1 #include "petscvec.h" 2 #include "taosolver.h" 3 #include "tao-private/taolinesearch_impl.h" 4 #include "owarmijo.h" 5 6 #define REPLACE_FIFO 1 7 #define REPLACE_MRU 2 8 9 #define REFERENCE_MAX 1 10 #define REFERENCE_AVE 2 11 #define REFERENCE_MEAN 3 12 13 14 15 static PetscErrorCode ProjWork_OWLQN(Vec w,Vec x,Vec gv,PetscReal *gdx) 16 { 17 PetscReal *xptr,*wptr,*gptr; 18 PetscErrorCode ierr; 19 PetscInt low,high,low1,high1,low2,high2,i; 20 21 PetscFunctionBegin; 22 23 ierr=VecGetOwnershipRange(w,&low,&high);CHKERRQ(ierr); 24 ierr=VecGetOwnershipRange(x,&low1,&high1);CHKERRQ(ierr); 25 ierr=VecGetOwnershipRange(gv,&low2,&high2);CHKERRQ(ierr); 26 27 *gdx=0.0; 28 ierr = VecGetArray(x,&xptr);CHKERRQ(ierr); 29 ierr = VecGetArray(w,&wptr);CHKERRQ(ierr); 30 ierr = VecGetArray(gv,&gptr);CHKERRQ(ierr); 31 32 for (i=0;i<high-low;i++) 33 { 34 if (xptr[i]*wptr[i]<0.0) 35 wptr[i]=0.0; 36 37 *gdx = *gdx + gptr[i]*(wptr[i]-xptr[i]); 38 } 39 ierr = VecRestoreArray(w,&wptr);CHKERRQ(ierr); 40 ierr = VecRestoreArray(x,&xptr);CHKERRQ(ierr); 41 ierr = VecRestoreArray(gv,&gptr);CHKERRQ(ierr); 42 43 PetscFunctionReturn(0); 44 } 45 46 #undef __FUNCT__ 47 #define __FUNCT__ "TaoLineSearchDestroy_OWArmijo" 48 static PetscErrorCode TaoLineSearchDestroy_OWArmijo(TaoLineSearch ls) 49 { 50 TAOLINESEARCH_OWARMIJO_CTX *armP = (TAOLINESEARCH_OWARMIJO_CTX *)ls->data; 51 PetscErrorCode ierr; 52 53 PetscFunctionBegin; 54 55 if (armP->memory != PETSC_NULL) { 56 ierr = PetscFree(armP->memory); CHKERRQ(ierr); 57 armP->memory = PETSC_NULL; 58 } 59 if (armP->x) { 60 ierr = PetscObjectDereference((PetscObject)armP->x); CHKERRQ(ierr); 61 } 62 ierr = VecDestroy(&armP->work); CHKERRQ(ierr); 63 ierr = PetscFree(ls->data); CHKERRQ(ierr); 64 ls->data = PETSC_NULL; 65 PetscFunctionReturn(0); 66 } 67 68 #undef __FUNCT__ 69 #define __FUNCT__ "TaoLineSearchSetFromOptions_OWArmijo" 70 static PetscErrorCode TaoLineSearchSetFromOptions_OWArmijo(TaoLineSearch ls) 71 { 72 TAOLINESEARCH_OWARMIJO_CTX *armP = (TAOLINESEARCH_OWARMIJO_CTX *)ls->data; 73 PetscErrorCode ierr; 74 75 PetscFunctionBegin; 76 ierr = PetscOptionsHead("OWArmijo linesearch options");CHKERRQ(ierr); 77 ierr = PetscOptionsReal("-tao_ls_OWArmijo_alpha", "initial reference constant", "", armP->alpha, &armP->alpha, 0); CHKERRQ(ierr); 78 ierr = PetscOptionsReal("-tao_ls_OWArmijo_beta_inf", "decrease constant one", "", armP->beta_inf, &armP->beta_inf, 0); CHKERRQ(ierr); 79 ierr = PetscOptionsReal("-tao_ls_OWArmijo_beta", "decrease constant", "", armP->beta, &armP->beta, 0); CHKERRQ(ierr); 80 ierr = PetscOptionsReal("-tao_ls_OWArmijo_sigma", "acceptance constant", "", armP->sigma, &armP->sigma, 0); CHKERRQ(ierr); 81 ierr = PetscOptionsInt("-tao_ls_OWArmijo_memory_size", "number of historical