Actual source code: gpcglinesearch.c
petsc-3.11.4 2019-09-28
1: #include <petsc/private/taolinesearchimpl.h>
2: #include <../src/tao/linesearch/impls/gpcglinesearch/gpcglinesearch.h>
4: /* ---------------------------------------------------------- */
6: static PetscErrorCode TaoLineSearchDestroy_GPCG(TaoLineSearch ls)
7: {
8: PetscErrorCode ierr;
9: TaoLineSearch_GPCG *ctx = (TaoLineSearch_GPCG *)ls->data;
12: VecDestroy(&ctx->W1);
13: VecDestroy(&ctx->W2);
14: VecDestroy(&ctx->Gold);
15: VecDestroy(&ctx->x);
16: PetscFree(ls->data);
17: return(0);
18: }
21: /*------------------------------------------------------------*/
22: static PetscErrorCode TaoLineSearchView_GPCG(TaoLineSearch ls, PetscViewer viewer)
23: {
24: PetscBool isascii;
28: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&isascii);
29: if (isascii) {
30: PetscViewerASCIIPrintf(viewer," GPCG Line search");
31: }
32: return(0);
33: }
35: /*------------------------------------------------------------*/
36: static PetscErrorCode TaoLineSearchApply_GPCG(TaoLineSearch ls, Vec x, PetscReal *f, Vec g, Vec s)
37: {
38: TaoLineSearch_GPCG *neP = (TaoLineSearch_GPCG *)ls->data;
39: PetscErrorCode ierr;
40: PetscInt i;
41: PetscBool g_computed=PETSC_FALSE; /* to prevent extra gradient computation */
42: PetscReal d1,finit,actred,prered,rho, gdx;
45: /* ls->stepmin - lower bound for step */
46: /* ls->stepmax - upper bound for step */
47: /* ls->rtol - relative tolerance for an acceptable step */
48: /* ls->ftol - tolerance for sufficient decrease condition */
49: /* ls->gtol - tolerance for curvature condition */
50: /* ls->nfeval - number of function evaluations */
51: /* ls->nfeval - number of function/gradient evaluations */
52: /* ls->max_funcs - maximum number of function evaluations */
53:
54: TaoLineSearchMonitor(ls, 0, *f, 0.0);
56: ls->reason = TAOLINESEARCH_CONTINUE_ITERATING;
57: ls->step = ls->initstep;
58: if (!neP->W2) {
59: VecDuplicate(x,&neP->W2);
60: VecDuplicate(x,&neP->W1);
61: VecDuplicate(x,&neP->Gold);
62: neP->x = x;
63: PetscObjectReference((PetscObject)neP->x);
64: } else if (x != neP->x) {
65: VecDestroy(&neP->x);
66: VecDestroy(&neP->W1);
67: VecDestroy(&neP->W2);
68: VecDestroy(&neP->Gold);
69: VecDuplicate(x,&neP->W1);
70: VecDuplicate(x,&neP->W2);
71: VecDuplicate(x,&neP->Gold);
72: PetscObjectDereference((PetscObject)neP->x);
73: neP->x = x;
74: PetscObjectReference((PetscObject)neP->x);
75: }
77: VecDot(g,s,&gdx);
78: if (gdx > 0) {
79: PetscInfo1(ls,"Line search error: search direction is not descent direction. dot(g,s) = %g\n",(double)gdx);
80: ls->reason = TAOLINESEARCH_FAILED_ASCENT;
81: return(0);
82: }
83: VecCopy(x,neP->W2);
84: VecCopy(g,neP->Gold);
85: if (ls->bounded) {
86: /* Compute the smallest steplength that will make one nonbinding variable equal the bound */
87: VecStepBoundInfo(x,s,ls->lower,ls->upper,&rho,&actred,&d1);
88: ls->step = PetscMin(ls->step,d1);
89: }
90: rho=0; actred=0;
92: if (ls->step < 0) {
93: PetscInfo1(ls,"Line search error: initial step parameter %g< 0\n",(double)ls->step);
94: ls->reason = TAOLINESEARCH_HALTED_OTHER;
95: return(0);
96: }
98: /* Initialization */
99: finit = *f;
100: for (i=0; i< ls->max_funcs; i++) {
101: /* Force the step to be within the bounds */
102: ls->step = PetscMax(ls->step,ls->stepmin);
103: ls->step = PetscMin(ls->step,ls->stepmax);
105: VecCopy(x,neP->W2);
106: VecAXPY(neP->W2,ls->step,s);
107: if (ls->bounded) {
108: /* Make sure new vector is numerically within bounds */
109: VecMedian(neP->W2,ls->lower,ls->upper,neP->W2);
110: }
112: /* Gradient is not needed here. Unless there is a separate
113: gradient routine, compute it here anyway to prevent recomputing at
114: the end of the line search */
115: if (ls->hasobjective) {
116: TaoLineSearchComputeObjective(ls,neP->W2,f);
117: g_computed=PETSC_FALSE;
118: } else if (ls->usegts){
119: TaoLineSearchComputeObjectiveAndGTS(ls,neP->W2,f,&gdx);
120: g_computed=PETSC_FALSE;
121: } else {
122: TaoLineSearchComputeObjectiveAndGradient(ls,neP->W2,f,g);
123: g_computed=PETSC_TRUE;
124: }
125:
126: TaoLineSearchMonitor(ls, i+1, *f, ls->step);
128: if (0 == i) {
129: ls->f_fullstep = *f;
130: }
132: actred = *f - finit;
133: VecCopy(neP->W2,neP->W1);
134: VecAXPY(neP->W1,-1.0,x); /* W1 = W2 - X */
135: VecDot(neP->W1,neP->Gold,&prered);
137: if (PetscAbsReal(prered)<1.0e-100) prered=1.0e-12;
138: rho = actred/prered;
140: /*
141: If sufficient progress has been obtained, accept the
142: point. Otherwise, backtrack.
