/usr/include/opengm/inference/trws/smooth_nesterov.hxx is in libopengm-dev 2.3.6-2.
This file is owned by root:root, with mode 0o644.
The actual contents of the file can be viewed below.
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 | /*
* smooth_nesterov.hxx
*
* Created on: Dec 23, 2013
* Author: bsavchyn
*/
#ifndef SMOOTH_NESTEROV_HXX_
#define SMOOTH_NESTEROV_HXX_
#include <opengm/inference/inference.hxx>
#include <opengm/inference/trws/trws_base.hxx>
#include <opengm/inference/auxiliary/primal_lpbound.hxx>
#include <opengm/inference/trws/smoothing_strategy.hxx>
namespace opengm{
template<class GM>
struct Nesterov_Parameter : public trws_base::SmoothingBasedInference_Parameter<GM>
{
typedef typename GM::ValueType ValueType;
typedef trws_base::DecompositionStorage<GM> Storage;
typedef typename trws_base::SmoothingBasedInference_Parameter<GM> parent;
typedef typename parent::SmoothingParametersType SmoothingParametersType;
typedef typename parent::SumProdSolverParametersType SumProdSolverParametersType;
typedef typename parent::MaxSumSolverParametersType MaxSumSolverParametersType;
typedef typename parent::PrimalLPEstimatorParametersType PrimalLPEstimatorParametersType;
typedef typename parent::SmoothingStrategyType SmoothingStrategyType;
typedef enum {ADAPTIVE_STEP,WC_STEP,JOJIC_STEP} GradientStepType;
static GradientStepType getGradientStepType(const std::string& name)
{
if (name.compare("WC_STEP")==0) return WC_STEP;
else if (name.compare("JOJIC_STEP")==0) return JOJIC_STEP;
else return ADAPTIVE_STEP;
}
static std::string getString(GradientStepType steptype)
{
switch (steptype)
{
case ADAPTIVE_STEP: return std::string("ADAPTIVE_STEP");
case WC_STEP : return std::string("WC_STEP");
case JOJIC_STEP : return std::string("JOJIC_STEP");
default: return std::string("UNKNOWN");
}
}
Nesterov_Parameter(size_t numOfExternalIterations=0,
ValueType precision=1.0,
bool absolutePrecision=true,
size_t numOfInternalIterations=3,
typename Storage::StructureType decompositionType=Storage::GENERALSTRUCTURE,
ValueType smoothingGapRatio=4,
ValueType startSmoothingValue=0.0,
ValueType primalBoundPrecision=std::numeric_limits<ValueType>::epsilon(),
size_t maxPrimalBoundIterationNumber=100,
size_t presolveMaxIterNumber=100,
bool canonicalNormalization=true,
ValueType presolveMinRelativeDualImprovement=0.01,
bool lazyLPPrimalBoundComputation=true,
ValueType smoothingDecayMultiplier=-1.0,
SmoothingStrategyType smoothingStrategy=SmoothingParametersType::ADAPTIVE_DIMINISHING,
bool fastComputations=true,
bool verbose=false,
GradientStepType gradientStep=ADAPTIVE_STEP,
ValueType gamma0=1e6,
bool plainGradient=false
)
:parent(numOfExternalIterations,
precision,
absolutePrecision,
numOfInternalIterations,
decompositionType,
smoothingGapRatio,
startSmoothingValue,
primalBoundPrecision,
maxPrimalBoundIterationNumber,
presolveMaxIterNumber,
canonicalNormalization,
presolveMinRelativeDualImprovement,
lazyLPPrimalBoundComputation,
smoothingDecayMultiplier,
smoothingStrategy,
fastComputations,
verbose
),
gradientStep_(gradientStep),
gamma0_(gamma0),
plainGradient_(plainGradient)
{};
GradientStepType gradientStep_;
ValueType gamma0_;
bool plainGradient_;
#ifdef TRWS_DEBUG_OUTPUT
void print(std::ostream& fout)const
{
parent::print(fout);
fout << "gradientStep_="<<getString(gradientStep_)<<std::endl;
fout << "gamma0_=" << gamma0_ <<std::endl;
fout << "plainGradient_=" << plainGradient_ <<std::endl;
