xref: /petsc/src/tao/unconstrained/impls/lmvm/lmvm.c (revision 763847b449aff3e097428fd9cc317d0447b1751d)
1 #include <petsctaolinesearch.h>
2 #include <../src/tao/matrix/lmvmmat.h>
3 #include <../src/tao/unconstrained/impls/lmvm/lmvm.h>
4 
5 #define LMVM_BFGS                0
6 #define LMVM_SCALED_GRADIENT     1
7 #define LMVM_GRADIENT            2
8 
9 static PetscErrorCode TaoSolve_LMVM(Tao tao)
10 {
11   TAO_LMVM                     *lmP = (TAO_LMVM *)tao->data;
12   PetscReal                    f, fold, gdx, gnorm;
13   PetscReal                    step = 1.0;
14   PetscReal                    delta;
15   PetscErrorCode               ierr;
16   PetscInt                     stepType;
17   TaoLineSearchConvergedReason ls_status = TAOLINESEARCH_CONTINUE_ITERATING;
18 
19   PetscFunctionBegin;
20 
21   if (tao->XL || tao->XU || tao->ops->computebounds) {
22     ierr = PetscPrintf(((PetscObject)tao)->comm,"WARNING: Variable bounds have been set but will be ignored by lmvm algorithm\n");CHKERRQ(ierr);
23   }
24 
25   /*  Check convergence criteria */
26   ierr = TaoComputeObjectiveAndGradient(tao, tao->solution, &f, tao->gradient);CHKERRQ(ierr);
27   ierr = TaoGradientNorm(tao, tao->gradient,NORM_2,&gnorm);CHKERRQ(ierr);
28 
29   if (PetscIsInfOrNanReal(f) || PetscIsInfOrNanReal(gnorm)) SETERRQ(PETSC_COMM_SELF,1, "User provided compute function generated Inf or NaN");
30 
31   tao->reason = TAO_CONTINUE_ITERATING;
32   ierr = TaoLogConvergenceHistory(tao,f,gnorm,0.0,tao->ksp_its);CHKERRQ(ierr);
33   ierr = TaoMonitor(tao,tao->niter,f,gnorm,0.0,step);CHKERRQ(ierr);
34   ierr = (*tao->ops->convergencetest)(tao,tao->cnvP);CHKERRQ(ierr);
35   if (tao->reason != TAO_CONTINUE_ITERATING) PetscFunctionReturn(0);
36 
37   /*  Set initial scaling for the function */
38   if (f != 0.0) {
39     delta = 2.0 * PetscAbsScalar(f) / (gnorm*gnorm);
40   } else {
41     delta = 2.0 / (gnorm*gnorm);
42   }
43   ierr = MatLMVMSetDelta(lmP->M,delta);CHKERRQ(ierr);
44 
45   /*  Set counter for gradient/reset steps */
46   ierr = MatLMVMGetRecycleFlag(lmP->M, &recycle);CHKERRQ(ierr);
47   if (!recycle) {
48     lmP->bfgs = 0;
49     lmP->sgrad = 0;
50     lmP->grad = 0;
51     ierr = MatLMVMReset(lmP->M); CHKERRQ(ierr);
52   }
53 
54   /*  Have not converged; continue with Newton method */
55   while (tao->reason == TAO_CONTINUE_ITERATING) {
56     /*  Compute direction */
57     ierr = MatLMVMUpdate(lmP->M,tao->solution,tao->gradient);CHKERRQ(ierr);
58     ierr = MatLMVMSolve(lmP->M, tao->gradient, lmP->D);CHKERRQ(ierr);
59 
60     /*  Check for success (descent direction) */
61     ierr = VecDot(lmP->D, tao->gradient, &gdx);CHKERRQ(ierr);
62     if ((gdx <= 0.0) || PetscIsInfOrNanReal(gdx)) {
63       /* Step is not descent or direction produced not a number
64          We can assert bfgsUpdates > 1 in this case because
65          the first solve produces the scaled gradient direction,
66          which is guaranteed to be descent
67 
68          Use steepest descent direction (scaled)
