xref: /petsc/src/tao/unconstrained/impls/lmvm/lmvm.c (revision 4d6623b4ac38e61a86b5c0cf094187b5a5ae29d2)
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   TaoConvergedReason           reason = TAO_CONTINUE_ITERATING;
19   TaoLineSearchConvergedReason ls_status = TAOLINESEARCH_CONTINUE_ITERATING;
20 
21   PetscFunctionBegin;
22 
23   if (tao->XL || tao->XU || tao->ops->computebounds) {
24     ierr = PetscPrintf(((PetscObject)tao)->comm,"WARNING: Variable bounds have been set but will be ignored by lmvm algorithm\n");CHKERRQ(ierr);
25   }
26 
27   /*  Check convergence criteria */
28   ierr = TaoComputeObjectiveAndGradient(tao, tao->solution, &f, tao->gradient);CHKERRQ(ierr);
29   ierr = TaoGradientNorm(tao, tao->gradient,NORM_2,&gnorm);CHKERRQ(ierr);
30 
31   if (PetscIsInfOrNanReal(f) || PetscIsInfOrNanReal(gnorm)) SETERRQ(PETSC_COMM_SELF,1, "User provided compute function generated Inf or NaN");
32 
33   ierr = TaoMonitor(tao, tao->niter, f, gnorm, 0.0, step, &reason);CHKERRQ(ierr);
34   if (reason != TAO_CONTINUE_ITERATING) PetscFunctionReturn(0);
35 
36   /*  Set initial scaling for the function */
37   if (f != 0.0) {
38     delta = 2.0 * PetscAbsScalar(f) / (gnorm*gnorm);
39   } else {
40     delta = 2.0 / (gnorm*gnorm);
41   }
42   ierr = MatLMVMSetDelta(lmP->M,delta);CHKERRQ(ierr);
43 
44   /*  Set counter for gradient/reset steps */
45   ierr = MatLMVMGetRecycleFlag(lmP->M, &recycle);CHKERRQ(ierr);
46   if (!recycle) {
47     lmP->bfgs = 0;
48     lmP->sgrad = 0;
49     lmP->grad = 0;
50   }
51 
52   /*  Have not converged; continue with Newton method */
53   while (reason == TAO_CONTINUE_ITERATING) {
54     /*  Compute direction */
55     ierr = MatLMVMUpdate(lmP->M,tao->solution,tao->gradient);CHKERRQ(ierr);
56     ierr = MatLMVMSolve(lmP->M, tao->gradient, lmP->D);CHKERRQ(ierr);
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       if (f != 0.0) {
70         delta = 2.0 * PetscAbsScalar(f) / (gnorm*gnorm);
71       } else {
72         delta = 2.0 / (gnorm*gnorm);
73       }
74       ierr = MatLMVMSetDelta(lmP->M, delta);CHKERRQ(ierr);
75       ierr = MatLMVMReset(lmP->M);CHKERRQ(ierr);
76       ierr = MatLMVMUpdate(lmP->M, tao->solution, tao->gradient);CHKERRQ(ierr);
77       ierr = MatLMVMSolve(lmP->M,tao->gradient, lmP->D);CHKERRQ(ierr);
78 
79       /* On a reset, the direction cannot be not a number; it is a
80          scaled gradient step.  No need to check for this condition. */
81       ++lmP->sgrad;
82       stepType = LMVM_SCALED_GRADIENT;
83     } else {
84       ierr = MatLMVMGetUpdates(lmP->M, &nupdates); CHKERRQ(ierr);
85       if (1 == nupdates) {
86         /*  The first BFGS direction is always the scaled gradient */
87         ++lmP->sgrad;
88         stepType = LMVM_SCALED_GRADIENT;
89       } else {
90         ++lmP->bfgs;
91         stepType = LMVM_BFGS;
92       }
93     }
94     ierr = VecScale(lmP->D, -1.0);CHKERRQ(ierr);
95 
96     /*  Perform the linesearch */
97     fold = f;