}
#endif
};
template<class GM, class ACC>
class NesterovAcceleratedGradient : public trws_base::SmoothingBasedInference<GM, ACC> //Inference<GM, ACC>
{
public:
typedef trws_base::SmoothingBasedInference<GM, ACC> parent;
typedef ACC AccumulationType;
typedef GM GraphicalModelType;
OPENGM_GM_TYPE_TYPEDEFS;
//typedef std::vector<typename GM::ValueType> DDVectorType;
typedef typename parent::Storage Storage;
typedef typename Storage::DDVectorType DDVectorType;
typedef typename parent::SumProdSolver SumProdSolver;
typedef typename parent::MaxSumSolver MaxSumSolver;
typedef typename parent::PrimalBoundEstimator PrimalBoundEstimator;
typedef Nesterov_Parameter<GM> Parameter;
// typedef visitors::ExplicitVerboseVisitor<NesterovAcceleratedGradient<GM, ACC> > VerboseVisitorType;
// typedef visitors::ExplicitTimingVisitor <NesterovAcceleratedGradient<GM, ACC> > TimingVisitorType;
// typedef visitors::ExplicitEmptyVisitor <NesterovAcceleratedGradient<GM, ACC> > EmptyVisitorType;
typedef visitors::VerboseVisitor<NesterovAcceleratedGradient<GM, ACC> > VerboseVisitorType;
typedef visitors::TimingVisitor <NesterovAcceleratedGradient<GM, ACC> > TimingVisitorType;
typedef visitors::EmptyVisitor <NesterovAcceleratedGradient<GM, ACC> > EmptyVisitorType;
NesterovAcceleratedGradient(const GraphicalModelType& gm,const Parameter& param
#ifdef TRWS_DEBUG_OUTPUT
,std::ostream& fout=std::cout
#endif
)
:
parent(gm,param
#ifdef TRWS_DEBUG_OUTPUT
,(param.verbose_ ? fout : *OUT::nullstream::Instance())
#endif
),
_parameters(param),
_currentDualVector(_getDualVectorSize(),0.0)
{
#ifdef TRWS_DEBUG_OUTPUT
parent::_fout << "Parameters of the "<< name() <<" algorithm:"<<std::endl;
param.print(parent::_fout);
#endif
if (param.numOfExternalIterations_==0) throw std::runtime_error("NEST: a strictly positive number of iterations must be provided!");
};
template<class VISITOR>
InferenceTermination infer(VISITOR & visitor);
std::string name() const{ return "NEST"; }
InferenceTermination infer(){EmptyVisitorType visitor; return infer(visitor);};
// InferenceTermination marginal(const IndexType varID, IndependentFactorType& out) //const
// {
// parent::_marginalsTemp.resize(parent::_storage.numberOfLabels(varID));
// parent::_sumprodsolver.GetMarginals(varID, parent::_marginalsTemp.begin());
// OPENGM_ASSERT(parent::_marginalsTemp.size() == out.size());
// out.assign(parent::_storage.masterModel(), &varID, &varID+1, ACC::template neutral<ValueType>());
// for (LabelType i=0;i<out.size();++i)
// out(i)=parent::_marginalsTemp[i];
// }
private:
ValueType _evaluateGradient(const DDVectorType& point,DDVectorType* pgradient);
ValueType _evaluateSmoothObjective(const DDVectorType& point);
size_t _getDualVectorSize()const{return parent::_storage.getDDVectorSize();}
void _SetDualVariables(const DDVectorType& lambda);
ValueType _estimateOmega0()const{return 1;};//TODO: exchange with a reasonable value
void _InitSmoothing();//TODO: refactor me
void _GradientStep(const DDVectorType& gradient, const DDVectorType& startPoint, DDVectorType& endPoint, ValueType stepsize);
ValueType getLipschitzConstant()const;
Parameter _parameters;
DDVectorType _currentDualVector;
};
template<class GM, class ACC>
typename NesterovAcceleratedGradient<GM,ACC>::ValueType
NesterovAcceleratedGradient<GM,ACC>::getLipschitzConstant()const
{
ValueType result=0;
for (IndexType modelId=0;modelId<parent::_storage.numberOfModels();++modelId)
