xref: /petsc/src/tao/unconstrained/impls/lmvm/lmvm.c (revision ccb4e88a40f0b86eaeca07ff64c64e4de2fae686)
1 #include <petsctaolinesearch.h>
2 #include <../src/tao/unconstrained/impls/lmvm/lmvm.h>
3 
4 #define LMVM_STEP_BFGS     0
5 #define LMVM_STEP_GRAD     1
6 
7 static PetscErrorCode TaoSolve_LMVM(Tao tao)
8 {
9   TAO_LMVM                     *lmP = (TAO_LMVM *)tao->data;
10   PetscReal                    f, fold, gdx, gnorm;
11   PetscReal                    step = 1.0;
12   PetscErrorCode               ierr;
13   PetscInt                     stepType = LMVM_STEP_GRAD, nupdates;
14   TaoLineSearchConvergedReason ls_status = TAOLINESEARCH_CONTINUE_ITERATING;
15 
16   PetscFunctionBegin;
17 
18   if (tao->XL || tao->XU || tao->ops->computebounds) {
19     ierr = PetscInfo(tao,"WARNING: Variable bounds have been set but will be ignored by lmvm algorithm\n");CHKERRQ(ierr);
20   }
21 
22   /*  Check convergence criteria */
23   ierr = TaoComputeObjectiveAndGradient(tao, tao->solution, &f, tao->gradient);CHKERRQ(ierr);
24   ierr = TaoGradientNorm(tao, tao->gradient,NORM_2,&gnorm);CHKERRQ(ierr);
25 
26   if (PetscIsInfOrNanReal(f) || PetscIsInfOrNanReal(gnorm)) SETERRQ(PetscObjectComm((PetscObject)tao),PETSC_ERR_USER, "User provided compute function generated Inf or NaN");
27 
28   tao->reason = TAO_CONTINUE_ITERATING;
29   ierr = TaoLogConvergenceHistory(tao,f,gnorm,0.0,tao->ksp_its);CHKERRQ(ierr);
30   ierr = TaoMonitor(tao,tao->niter,f,gnorm,0.0,step);CHKERRQ(ierr);
31   ierr = (*tao->ops->convergencetest)(tao,tao->cnvP);CHKERRQ(ierr);
32   if (tao->reason != TAO_CONTINUE_ITERATING) PetscFunctionReturn(0);
33 
34   /*  Set counter for gradient/reset steps */
35   if (!lmP->recycle) {
36     lmP->bfgs = 0;
37     lmP->grad = 0;
38     ierr = MatLMVMReset(lmP->M, PETSC_FALSE);CHKERRQ(ierr);
39   }
40 
41   /*  Have not converged; continue with Newton method */
42   while (tao->reason == TAO_CONTINUE_ITERATING) {
43     /* Call general purpose update function */
44     if (tao->ops->update) {
45       ierr = (*tao->ops->update)(tao, tao->niter, tao->user_update);CHKERRQ(ierr);
46     }
47 
48     /*  Compute direction */
49     if (lmP->H0) {
50       ierr = MatLMVMSetJ0(lmP->M, lmP->H0);CHKERRQ(ierr);
51       stepType = LMVM_STEP_BFGS;
52     }
53     ierr = MatLMVMUpdate(lmP->M,tao->solution,tao->gradient);CHKERRQ(ierr);
54     ierr = MatSolve(lmP->M, tao->gradient, lmP->D);CHKERRQ(ierr);
55     ierr = MatLMVMGetUpdateCount(lmP->M, &nupdates);CHKERRQ(ierr);
56     if (nupdates > 0) stepType = LMVM_STEP_BFGS;
57 
58     /*  Check for success (descent direction) */
59     ierr = VecDot(lmP->D, tao->gradient, &gdx);CHKERRQ(ierr);
60     if ((gdx <= 0.0) || PetscIsInfOrNanReal(gdx)) {
61       /* Step is not descent or direction produced not a number
62          We can assert bfgsUpdates > 1 in this case because
63          the first solve produces the scaled gradient direction,
64          which is guaranteed to be descent
65 
66          Use steepest descent direction (scaled)
67       */
68 
69       ierr = MatLMVMReset(lmP->M, PETSC_FALSE);CHKERRQ(ierr);
70       ierr = MatLMVMClearJ0(lmP->M);CHKERRQ(ierr);