elements", "", armP->memorySize, &armP->memorySize, 0); CHKERRQ(ierr); 82 ierr = PetscOptionsInt("-tao_ls_OWArmijo_reference_policy", "policy for updating reference value", "", armP->referencePolicy, &armP->referencePolicy, 0); CHKERRQ(ierr); 83 ierr = PetscOptionsInt("-tao_ls_OWArmijo_replacement_policy", "policy for updating memory", "", armP->replacementPolicy, &armP->replacementPolicy, 0); CHKERRQ(ierr); 84 ierr = PetscOptionsBool("-tao_ls_OWArmijo_nondescending","Use nondescending OWArmijo algorithm","",armP->nondescending,&armP->nondescending, 0); CHKERRQ(ierr); 85 ierr = PetscOptionsTail();CHKERRQ(ierr); 86 PetscFunctionReturn(0); 87 } 88 89 #undef __FUNCT__ 90 #define __FUNCT__ "TaoLineSearchView_OWArmijo" 91 static PetscErrorCode TaoLineSearchView_OWArmijo(TaoLineSearch ls, PetscViewer pv) 92 { 93 TAOLINESEARCH_OWARMIJO_CTX *armP = (TAOLINESEARCH_OWARMIJO_CTX *)ls->data; 94 PetscBool isascii; 95 PetscErrorCode ierr; 96 97 PetscFunctionBegin; 98 ierr = PetscObjectTypeCompare((PetscObject)pv, PETSCVIEWERASCII, &isascii); CHKERRQ(ierr); 99 if (isascii) { 100 ierr = PetscViewerASCIIPrintf(pv," maxf=%d, ftol=%g, gtol=%g\n",ls->max_funcs, ls->rtol, ls->ftol); CHKERRQ(ierr); 101 ierr=PetscViewerASCIIPrintf(pv," OWArmijo linesearch",armP->alpha);CHKERRQ(ierr); 102 if (armP->nondescending) { 103 ierr = PetscViewerASCIIPrintf(pv, " (nondescending)"); CHKERRQ(ierr); 104 } 105 ierr=PetscViewerASCIIPrintf(pv,": alpha=%g beta=%g ",armP->alpha,armP->beta);CHKERRQ(ierr); 106 ierr=PetscViewerASCIIPrintf(pv,"sigma=%g ",armP->sigma);CHKERRQ(ierr); 107 ierr=PetscViewerASCIIPrintf(pv,"memsize=%d\n",armP->memorySize);CHKERRQ(ierr); 108 } else { 109 SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"Viewer type %s not supported for OWArmijo TaoLineSearch",((PetscObject)pv)->type_name); 110 } 111 PetscFunctionReturn(0); 112 } 113 114 #undef __FUNCT__ 115 #define __FUNCT__ "TaoLineSearchApply_OWArmijo" 116 /* @ TaoApply_OWArmijo - This routine performs a linesearch. It 117 backtracks until the (nonmonotone) OWArmijo conditions are satisfied. 118 119 Input Parameters: 120 + tao - TAO_SOLVER context 121 . X - current iterate (on output X contains new iterate, X + step*S) 122 . S - search direction 123 . f - merit function evaluated at X 124 . G - gradient of merit function evaluated at X 125 . W - work vector 126 - step - initial estimate of step length 127 128 Output parameters: 129 + f - merit function evaluated at new iterate, X + step*S 130 . G - gradient of merit function evaluated at new iterate, X + step*S 131 . X - new iterate 132 - step - final step length 133 134 Info is set to one of: 135 . 0 - the line search succeeds; the sufficient decrease 136 condition and the directional derivative condition hold 137 138 negative number if an input parameter is invalid 139 - -1 - step < 0 140 141 positive number > 1 if the line search otherwise terminates 142 + 1 - Step is at the lower bound, stepmin. 