69       */
70 
71       if (f != 0.0) {
72         delta = 2.0 * PetscAbsScalar(f) / (gnorm*gnorm);
73       } else {
74         delta = 2.0 / (gnorm*gnorm);
75       }
76       ierr = MatLMVMSetDelta(lmP->M, delta);CHKERRQ(ierr);
77       ierr = MatLMVMReset(lmP->M);CHKERRQ(ierr);
78       ierr = MatLMVMUpdate(lmP->M, tao->solution, tao->gradient);CHKERRQ(ierr);
79       ierr = MatLMVMSolve(lmP->M,tao->gradient, lmP->D);CHKERRQ(ierr);
80 
81       /* On a reset, the direction cannot be not a number; it is a
82          scaled gradient step.  No need to check for this condition. */
83       ++lmP->sgrad;
84       stepType = LMVM_SCALED_GRADIENT;
85     } else {
86       ierr = MatLMVMGetUpdates(lmP->M, &nupdates); CHKERRQ(ierr);
87       if (1 == nupdates) {
88         /*  The first BFGS direction is always the scaled gradient */
89         ++lmP->sgrad;
90         stepType = LMVM_SCALED_GRADIENT;
91       } else {
92         ++lmP->bfgs;
93         stepType = LMVM_BFGS;
94       }
95     }
96     ierr = VecScale(lmP->D, -1.0);CHKERRQ(ierr);
97 
98     /*  Perform the linesearch */
99     fold = f;
100     ierr = VecCopy(tao->solution, lmP->Xold);CHKERRQ(ierr);
101     ierr = VecCopy(tao->gradient, lmP->Gold);CHKERRQ(ierr);
102 
103     ierr = TaoLineSearchApply(tao->linesearch, tao->solution, &f, tao->gradient, lmP->D, &step,&ls_status);CHKERRQ(ierr);
104     ierr = TaoAddLineSearchCounts(tao);CHKERRQ(ierr);
105 
106     while (ls_status != TAOLINESEARCH_SUCCESS && ls_status != TAOLINESEARCH_SUCCESS_USER && (stepType != LMVM_GRADIENT)) {
107       /*  Linesearch failed */
108       ierr = PetscInfo(lmP, "WARNING: Linesearch failed!\n"); CHKERRQ(ierr);
109       /*  Reset factors and use scaled gradient step */
110       f = fold;
111       ierr = VecCopy(lmP->Xold, tao->solution);CHKERRQ(ierr);
112       ierr = VecCopy(lmP->Gold, tao->gradient);CHKERRQ(ierr);
113 
114       switch(stepType) {
115       case LMVM_BFGS:
116         /*  Failed to obtain acceptable iterate with BFGS step */
117         /*  Attempt to use the scaled gradient direction */
118 
119         if (f != 0.0) {
120           delta = 2.0 * PetscAbsScalar(f) / (gnorm*gnorm);
121         } else {
122           delta = 2.0 / (gnorm*gnorm);
123         }
124         ierr = MatLMVMSetDelta(lmP->M, delta);CHKERRQ(ierr);
125         ierr = MatLMVMReset(lmP->M);CHKERRQ(ierr);
126         ierr = MatLMVMUpdate(lmP->M, tao->solution, tao->gradient);CHKERRQ(ierr);
127         ierr = MatLMVMSolve(lmP->M, tao->gradient, lmP->D);CHKERRQ(ierr);
128 
129         /* On a reset, the direction cannot be not a number; it is a
130            scaled gradient step.  No need to check for this condition. */
131         --lmP->bfgs;
132         ++lmP->sgrad;
133         stepType = LMVM_SCALED_GRADIENT;
134         break;
135 
136       case LMVM_SCALED_GRADIENT:
137         /* The scaled gradient step did not produce a new iterate;
138            attempt to use the gradient direction.