71       ierr = MatLMVMUpdate(lmP->M, tao->solution, tao->gradient);CHKERRQ(ierr);
72       ierr = MatSolve(lmP->M,tao->gradient, lmP->D);CHKERRQ(ierr);
73 
74       /* On a reset, the direction cannot be not a number; it is a
75          scaled gradient step.  No need to check for this condition. */
76       stepType = LMVM_STEP_GRAD;
77     }
78     ierr = VecScale(lmP->D, -1.0);CHKERRQ(ierr);
79 
80     /*  Perform the linesearch */
81     fold = f;
82     ierr = VecCopy(tao->solution, lmP->Xold);CHKERRQ(ierr);
83     ierr = VecCopy(tao->gradient, lmP->Gold);CHKERRQ(ierr);
84 
85     ierr = TaoLineSearchApply(tao->linesearch, tao->solution, &f, tao->gradient, lmP->D, &step,&ls_status);CHKERRQ(ierr);
86     ierr = TaoAddLineSearchCounts(tao);CHKERRQ(ierr);
87 
88     if (ls_status != TAOLINESEARCH_SUCCESS && ls_status != TAOLINESEARCH_SUCCESS_USER && (stepType != LMVM_STEP_GRAD)) {
89       /*  Reset factors and use scaled gradient step */
90       f = fold;
91       ierr = VecCopy(lmP->Xold, tao->solution);CHKERRQ(ierr);
92       ierr = VecCopy(lmP->Gold, tao->gradient);CHKERRQ(ierr);
93 
94       /*  Failed to obtain acceptable iterate with BFGS step */
95       /*  Attempt to use the scaled gradient direction */
96 
97       ierr = MatLMVMReset(lmP->M, PETSC_FALSE);CHKERRQ(ierr);
98       ierr = MatLMVMClearJ0(lmP->M);CHKERRQ(ierr);
99       ierr = MatLMVMUpdate(lmP->M, tao->solution, tao->gradient);CHKERRQ(ierr);
100       ierr = MatSolve(lmP->M, tao->solution, tao->gradient);CHKERRQ(ierr);
101 
102       /* On a reset, the direction cannot be not a number; it is a
103           scaled gradient step.  No need to check for this condition. */
104       stepType = LMVM_STEP_GRAD;
105       ierr = VecScale(lmP->D, -1.0);CHKERRQ(ierr);
106 
107       /*  Perform the linesearch */
108       ierr = TaoLineSearchApply(tao->linesearch, tao->solution, &f, tao->gradient, lmP->D, &step, &ls_status);CHKERRQ(ierr);
109       ierr = TaoAddLineSearchCounts(tao);CHKERRQ(ierr);
110     }
111 
112     if (ls_status != TAOLINESEARCH_SUCCESS && ls_status != TAOLINESEARCH_SUCCESS_USER) {
113       /*  Failed to find an improving point */
114       f = fold;
115       ierr = VecCopy(lmP->Xold, tao->solution);CHKERRQ(ierr);
116       ierr = VecCopy(lmP->Gold, tao->gradient);CHKERRQ(ierr);
117       step = 0.0;
118       tao->reason = TAO_DIVERGED_LS_FAILURE;
119     } else {
120       /* LS found valid step, so tally up step type */
121       switch (stepType) {
122       case LMVM_STEP_BFGS:
123         ++lmP->bfgs;
124         break;
125       case LMVM_STEP_GRAD:
126         ++lmP->grad;
127         break;
128       default:
129         break;
130       }
131       /*  Compute new gradient norm */
132       ierr = TaoGradientNorm(tao, tao->gradient,NORM_2,&gnorm);CHKERRQ(ierr);
133     }
134 
135     /* Check convergence */
136     tao->niter++;
137     ierr = TaoLogConvergenceHistory(tao,f,gnorm,0.0,tao->ksp_its);CHKERRQ(ierr);
138     ierr = TaoMonitor(tao,tao->niter,f,gnorm,0.0,step);CHKERRQ(ierr);
139     ierr = (*tao->ops->convergencetest)(tao,tao->cnvP);CHKERRQ(ierr);
140   }
141   PetscFunctionReturn(0);