143 @ */ 144 static PetscErrorCode TaoLineSearchApply_OWArmijo(TaoLineSearch ls, Vec x, PetscReal *f, Vec g, Vec s) 145 { 146 TAOLINESEARCH_OWARMIJO_CTX *armP = (TAOLINESEARCH_OWARMIJO_CTX *)ls->data; 147 PetscErrorCode ierr; 148 PetscInt i; 149 PetscReal fact, ref, gdx; 150 PetscInt idx; 151 PetscBool g_computed=PETSC_FALSE; /* to prevent extra gradient computation */ 152 153 Vec g_old; 154 PetscReal owlqn_minstep=0.005; 155 PetscReal partgdx; 156 157 PetscFunctionBegin; 158 159 fact = 0.0; 160 ls->nfeval=0; 161 ls->reason = TAOLINESEARCH_CONTINUE_ITERATING; 162 if (!armP->work) { 163 ierr = VecDuplicate(x,&armP->work); CHKERRQ(ierr); 164 armP->x = x; 165 ierr = PetscObjectReference((PetscObject)armP->x); CHKERRQ(ierr); 166 } 167 /* If x has changed, then recreate work */ 168 else if (x != armP->x) { 169 ierr = VecDestroy(&armP->work); CHKERRQ(ierr); 170 ierr = VecDuplicate(x,&armP->work); CHKERRQ(ierr); 171 ierr = PetscObjectDereference((PetscObject)armP->x); CHKERRQ(ierr); 172 armP->x = x; 173 ierr = PetscObjectReference((PetscObject)armP->x); CHKERRQ(ierr); 174 } 175 176 /* Check linesearch parameters */ 177 if (armP->alpha < 1) { 178 ierr = PetscInfo1(ls,"OWArmijo line search error: alpha (%g) < 1\n", armP->alpha); CHKERRQ(ierr); 179 ls->reason=TAOLINESEARCH_FAILED_BADPARAMETER; 180 } 181 182 else if ((armP->beta <= 0) || (armP->beta >= 1)) { 183 ierr = PetscInfo1(ls,"OWArmijo line search error: beta (%g) invalid\n", armP->beta); CHKERRQ(ierr); 184 ls->reason=TAOLINESEARCH_FAILED_BADPARAMETER; 185 186 } 187 188 else if ((armP->beta_inf <= 0) || (armP->beta_inf >= 1)) { 189 ierr = PetscInfo1(ls,"OWArmijo line search error: beta_inf (%g) invalid\n", armP->beta_inf); CHKERRQ(ierr); 190 ls->reason=TAOLINESEARCH_FAILED_BADPARAMETER; 191 } 192 193 else if ((armP->sigma <= 0) || (armP->sigma >= 0.5)) { 194 ierr = PetscInfo1(ls,"OWArmijo line search error: sigma (%g) invalid\n", armP->sigma); CHKERRQ(ierr); 195 ls->reason=TAOLINESEARCH_FAILED_BADPARAMETER; 196 } 197 198 else if (armP->memorySize < 1) { 199 ierr = PetscInfo1(ls,"OWArmijo line search error: memory_size (%d) < 1\n", armP->memorySize); CHKERRQ(ierr); 200 ls->reason=TAOLINESEARCH_FAILED_BADPARAMETER; 201 } 202 203 else if ((armP->referencePolicy != REFERENCE_MAX) && 204 (armP->referencePolicy != REFERENCE_AVE) && 205 (armP->referencePolicy != REFERENCE_MEAN)) { 206 ierr = PetscInfo(ls,"OWArmijo line search error: reference_policy invalid\n"); CHKERRQ(ierr); 207 ls->reason=TAOLINESEARCH_FAILED_BADPARAMETER; 208 209 } 210 211 else if ((armP->replacementPolicy != REPLACE_FIFO) && 212 (armP->replacementPolicy != REPLACE_MRU)) { 