143: */
145: if (actred > 0) {
146: PetscInfo(ls,"Step resulted in ascent, rejecting.\n");
147: ls->step = (ls->step)/2;
148: } else if (rho > ls->ftol){
149: break;
150: } else{
151: ls->step = (ls->step)/2;
152: }
154: /* Convergence testing */
156: if (ls->step <= ls->stepmin || ls->step >= ls->stepmax) {
157: ls->reason = TAOLINESEARCH_HALTED_OTHER;
158: PetscInfo(ls,"Rounding errors may prevent further progress. May not be a step satisfying\n");
159: PetscInfo(ls,"sufficient decrease and curvature conditions. Tolerances may be too small.\n");
160: break;
161: }
162: if (ls->step == ls->stepmax) {
163: PetscInfo1(ls,"Step is at the upper bound, stepmax (%g)\n",(double)ls->stepmax);
164: ls->reason = TAOLINESEARCH_HALTED_UPPERBOUND;
165: break;
166: }
167: if (ls->step == ls->stepmin) {
168: PetscInfo1(ls,"Step is at the lower bound, stepmin (%g)\n",(double)ls->stepmin);
169: ls->reason = TAOLINESEARCH_HALTED_LOWERBOUND;
170: break;
171: }
172: if ((ls->nfeval+ls->nfgeval) >= ls->max_funcs) {
173: PetscInfo2(ls,"Number of line search function evals (%D) > maximum (%D)\n",ls->nfeval+ls->nfgeval,ls->max_funcs);
174: ls->reason = TAOLINESEARCH_HALTED_MAXFCN;
175: break;
176: }
177: if ((neP->bracket) && (ls->stepmax - ls->stepmin <= ls->rtol*ls->stepmax)){
178: PetscInfo1(ls,"Relative width of interval of uncertainty is at most rtol (%g)\n",(double)ls->rtol);
179: ls->reason = TAOLINESEARCH_HALTED_RTOL;
180: break;
181: }
182: }
183: PetscInfo2(ls,"%D function evals in line search, step = %g\n",ls->nfeval+ls->nfgeval,(double)ls->step);
184: /* set new solution vector and compute gradient if necessary */
185: VecCopy(neP->W2, x);
186: if (ls->reason == TAOLINESEARCH_CONTINUE_ITERATING) {
187: ls->reason = TAOLINESEARCH_SUCCESS;
188: }
189: if (!g_computed) {
190: TaoLineSearchComputeGradient(ls,x,g);
191: }
192: return(0);
193: }
195: /* ---------------------------------------------------------- */
197: /*MC
198: TAOLINESEARCHGPCG - Special line-search method for the Gradient-Projected Conjugate Gradient (TAOGPCG) algorithm.
199: Should not be used with any other algorithm.
201: Level: developer
203: .keywords: Tao, linesearch
204: M*/
205: PETSC_EXTERN PetscErrorCode TaoLineSearchCreate_GPCG(TaoLineSearch ls)
206: {
207: PetscErrorCode ierr;
208: TaoLineSearch_GPCG *neP;
211: ls->ftol = 0.05;
212: ls->rtol = 0.0;
213: ls->gtol = 0.0;
214: ls->stepmin = 1.0e-20;
215: ls->stepmax = 1.0e+20;
216: ls->nfeval = 0;
217: ls->max_funcs = 30;
218: ls->step = 1.0;
220: PetscNewLog(ls,&neP);
221: neP->bracket = 0;
222: neP->infoc = 1;
223: ls->data = (void*)neP;
225: ls->ops->setup = 0;
226: ls->ops->reset = 0;
227: ls->ops->apply=TaoLineSearchApply_GPCG;
228: ls->ops->view =TaoLineSearchView_GPCG;
229: ls->ops->destroy=TaoLineSearchDestroy_GPCG;
230: ls->ops->setfromoptions=0;
231: return(0);
232: }