLearn::OnlineLearningModule.
Definition at line 285 of file NLLErrModule.cc.
{ }
given the input, compute the output (possibly resize it appropriately) SOON TO BE DEPRECATED, USE fprop(const TVec<Mat*>& ports_value)
Reimplemented from PLearn::OnlineLearningModule.
Definition at line 111 of file NLLErrModule.cc.
References PLearn::argmax(), fp_sm, getTarget(), PLearn::OnlineLearningModule::input_size, PLearn::OnlineLearningModule::output_size, pl_log, PLearn::TVec< T >::resize(), PLearn::softmax(), and PLearn::TVec< T >::subVec().
{ int target = getTarget( input ); // size check is done in getTarget() Vec input_ = input.subVec( 0, input_size ); output.resize( output_size ); fp_sm = softmax( input_ ); output[0] = - pl_log( fp_sm[target] ); if( output_size > 1 ) output[1] = ( argmax( input_ ) == target ) ? 0 : 1; }
OptionList & PLearn::NLLErrModule::getOptionList | ( | ) | const [virtual] |
Reimplemented from PLearn::Object.
Definition at line 56 of file NLLErrModule.cc.
OptionMap & PLearn::NLLErrModule::getOptionMap | ( | ) | const [virtual] |
Reimplemented from PLearn::Object.
Definition at line 56 of file NLLErrModule.cc.
RemoteMethodMap & PLearn::NLLErrModule::getRemoteMethodMap | ( | ) | const [virtual] |
Reimplemented from PLearn::Object.
Definition at line 56 of file NLLErrModule.cc.
Definition at line 67 of file NLLErrModule.cc.
References PLearn::argmax(), i, PLearn::OnlineLearningModule::input_size, PLearn::is_equal(), PLASSERT, PLERROR, PLearn::TVec< T >::size(), and PLearn::TVec< T >::subVec().
Referenced by bbpropUpdate(), bpropUpdate(), and fprop().
{ int t_size = input.size() - input_size; int target = -1; // size check if( t_size == 1 ) { target = (int) input[ input_size ]; } else if( t_size == input_size ) { /* PLWARNING("NLLErrModule::getTarget: You're giving a target the same\n" "size as the input, instead of an integer. Checking if\n" "this is a one-hot vector from this integer.\n"); */ Vec the_target = input.subVec( input_size, t_size ); // get position of '1' target = argmax( the_target ); #ifdef BOUNDCHECK // check if vector matches with a one-hot one PLASSERT( is_equal( the_target[target], 1. ) ) ; for( int i=0 ; i<input_size ; i++ ) PLASSERT( is_equal( the_target[i], 0. ) || i == target ); #endif } else { PLERROR("NLLErrModule::getTarget: target.size() is %i,\n" " but should be 1. 'target' should contain an integer.\n", t_size); } if( target < 0 || target >= input_size ) PLERROR("NLLErrModule::getTarget: target should be between 0 and" "input_size (%i).\n", input_size); return target; }
void PLearn::NLLErrModule::makeDeepCopyFromShallowCopy | ( | CopiesMap & | copies | ) | [virtual] |
Transforms a shallow copy into a deep copy.
Reimplemented from PLearn::OnlineLearningModule.
Definition at line 297 of file NLLErrModule.cc.
References PLearn::OnlineLearningModule::makeDeepCopyFromShallowCopy().
{ inherited::makeDeepCopyFromShallowCopy(copies); }
Reimplemented from PLearn::OnlineLearningModule.
Definition at line 106 of file NLLErrModule.h.
Vec PLearn::NLLErrModule::fp_sm [mutable, private] |
Definition at line 135 of file NLLErrModule.h.
Referenced by bpropUpdate(), build_(), and fprop().
int PLearn::NLLErrModule::target_size [protected] |
Definition at line 117 of file NLLErrModule.h.
Referenced by build_().