213 ierr = PetscInfo(ls,"OWArmijo line search error: replacement_policy invalid\n"); CHKERRQ(ierr); 214 ls->reason=TAOLINESEARCH_FAILED_BADPARAMETER; 215 } 216 217 else if (PetscIsInfOrNanReal(*f)) { 218 ierr = PetscInfo(ls,"OWArmijo line search error: initial function inf or nan\n"); CHKERRQ(ierr); 219 ls->reason=TAOLINESEARCH_FAILED_BADPARAMETER; 220 } 221 222 if (ls->reason != TAOLINESEARCH_CONTINUE_ITERATING) { 223 PetscFunctionReturn(0); 224 } 225 226 /* Check to see of the memory has been allocated. If not, allocate 227 the historical array and populate it with the initial function 228 values. */ 229 if (armP->memory == PETSC_NULL) { 230 ierr = PetscMalloc(sizeof(PetscReal)*armP->memorySize, &armP->memory ); CHKERRQ(ierr); 231 } 232 233 if (!armP->memorySetup) { 234 for (i = 0; i < armP->memorySize; i++) { 235 armP->memory[i] = armP->alpha*(*f); 236 } 237 238 armP->current = 0; 239 armP->lastReference = armP->memory[0]; 240 armP->memorySetup=PETSC_TRUE; 241 } 242 243 /* Calculate reference value (MAX) */ 244 ref = armP->memory[0]; 245 idx = 0; 246 247 for (i = 1; i < armP->memorySize; i++) { 248 if (armP->memory[i] > ref) { 249 ref = armP->memory[i]; 250 idx = i; 251 } 252 } 253 254 if (armP->referencePolicy == REFERENCE_AVE) { 255 ref = 0; 256 for (i = 0; i < armP->memorySize; i++) { 257 ref += armP->memory[i]; 258 } 259 ref = ref / armP->memorySize; 260 ref = PetscMax(ref, armP->memory[armP->current]); 261 } 262 else if (armP->referencePolicy == REFERENCE_MEAN) { 263 ref = PetscMin(ref, 0.5*(armP->lastReference + armP->memory[armP->current])); 264 } 265 266 267 if (armP->nondescending) { 268 fact = armP->sigma; 269 } 270 271 VecDuplicate(g,&g_old); 272 VecCopy(g,g_old); 273 274 ls->step = ls->initstep; 275 276 while (ls->step >= owlqn_minstep && ls->nfeval < ls->max_funcs) { 277 /* Calculate iterate */ 278 ierr = VecCopy(x,armP->work); CHKERRQ(ierr); 279 ierr = VecAXPY(armP->work,ls->step,s); CHKERRQ(ierr); 280 281 partgdx=0.0; 282 ierr = ProjWork_OWLQN(armP->work,x,g_old,&partgdx); 283 ierr = (PetscErrorCode)MPI_Allreduce(&partgdx,&gdx,1,MPI_DOUBLE,MPI_SUM,PETSC_COMM_WORLD); CHKERRQ(ierr); 284 285 /* Check the condition of gdx */ 286 if (PetscIsInfOrNanReal(gdx)) { 287 ierr = PetscInfo1(ls,"Initial Line Search step * g is Inf or Nan (%g)\n",gdx); CHKERRQ(ierr); 288 ls->reason=TAOLINESEARCH_FAILED_INFORNAN; 289 PetscFunctionReturn(0); 290 } 291 if (gdx >= 0.0) { 292 ierr = PetscInfo1(ls,"Initial Line Search step is not descent direction (g's=%g)\n",gdx); CHKERRQ(ierr); 293 ls->reason = TAOLINESEARCH_FAILED_ASCENT; 294 PetscFunctionReturn(0); 295 } 296 297 /* Calculate function at new iterate */ 298 ierr = TaoLineSearchComputeObjectiveAndGradient(ls,armP->work,f,g); CHKERRQ(ierr); 299 g_computed=PETSC_TRUE; 