xref: /petsc/src/tao/unconstrained/impls/lmvm/lmvm.c (revision e7f46db8d62cb2e4e59111bf21061e64ea60daab)
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, nupdates;
17   PetscBool                    recycle;
18   TaoLineSearchConvergedReason ls_status = TAOLINESEARCH_CONTINUE_ITERATING;
19 
20   PetscFunctionBegin;
21 
22   if (tao->XL || tao->XU || tao->ops->computebounds) {
23     ierr = PetscPrintf(((PetscObject)tao)->comm,"WARNING: Variable bounds have been set but will be ignored by lmvm algorithm\n");CHKERRQ(ierr);
24   }
25 
26   /*  Check convergence criteria */
27   ierr = TaoComputeObjectiveAndGradient(tao, tao->solution, &f, tao->gradient);CHKERRQ(ierr);
28   ierr = TaoGradientNorm(tao, tao->gradient,NORM_2,&gnorm);CHKERRQ(ierr);
29 
30   if (PetscIsInfOrNanReal(f) || PetscIsInfOrNanReal(gnorm)) SETERRQ(PETSC_COMM_SELF,1, "User provided compute function generated Inf or NaN");
31 
32   tao->reason = TAO_CONTINUE_ITERATING;
33   ierr = TaoLogConvergenceHistory(tao,f,gnorm,0.0,tao->ksp_its);CHKERRQ(ierr);
34   ierr = TaoMonitor(tao,tao->niter,f,gnorm,0.0,step);CHKERRQ(ierr);
35   ierr = (*tao->ops->convergencetest)(tao,tao->cnvP);CHKERRQ(ierr);
36   if (tao->reason != TAO_CONTINUE_ITERATING) PetscFunctionReturn(0);
37 
38   /*  Set initial scaling for the function */
39   if (f != 0.0) {
40     delta = 2.0 * PetscAbsScalar(f) / (gnorm*gnorm);
41   } else {
42     delta = 2.0 / (gnorm*gnorm);
43   }
44   ierr = MatLMVMSetDelta(lmP->M,delta);CHKERRQ(ierr);
45 
46   /*  Set counter for gradient/reset steps */
47   ierr = MatLMVMGetRecycleFlag(lmP->M, &recycle);CHKERRQ(ierr);
48   if (!recycle) {
49     lmP->bfgs = 0;
50     lmP->sgrad = 0;
51     lmP->grad = 0;
52     ierr = MatLMVMReset(lmP->M); CHKERRQ(ierr);
53   }
54 
55   /*  Have not converged; continue with Newton method */
56   while (tao->reason == TAO_CONTINUE_ITERATING) {
57     /*  Compute direction */
58     ierr = MatLMVMUpdate(lmP->M,tao->solution,tao->gradient);CHKERRQ(ierr);
59     ierr = MatLMVMSolve(lmP->M, tao->gradient, lmP->D);CHKERRQ(ierr);
60 
61     /*  Check for success (descent direction) */
62     ierr = VecDot(lmP->D, tao->gradient, &gdx);CHKERRQ(ierr);
63     if ((gdx <= 0.0) || PetscIsInfOrNanReal(gdx)) {
64       /* Step is not descent or direction produced not a number
65          We can assert bfgsUpdates > 1 in this case because
66          the first solve produces the scaled gradient direction,
67          which is guaranteed to be descent
68 
69          Use steepest descent direction (scaled)
70       */
71 
72       if (f != 0.0) {
73         delta = 2.0 * PetscAbsScalar(f) / (gnorm*gnorm);
74       } else {
75         delta = 2.0 / (gnorm*gnorm);
76       }
77       ierr = MatLMVMSetDelta(lmP->M, delta);CHKERRQ(ierr);
