Actual source code: owarmijo.c
petsc-3.9.4 2018-09-11
2: #include <petsc/private/taolinesearchimpl.h>
3: #include <../src/tao/linesearch/impls/owarmijo/owarmijo.h>
5: #define REPLACE_FIFO 1
6: #define REPLACE_MRU 2
8: #define REFERENCE_MAX 1
9: #define REFERENCE_AVE 2
10: #define REFERENCE_MEAN 3
12: static PetscErrorCode ProjWork_OWLQN(Vec w,Vec x,Vec gv,PetscReal *gdx)
13: {
14: const PetscReal *xptr,*gptr;
15: PetscReal *wptr;
16: PetscErrorCode ierr;
17: PetscInt low,high,low1,high1,low2,high2,i;
20: ierr=VecGetOwnershipRange(w,&low,&high);
21: ierr=VecGetOwnershipRange(x,&low1,&high1);
22: ierr=VecGetOwnershipRange(gv,&low2,&high2);
24: *gdx=0.0;
25: VecGetArray(w,&wptr);
26: VecGetArrayRead(x,&xptr);
27: VecGetArrayRead(gv,&gptr);
29: for (i=0;i<high-low;i++) {
30: if (xptr[i]*wptr[i]<0.0) wptr[i]=0.0;
31: *gdx = *gdx + gptr[i]*(wptr[i]-xptr[i]);
32: }
33: VecRestoreArray(w,&wptr);
34: VecRestoreArrayRead(x,&xptr);
35: VecRestoreArrayRead(gv,&gptr);
36: return(0);
37: }
39: static PetscErrorCode TaoLineSearchDestroy_OWArmijo(TaoLineSearch ls)
40: {
41: TaoLineSearch_OWARMIJO *armP = (TaoLineSearch_OWARMIJO *)ls->data;
42: PetscErrorCode ierr;
45: PetscFree(armP->memory);
46: if (armP->x) {
47: PetscObjectDereference((PetscObject)armP->x);
48: }
49: VecDestroy(&armP->work);
50: PetscFree(ls->data);
51: return(0);
52: }
54: static PetscErrorCode TaoLineSearchSetFromOptions_OWArmijo(PetscOptionItems *PetscOptionsObject,TaoLineSearch ls)
55: {
56: TaoLineSearch_OWARMIJO *armP = (TaoLineSearch_OWARMIJO *)ls->data;
57: PetscErrorCode ierr;
60: PetscOptionsHead(PetscOptionsObject,"OWArmijo linesearch options");
61: PetscOptionsReal("-tao_ls_OWArmijo_alpha", "initial reference constant", "", armP->alpha, &armP->alpha,NULL);
62: PetscOptionsReal("-tao_ls_OWArmijo_beta_inf", "decrease constant one", "", armP->beta_inf, &armP->beta_inf,NULL);
63: PetscOptionsReal("-tao_ls_OWArmijo_beta", "decrease constant", "", armP->beta, &armP->beta,NULL);
64: PetscOptionsReal("-tao_ls_OWArmijo_sigma", "acceptance constant", "", armP->sigma, &armP->sigma,NULL);
65: PetscOptionsInt("-tao_ls_OWArmijo_memory_size", "number of historical elements", "", armP->memorySize, &armP->memorySize,NULL);
66: PetscOptionsInt("-tao_ls_OWArmijo_reference_policy", "policy for updating reference value", "", armP->referencePolicy, &armP->referencePolicy,NULL);
67: PetscOptionsInt("-tao_ls_OWArmijo_replacement_policy", "policy for updating memory", "", armP->replacementPolicy, &armP->replacementPolicy,NULL);
68: PetscOptionsBool("-tao_ls_OWArmijo_nondescending","Use nondescending OWArmijo algorithm","",armP->nondescending,&armP->nondescending,NULL);
69: PetscOptionsTail();
70: return(0);
71: }
73: static PetscErrorCode TaoLineSearchView_OWArmijo(TaoLineSearch ls, PetscViewer pv)
74: {
75: TaoLineSearch_OWARMIJO *armP = (TaoLineSearch_OWARMIJO *)ls->data;
76: PetscBool isascii;
77: PetscErrorCode ierr;
