1 #include <petsc/private/taolinesearchimpl.h> 2 #include <../src/tao/linesearch/impls/armijo/armijo.h> 3 4 #define REPLACE_FIFO 1 5 #define REPLACE_MRU 2 6 7 #define REFERENCE_MAX 1 8 #define REFERENCE_AVE 2 9 #define REFERENCE_MEAN 3 10 11 static PetscErrorCode TaoLineSearchDestroy_Armijo(TaoLineSearch ls) 12 { 13 TaoLineSearch_ARMIJO *armP = (TaoLineSearch_ARMIJO *)ls->data; 14 15 PetscFunctionBegin; 16 PetscCall(PetscFree(armP->memory)); 17 PetscCall(VecDestroy(&armP->x)); 18 PetscCall(VecDestroy(&armP->work)); 19 PetscCall(PetscFree(ls->data)); 20 PetscFunctionReturn(PETSC_SUCCESS); 21 } 22 23 static PetscErrorCode TaoLineSearchReset_Armijo(TaoLineSearch ls) 24 { 25 TaoLineSearch_ARMIJO *armP = (TaoLineSearch_ARMIJO *)ls->data; 26 27 PetscFunctionBegin; 28 PetscCall(PetscFree(armP->memory)); 29 armP->memorySetup = PETSC_FALSE; 30 PetscFunctionReturn(PETSC_SUCCESS); 31 } 32 33 static PetscErrorCode TaoLineSearchSetFromOptions_Armijo(TaoLineSearch ls, PetscOptionItems PetscOptionsObject) 34 { 35 TaoLineSearch_ARMIJO *armP = (TaoLineSearch_ARMIJO *)ls->data; 36 37 PetscFunctionBegin; 38 PetscOptionsHeadBegin(PetscOptionsObject, "Armijo linesearch options"); 39 PetscCall(PetscOptionsReal("-tao_ls_armijo_alpha", "initial reference constant", "", armP->alpha, &armP->alpha, NULL)); 40 PetscCall(PetscOptionsReal("-tao_ls_armijo_beta_inf", "decrease constant one", "", armP->beta_inf, &armP->beta_inf, NULL)); 41 PetscCall(PetscOptionsReal("-tao_ls_armijo_beta", "decrease constant", "", armP->beta, &armP->beta, NULL)); 42 PetscCall(PetscOptionsReal("-tao_ls_armijo_sigma", "acceptance constant", "", armP->sigma, &armP->sigma, NULL)); 43 PetscCall(PetscOptionsInt("-tao_ls_armijo_memory_size", "number of historical elements", "", armP->memorySize, &armP->memorySize, NULL)); 44 PetscCall(PetscOptionsInt("-tao_ls_armijo_reference_policy", "policy for updating reference value", "", armP->referencePolicy, &armP->referencePolicy, NULL)); 45 PetscCall(PetscOptionsInt("-tao_ls_armijo_replacement_policy", "policy for updating memory", "", armP->replacementPolicy, &armP->replacementPolicy, NULL)); 46 PetscCall(PetscOptionsBool("-tao_ls_armijo_nondescending", "Use nondescending armijo algorithm", "", armP->nondescending, &armP->nondescending, NULL)); 47 PetscOptionsHeadEnd(); 48 PetscFunctionReturn(PETSC_SUCCESS); 49 } 50 51 static PetscErrorCode TaoLineSearchView_Armijo(TaoLineSearch ls, PetscViewer pv) 52 { 53 TaoLineSearch_ARMIJO *armP = (TaoLineSearch_ARMIJO *)ls->data; 54 PetscBool isascii; 55 56 PetscFunctionBegin; 57 PetscCall(PetscObjectTypeCompare((PetscObject)pv, PETSCVIEWERASCII, &isascii)); 58 if (isascii) { 59 PetscCall(PetscViewerASCIIPrintf(pv, " Armijo linesearch")); 60 if (armP->nondescending) PetscCall(PetscViewerASCIIPrintf(pv, " (nondescending)")); 61 if (ls->bounded) PetscCall(PetscViewerASCIIPrintf(pv, " (projected)")); 62 PetscCall(PetscViewerASCIIPrintf(pv, ": alpha=%g beta=%g ", (double)armP->alpha, (double)armP->beta)); 63 PetscCall(PetscViewerASCIIPrintf(pv, "sigma=%g ", (double)armP->sigma)); 64 PetscCall(PetscViewerASCIIPrintf(pv, "memsize=%" PetscInt_FMT "\n", armP->memorySize)); 65 } 66 PetscFunctionReturn(PETSC_SUCCESS); 67 } 68 69 /* @ TaoApply_Armijo - This routine performs a linesearch. It 70 backtracks until the (nonmonotone) Armijo conditions are satisfied. 