78       ierr = MatLMVMReset(lmP->M);CHKERRQ(ierr);
79       ierr = MatLMVMUpdate(lmP->M, tao->solution, tao->gradient);CHKERRQ(ierr);
80       ierr = MatLMVMSolve(lmP->M,tao->gradient, lmP->D);CHKERRQ(ierr);
81 
82       /* On a reset, the direction cannot be not a number; it is a
83          scaled gradient step.  No need to check for this condition. */
84       ++lmP->sgrad;
85       stepType = LMVM_SCALED_GRADIENT;
86     } else {
87       ierr = MatLMVMGetUpdates(lmP->M, &nupdates); CHKERRQ(ierr);
88       if (1 == nupdates) {
89         /*  The first BFGS direction is always the scaled gradient */
90         ++lmP->sgrad;
91         stepType = LMVM_SCALED_GRADIENT;
92       } else {
93         ++lmP->bfgs;
94         stepType = LMVM_BFGS;
95       }
96     }
97     ierr = VecScale(lmP->D, -1.0);CHKERRQ(ierr);
98 
99     /*  Perform the linesearch */
100     fold = f;
101     ierr = VecCopy(tao->solution, lmP->Xold);CHKERRQ(ierr);
102     ierr = VecCopy(tao->gradient, lmP->Gold);CHKERRQ(ierr);
103 
104     ierr = TaoLineSearchApply(tao->linesearch, tao->solution, &f, tao->gradient, lmP->D, &step,&ls_status);CHKERRQ(ierr);
105     ierr = TaoAddLineSearchCounts(tao);CHKERRQ(ierr);
106 
107     while (ls_status != TAOLINESEARCH_SUCCESS && ls_status != TAOLINESEARCH_SUCCESS_USER && (stepType != LMVM_GRADIENT)) {
108       /*  Linesearch failed */
109       ierr = PetscInfo(lmP, "WARNING: Linesearch failed!\n"); CHKERRQ(ierr);
110       /*  Reset factors and use scaled gradient step */
111       f = fold;
112       ierr = VecCopy(lmP->Xold, tao->solution);CHKERRQ(ierr);
113       ierr = VecCopy(lmP->Gold, tao->gradient);CHKERRQ(ierr);
114 
115       switch(stepType) {
116       case LMVM_BFGS:
117         /*  Failed to obtain acceptable iterate with BFGS step */
118         /*  Attempt to use the scaled gradient direction */
119 
120         if (f != 0.0) {
121           delta = 2.0 * PetscAbsScalar(f) / (gnorm*gnorm);
122         } else {
123           delta = 2.0 / (gnorm*gnorm);
124         }
125         ierr = MatLMVMSetDelta(lmP->M, delta);CHKERRQ(ierr);
126         ierr = MatLMVMReset(lmP->M);CHKERRQ(ierr);
127         ierr = MatLMVMUpdate(lmP->M, tao->solution, tao->gradient);CHKERRQ(ierr);
128         ierr = MatLMVMSolve(lmP->M, tao->gradient, lmP->D);CHKERRQ(ierr);
129 
130         /* On a reset, the direction cannot be not a number; it is a
131            scaled gradient step.  No need to check for this condition. */
132         --lmP->bfgs;
133         ++lmP->sgrad;
134         stepType = LMVM_SCALED_GRADIENT;
135         break;
136 
137       case LMVM_SCALED_GRADIENT:
138         /* The scaled gradient step did not produce a new iterate;
139            attempt to use the gradient direction.
