xref: /petsc/src/tao/linesearch/impls/owarmijo/owarmijo.c (revision 7efe37a1cedd385a2f501b843d47cdf14dfb49ea)
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   PetscInt         low, high, low1, high1, low2, high2, i;
17 
18   PetscFunctionBegin;
19   PetscCall(VecGetOwnershipRange(w, &low, &high));
20   PetscCall(VecGetOwnershipRange(x, &low1, &high1));
21   PetscCall(VecGetOwnershipRange(gv, &low2, &high2));
22 
23   *gdx = 0.0;
24   PetscCall(VecGetArray(w, &wptr));
25   PetscCall(VecGetArrayRead(x, &xptr));
26   PetscCall(VecGetArrayRead(gv, &gptr));
27 
28   for (i = 0; i < high - low; i++) {
29     if (xptr[i] * wptr[i] < 0.0) wptr[i] = 0.0;
30     *gdx = *gdx + gptr[i] * (wptr[i] - xptr[i]);
31   }
32   PetscCall(VecRestoreArray(w, &wptr));
33   PetscCall(VecRestoreArrayRead(x, &xptr));
34   PetscCall(VecRestoreArrayRead(gv, &gptr));
35   PetscFunctionReturn(0);
36 }
37 
38 static PetscErrorCode TaoLineSearchDestroy_OWArmijo(TaoLineSearch ls)
39 {
40   TaoLineSearch_OWARMIJO *armP = (TaoLineSearch_OWARMIJO *)ls->data;
41 
42   PetscFunctionBegin;
43   PetscCall(PetscFree(armP->memory));
44   if (armP->x) PetscCall(PetscObjectDereference((PetscObject)armP->x));
45   PetscCall(VecDestroy(&armP->work));
46   PetscCall(PetscFree(ls->data));
47   PetscFunctionReturn(0);
48 }
49 
50 static PetscErrorCode TaoLineSearchSetFromOptions_OWArmijo(TaoLineSearch ls, PetscOptionItems *PetscOptionsObject)
51 {
52   TaoLineSearch_OWARMIJO *armP = (TaoLineSearch_OWARMIJO *)ls->data;
53 
54   PetscFunctionBegin;
55   PetscOptionsHeadBegin(PetscOptionsObject, "OWArmijo linesearch options");
56   PetscCall(PetscOptionsReal("-tao_ls_OWArmijo_alpha", "initial reference constant", "", armP->alpha, &armP->alpha, NULL));
57   PetscCall(PetscOptionsReal("-tao_ls_OWArmijo_beta_inf", "decrease constant one", "", armP->beta_inf, &armP->beta_inf, NULL));
58   PetscCall(PetscOptionsReal("-tao_ls_OWArmijo_beta", "decrease constant", "", armP->beta, &armP->beta, NULL));
59   PetscCall(PetscOptionsReal("-tao_ls_OWArmijo_sigma", "acceptance constant", "", armP->sigma, &armP->sigma, NULL));
60   PetscCall(PetscOptionsInt("-tao_ls_OWArmijo_memory_size", "number of historical elements", "", armP->memorySize, &armP->memorySize, NULL));
61   PetscCall(PetscOptionsInt("-tao_ls_OWArmijo_reference_policy", "policy for updating reference value", "", armP->referencePolicy, &armP->referencePolicy, NULL));
62   PetscCall(PetscOptionsInt("-tao_ls_OWArmijo_replacement_policy", "policy for updating memory", "", armP->replacementPolicy, &armP->replacementPolicy, NULL));
63   PetscCall(PetscOptionsBool("-tao_ls_OWArmijo_nondescending", "Use nondescending OWArmijo algorithm", "", armP->nondescending, &armP->nondescending, NULL));
64   PetscOptionsHeadEnd();
65   PetscFunctionReturn(0);
66 }
67 
68 static PetscErrorCode TaoLineSearchView_OWArmijo(TaoLineSearch ls, PetscViewer pv)
69 {
70   TaoLineSearch_OWARMIJO *armP = (TaoLineSearch_OWARMIJO *)ls->data;
71   PetscBool               isascii;
72 
73   PetscFunctionBegin;
74   PetscCall(PetscObjectTypeCompare((PetscObject)pv, PETSCVIEWERASCII, &isascii));
75   if (isascii) {
76     PetscCall(PetscViewerASCIIPrintf(pv, "  OWArmijo linesearch"));
77     if (armP->nondescending) PetscCall(PetscViewerASCIIPrintf(pv, " (nondescending)"));
78     PetscCall(PetscViewerASCIIPrintf(pv, ": alpha=%g beta=%g ", (double)armP->alpha, (double)armP->beta));
79     PetscCall(PetscViewerASCIIPrintf(pv, "sigma=%g ", (double)armP->sigma));
80     PetscCall(PetscViewerASCIIPrintf(pv, "memsize=%" PetscInt_FMT "\n", armP->memorySize));
81   }
82   PetscFunctionReturn(0);
83 }
84 
85 /* @ TaoApply_OWArmijo - This routine performs a linesearch. It
86    backtracks until the (nonmonotone) OWArmijo conditions are satisfied.
