xref: /petsc/src/tao/linesearch/impls/owarmijo/owarmijo.c (revision 65f8aed5f7eaa1e2ef2ddeffe666264e0669c876)
1 
2 #include <petsc/private/taolinesearchimpl.h>
3 #include <../src/tao/linesearch/impls/owarmijo/owarmijo.h>
4 
5 #define REPLACE_FIFO 1
6 #define REPLACE_MRU  2
7 
8 #define REFERENCE_MAX  1
9 #define REFERENCE_AVE  2
10 #define REFERENCE_MEAN 3
11 
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;
18 
19   PetscFunctionBegin;
20   ierr=VecGetOwnershipRange(w,&low,&high);CHKERRQ(ierr);
21   ierr=VecGetOwnershipRange(x,&low1,&high1);CHKERRQ(ierr);
22   ierr=VecGetOwnershipRange(gv,&low2,&high2);CHKERRQ(ierr);
23 
24   *gdx=0.0;
25   ierr = VecGetArray(w,&wptr);CHKERRQ(ierr);
26   ierr = VecGetArrayRead(x,&xptr);CHKERRQ(ierr);
27   ierr = VecGetArrayRead(gv,&gptr);CHKERRQ(ierr);
28 
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   ierr = VecRestoreArray(w,&wptr);CHKERRQ(ierr);
34   ierr = VecRestoreArrayRead(x,&xptr);CHKERRQ(ierr);
35   ierr = VecRestoreArrayRead(gv,&gptr);CHKERRQ(ierr);
36   PetscFunctionReturn(0);
37 }
38 
39 static PetscErrorCode TaoLineSearchDestroy_OWArmijo(TaoLineSearch ls)
40 {
41   TaoLineSearch_OWARMIJO *armP = (TaoLineSearch_OWARMIJO *)ls->data;
42   PetscErrorCode         ierr;
43 
44   PetscFunctionBegin;
45   ierr = PetscFree(armP->memory);CHKERRQ(ierr);
46   if (armP->x) {
47     ierr = PetscObjectDereference((PetscObject)armP->x);CHKERRQ(ierr);
48   }
49   ierr = VecDestroy(&armP->work);CHKERRQ(ierr);
50   ierr = PetscFree(ls->data);CHKERRQ(ierr);
51   PetscFunctionReturn(0);
52 }
53 
54 static PetscErrorCode TaoLineSearchSetFromOptions_OWArmijo(PetscOptionItems *PetscOptionsObject,TaoLineSearch ls)
55 {
56   TaoLineSearch_OWARMIJO *armP = (TaoLineSearch_OWARMIJO *)ls->data;
57   PetscErrorCode         ierr;
58 
59   PetscFunctionBegin;
60   ierr = PetscOptionsHead(PetscOptionsObject,"OWArmijo linesearch options");CHKERRQ(ierr);
61   ierr = PetscOptionsReal("-tao_ls_OWArmijo_alpha", "initial reference constant", "", armP->alpha, &armP->alpha,NULL);CHKERRQ(ierr);
62   ierr = PetscOptionsReal("-tao_ls_OWArmijo_beta_inf", "decrease constant one", "", armP->beta_inf, &armP->beta_inf,NULL);CHKERRQ(ierr);
63   ierr = PetscOptionsReal("-tao_ls_OWArmijo_beta", "decrease constant", "", armP->beta, &armP->beta,NULL);CHKERRQ(ierr);
64   ierr = PetscOptionsReal("-tao_ls_OWArmijo_sigma", "acceptance constant", "", armP->sigma, &armP->sigma,NULL);CHKERRQ(ierr);
65   ierr = PetscOptionsInt("-tao_ls_OWArmijo_memory_size", "number of historical elements", "", armP->memorySize, &armP->memorySize,NULL);CHKERRQ(ierr);
66   ierr = PetscOptionsInt("-tao_ls_OWArmijo_reference_policy", "policy for updating reference value", "", armP->referencePolicy, &armP->referencePolicy,NULL);CHKERRQ(ierr);
67   ierr = PetscOptionsInt("-tao_ls_OWArmijo_replacement_policy", "policy for updating memory", "", armP->replacementPolicy, &armP->replacementPolicy,NULL);CHKERRQ(ierr);
68   ierr = PetscOptionsBool("-tao_ls_OWArmijo_nondescending","Use nondescending OWArmijo algorithm","",armP->nondescending,&armP->nondescending,NULL);CHKERRQ(ierr);
69   ierr = PetscOptionsTail();CHKERRQ(ierr);
70   PetscFunctionReturn(0);
71 }
72 
73 static PetscErrorCode TaoLineSearchView_OWArmijo(TaoLineSearch ls, PetscViewer pv)