140            Need to make sure we are not using a different diagonal scaling */
141         ierr = MatLMVMSetDelta(lmP->M, 1.0);CHKERRQ(ierr);
142         ierr = MatLMVMReset(lmP->M);CHKERRQ(ierr);
143         ierr = MatLMVMUpdate(lmP->M, tao->solution, tao->gradient);CHKERRQ(ierr);
144         ierr = MatLMVMSolve(lmP->M, tao->gradient, lmP->D);CHKERRQ(ierr);
145 
146         --lmP->sgrad;
147         ++lmP->grad;
148         stepType = LMVM_GRADIENT;
149         break;
150       }
151       ierr = VecScale(lmP->D, -1.0);CHKERRQ(ierr);
152 
153       /*  Perform the linesearch */
154       ierr = TaoLineSearchApply(tao->linesearch, tao->solution, &f, tao->gradient, lmP->D, &step, &ls_status);CHKERRQ(ierr);
155       ierr = TaoAddLineSearchCounts(tao);CHKERRQ(ierr);
156     }
157 
158     if (ls_status != TAOLINESEARCH_SUCCESS && ls_status != TAOLINESEARCH_SUCCESS_USER) {
159       /*  Failed to find an improving point */
160       f = fold;
161       ierr = VecCopy(lmP->Xold, tao->solution);CHKERRQ(ierr);
162       ierr = VecCopy(lmP->Gold, tao->gradient);CHKERRQ(ierr);
163       step = 0.0;
164       tao->reason = TAO_DIVERGED_LS_FAILURE;
165     }
166 
167     /*  Check for termination */
168     ierr = TaoGradientNorm(tao, tao->gradient,NORM_2,&gnorm);CHKERRQ(ierr);
169 
170     tao->niter++;
171     ierr = TaoLogConvergenceHistory(tao,f,gnorm,0.0,tao->ksp_its);CHKERRQ(ierr);
172     ierr = TaoMonitor(tao,tao->niter,f,gnorm,0.0,step);CHKERRQ(ierr);
173     ierr = (*tao->ops->convergencetest)(tao,tao->cnvP);CHKERRQ(ierr);
174   }
175   PetscFunctionReturn(0);
176 }
177 
178 static PetscErrorCode TaoSetUp_LMVM(Tao tao)
179 {
180   TAO_LMVM       *lmP = (TAO_LMVM *)tao->data;
181   PetscInt       n,N;
182   PetscErrorCode ierr;
183   KSP            H0ksp;
184 
185   PetscFunctionBegin;
186   /* Existence of tao->solution checked in TaoSetUp() */
187   if (!tao->gradient) {ierr = VecDuplicate(tao->solution,&tao->gradient);CHKERRQ(ierr);  }
188   if (!tao->stepdirection) {ierr = VecDuplicate(tao->solution,&tao->stepdirection);CHKERRQ(ierr);  }
189   if (!lmP->D) {ierr = VecDuplicate(tao->solution,&lmP->D);CHKERRQ(ierr);  }
190   if (!lmP->Xold) {ierr = VecDuplicate(tao->solution,&lmP->Xold);CHKERRQ(ierr);  }
191   if (!lmP->Gold) {ierr = VecDuplicate(tao->solution,&lmP->Gold);CHKERRQ(ierr);  }
192 
193   /*  Create matrix for the limited memory approximation */
194   ierr = VecGetLocalSize(tao->solution,&n);CHKERRQ(ierr);
195   ierr = VecGetSize(tao->solution,&N);CHKERRQ(ierr);
196   ierr = MatCreateLMVM(((PetscObject)tao)->comm,n,N,&lmP->M);CHKERRQ(ierr);
197   ierr = MatLMVMAllocateVectors(lmP->M,tao->solution);CHKERRQ(ierr);
198 
199   /* If the user has set a matrix to solve as the initial H0, set the options prefix here, and set up the KSP */
200   if (lmP->H0) {
201     const char *prefix;
202     PC H0pc;
203 
204     ierr = MatLMVMSetH0(lmP->M, lmP->H0);CHKERRQ(ierr);
205     ierr = MatLMVMGetH0KSP(lmP->M, &H0ksp);CHKERRQ(ierr);
206 
207     ierr = TaoGetOptionsPrefix(tao, &prefix);CHKERRQ(ierr);