result+=(ValueType)parent::_storage.size(modelId);
return result/parent::_sumprodsolver.GetSmoothing();
}
template<class GM, class ACC>
typename NesterovAcceleratedGradient<GM,ACC>::ValueType
NesterovAcceleratedGradient<GM,ACC>::_evaluateGradient(const DDVectorType& point,DDVectorType* pgradient)
{
ValueType bound=_evaluateSmoothObjective(point);
parent::_sumprodsolver.GetMarginalsMove();
std::vector<ValueType> buffer1st;
std::vector<ValueType> buffer;
//transform marginals to dual vector
pgradient->resize(_currentDualVector.size());
typename DDVectorType::iterator gradientIt=pgradient->begin();
for (IndexType varId=0;varId<parent::_storage.masterModel().numberOfVariables();++varId)// all variables
{
const typename Storage::SubVariableListType& varList=parent::_storage.getSubVariableList(varId);
if (varList.size()==1) continue;
typename Storage::SubVariableListType::const_iterator modelIt=varList.begin();
IndexType firstModelId=modelIt->subModelId_;
IndexType firstModelVariableId=modelIt->subVariableId_;
buffer1st.resize(parent::_storage.masterModel().numberOfLabels(varId));
buffer.resize(parent::_storage.masterModel().numberOfLabels(varId));
parent::_sumprodsolver.GetMarginalsForSubModel(firstModelId,firstModelVariableId,buffer1st.begin());
++modelIt;
for(;modelIt!=varList.end();++modelIt) //all related models
{
parent::_sumprodsolver.GetMarginalsForSubModel(modelIt->subModelId_,modelIt->subVariableId_,buffer.begin());
gradientIt=std::transform(buffer.begin(),buffer.end(),buffer1st.begin(),gradientIt,std::minus<ValueType>());
}
}
return bound;
}
template<class GM, class ACC>
typename NesterovAcceleratedGradient<GM,ACC>::ValueType
NesterovAcceleratedGradient<GM,ACC>::_evaluateSmoothObjective(const DDVectorType& point)
{
_SetDualVariables(point);
parent::_sumprodsolver.ForwardMove();
return parent::_sumprodsolver.bound();
}
template<class GM, class ACC>
void NesterovAcceleratedGradient<GM,ACC>::_SetDualVariables(const DDVectorType& lambda)
{
DDVectorType delta(_currentDualVector.size());
std::transform(lambda.begin(),lambda.end(),_currentDualVector.begin(),delta.begin(),std::minus<ValueType>());
_currentDualVector=lambda;
parent::_storage.addDDvector(delta);
};
template<class GM, class ACC>
void NesterovAcceleratedGradient<GM,ACC>::_InitSmoothing()
{
if (_parameters.smoothing_ > 0.0)
parent::_sumprodsolver.SetSmoothing(_parameters.smoothing_);
else
throw std::runtime_error("NesterovAcceleratedGradient::_InitSmoothing(): Error! Automatic smoothing selection is not implemented yet.");
};
template<class GM, class ACC>
void NesterovAcceleratedGradient<GM,ACC>::_GradientStep(const DDVectorType& gradient, const DDVectorType& startPoint, DDVectorType& endPoint, ValueType stepsize)
{
//!>lambda=y+gradient*stepsize;
std::transform(gradient.begin(),gradient.end(),endPoint.begin(),std::bind1st(std::multiplies<ValueType>(),stepsize));
std::transform(startPoint.begin(),startPoint.end(),endPoint.begin(),endPoint.begin(),std::plus<ValueType>());
}
template<class GM, class ACC>
template<class VISITOR>
InferenceTermination NesterovAcceleratedGradient<GM,ACC>::infer(VISITOR & vis)
{
visitors::VisitorWrapper<VISITOR,NesterovAcceleratedGradient<GM, ACC> > visitor(&vis,this);
size_t counter=0;//!> oracle calls counter
visitor.addLog("primalLPbound");
visitor.addLog("oracleCalls");
visitor.begin();
if (parent::_sumprodsolver.GetSmoothing()<=0.0)
{