74 {
75   TaoLineSearch_OWARMIJO *armP = (TaoLineSearch_OWARMIJO *)ls->data;
76   PetscBool              isascii;
77   PetscErrorCode         ierr;
78 
79   PetscFunctionBegin;
80   ierr = PetscObjectTypeCompare((PetscObject)pv, PETSCVIEWERASCII, &isascii);CHKERRQ(ierr);
81   if (isascii) {
82     ierr=PetscViewerASCIIPrintf(pv,"  OWArmijo linesearch",armP->alpha);CHKERRQ(ierr);
83     if (armP->nondescending) {
84       ierr = PetscViewerASCIIPrintf(pv, " (nondescending)");CHKERRQ(ierr);
85     }
86     ierr=PetscViewerASCIIPrintf(pv,": alpha=%g beta=%g ",(double)armP->alpha,(double)armP->beta);CHKERRQ(ierr);
87     ierr=PetscViewerASCIIPrintf(pv,"sigma=%g ",(double)armP->sigma);CHKERRQ(ierr);
88     ierr=PetscViewerASCIIPrintf(pv,"memsize=%D\n",armP->memorySize);CHKERRQ(ierr);
89   }
90   PetscFunctionReturn(0);
91 }
92 
93 /* @ TaoApply_OWArmijo - This routine performs a linesearch. It
94    backtracks until the (nonmonotone) OWArmijo conditions are satisfied.
95 
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
104 
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
110 
111    Info is set to one of:
112 .   0 - the line search succeeds; the sufficient decrease
113    condition and the directional derivative condition hold
114 
115    negative number if an input parameter is invalid
116 -   -1 -  step < 0
117 
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, its=0;
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;
133 
134   PetscFunctionBegin;
135   ierr = PetscObjectGetComm((PetscObject)ls,&comm);CHKERRQ(ierr);
136   fact = 0.0;
137   ls->nfeval=0;
138   ls->reason = TAOLINESEARCH_CONTINUE_ITERATING;
139   if (!armP->work) {
140     ierr = VecDuplicate(x,&armP->work);CHKERRQ(ierr);
141     armP->x = x;
142     ierr = PetscObjectReference((PetscObject)armP->x);CHKERRQ(ierr);
143   } else if (x != armP->x) {
144     ierr = VecDestroy(&armP->work);CHKERRQ(ierr);
145     ierr = VecDuplicate(x,&armP->work);CHKERRQ(ierr);
146     ierr = PetscObjectDereference((PetscObject)armP->x);CHKERRQ(ierr);
147     armP->x = x;
148     ierr = PetscObjectReference((PetscObject)armP->x);CHKERRQ(ierr);
149   }
150 
151   ierr = TaoLineSearchMonitor(ls, 0, *f, 0.0);CHKERRQ(ierr);
152 
153   /* Check linesearch parameters */
154   if (armP->alpha < 1) {
155     ierr = PetscInfo1(ls,"OWArmijo line search error: alpha (%g) < 1\n", (double)armP->alpha);CHKERRQ(ierr);
156     ls->reason=TAOLINESEARCH_FAILED_BADPARAMETER;
157   } else if ((armP->beta <= 0) || (armP->beta >= 1)) {
158     ierr = PetscInfo1(ls,"OWArmijo line search error: beta (%g) invalid\n", (double)armP->beta);CHKERRQ(ierr);
159     ls->reason=TAOLINESEARCH_FAILED_BADPARAMETER;
160   } else if ((armP->beta_inf <= 0) || (armP->beta_inf >= 1)) {
161     ierr = PetscInfo1(ls,"OWArmijo line search error: beta_inf (%g) invalid\n", (double)armP->beta_inf);CHKERRQ(ierr);
162     ls->reason=TAOLINESEARCH_FAILED_BADPARAMETER;
163   } else if ((armP->sigma <= 0) || (armP->sigma >= 0.5)) {
164     ierr = PetscInfo1(ls,"OWArmijo line search error: sigma (%g) invalid\n", (double)armP->sigma);CHKERRQ(ierr);
165     ls->reason=TAOLINESEARCH_FAILED_BADPARAMETER;
166   } else if (armP->memorySize < 1) {