208     ierr = KSPSetOptionsPrefix(H0ksp, prefix);CHKERRQ(ierr);
209     ierr = PetscObjectAppendOptionsPrefix((PetscObject)H0ksp, "tao_h0_");CHKERRQ(ierr);
210     ierr = KSPGetPC(H0ksp, &H0pc);CHKERRQ(ierr);
211     ierr = PetscObjectAppendOptionsPrefix((PetscObject)H0pc,  "tao_h0_");CHKERRQ(ierr);
212 
213     ierr = KSPSetFromOptions(H0ksp);CHKERRQ(ierr);
214     ierr = KSPSetUp(H0ksp);CHKERRQ(ierr);
215   }
216 
217   PetscFunctionReturn(0);
218 }
219 
220 /* ---------------------------------------------------------- */
221 static PetscErrorCode TaoDestroy_LMVM(Tao tao)
222 {
223   TAO_LMVM       *lmP = (TAO_LMVM *)tao->data;
224   PetscErrorCode ierr;
225   PetscBool      recycle;
226 
227   PetscFunctionBegin;
228   if (lmP->M) {
229     ierr = MatLMVMGetRecycleFlag(lmP->M, &recycle); CHKERRQ(ierr);
230     if (recycle) {
231       ierr = PetscInfo(lmP->M, "WARNING: TaoDestroy() called when LMVM recycling is enabled!\n"); CHKERRQ(ierr);
232     }
233   }
234 
235   if (tao->setupcalled) {
236     ierr = VecDestroy(&lmP->Xold);CHKERRQ(ierr);
237     ierr = VecDestroy(&lmP->Gold);CHKERRQ(ierr);
238     ierr = VecDestroy(&lmP->D);CHKERRQ(ierr);
239     ierr = MatDestroy(&lmP->M);CHKERRQ(ierr);
240   }
241 
242   if (lmP->H0) {
243     ierr = PetscObjectDereference((PetscObject)lmP->H0);CHKERRQ(ierr);
244   }
245 
246   ierr = PetscFree(tao->data);CHKERRQ(ierr);
247 
248   PetscFunctionReturn(0);
249 }
250 
251 /*------------------------------------------------------------*/
252 static PetscErrorCode TaoSetFromOptions_LMVM(PetscOptionItems *PetscOptionsObject,Tao tao)
253 {
254   PetscErrorCode ierr;
255 
256   PetscFunctionBegin;
257   ierr = PetscOptionsHead(PetscOptionsObject,"Limited-memory variable-metric method for unconstrained optimization");CHKERRQ(ierr);
258   ierr = TaoLineSearchSetFromOptions(tao->linesearch);CHKERRQ(ierr);
259   ierr = PetscOptionsTail();CHKERRQ(ierr);
260   PetscFunctionReturn(0);
261 }
262 
263 /*------------------------------------------------------------*/
264 static PetscErrorCode TaoView_LMVM(Tao tao, PetscViewer viewer)
265 {
266   TAO_LMVM       *lm = (TAO_LMVM *)tao->data;
267   PetscBool      isascii, recycle;
268   PetscInt       recycled_its;
269   PetscErrorCode ierr;
270 
271   PetscFunctionBegin;
272   ierr = PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERASCII, &isascii);CHKERRQ(ierr);
273   if (isascii) {
274     ierr = PetscViewerASCIIPushTab(viewer);CHKERRQ(ierr);
275     ierr = PetscViewerASCIIPrintf(viewer, "BFGS steps: %D\n", lm->bfgs);CHKERRQ(ierr);
276     ierr = PetscViewerASCIIPrintf(viewer, "Scaled gradient steps: %D\n", lm->sgrad);CHKERRQ(ierr);
277     ierr = PetscViewerASCIIPrintf(viewer, "Gradient steps: %D\n", lm->grad);CHKERRQ(ierr);
278     ierr = MatLMVMGetRecycleFlag(lm->M, &recycle);CHKERRQ(ierr);
279     if (recycle) {
280       ierr = PetscViewerASCIIPrintf(viewer, "Recycle: on\n");CHKERRQ(ierr);
281       recycled_its = lm->bfgs + lm->sgrad + lm->grad;