139            Need to make sure we are not using a different diagonal scaling */
140         ierr = MatLMVMSetDelta(lmP->M, 1.0);CHKERRQ(ierr);
141         ierr = MatLMVMReset(lmP->M);CHKERRQ(ierr);
142         ierr = MatLMVMUpdate(lmP->M, tao->solution, tao->gradient);CHKERRQ(ierr);
143         ierr = MatLMVMSolve(lmP->M, tao->gradient, lmP->D);CHKERRQ(ierr);
144 
145         --lmP->sgrad;
146         ++lmP->grad;
147         stepType = LMVM_GRADIENT;
148         break;
149       }
150       ierr = VecScale(lmP->D, -1.0);CHKERRQ(ierr);
151 
152       /*  Perform the linesearch */
153       ierr = TaoLineSearchApply(tao->linesearch, tao->solution, &f, tao->gradient, lmP->D, &step, &ls_status);CHKERRQ(ierr);
154       ierr = TaoAddLineSearchCounts(tao);CHKERRQ(ierr);
155     }
156 
157     if (ls_status != TAOLINESEARCH_SUCCESS && ls_status != TAOLINESEARCH_SUCCESS_USER) {
158       /*  Failed to find an improving point */
159       f = fold;
160       ierr = VecCopy(lmP->Xold, tao->solution);CHKERRQ(ierr);
161       ierr = VecCopy(lmP->Gold, tao->gradient);CHKERRQ(ierr);
162       step = 0.0;
163       tao->reason = TAO_DIVERGED_LS_FAILURE;
164     }
165 
166     /*  Check for termination */
167     ierr = TaoGradientNorm(tao, tao->gradient,NORM_2,&gnorm);CHKERRQ(ierr);
168 
169     tao->niter++;
170     ierr = TaoLogConvergenceHistory(tao,f,gnorm,0.0,tao->ksp_its);CHKERRQ(ierr);
171     ierr = TaoMonitor(tao,tao->niter,f,gnorm,0.0,step);CHKERRQ(ierr);
172     ierr = (*tao->ops->convergencetest)(tao,tao->cnvP);CHKERRQ(ierr);
173   }
174   PetscFunctionReturn(0);
175 }
176 
177 static PetscErrorCode TaoSetUp_LMVM(Tao tao)
178 {
179   TAO_LMVM       *lmP = (TAO_LMVM *)tao->data;
180   PetscInt       n,N;
181   PetscErrorCode ierr;
182   KSP            H0ksp;
183 
184   PetscFunctionBegin;
185   /* Existence of tao->solution checked in TaoSetUp() */
186   if (!tao->gradient) {ierr = VecDuplicate(tao->solution,&tao->gradient);CHKERRQ(ierr);  }
187   if (!tao->stepdirection) {ierr = VecDuplicate(tao->solution,&tao->stepdirection);CHKERRQ(ierr);  }
188   if (!lmP->D) {ierr = VecDuplicate(tao->solution,&lmP->D);CHKERRQ(ierr);  }
189   if (!lmP->Xold) {ierr = VecDuplicate(tao->solution,&lmP->Xold);CHKERRQ(ierr);  }
190   if (!lmP->Gold) {ierr = VecDuplicate(tao->solution,&lmP->Gold);CHKERRQ(ierr);  }
191 
192   /*  Create matrix for the limited memory approximation */
193   ierr = VecGetLocalSize(tao->solution,&n);CHKERRQ(ierr);
194   ierr = VecGetSize(tao->solution,&N);CHKERRQ(ierr);
195   ierr = MatCreateLMVM(((PetscObject)tao)->comm,n,N,&lmP->M);CHKERRQ(ierr);
196   ierr = MatLMVMAllocateVectors(lmP->M,tao->solution);CHKERRQ(ierr);
197 
198   /* If the user has set a matrix to solve as the initial H0, set the options prefix here, and set up the KSP */
199   if (lmP->H0) {
200     const char *prefix;
201     PC H0pc;
202 
203     ierr = MatLMVMSetH0(lmP->M, lmP->H0);CHKERRQ(ierr);
204     ierr = MatLMVMGetH0KSP(lmP->M, &H0ksp);CHKERRQ(ierr);
205 
206     ierr = TaoGetOptionsPrefix(tao, &prefix);CHKERRQ(ierr);