167     ierr = PetscInfo1(ls,"OWArmijo line search error: memory_size (%D) < 1\n", armP->memorySize);CHKERRQ(ierr);
168     ls->reason=TAOLINESEARCH_FAILED_BADPARAMETER;
169   }  else if ((armP->referencePolicy != REFERENCE_MAX) && (armP->referencePolicy != REFERENCE_AVE) && (armP->referencePolicy != REFERENCE_MEAN)) {
170     ierr = PetscInfo(ls,"OWArmijo line search error: reference_policy invalid\n");CHKERRQ(ierr);
171     ls->reason=TAOLINESEARCH_FAILED_BADPARAMETER;
172   } else if ((armP->replacementPolicy != REPLACE_FIFO) && (armP->replacementPolicy != REPLACE_MRU)) {
173     ierr = PetscInfo(ls,"OWArmijo line search error: replacement_policy invalid\n");CHKERRQ(ierr);
174     ls->reason=TAOLINESEARCH_FAILED_BADPARAMETER;
175   } else if (PetscIsInfOrNanReal(*f)) {
176     ierr = PetscInfo(ls,"OWArmijo line search error: initial function inf or nan\n");CHKERRQ(ierr);
177     ls->reason=TAOLINESEARCH_FAILED_BADPARAMETER;
178   }
179 
180   if (ls->reason != TAOLINESEARCH_CONTINUE_ITERATING) PetscFunctionReturn(0);
181 
182   /* Check to see of the memory has been allocated.  If not, allocate
183      the historical array and populate it with the initial function
184      values. */
185   if (!armP->memory) {
186     ierr = PetscMalloc1(armP->memorySize, &armP->memory );CHKERRQ(ierr);
187   }
188 
189   if (!armP->memorySetup) {
190     for (i = 0; i < armP->memorySize; i++) {
191       armP->memory[i] = armP->alpha*(*f);
192     }
193     armP->current = 0;
194     armP->lastReference = armP->memory[0];
195     armP->memorySetup=PETSC_TRUE;
196   }
197 
198   /* Calculate reference value (MAX) */
199   ref = armP->memory[0];
200   idx = 0;
201 
202   for (i = 1; i < armP->memorySize; i++) {
203     if (armP->memory[i] > ref) {
204       ref = armP->memory[i];
205       idx = i;
206     }
207   }
208 
209   if (armP->referencePolicy == REFERENCE_AVE) {
210     ref = 0;
211     for (i = 0; i < armP->memorySize; i++) {
212       ref += armP->memory[i];
213     }
214     ref = ref / armP->memorySize;
215     ref = PetscMax(ref, armP->memory[armP->current]);
216   } else if (armP->referencePolicy == REFERENCE_MEAN) {
217     ref = PetscMin(ref, 0.5*(armP->lastReference + armP->memory[armP->current]));
218   }
219 
220   if (armP->nondescending) {
221     fact = armP->sigma;
222   }
223 
224   ierr = VecDuplicate(g,&g_old);CHKERRQ(ierr);
225   ierr = VecCopy(g,g_old);CHKERRQ(ierr);
226 
227   ls->step = ls->initstep;
228   while (ls->step >= owlqn_minstep && ls->nfeval < ls->max_funcs) {
229     /* Calculate iterate */
230     ++its;
231     ierr = VecCopy(x,armP->work);CHKERRQ(ierr);
232     ierr = VecAXPY(armP->work,ls->step,s);CHKERRQ(ierr);
233 
234     partgdx=0.0;
235     ierr = ProjWork_OWLQN(armP->work,x,g_old,&partgdx);CHKERRQ(ierr);
236     ierr = MPIU_Allreduce(&partgdx,&gdx,1,MPIU_REAL,MPIU_SUM,comm);CHKERRQ(ierr);
237 
238     /* Check the condition of gdx */
239     if (PetscIsInfOrNanReal(gdx)) {
240       ierr = PetscInfo1(ls,"Initial Line Search step * g is Inf or Nan (%g)\n",(double)gdx);CHKERRQ(ierr);
241       ls->reason=TAOLINESEARCH_FAILED_INFORNAN;
242       PetscFunctionReturn(0);
243     }
244     if (gdx >= 0.0) {
245       ierr = PetscInfo1(ls,"Initial Line Search step is not descent direction (g's=%g)\n",(double)gdx);CHKERRQ(ierr);