142 }
143 
144 static PetscErrorCode TaoSetUp_LMVM(Tao tao)
145 {
146   TAO_LMVM       *lmP = (TAO_LMVM *)tao->data;
147   PetscInt       n,N;
148   PetscErrorCode ierr;
149   PetscBool      is_spd;
150 
151   PetscFunctionBegin;
152   /* Existence of tao->solution checked in TaoSetUp() */
153   if (!tao->gradient) {ierr = VecDuplicate(tao->solution,&tao->gradient);CHKERRQ(ierr);  }
154   if (!tao->stepdirection) {ierr = VecDuplicate(tao->solution,&tao->stepdirection);CHKERRQ(ierr);  }
155   if (!lmP->D) {ierr = VecDuplicate(tao->solution,&lmP->D);CHKERRQ(ierr);  }
156   if (!lmP->Xold) {ierr = VecDuplicate(tao->solution,&lmP->Xold);CHKERRQ(ierr);  }
157   if (!lmP->Gold) {ierr = VecDuplicate(tao->solution,&lmP->Gold);CHKERRQ(ierr);  }
158 
159   /*  Create matrix for the limited memory approximation */
160   ierr = VecGetLocalSize(tao->solution,&n);CHKERRQ(ierr);
161   ierr = VecGetSize(tao->solution,&N);CHKERRQ(ierr);
162   ierr = MatSetSizes(lmP->M, n, n, N, N);CHKERRQ(ierr);
163   ierr = MatLMVMAllocate(lmP->M,tao->solution,tao->gradient);CHKERRQ(ierr);
164   ierr = MatGetOption(lmP->M, MAT_SPD, &is_spd);CHKERRQ(ierr);
165   if (!is_spd) SETERRQ(PetscObjectComm((PetscObject)tao), PETSC_ERR_ARG_INCOMP, "LMVM matrix is not symmetric positive-definite.");
166 
167   /* If the user has set a matrix to solve as the initial H0, set the options prefix here, and set up the KSP */
168   if (lmP->H0) {
169     ierr = MatLMVMSetJ0(lmP->M, lmP->H0);CHKERRQ(ierr);
170   }
171 
172   PetscFunctionReturn(0);
173 }
174 
175 /* ---------------------------------------------------------- */
176 static PetscErrorCode TaoDestroy_LMVM(Tao tao)
177 {
178   TAO_LMVM       *lmP = (TAO_LMVM *)tao->data;
179   PetscErrorCode ierr;
180 
181   PetscFunctionBegin;
182   if (tao->setupcalled) {
183     ierr = VecDestroy(&lmP->Xold);CHKERRQ(ierr);
184     ierr = VecDestroy(&lmP->Gold);CHKERRQ(ierr);
185     ierr = VecDestroy(&lmP->D);CHKERRQ(ierr);
186   }
187   ierr = MatDestroy(&lmP->M);CHKERRQ(ierr);
188   if (lmP->H0) {
189     ierr = PetscObjectDereference((PetscObject)lmP->H0);CHKERRQ(ierr);
190   }
191   ierr = PetscFree(tao->data);CHKERRQ(ierr);
192 
193   PetscFunctionReturn(0);
194 }
195 
196 /*------------------------------------------------------------*/
197 static PetscErrorCode TaoSetFromOptions_LMVM(PetscOptionItems *PetscOptionsObject,Tao tao)
198 {
199   TAO_LMVM       *lm = (TAO_LMVM *)tao->data;
200   PetscErrorCode ierr;
201 
202   PetscFunctionBegin;
203   ierr = PetscOptionsHead(PetscOptionsObject,"Limited-memory variable-metric method for unconstrained optimization");CHKERRQ(ierr);
204   ierr = PetscOptionsBool("-tao_lmvm_recycle","enable recycling of the BFGS matrix between subsequent TaoSolve() calls","",lm->recycle,&lm->recycle,NULL);CHKERRQ(ierr);
205   ierr = TaoLineSearchSetFromOptions(tao->linesearch);CHKERRQ(ierr);
206   ierr = MatSetFromOptions(lm->M);CHKERRQ(ierr);
207   ierr = PetscOptionsTail();CHKERRQ(ierr);
208   PetscFunctionReturn(0);
209 }
210 
211 /*------------------------------------------------------------*/
212 static PetscErrorCode TaoView_LMVM(Tao tao, PetscViewer viewer)