parent::_maxsumsolver.ForwardMove();
parent::_maxsumsolver.EstimateIntegerLabelingAndBound();
parent::_SelectOptimalBoundsAndLabeling();
++counter;
if (parent::_sumprodsolver.CheckDualityGap(parent::value(),parent::bound()))
{
#ifdef TRWS_DEBUG_OUTPUT
parent::_fout << "NesterovAcceleratedGradient::_CheckStoppingCondition(): Precision attained! Problem solved!"<<std::endl;
#endif
return NORMAL;
}
counter+=parent::_EstimateStartingSmoothing(visitor);
}else
{
parent::_sumprodsolver.SetSmoothing(_parameters.startSmoothingValue());
}
DDVectorType gradient(_currentDualVector.size()),
lambda(_currentDualVector.size()),
y(_currentDualVector),
v(_currentDualVector);
DDVectorType w(_currentDualVector.size());//temp variable
ValueType alpha,
gamma= _parameters.gamma0_,
omega=_estimateOmega0();
omega=omega/2.0;
for (size_t i=0;i<_parameters.maxNumberOfIterations();++i)
{
#ifdef TRWS_DEBUG_OUTPUT
parent::_fout <<"i="<<i<<std::endl;
#endif
//gradient step with approximate linear search:
ValueType doubledLipschitzConstant=2*getLipschitzConstant();//depends on a smoothing value
//===================== begin of internal loop ===========================================
for (size_t j=0;j<_parameters.numberOfInternalIterations();++j)
{
ValueType mul=1.0;
counter+=2; ValueType oldObjVal=_evaluateGradient(y,&gradient);
#ifdef TRWS_DEBUG_OUTPUT
parent::_fout <<"Dual smooth objective ="<<oldObjVal<<std::endl;
#endif
switch (_parameters.gradientStep_)
{
case Parameter::ADAPTIVE_STEP:
{
ValueType norm2=std::inner_product(gradient.begin(),gradient.end(),gradient.begin(),(ValueType)0);//squared L2 norm
ValueType newObjVal;
omega/=4.0;
do
{
omega*=2.0;
ACC::iop(-1.0,1.0,mul);
_GradientStep(gradient,y,lambda,mul/omega);//!>lambda=y+gradient/omega; plus/minus depending on ACC
//newObjVal=_evaluateSmoothObjective(lambda,((j+1)==_parameters.numberOfInternalIterations()));
++counter; newObjVal=_evaluateSmoothObjective(lambda);
}
while ( ACC::bop(newObjVal,(ValueType)(oldObjVal+mul*norm2/2.0/omega)) && (omega < doubledLipschitzConstant));//TODO: +/- and >/< depending on ACC
++counter; parent::_sumprodsolver.GetMarginalsAndDerivativeMove();
if (omega >= doubledLipschitzConstant)
{
#ifdef TRWS_DEBUG_OUTPUT
parent::_fout << "Step size is smaller then the inverse Lipschitz constant. Passing to smoothing update." <<std::endl;
#endif
}
}
break;
case Parameter::WC_STEP:
omega=doubledLipschitzConstant/2.0;
_GradientStep(gradient,y,lambda,mul/omega);
break;
case Parameter::JOJIC_STEP:
omega=1.0/parent::_sumprodsolver.GetSmoothing();
_GradientStep(gradient,y,lambda,mul/omega);
break;
default:
std::runtime_error("NesterovAcceleratedGradient::infer():Error! Unknown value of the step size selector _parameters.gradientStep_.");
break;
}//switch
if (!_parameters.plainGradient_)
{
//updating parameters
alpha=(sqrt(gamma*gamma+4*omega*gamma)-gamma)/omega/2.0;
gamma*=(1-alpha);
//v+=(alpha/gamma)*gradient;
trws_base::transform_inplace(gradient.begin(),gradient.end(),std::bind1st(std::multiplies<ValueType>(),mul*alpha/gamma));//!> plus/minus depending on ACC
std::transform(v.begin(),v.end(),gradient.begin(),v.begin(),std::plus<ValueType>());
//y=alpha*v+(1-alpha)*lambda;
trws_base::transform_inplace(lambda.begin(),lambda.end(),std::bind1st(std::multiplies<ValueType>(),(1-alpha)));
std::transform(v.begin(),v.end(),w.begin(),std::bind1st(std::multiplies<ValueType>(),alpha));
std::transform(w.begin(),w.end(),lambda.begin(),y.begin(),std::plus<ValueType>());