87 
88    Input Parameters:
89 +  tao - TAO_SOLVER context
90 .  X - current iterate (on output X contains new iterate, X + step*S)
91 .  S - search direction
92 .  f - merit function evaluated at X
93 .  G - gradient of merit function evaluated at X
94 .  W - work vector
95 -  step - initial estimate of step length
96 
97    Output parameters:
98 +  f - merit function evaluated at new iterate, X + step*S
99 .  G - gradient of merit function evaluated at new iterate, X + step*S
100 .  X - new iterate
101 -  step - final step length
102 
103    Info is set to one of:
104 .   0 - the line search succeeds; the sufficient decrease
105    condition and the directional derivative condition hold
106 
107    negative number if an input parameter is invalid
108 -   -1 -  step < 0
109 
110    positive number > 1 if the line search otherwise terminates
111 +    1 -  Step is at the lower bound, stepmin.
112 @ */
113 static PetscErrorCode TaoLineSearchApply_OWArmijo(TaoLineSearch ls, Vec x, PetscReal *f, Vec g, Vec s)
114 {
115   TaoLineSearch_OWARMIJO *armP = (TaoLineSearch_OWARMIJO *)ls->data;
116   PetscInt                i, its = 0;
117   PetscReal               fact, ref, gdx;
118   PetscInt                idx;
119   PetscBool               g_computed = PETSC_FALSE; /* to prevent extra gradient computation */
120   Vec                     g_old;
121   PetscReal               owlqn_minstep = 0.005;
122   PetscReal               partgdx;
123   MPI_Comm                comm;
124 
125   PetscFunctionBegin;
126   PetscCall(PetscObjectGetComm((PetscObject)ls, &comm));
127   fact       = 0.0;
128   ls->nfeval = 0;
129   ls->reason = TAOLINESEARCH_CONTINUE_ITERATING;
130   if (!armP->work) {
131     PetscCall(VecDuplicate(x, &armP->work));
132     armP->x = x;
133     PetscCall(PetscObjectReference((PetscObject)armP->x));
134   } else if (x != armP->x) {
135     PetscCall(VecDestroy(&armP->work));
136     PetscCall(VecDuplicate(x, &armP->work));
137     PetscCall(PetscObjectDereference((PetscObject)armP->x));
138     armP->x = x;
139     PetscCall(PetscObjectReference((PetscObject)armP->x));
140   }
141 
142   PetscCall(TaoLineSearchMonitor(ls, 0, *f, 0.0));
143 
144   /* Check linesearch parameters */
145   if (armP->alpha < 1) {
146     PetscCall(PetscInfo(ls, "OWArmijo line search error: alpha (%g) < 1\n", (double)armP->alpha));
147     ls->reason = TAOLINESEARCH_FAILED_BADPARAMETER;
148   } else if ((armP->beta <= 0) || (armP->beta >= 1)) {
149     PetscCall(PetscInfo(ls, "OWArmijo line search error: beta (%g) invalid\n", (double)armP->beta));
150     ls->reason = TAOLINESEARCH_FAILED_BADPARAMETER;
151   } else if ((armP->beta_inf <= 0) || (armP->beta_inf >= 1)) {
152     PetscCall(PetscInfo(ls, "OWArmijo line search error: beta_inf (%g) invalid\n", (double)armP->beta_inf));
153     ls->reason = TAOLINESEARCH_FAILED_BADPARAMETER;
154   } else if ((armP->sigma <= 0) || (armP->sigma >= 0.5)) {
155     PetscCall(PetscInfo(ls, "OWArmijo line search error: sigma (%g) invalid\n", (double)armP->sigma));
156     ls->reason = TAOLINESEARCH_FAILED_BADPARAMETER;
157   } else if (armP->memorySize < 1) {
158     PetscCall(PetscInfo(ls, "OWArmijo line search error: memory_size (%" PetscInt_FMT ") < 1\n", armP->memorySize));