80: PetscObjectTypeCompare((PetscObject)pv, PETSCVIEWERASCII, &isascii);
81: if (isascii) {
82: ierr=PetscViewerASCIIPrintf(pv," OWArmijo linesearch",armP->alpha);
83: if (armP->nondescending) {
84: PetscViewerASCIIPrintf(pv, " (nondescending)");
85: }
86: ierr=PetscViewerASCIIPrintf(pv,": alpha=%g beta=%g ",(double)armP->alpha,(double)armP->beta);
87: ierr=PetscViewerASCIIPrintf(pv,"sigma=%g ",(double)armP->sigma);
88: ierr=PetscViewerASCIIPrintf(pv,"memsize=%D\n",armP->memorySize);
89: }
90: return(0);
91: }
93: /* @ TaoApply_OWArmijo - This routine performs a linesearch. It
94: backtracks until the (nonmonotone) OWArmijo conditions are satisfied.
96: Input Parameters:
97: + tao - TAO_SOLVER context
98: . X - current iterate (on output X contains new iterate, X + step*S)
99: . S - search direction
100: . f - merit function evaluated at X
101: . G - gradient of merit function evaluated at X
102: . W - work vector
103: - step - initial estimate of step length
105: Output parameters:
106: + f - merit function evaluated at new iterate, X + step*S
107: . G - gradient of merit function evaluated at new iterate, X + step*S
108: . X - new iterate
109: - step - final step length
111: Info is set to one of:
112: . 0 - the line search succeeds; the sufficient decrease
113: condition and the directional derivative condition hold
115: negative number if an input parameter is invalid
116: - -1 - step < 0
118: positive number > 1 if the line search otherwise terminates
119: + 1 - Step is at the lower bound, stepmin.
120: @ */
121: static PetscErrorCode TaoLineSearchApply_OWArmijo(TaoLineSearch ls, Vec x, PetscReal *f, Vec g, Vec s)
122: {
123: TaoLineSearch_OWARMIJO *armP = (TaoLineSearch_OWARMIJO *)ls->data;
124: PetscErrorCode ierr;
125: PetscInt i;
126: PetscReal fact, ref, gdx;
127: PetscInt idx;
128: PetscBool g_computed=PETSC_FALSE; /* to prevent extra gradient computation */
129: Vec g_old;
130: PetscReal owlqn_minstep=0.005;
131: PetscReal partgdx;
132: MPI_Comm comm;
135: PetscObjectGetComm((PetscObject)ls,&comm);
136: fact = 0.0;
137: ls->nfeval=0;
138: ls->reason = TAOLINESEARCH_CONTINUE_ITERATING;
139: if (!armP->work) {
140: VecDuplicate(x,&armP->work);
141: armP->x = x;
142: PetscObjectReference((PetscObject)armP->x);
143: } else if (x != armP->x) {
144: VecDestroy(&armP->work);
145: VecDuplicate(x,&armP->work);
146: PetscObjectDereference((PetscObject)armP->x);
147: armP->x = x;
148: PetscObjectReference((PetscObject)armP->x);
149: }
151: /* Check linesearch parameters */
152: if (armP->alpha < 1) {
153: PetscInfo1(ls,"OWArmijo line search error: alpha (%g) < 1\n", (double)armP->alpha);
154: ls->reason=TAOLINESEARCH_FAILED_BADPARAMETER;
155: } else if ((armP->beta <= 0) || (armP->beta >= 1)) {
156: PetscInfo1(ls,"OWArmijo line search error: beta (%g) invalid\n", (double)armP->beta);
157: ls->reason=TAOLINESEARCH_FAILED_BADPARAMETER;