}else //plain gradient algorithm
{
std::copy(lambda.begin(),lambda.end(),y.begin());
}
}
//=================================== end of internal loop ===============================================
parent::_maxsumsolver.ForwardMove();//initializes a move, makes a forward move and computes the dual bound, is used also in derivative computation in the next line
parent::_maxsumsolver.EstimateIntegerLabelingAndBound();
++counter;
#ifdef TRWS_DEBUG_OUTPUT
parent::_fout << "_maxsumsolver.bound()=" <<parent::_maxsumsolver.bound()<<", _maxsumsolver.value()=" <<parent::_maxsumsolver.value() <<std::endl;
#endif
ValueType derivative=parent::_EstimateRhoDerivative();
#ifdef TRWS_DEBUG_OUTPUT
parent::_fout << "derivative="<<derivative<<std::endl;
#endif
InferenceTermination returncode;
if ( parent::_CheckStoppingCondition(&returncode))
{
visitor();
//std::cout << "counter=" << counter <<std::endl;
visitor.log("oracleCalls",(double)counter);
visitor.log("primalLPbound",(double)parent::_bestPrimalLPbound);
visitor.end();
return NORMAL;
}
size_t flag=visitor();
//std::cout << "counter=" << counter <<std::endl;
visitor.log("oracleCalls",(double)counter);
visitor.log("primalLPbound",(double)parent::_bestPrimalLPbound);
if( flag != visitors::VisitorReturnFlag::ContinueInf ){
break;
}
parent::_UpdateSmoothing(parent::_bestPrimalBound,parent::_maxsumsolver.bound(),parent::_sumprodsolver.bound(),derivative,i+1);
}
//update smoothing
parent::_SelectOptimalBoundsAndLabeling();
visitor();
visitor.log("oracleCalls",(double)counter);
visitor.log("primalLPbound",(double)parent::_bestPrimalLPbound);
visitor.end();
return NORMAL;
}
//template<class GM>
//class NesterovAcceleratedGradient<GM, opengm::Integrator>
//{
//public:
// typedef GM GraphicalModelType;
// OPENGM_GM_TYPE_TYPEDEFS;
// typedef NesterovAcceleratedGradient <GM, opengm::Maximizer> parent;
// typedef typename parent::AccumulationType AccumulationType;
//
// typedef typename parent::Parameter Parameter;
// typedef typename parent::Storage Storage;
// typedef typename Storage::DDVectorType DDVectorType;
// typedef typename parent::SumProdSolver SumProdSolver;
// typedef typename parent::MaxSumSolver MaxSumSolver;
// typedef typename parent::PrimalBoundEstimator PrimalBoundEstimator;
//
// typedef typename parent::VerboseVisitorType VerboseVisitorType;
// typedef typename parent::TimingVisitorType TimingVisitorType;
// typedef typename parent::EmptyVisitorType EmptyVisitorType;
//
// NesterovAcceleratedGradient(const GraphicalModelType& gm, const Parameter& param
// #ifdef TRWS_DEBUG_OUTPUT
// ,std::ostream& fout=std::cout
// #endif
// )
// {
// Parameter param1=param;
// param1.smoothingStrategy() = Parameter::SmoothingParametersType::FIXED;
// param1.setStartSmoothingValue(1.0);
//
// _pparent = new parent(gm,param1
// #ifdef TRWS_DEBUG_OUTPUT
// ,fout
// #endif
// );
// }
// virtual ~NesterovAcceleratedGradient(){delete _pparent;}
//
// std::string name() const{ return "NEST"; }
//
// InferenceTermination marginal(const IndexType varID, IndependentFactorType& out) const
// {
// return _pparent->marginal(varID,out);
// }
//
// template<class VISITOR>
// InferenceTermination infer(VISITOR & visitor){return _pparent->infer(visitor);};
//
// InferenceTermination infer(){return _pparent->infer();}
//
//private:
// parent* _pparent;
//};
}//namespace opengm
#endif /* SMOOTH_NESTEROV_HXX_ */
|