213 {
214   TAO_LMVM       *lm = (TAO_LMVM *)tao->data;
215   PetscBool      isascii;
216   PetscInt       recycled_its;
217   PetscErrorCode ierr;
218 
219   PetscFunctionBegin;
220   ierr = PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERASCII, &isascii);CHKERRQ(ierr);
221   if (isascii) {
222     ierr = PetscViewerASCIIPrintf(viewer, "  Gradient steps: %D\n", lm->grad);CHKERRQ(ierr);
223     if (lm->recycle) {
224       ierr = PetscViewerASCIIPrintf(viewer, "  Recycle: on\n");CHKERRQ(ierr);
225       recycled_its = lm->bfgs + lm->grad;
226       ierr = PetscViewerASCIIPrintf(viewer, "  Total recycled iterations: %D\n", recycled_its);CHKERRQ(ierr);
227     }
228   }
229   PetscFunctionReturn(0);
230 }
231 
232 /* ---------------------------------------------------------- */
233 
234 /*MC
235   TAOLMVM - Limited Memory Variable Metric method is a quasi-Newton
236   optimization solver for unconstrained minimization. It solves
237   the Newton step
238           Hkdk = - gk
239 
240   using an approximation Bk in place of Hk, where Bk is composed using
241   the BFGS update formula. A More-Thuente line search is then used
242   to computed the steplength in the dk direction
243 
244   Options Database Keys:
245 +   -tao_lmvm_recycle - enable recycling LMVM updates between TaoSolve() calls
246 -   -tao_lmvm_no_scale - (developer) disables diagonal Broyden scaling on the LMVM approximation
247 
248   Level: beginner
249 M*/
250 
251 PETSC_EXTERN PetscErrorCode TaoCreate_LMVM(Tao tao)
252 {
253   TAO_LMVM       *lmP;
254   const char     *morethuente_type = TAOLINESEARCHMT;
255   PetscErrorCode ierr;
256 
257   PetscFunctionBegin;
258   tao->ops->setup = TaoSetUp_LMVM;
259   tao->ops->solve = TaoSolve_LMVM;
260   tao->ops->view = TaoView_LMVM;
261   tao->ops->setfromoptions = TaoSetFromOptions_LMVM;
262   tao->ops->destroy = TaoDestroy_LMVM;
263 
264   ierr = PetscNewLog(tao,&lmP);CHKERRQ(ierr);
265   lmP->D = NULL;
266   lmP->M = NULL;
267   lmP->Xold = NULL;
268   lmP->Gold = NULL;
269   lmP->H0   = NULL;
270   lmP->recycle = PETSC_FALSE;
271 
272   tao->data = (void*)lmP;
273   /* Override default settings (unless already changed) */
274   if (!tao->max_it_changed) tao->max_it = 2000;
275   if (!tao->max_funcs_changed) tao->max_funcs = 4000;
276 
277   ierr = TaoLineSearchCreate(((PetscObject)tao)->comm,&tao->linesearch);CHKERRQ(ierr);
278   ierr = PetscObjectIncrementTabLevel((PetscObject)tao->linesearch, (PetscObject)tao, 1);CHKERRQ(ierr);
279   ierr = TaoLineSearchSetType(tao->linesearch,morethuente_type);CHKERRQ(ierr);
280   ierr = TaoLineSearchUseTaoRoutines(tao->linesearch,tao);CHKERRQ(ierr);
281   ierr = TaoLineSearchSetOptionsPrefix(tao->linesearch,tao->hdr.prefix);CHKERRQ(ierr);
282 
283   ierr = KSPInitializePackage();CHKERRQ(ierr);
284   ierr = MatCreate(((PetscObject)tao)->comm, &lmP->M);CHKERRQ(ierr);
285   ierr = PetscObjectIncrementTabLevel((PetscObject)lmP->M, (PetscObject)tao, 1);CHKERRQ(ierr);
286   ierr = MatSetType(lmP->M, MATLMVMBFGS);CHKERRQ(ierr);
287   ierr = MatSetOptionsPrefix(lmP->M, "tao_lmvm_");CHKERRQ(ierr);
288   PetscFunctionReturn(0);
289 }
290