282       ierr = PetscViewerASCIIPrintf(viewer, "Total recycled iterations: %D\n", recycled_its);CHKERRQ(ierr);
283     }
284     ierr = PetscViewerASCIIPopTab(viewer);CHKERRQ(ierr);
285   }
286   PetscFunctionReturn(0);
287 }
288 
289 /* ---------------------------------------------------------- */
290 
291 /*MC
292      TAOLMVM - Limited Memory Variable Metric method is a quasi-Newton
293      optimization solver for unconstrained minimization. It solves
294      the Newton step
295               Hkdk = - gk
296 
297      using an approximation Bk in place of Hk, where Bk is composed using
298      the BFGS update formula. A More-Thuente line search is then used
299      to computed the steplength in the dk direction
300   Options Database Keys:
301 +     -tao_lmm_vectors - number of vectors to use for approximation
302 .     -tao_lmm_scale_type - "none","scalar","broyden"
303 .     -tao_lmm_limit_type - "none","average","relative","absolute"
304 .     -tao_lmm_rescale_type - "none","scalar","gl"
305 .     -tao_lmm_limit_mu - mu limiting factor
306 .     -tao_lmm_limit_nu - nu limiting factor
307 .     -tao_lmm_delta_min - minimum delta value
308 .     -tao_lmm_delta_max - maximum delta value
309 .     -tao_lmm_broyden_phi - phi factor for Broyden scaling
310 .     -tao_lmm_scalar_alpha - alpha factor for scalar scaling
311 .     -tao_lmm_rescale_alpha - alpha factor for rescaling diagonal
312 .     -tao_lmm_rescale_beta - beta factor for rescaling diagonal
313 .     -tao_lmm_scalar_history - amount of history for scalar scaling
314 .     -tao_lmm_rescale_history - amount of history for rescaling diagonal
315 .     -tao_lmm_eps - rejection tolerance
316 -     -tao_lmm_recycle - enable recycling LMVM updates between TaoSolve() calls
317 
318   Level: beginner
319 M*/
320 
321 PETSC_EXTERN PetscErrorCode TaoCreate_LMVM(Tao tao)
322 {
323   TAO_LMVM       *lmP;
324   const char     *morethuente_type = TAOLINESEARCHMT;
325   PetscErrorCode ierr;
326 
327   PetscFunctionBegin;
328   tao->ops->setup = TaoSetUp_LMVM;
329   tao->ops->solve = TaoSolve_LMVM;
330   tao->ops->view = TaoView_LMVM;
331   tao->ops->setfromoptions = TaoSetFromOptions_LMVM;
332   tao->ops->destroy = TaoDestroy_LMVM;
333 
334   ierr = PetscNewLog(tao,&lmP);CHKERRQ(ierr);
335   lmP->D = 0;
336   lmP->M = 0;
337   lmP->Xold = 0;
338   lmP->Gold = 0;
339   lmP->H0   = NULL;
340 
341   tao->data = (void*)lmP;
342   /* Override default settings (unless already changed) */
343   if (!tao->max_it_changed) tao->max_it = 2000;
344   if (!tao->max_funcs_changed) tao->max_funcs = 4000;
345 
346   ierr = TaoLineSearchCreate(((PetscObject)tao)->comm,&tao->linesearch);CHKERRQ(ierr);
347   ierr = PetscObjectIncrementTabLevel((PetscObject)tao->linesearch, (PetscObject)tao, 1);CHKERRQ(ierr);
348   ierr = TaoLineSearchSetType(tao->linesearch,morethuente_type);CHKERRQ(ierr);
349   ierr = TaoLineSearchUseTaoRoutines(tao->linesearch,tao);CHKERRQ(ierr);
350   ierr = TaoLineSearchSetOptionsPrefix(tao->linesearch,tao->hdr.prefix);CHKERRQ(ierr);
351   PetscFunctionReturn(0);
352 }
353