71 72 Input Parameters: 73 + tao - Tao context 74 . X - current iterate (on output X contains new iterate, X + step*S) 75 . S - search direction 76 . f - merit function evaluated at X 77 . G - gradient of merit function evaluated at X 78 . W - work vector 79 - step - initial estimate of step length 80 81 Output parameters: 82 + f - merit function evaluated at new iterate, X + step*S 83 . G - gradient of merit function evaluated at new iterate, X + step*S 84 . X - new iterate 85 - step - final step length 86 87 @ */ 88 static PetscErrorCode TaoLineSearchApply_Armijo(TaoLineSearch ls, Vec x, PetscReal *f, Vec g, Vec s) 89 { 90 TaoLineSearch_ARMIJO *armP = (TaoLineSearch_ARMIJO *)ls->data; 91 PetscInt i, its = 0; 92 PetscReal fact, ref, gdx; 93 PetscInt idx; 94 PetscBool g_computed = PETSC_FALSE; /* to prevent extra gradient computation */ 95 96 PetscFunctionBegin; 97 PetscCall(TaoLineSearchMonitor(ls, 0, *f, 0.0)); 98 99 ls->reason = TAOLINESEARCH_CONTINUE_ITERATING; 100 if (!armP->work) { 101 PetscCall(VecDuplicate(x, &armP->work)); 102 armP->x = x; 103 PetscCall(PetscObjectReference((PetscObject)armP->x)); 104 } else if (x != armP->x) { 105 /* If x has changed, then recreate work */ 106 PetscCall(VecDestroy(&armP->work)); 107 PetscCall(VecDuplicate(x, &armP->work)); 108 PetscCall(PetscObjectDereference((PetscObject)armP->x)); 109 armP->x = x; 110 PetscCall(PetscObjectReference((PetscObject)armP->x)); 111 } 112 113 /* Check linesearch parameters */ 114 if (armP->alpha < 1) { 115 PetscCall(PetscInfo(ls, "Armijo line search error: alpha (%g) < 1\n", (double)armP->alpha)); 116 ls->reason = TAOLINESEARCH_FAILED_BADPARAMETER; 117 } else if ((armP->beta <= 0) || (armP->beta >= 1)) { 118 PetscCall(PetscInfo(ls, "Armijo line search error: beta (%g) invalid\n", (double)armP->beta)); 119 ls->reason = TAOLINESEARCH_FAILED_BADPARAMETER; 120 } else if ((armP->beta_inf <= 0) || (armP->beta_inf >= 1)) { 121 PetscCall(PetscInfo(ls, "Armijo line search error: beta_inf (%g) invalid\n", (double)armP->beta_inf)); 122 ls->reason = TAOLINESEARCH_FAILED_BADPARAMETER; 123 } else if ((armP->sigma <= 0) || (armP->sigma >= 0.5)) { 124 PetscCall(PetscInfo(ls, "Armijo line search error: sigma (%g) invalid\n", (double)armP->sigma)); 125 ls->reason = TAOLINESEARCH_FAILED_BADPARAMETER; 126 } else if (armP->memorySize < 1) { 127 PetscCall(PetscInfo(ls, "Armijo line search error: memory_size (%" PetscInt_FMT ") < 1\n", armP->memorySize)); 128 ls->reason = TAOLINESEARCH_FAILED_BADPARAMETER; 129 } else if ((armP->referencePolicy != REFERENCE_MAX) && (armP->referencePolicy != REFERENCE_AVE) && (armP->referencePolicy != REFERENCE_MEAN)) { 130 PetscCall(PetscInfo(ls, "Armijo line search error: reference_policy invalid\n")); 131 