246       ls->reason = TAOLINESEARCH_FAILED_ASCENT;
247       PetscFunctionReturn(0);
248     }
249 
250     /* Calculate function at new iterate */
251     ierr = TaoLineSearchComputeObjectiveAndGradient(ls,armP->work,f,g);CHKERRQ(ierr);
252     g_computed=PETSC_TRUE;
253 
254     ierr = TaoLineSearchMonitor(ls, its, *f, ls->step);CHKERRQ(ierr);
255 
256     if (ls->step == ls->initstep) {
257       ls->f_fullstep = *f;
258     }
259 
260     if (PetscIsInfOrNanReal(*f)) {
261       ls->step *= armP->beta_inf;
262     } else {
263       /* Check descent condition */
264       if (armP->nondescending && *f <= ref - ls->step*fact*ref) break;
265       if (!armP->nondescending && *f <= ref + armP->sigma * gdx) break;
266       ls->step *= armP->beta;
267     }
268   }
269   ierr = VecDestroy(&g_old);CHKERRQ(ierr);
270 
271   /* Check termination */
272   if (PetscIsInfOrNanReal(*f)) {
273     ierr = PetscInfo(ls, "Function is inf or nan.\n");CHKERRQ(ierr);
274     ls->reason = TAOLINESEARCH_FAILED_BADPARAMETER;
275   } else if (ls->step < owlqn_minstep) {
276     ierr = PetscInfo(ls, "Step length is below tolerance.\n");CHKERRQ(ierr);
277     ls->reason = TAOLINESEARCH_HALTED_RTOL;
278   } else if (ls->nfeval >= ls->max_funcs) {
279     ierr = PetscInfo2(ls, "Number of line search function evals (%D) > maximum allowed (%D)\n",ls->nfeval, ls->max_funcs);CHKERRQ(ierr);
280     ls->reason = TAOLINESEARCH_HALTED_MAXFCN;
281   }
282   if (ls->reason) PetscFunctionReturn(0);
283 
284   /* Successful termination, update memory */
285   ls->reason = TAOLINESEARCH_SUCCESS;
286   armP->lastReference = ref;
287   if (armP->replacementPolicy == REPLACE_FIFO) {
288     armP->memory[armP->current++] = *f;
289     if (armP->current >= armP->memorySize) {
290       armP->current = 0;
291     }
292   } else {
293     armP->current = idx;
294     armP->memory[idx] = *f;
295   }
296 
297   /* Update iterate and compute gradient */
298   ierr = VecCopy(armP->work,x);CHKERRQ(ierr);
299   if (!g_computed) {
300     ierr = TaoLineSearchComputeGradient(ls, x, g);CHKERRQ(ierr);
301   }
302   ierr = PetscInfo2(ls, "%D function evals in line search, step = %10.4f\n",ls->nfeval, (double)ls->step);CHKERRQ(ierr);
303   PetscFunctionReturn(0);
304 }
305 
306 PETSC_EXTERN PetscErrorCode TaoLineSearchCreate_OWArmijo(TaoLineSearch ls)
307 {
308   TaoLineSearch_OWARMIJO *armP;
309   PetscErrorCode         ierr;
310 
311   PetscFunctionBegin;
312   PetscValidHeaderSpecific(ls,TAOLINESEARCH_CLASSID,1);
313   ierr = PetscNewLog(ls,&armP);CHKERRQ(ierr);
314 
315   armP->memory = NULL;
316   armP->alpha = 1.0;
317   armP->beta = 0.25;
318   armP->beta_inf = 0.25;
319   armP->sigma = 1e-4;
320   armP->memorySize = 1;
321   armP->referencePolicy = REFERENCE_MAX;
322   armP->replacementPolicy = REPLACE_MRU;
323   armP->nondescending=PETSC_FALSE;
324   ls->data = (void*)armP;
325   ls->initstep=0.1;
326   ls->ops->setup=0;
327   ls->ops->reset=0;
328   ls->ops->apply=TaoLineSearchApply_OWArmijo;
329   ls->ops->view = TaoLineSearchView_OWArmijo;
330   ls->ops->destroy = TaoLineSearchDestroy_OWArmijo;
331   ls->ops->setfromoptions = TaoLineSearchSetFromOptions_OWArmijo;
332   PetscFunctionReturn(0);
333 }
334 
335