158: } else if ((armP->beta_inf <= 0) || (armP->beta_inf >= 1)) {
159: PetscInfo1(ls,"OWArmijo line search error: beta_inf (%g) invalid\n", (double)armP->beta_inf);
160: ls->reason=TAOLINESEARCH_FAILED_BADPARAMETER;
161: } else if ((armP->sigma <= 0) || (armP->sigma >= 0.5)) {
162: PetscInfo1(ls,"OWArmijo line search error: sigma (%g) invalid\n", (double)armP->sigma);
163: ls->reason=TAOLINESEARCH_FAILED_BADPARAMETER;
164: } else if (armP->memorySize < 1) {
165: PetscInfo1(ls,"OWArmijo line search error: memory_size (%D) < 1\n", armP->memorySize);
166: ls->reason=TAOLINESEARCH_FAILED_BADPARAMETER;
167: } else if ((armP->referencePolicy != REFERENCE_MAX) && (armP->referencePolicy != REFERENCE_AVE) && (armP->referencePolicy != REFERENCE_MEAN)) {
168: PetscInfo(ls,"OWArmijo line search error: reference_policy invalid\n");
169: ls->reason=TAOLINESEARCH_FAILED_BADPARAMETER;
170: } else if ((armP->replacementPolicy != REPLACE_FIFO) && (armP->replacementPolicy != REPLACE_MRU)) {
171: PetscInfo(ls,"OWArmijo line search error: replacement_policy invalid\n");
172: ls->reason=TAOLINESEARCH_FAILED_BADPARAMETER;
173: } else if (PetscIsInfOrNanReal(*f)) {
174: PetscInfo(ls,"OWArmijo line search error: initial function inf or nan\n");
175: ls->reason=TAOLINESEARCH_FAILED_BADPARAMETER;
176: }
178: if (ls->reason != TAOLINESEARCH_CONTINUE_ITERATING) return(0);
180: /* Check to see of the memory has been allocated. If not, allocate
181: the historical array and populate it with the initial function
182: values. */
183: if (!armP->memory) {
184: PetscMalloc1(armP->memorySize, &armP->memory );
185: }
187: if (!armP->memorySetup) {
188: for (i = 0; i < armP->memorySize; i++) {
189: armP->memory[i] = armP->alpha*(*f);
190: }
191: armP->current = 0;
192: armP->lastReference = armP->memory[0];
193: armP->memorySetup=PETSC_TRUE;
194: }
196: /* Calculate reference value (MAX) */
197: ref = armP->memory[0];
198: idx = 0;
200: for (i = 1; i < armP->memorySize; i++) {
201: if (armP->memory[i] > ref) {
202: ref = armP->memory[i];
203: idx = i;
204: }
205: }
207: if (armP->referencePolicy == REFERENCE_AVE) {
208: ref = 0;
209: for (i = 0; i < armP->memorySize; i++) {
210: ref += armP->memory[i];
211: }
212: ref = ref / armP->memorySize;
213: ref = PetscMax(ref, armP->memory[armP->current]);
214: } else if (armP->referencePolicy == REFERENCE_MEAN) {
215: ref = PetscMin(ref, 0.5*(armP->lastReference + armP->memory[armP->current]));
216: }
218: if (armP->nondescending) {
219: fact = armP->sigma;
220: }
222: VecDuplicate(g,&g_old);
223: VecCopy(g,g_old);
225: ls->step = ls->initstep;
226: while (ls->step >= owlqn_minstep && ls->nfeval < ls->max_funcs) {
227: /* Calculate iterate */
228: VecCopy(x,armP->work);
229: VecAXPY(armP->work,ls->step,s);
231: partgdx=0.0;
232: ProjWork_OWLQN(armP->work,x,g_old,&partgdx);
233: MPIU_Allreduce(&partgdx,&gdx,1,MPIU_REAL,MPIU_SUM,comm);
235: /* Check the condition of gdx */