207     ierr = KSPSetOptionsPrefix(H0ksp, prefix);CHKERRQ(ierr);
208     ierr = PetscObjectAppendOptionsPrefix((PetscObject)H0ksp, "tao_h0_");CHKERRQ(ierr);
209     ierr = KSPGetPC(H0ksp, &H0pc);CHKERRQ(ierr);
210     ierr = PetscObjectAppendOptionsPrefix((PetscObject)H0pc,  "tao_h0_");CHKERRQ(ierr);
211 
212     ierr = KSPSetFromOptions(H0ksp);CHKERRQ(ierr);
213     ierr = KSPSetUp(H0ksp);CHKERRQ(ierr);
214   }
215 
216   PetscFunctionReturn(0);
217 }
218 
219 /* ---------------------------------------------------------- */
220 static PetscErrorCode TaoDestroy_LMVM(Tao tao)
221 {
222   TAO_LMVM       *lmP = (TAO_LMVM *)tao->data;
223   PetscErrorCode ierr;
224   PetscBool      recycle;
225 
226   PetscFunctionBegin;
227   if (lmP->M) {
228     ierr = MatLMVMGetRecycleFlag(lmP->M, &recycle); CHKERRQ(ierr);
229     if (recycle) {
230       ierr = PetscInfo(lmP->M, "WARNING: TaoDestroy() called when LMVM recycling is enabled!\n"); CHKERRQ(ierr);
231     }
232   }
233 
234   if (tao->setupcalled) {
235     ierr = VecDestroy(&lmP->Xold);CHKERRQ(ierr);
236     ierr = VecDestroy(&lmP->Gold);CHKERRQ(ierr);
237     ierr = VecDestroy(&lmP->D);CHKERRQ(ierr);
238     ierr = MatDestroy(&lmP->M);CHKERRQ(ierr);
239   }
240 
241   if (lmP->H0) {
242     ierr = PetscObjectDereference((PetscObject)lmP->H0);CHKERRQ(ierr);
243   }
244 
245   ierr = PetscFree(tao->data);CHKERRQ(ierr);
246 
247   PetscFunctionReturn(0);
248 }
249 
250 /*------------------------------------------------------------*/
251 static PetscErrorCode TaoSetFromOptions_LMVM(PetscOptionItems *PetscOptionsObject,Tao tao)
252 {
253   PetscErrorCode ierr;
254 
255   PetscFunctionBegin;
256   ierr = PetscOptionsHead(PetscOptionsObject,"Limited-memory variable-metric method for unconstrained optimization");CHKERRQ(ierr);
257   ierr = TaoLineSearchSetFromOptions(tao->linesearch);CHKERRQ(ierr);
258   ierr = PetscOptionsTail();CHKERRQ(ierr);
259   PetscFunctionReturn(0);
260 }
261 
262 /*------------------------------------------------------------*/
263 static PetscErrorCode TaoView_LMVM(Tao tao, PetscViewer viewer)
264 {
265   TAO_LMVM       *lm = (TAO_LMVM *)tao->data;
266   PetscBool      isascii, recycle;
267   PetscInt       recycled_its;
268   PetscErrorCode ierr;
269 
270   PetscFunctionBegin;
271   ierr = PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERASCII, &isascii);CHKERRQ(ierr);
272   if (isascii) {
273     ierr = PetscViewerASCIIPushTab(viewer);CHKERRQ(ierr);
274     ierr = PetscViewerASCIIPrintf(viewer, "BFGS steps: %D\n", lm->bfgs);CHKERRQ(ierr);
275     ierr = PetscViewerASCIIPrintf(viewer, "Scaled gradient steps: %D\n", lm->sgrad);CHKERRQ(ierr);
276     ierr = PetscViewerASCIIPrintf(viewer, "Gradient steps: %D\n", lm->grad);CHKERRQ(ierr);
277     ierr = MatLMVMGetRecycleFlag(lm->M, &recycle);CHKERRQ(ierr);
278     if (recycle) {
279       ierr = PetscViewerASCIIPrintf(viewer, "Recycle: on\n");CHKERRQ(ierr);
280       recycled_its = lm->bfgs + lm->sgrad + lm->grad;