159     ls->reason = TAOLINESEARCH_FAILED_BADPARAMETER;
160   } else if ((armP->referencePolicy != REFERENCE_MAX) && (armP->referencePolicy != REFERENCE_AVE) && (armP->referencePolicy != REFERENCE_MEAN)) {
161     PetscCall(PetscInfo(ls, "OWArmijo line search error: reference_policy invalid\n"));
162     ls->reason = TAOLINESEARCH_FAILED_BADPARAMETER;
163   } else if ((armP->replacementPolicy != REPLACE_FIFO) && (armP->replacementPolicy != REPLACE_MRU)) {
164     PetscCall(PetscInfo(ls, "OWArmijo line search error: replacement_policy invalid\n"));
165     ls->reason = TAOLINESEARCH_FAILED_BADPARAMETER;
166   } else if (PetscIsInfOrNanReal(*f)) {
167     PetscCall(PetscInfo(ls, "OWArmijo line search error: initial function inf or nan\n"));
168     ls->reason = TAOLINESEARCH_FAILED_BADPARAMETER;
169   }
170 
171   if (ls->reason != TAOLINESEARCH_CONTINUE_ITERATING) PetscFunctionReturn(0);
172 
173   /* Check to see of the memory has been allocated.  If not, allocate
174      the historical array and populate it with the initial function
175      values. */
176   if (!armP->memory) PetscCall(PetscMalloc1(armP->memorySize, &armP->memory));
177 
178   if (!armP->memorySetup) {
179     for (i = 0; i < armP->memorySize; i++) armP->memory[i] = armP->alpha * (*f);
180     armP->current       = 0;
181     armP->lastReference = armP->memory[0];
182     armP->memorySetup   = PETSC_TRUE;
183   }
184 
185   /* Calculate reference value (MAX) */
186   ref = armP->memory[0];
187   idx = 0;
188 
189   for (i = 1; i < armP->memorySize; i++) {
190     if (armP->memory[i] > ref) {
191       ref = armP->memory[i];
192       idx = i;
193     }
194   }
195 
196   if (armP->referencePolicy == REFERENCE_AVE) {
197     ref = 0;
198     for (i = 0; i < armP->memorySize; i++) ref += armP->memory[i];
199     ref = ref / armP->memorySize;
200     ref = PetscMax(ref, armP->memory[armP->current]);
201   } else if (armP->referencePolicy == REFERENCE_MEAN) {
202     ref = PetscMin(ref, 0.5 * (armP->lastReference + armP->memory[armP->current]));
203   }
204 
205   if (armP->nondescending) fact = armP->sigma;
206 
207   PetscCall(VecDuplicate(g, &g_old));
208   PetscCall(VecCopy(g, g_old));
209 
210   ls->step = ls->initstep;
211   while (ls->step >= owlqn_minstep && ls->nfeval < ls->max_funcs) {
212     /* Calculate iterate */
213     ++its;
214     PetscCall(VecWAXPY(armP->work, ls->step, s, x));
215 
216     partgdx = 0.0;
217     PetscCall(ProjWork_OWLQN(armP->work, x, g_old, &partgdx));
218     PetscCall(MPIU_Allreduce(&partgdx, &gdx, 1, MPIU_REAL, MPIU_SUM, comm));
219 
220     /* Check the condition of gdx */
221     if (PetscIsInfOrNanReal(gdx)) {
222       PetscCall(PetscInfo(ls, "Initial Line Search step * g is Inf or Nan (%g)\n", (double)gdx));
223       ls->reason = TAOLINESEARCH_FAILED_INFORNAN;
224       PetscFunctionReturn(0);
225     }
226     if (gdx >= 0.0) {
227       PetscCall(PetscInfo(ls, "Initial Line Search step is not descent direction (g's=%g)\n", (double)gdx));
228       ls->reason = TAOLINESEARCH_FAILED_ASCENT;
229       PetscFunctionReturn(0);
230     }
231 
232     /* Calculate function at new iterate */
233     PetscCall(TaoLineSearchComputeObjectiveAndGradient(ls, armP->work, f, g));
234     g_computed = PETSC_TRUE;
235 