236: if (PetscIsInfOrNanReal(gdx)) {
237: PetscInfo1(ls,"Initial Line Search step * g is Inf or Nan (%g)\n",(double)gdx);
238: ls->reason=TAOLINESEARCH_FAILED_INFORNAN;
239: return(0);
240: }
241: if (gdx >= 0.0) {
242: PetscInfo1(ls,"Initial Line Search step is not descent direction (g's=%g)\n",(double)gdx);
243: ls->reason = TAOLINESEARCH_FAILED_ASCENT;
244: return(0);
245: }
247: /* Calculate function at new iterate */
248: TaoLineSearchComputeObjectiveAndGradient(ls,armP->work,f,g);
249: g_computed=PETSC_TRUE;
251: if (ls->step == ls->initstep) {
252: ls->f_fullstep = *f;
253: }
255: if (PetscIsInfOrNanReal(*f)) {
256: ls->step *= armP->beta_inf;
257: } else {
258: /* Check descent condition */
259: if (armP->nondescending && *f <= ref - ls->step*fact*ref) break;
260: if (!armP->nondescending && *f <= ref + armP->sigma * gdx) break;
261: ls->step *= armP->beta;
262: }
263: }
264: VecDestroy(&g_old);
266: /* Check termination */
267: if (PetscIsInfOrNanReal(*f)) {
268: PetscInfo(ls, "Function is inf or nan.\n");
269: ls->reason = TAOLINESEARCH_FAILED_BADPARAMETER;
270: } else if (ls->step < owlqn_minstep) {
271: PetscInfo(ls, "Step length is below tolerance.\n");
272: ls->reason = TAOLINESEARCH_HALTED_RTOL;
273: } else if (ls->nfeval >= ls->max_funcs) {
274: PetscInfo2(ls, "Number of line search function evals (%D) > maximum allowed (%D)\n",ls->nfeval, ls->max_funcs);
275: ls->reason = TAOLINESEARCH_HALTED_MAXFCN;
276: }
277: if (ls->reason) return(0);
279: /* Successful termination, update memory */
280: ls->reason = TAOLINESEARCH_SUCCESS;
281: armP->lastReference = ref;
282: if (armP->replacementPolicy == REPLACE_FIFO) {
283: armP->memory[armP->current++] = *f;
284: if (armP->current >= armP->memorySize) {
285: armP->current = 0;
286: }
287: } else {
288: armP->current = idx;
289: armP->memory[idx] = *f;
290: }
292: /* Update iterate and compute gradient */
293: VecCopy(armP->work,x);
294: if (!g_computed) {
295: TaoLineSearchComputeGradient(ls, x, g);
296: }
297: PetscInfo2(ls, "%D function evals in line search, step = %10.4f\n",ls->nfeval, (double)ls->step);
298: return(0);
299: }
301: PETSC_EXTERN PetscErrorCode TaoLineSearchCreate_OWArmijo(TaoLineSearch ls)
302: {
303: TaoLineSearch_OWARMIJO *armP;
304: PetscErrorCode ierr;
308: PetscNewLog(ls,&armP);
310: armP->memory = NULL;
311: armP->alpha = 1.0;
312: armP->beta = 0.25;
313: armP->beta_inf = 0.25;
314: armP->sigma = 1e-4;
315: armP->memorySize = 1;
316: armP->referencePolicy = REFERENCE_MAX;
317: armP->replacementPolicy = REPLACE_MRU;
318: armP->nondescending=PETSC_FALSE;
319: ls->data = (void*)armP;
320: ls->initstep=0.1;
321: ls->ops->setup=0;
322: ls->ops->reset=0;
323: ls->ops->apply=TaoLineSearchApply_OWArmijo;
324: ls->ops->view = TaoLineSearchView_OWArmijo;
325: ls->ops->destroy = TaoLineSearchDestroy_OWArmijo;
326: ls->ops->setfromoptions = TaoLineSearchSetFromOptions_OWArmijo;
327: return(0);
328: }