281       ierr = PetscViewerASCIIPrintf(viewer, "Total recycled iterations: %D\n", recycled_its);CHKERRQ(ierr);
282     }
283     ierr = PetscViewerASCIIPopTab(viewer);CHKERRQ(ierr);
284   }
285   PetscFunctionReturn(0);
286 }
287 
288 /* ---------------------------------------------------------- */
289 
290 /*MC
291      TAOLMVM - Limited Memory Variable Metric method is a quasi-Newton
292      optimization solver for unconstrained minimization. It solves
293      the Newton step
294               Hkdk = - gk
295 
296      using an approximation Bk in place of Hk, where Bk is composed using
297      the BFGS update formula. A More-Thuente line search is then used
298      to computed the steplength in the dk direction
299   Options Database Keys:
300 +     -tao_lmm_vectors - number of vectors to use for approximation
301 .     -tao_lmm_scale_type - "none","scalar","broyden"
302 .     -tao_lmm_limit_type - "none","average","relative","absolute"
303 .     -tao_lmm_rescale_type - "none","scalar","gl"
304 .     -tao_lmm_limit_mu - mu limiting factor
305 .     -tao_lmm_limit_nu - nu limiting factor
306 .     -tao_lmm_delta_min - minimum delta value
307 .     -tao_lmm_delta_max - maximum delta value
308 .     -tao_lmm_broyden_phi - phi factor for Broyden scaling
309 .     -tao_lmm_scalar_alpha - alpha factor for scalar scaling
310 .     -tao_lmm_rescale_alpha - alpha factor for rescaling diagonal
311 .     -tao_lmm_rescale_beta - beta factor for rescaling diagonal
312 .     -tao_lmm_scalar_history - amount of history for scalar scaling
313 .     -tao_lmm_rescale_history - amount of history for rescaling diagonal
314 .     -tao_lmm_eps - rejection tolerance
315 -     -tao_lmm_recycle - enable recycling LMVM updates between TaoSolve() calls
316 
317   Level: beginner
318 M*/
319 
320 PETSC_EXTERN PetscErrorCode TaoCreate_LMVM(Tao tao)
321 {
322   TAO_LMVM       *lmP;
323   const char     *morethuente_type = TAOLINESEARCHMT;
324   PetscErrorCode ierr;
325 
326   PetscFunctionBegin;
327   tao->ops->setup = TaoSetUp_LMVM;
328   tao->ops->solve = TaoSolve_LMVM;
329   tao->ops->view = TaoView_LMVM;
330   tao->ops->setfromoptions = TaoSetFromOptions_LMVM;
331   tao->ops->destroy = TaoDestroy_LMVM;
332 
333   ierr = PetscNewLog(tao,&lmP);CHKERRQ(ierr);
334   lmP->D = 0;
335   lmP->M = 0;
336   lmP->Xold = 0;
337   lmP->Gold = 0;
338   lmP->H0   = NULL;
339 
340   tao->data = (void*)lmP;
341   /* Override default settings (unless already changed) */
342   if (!tao->max_it_changed) tao->max_it = 2000;
343   if (!tao->max_funcs_changed) tao->max_funcs = 4000;
344 
345   ierr = TaoLineSearchCreate(((PetscObject)tao)->comm,&tao->linesearch);CHKERRQ(ierr);
346   ierr = PetscObjectIncrementTabLevel((PetscObject)tao->linesearch, (PetscObject)tao, 1);CHKERRQ(ierr);
347   ierr = TaoLineSearchSetType(tao->linesearch,morethuente_type);CHKERRQ(ierr);
348   ierr = TaoLineSearchUseTaoRoutines(tao->linesearch,tao);CHKERRQ(ierr);
349   ierr = TaoLineSearchSetOptionsPrefix(tao->linesearch,tao->hdr.prefix);CHKERRQ(ierr);
350   PetscFunctionReturn(0);
351 }
352