236     PetscCall(TaoLineSearchMonitor(ls, its, *f, ls->step));
237 
238     if (ls->step == ls->initstep) ls->f_fullstep = *f;
239 
240     if (PetscIsInfOrNanReal(*f)) {
241       ls->step *= armP->beta_inf;
242     } else {
243       /* Check descent condition */
244       if (armP->nondescending && *f <= ref - ls->step * fact * ref) break;
245       if (!armP->nondescending && *f <= ref + armP->sigma * gdx) break;
246       ls->step *= armP->beta;
247     }
248   }
249   PetscCall(VecDestroy(&g_old));
250 
251   /* Check termination */
252   if (PetscIsInfOrNanReal(*f)) {
253     PetscCall(PetscInfo(ls, "Function is inf or nan.\n"));
254     ls->reason = TAOLINESEARCH_FAILED_BADPARAMETER;
255   } else if (ls->step < owlqn_minstep) {
256     PetscCall(PetscInfo(ls, "Step length is below tolerance.\n"));
257     ls->reason = TAOLINESEARCH_HALTED_RTOL;
258   } else if (ls->nfeval >= ls->max_funcs) {
259     PetscCall(PetscInfo(ls, "Number of line search function evals (%" PetscInt_FMT ") > maximum allowed (%" PetscInt_FMT ")\n", ls->nfeval, ls->max_funcs));
260     ls->reason = TAOLINESEARCH_HALTED_MAXFCN;
261   }
262   if (ls->reason) PetscFunctionReturn(0);
263 
264   /* Successful termination, update memory */
265   ls->reason          = TAOLINESEARCH_SUCCESS;
266   armP->lastReference = ref;
267   if (armP->replacementPolicy == REPLACE_FIFO) {
268     armP->memory[armP->current++] = *f;
269     if (armP->current >= armP->memorySize) armP->current = 0;
270   } else {
271     armP->current     = idx;
272     armP->memory[idx] = *f;
273   }
274 
275   /* Update iterate and compute gradient */
276   PetscCall(VecCopy(armP->work, x));
277   if (!g_computed) PetscCall(TaoLineSearchComputeGradient(ls, x, g));
278   PetscCall(PetscInfo(ls, "%" PetscInt_FMT " function evals in line search, step = %10.4f\n", ls->nfeval, (double)ls->step));
279   PetscFunctionReturn(0);
280 }
281 
282 /*MC
283    TAOLINESEARCHOWARMIJO - Special line-search type for the Orthant-Wise Limited Quasi-Newton (TAOOWLQN) algorithm.
284    Should not be used with any other algorithm.
285 
286    Level: developer
287 
288 .keywords: Tao, linesearch
289 M*/
290 PETSC_EXTERN PetscErrorCode TaoLineSearchCreate_OWArmijo(TaoLineSearch ls)
291 {
292   TaoLineSearch_OWARMIJO *armP;
293 
294   PetscFunctionBegin;
295   PetscValidHeaderSpecific(ls, TAOLINESEARCH_CLASSID, 1);
296   PetscCall(PetscNew(&armP));
297 
298   armP->memory            = NULL;
299   armP->alpha             = 1.0;
300   armP->beta              = 0.25;
301   armP->beta_inf          = 0.25;
302   armP->sigma             = 1e-4;
303   armP->memorySize        = 1;
304   armP->referencePolicy   = REFERENCE_MAX;
305   armP->replacementPolicy = REPLACE_MRU;
306   armP->nondescending     = PETSC_FALSE;
307   ls->data                = (void *)armP;
308   ls->initstep            = 0.1;
309   ls->ops->monitor        = NULL;
310   ls->ops->setup          = NULL;
311   ls->ops->reset          = NULL;
312   ls->ops->apply          = TaoLineSearchApply_OWArmijo;
313   ls->ops->view           = TaoLineSearchView_OWArmijo;
314   ls->ops->destroy        = TaoLineSearchDestroy_OWArmijo;
315   ls->ops->setfromoptions = TaoLineSearchSetFromOptions_OWArmijo;
316   PetscFunctionReturn(0);
317 }
318