ls->reason = TAOLINESEARCH_FAILED_BADPARAMETER; 132 } else if ((armP->replacementPolicy != REPLACE_FIFO) && (armP->replacementPolicy != REPLACE_MRU)) { 133 PetscCall(PetscInfo(ls, "Armijo line search error: replacement_policy invalid\n")); 134 ls->reason = TAOLINESEARCH_FAILED_BADPARAMETER; 135 } else if (PetscIsInfOrNanReal(*f)) { 136 PetscCall(PetscInfo(ls, "Armijo line search error: initial function inf or nan\n")); 137 ls->reason = TAOLINESEARCH_FAILED_BADPARAMETER; 138 } 139 140 if (ls->reason != TAOLINESEARCH_CONTINUE_ITERATING) PetscFunctionReturn(PETSC_SUCCESS); 141 142 /* Check to see of the memory has been allocated. If not, allocate 143 the historical array and populate it with the initial function 144 values. */ 145 if (!armP->memory) PetscCall(PetscMalloc1(armP->memorySize, &armP->memory)); 146 147 if (!armP->memorySetup) { 148 for (i = 0; i < armP->memorySize; i++) armP->memory[i] = armP->alpha * (*f); 149 150 armP->current = 0; 151 armP->lastReference = armP->memory[0]; 152 armP->memorySetup = PETSC_TRUE; 153 } 154 155 /* Calculate reference value (MAX) */ 156 ref = *f; 157 idx = armP->current; 158 for (i = 0; i < armP->memorySize; i++) { 159 if (armP->memory[i] > ref) { 160 ref = armP->memory[i]; 161 idx = i; 162 } 163 } 164 165 if (armP->referencePolicy == REFERENCE_AVE) { 166 ref = 0; 167 for (i = 0; i < armP->memorySize; i++) ref += armP->memory[i]; 168 ref = ref / armP->memorySize; 169 ref = PetscMax(ref, armP->memory[armP->current]); 170 } else if (armP->referencePolicy == REFERENCE_MEAN) { 171 ref = PetscMin(ref, 0.5 * (armP->lastReference + armP->memory[armP->current])); 172 } 173 PetscCall(VecDot(g, s, &gdx)); 174 175 if (PetscIsInfOrNanReal(gdx)) { 176 PetscCall(PetscInfo(ls, "Initial Line Search step * g is infinity or NaN (%g)\n", (double)gdx)); 177 ls->reason = TAOLINESEARCH_FAILED_INFORNAN; 178 PetscFunctionReturn(PETSC_SUCCESS); 179 } 180 if (gdx >= 0.0) { 181 PetscCall(PetscInfo(ls, "Initial Line Search step is not descent direction (g's=%g)\n", (double)gdx)); 182 ls->reason = TAOLINESEARCH_FAILED_ASCENT; 183 PetscFunctionReturn(PETSC_SUCCESS); 184 } 185 186 if (armP->nondescending) { 187 fact = armP->sigma; 188 } else { 189 fact = armP->sigma * gdx; 190 } 191 ls->step = ls->initstep; 192 while (ls->step >= ls->stepmin && (ls->nfeval + ls->nfgeval) < ls->max_funcs) { 193 /* Calculate iterate */ 194 ++its; 195 PetscCall(VecWAXPY(armP->work, ls->step, s, x)); 196 if (ls->bounded) PetscCall(VecMedian(ls->lower, armP->work, ls->upper, armP->work)); 197 198 /* Calculate function at new iterate */ 199 PetscCall(VecLockReadPush(x)); 200 if (ls->hasobjective) { 201 PetscCall(TaoLineSearchComputeObjective(ls, armP->work, f)); 202 g_computed = PETSC_FALSE; 203 } else if (ls->usegts) { 204 PetscCall(TaoLineSearchComputeObjectiveAndGTS(ls, armP->work, f, &gdx)); 205 g_computed = PETSC_FALSE; 206 } else { 207 PetscCall(TaoLineSearchComputeObjectiveAndGradient(ls, armP->work, f, g)); 208 g_computed = PETSC_TRUE; 209 } 210 PetscCall(VecLockReadPop(x)); 211 if (ls->step == ls->initstep) ls->f_fullstep = *f; 212 213 PetscCall(TaoLineSearchMonitor(ls, its, *f, ls->step)); 214 215 if (PetscIsInfOrNanReal(*f)) { 216 ls->step *= armP->beta_inf; 217 } else { 218 /* Check descent condition */ 219 if (armP->nondescending && *f <= ref - ls->step * fact * ref) break; 220 if (!armP->nondescending && *f <= ref + ls->step * fact) break; 221 222 ls->step *= armP->beta; 223 } 224 } 225 226 /* Check termination */ 227 if (PetscIsInfOrNanReal(*f)) { 228 PetscCall(PetscInfo(ls, "Function is inf or nan.\n")); 229 ls->reason = TAOLINESEARCH_FAILED_INFORNAN; 230 } else if (ls->step < ls->stepmin) { 231 PetscCall(PetscInfo(ls, "Step length is below tolerance.\n")); 232 ls->reason = TAOLINESEARCH_HALTED_RTOL; 233 } else if ((ls->nfeval + ls->nfgeval) >= ls->max_funcs) { 234 PetscCall(PetscInfo(ls, "Number of line search function evals (%" PetscInt_FMT ") > maximum allowed (%" PetscInt_FMT ")\n", ls->nfeval + ls->nfgeval, ls->max_funcs)); 235 ls->reason = TAOLINESEARCH_HALTED_MAXFCN; 236 } 237 if (ls->reason) PetscFunctionReturn(PETSC_SUCCESS); 238 239 /* Successful termination, update memory */ 240 ls->reason = TAOLINESEARCH_SUCCESS; 241 armP->lastReference = ref; 242 if (armP->replacementPolicy == REPLACE_FIFO) { 243 armP->memory[armP->current++] = *f; 244 if (armP->current >= armP->memorySize) armP->current = 0; 245 } else { 246 armP->current = idx; 247 armP->memory[idx] = *f; 248 } 249 250 /* Update iterate and compute gradient */ 251 PetscCall(VecCopy(armP->work, x)); 252 if (!g_computed) PetscCall(TaoLineSearchComputeGradient(ls, x, g)); 253 PetscCall(PetscInfo(ls, "%" PetscInt_FMT " function evals in line search, step = %g\n", ls->nfeval, (double)ls->step)); 254 PetscFunctionReturn(PETSC_SUCCESS); 255 } 256 257 /*MC 258 TAOLINESEARCHARMIJO - Backtracking line-search that satisfies only the (nonmonotone) Armijo condition 259 (i.e., sufficient decrease). 260 261 Armijo line-search type can be selected with "-tao_ls_type armijo". 262 263 Level: developer 264 265 .seealso: `TaoLineSearchCreate()`, `TaoLineSearchSetType()`, `TaoLineSearchApply()` 266 M*/ 267 PETSC_EXTERN PetscErrorCode TaoLineSearchCreate_Armijo(TaoLineSearch ls) 268 { 269 TaoLineSearch_ARMIJO *armP; 270 271 PetscFunctionBegin; 272 PetscValidHeaderSpecific(ls, TAOLINESEARCH_CLASSID, 1); 273 PetscCall(PetscNew(&armP)); 274 275 armP->memory = NULL; 276 armP->alpha = 1.0; 277 armP->beta = 0.5; 278 armP->beta_inf = 0.5; 279 armP->sigma = 1e-4; 280 armP->memorySize = 1; 281 armP->referencePolicy = REFERENCE_MAX; 282 armP->replacementPolicy = REPLACE_MRU; 283 armP->nondescending = PETSC_FALSE; 284 ls->data = (void *)armP; 285 ls->initstep = 1.0; 286 ls->ops->setup = NULL; 287 ls->ops->monitor = NULL; 288 ls->ops->apply = TaoLineSearchApply_Armijo; 289 ls->ops->view = TaoLineSearchView_Armijo; 290 ls->ops->destroy = TaoLineSearchDestroy_Armijo; 291 ls->ops->reset = TaoLineSearchReset_Armijo; 292 ls->ops->setfromoptions = TaoLineSearchSetFromOptions_Armijo; 293 PetscFunctionReturn(PETSC_SUCCESS); 294 } 295