300 301 302 if (ls->step == ls->initstep) { 303 ls->f_fullstep = *f; 304 } 305 306 if (PetscIsInfOrNanReal(*f)) { 307 ls->step *= armP->beta_inf; 308 } 309 else { 310 /* Check descent condition */ 311 if (armP->nondescending && *f <= ref - ls->step*fact*ref) 312 break; 313 if (!armP->nondescending && *f <= ref + armP->sigma * gdx) { 314 break; 315 } 316 317 ls->step *= armP->beta; 318 } 319 } 320 321 VecDestroy(&g_old); 322 323 /* Check termination */ 324 if (PetscIsInfOrNanReal(*f)) { 325 ierr = PetscInfo(ls, "Function is inf or nan.\n"); CHKERRQ(ierr); 326 ls->reason = TAOLINESEARCH_FAILED_BADPARAMETER; 327 } else if (ls->step < owlqn_minstep) { 328 ierr = PetscInfo(ls, "Step length is below tolerance.\n"); CHKERRQ(ierr); 329 ls->reason = TAOLINESEARCH_HALTED_RTOL; 330 } else if (ls->nfeval >= ls->max_funcs) { 331 ierr = PetscInfo2(ls, "Number of line search function evals (%d) > maximum allowed (%d)\n",ls->nfeval, ls->max_funcs); CHKERRQ(ierr); 332 ls->reason = TAOLINESEARCH_HALTED_MAXFCN; 333 } 334 if (ls->reason) { 335 PetscFunctionReturn(0); 336 } 337 338 /* Successful termination, update memory */ 339 armP->lastReference = ref; 340 if (armP->replacementPolicy == REPLACE_FIFO) { 341 armP->memory[armP->current++] = *f; 342 if (armP->current >= armP->memorySize) { 343 armP->current = 0; 344 } 345 } else { 346 armP->current = idx; 347 armP->memory[idx] = *f; 348 } 349 350 /* Update iterate and compute gradient */ 351 ierr = VecCopy(armP->work,x); CHKERRQ(ierr); 352 if (!g_computed) { 353 ierr = TaoLineSearchComputeGradient(ls, x, g); CHKERRQ(ierr); 354 } 355 356 /* Finish computations */ 357 ierr = PetscInfo2(ls, "%d function evals in line search, step = %10.4f\n",ls->nfeval, ls->step); CHKERRQ(ierr); 358 359 PetscFunctionReturn(0); 360 } 361 362 EXTERN_C_BEGIN 363 #undef __FUNCT__ 364 #define __FUNCT__ "TaoLineSearchCreate_OWArmijo" 365 PetscErrorCode TaoLineSearchCreate_OWArmijo(TaoLineSearch ls) 366 { 367 TAOLINESEARCH_OWARMIJO_CTX *armP; 368 PetscErrorCode ierr; 369 370 PetscFunctionBegin; 371 PetscValidHeaderSpecific(ls,TAOLINESEARCH_CLASSID,1); 372 ierr = PetscNewLog(ls,&armP);CHKERRQ(ierr); 373 374 armP->memory = PETSC_NULL; 375 armP->alpha = 1.0; 376 armP->beta = 0.25; 377 armP->beta_inf = 0.25; 378 armP->sigma = 1e-4; 379 armP->memorySize = 1; 380 armP->referencePolicy = REFERENCE_MAX; 381 armP->replacementPolicy = REPLACE_MRU; 382 armP->nondescending=PETSC_FALSE; 383 ls->data = (void*)armP; 384 ls->initstep=0.1; 385 ls->ops->setup=0; 386 ls->ops->apply=TaoLineSearchApply_OWArmijo; 387 ls->ops->view = TaoLineSearchView_OWArmijo; 388 ls->ops->destroy = TaoLineSearchDestroy_OWArmijo; 389 ls->ops->setfromoptions = TaoLineSearchSetFromOptions_OWArmijo; 390 391 PetscFunctionReturn(0); 392 } 393 EXTERN_C_END 394