98     ierr = VecCopy(tao->solution, lmP->Xold);CHKERRQ(ierr);
99     ierr = VecCopy(tao->gradient, lmP->Gold);CHKERRQ(ierr);
100 
101     ierr = TaoLineSearchApply(tao->linesearch, tao->solution, &f, tao->gradient, lmP->D, &step,&ls_status);CHKERRQ(ierr);
102     ierr = TaoAddLineSearchCounts(tao);CHKERRQ(ierr);
103 
104     while (ls_status != TAOLINESEARCH_SUCCESS && ls_status != TAOLINESEARCH_SUCCESS_USER && (stepType != LMVM_GRADIENT)) {
105       /*  Linesearch failed */
106       /*  Reset factors and use scaled gradient step */
107       f = fold;
108       ierr = VecCopy(lmP->Xold, tao->solution);CHKERRQ(ierr);
109       ierr = VecCopy(lmP->Gold, tao->gradient);CHKERRQ(ierr);
110 
111       switch(stepType) {
112       case LMVM_BFGS:
113         /*  Failed to obtain acceptable iterate with BFGS step */
114         /*  Attempt to use the scaled gradient direction */
115 
116         if (f != 0.0) {
117           delta = 2.0 * PetscAbsScalar(f) / (gnorm*gnorm);
118         } else {
119           delta = 2.0 / (gnorm*gnorm);
120         }
121         ierr = MatLMVMSetDelta(lmP->M, delta);CHKERRQ(ierr);
122         ierr = MatLMVMReset(lmP->M);CHKERRQ(ierr);
123         ierr = MatLMVMUpdate(lmP->M, tao->solution, tao->gradient);CHKERRQ(ierr);
124         ierr = MatLMVMSolve(lmP->M, tao->gradient, lmP->D);CHKERRQ(ierr);
125 
126         /* On a reset, the direction cannot be not a number; it is a
127            scaled gradient step.  No need to check for this condition. */
128         --lmP->bfgs;
129         ++lmP->sgrad;
130         stepType = LMVM_SCALED_GRADIENT;
131         break;
132 
133       case LMVM_SCALED_GRADIENT:
134         /* The scaled gradient step did not produce a new iterate;
135            attempt to use the gradient direction.
136            Need to make sure we are not using a different diagonal scaling */
137         ierr = MatLMVMSetDelta(lmP->M, 1.0);CHKERRQ(ierr);
138         ierr = MatLMVMReset(lmP->M);CHKERRQ(ierr);
139         ierr = MatLMVMUpdate(lmP->M, tao->solution, tao->gradient);CHKERRQ(ierr);
140         ierr = MatLMVMSolve(lmP->M, tao->gradient, lmP->D);CHKERRQ(ierr);
141 
142         --lmP->sgrad;
143         ++lmP->grad;
144         stepType = LMVM_GRADIENT;
145         break;
146       }
147       ierr = VecScale(lmP->D, -1.0);CHKERRQ(ierr);
148 
149       /*  Perform the linesearch */
150       ierr = TaoLineSearchApply(tao->linesearch, tao->solution, &f, tao->gradient, lmP->D, &step, &ls_status);CHKERRQ(ierr);
151       ierr = TaoAddLineSearchCounts(tao);CHKERRQ(ierr);
152     }
153 
154     if (ls_status != TAOLINESEARCH_SUCCESS && ls_status != TAOLINESEARCH_SUCCESS_USER) {
155       /*  Failed to find an improving point */
156       f = fold;
157       ierr = VecCopy(lmP->Xold, tao->solution);CHKERRQ(ierr);
158       ierr = VecCopy(lmP->Gold, tao->gradient);CHKERRQ(ierr);
159       step = 0.0;
160       reason = TAO_DIVERGED_LS_FAILURE;
161       tao->reason = TAO_DIVERGED_LS_FAILURE;
162     }
163 
164     /*  Check for termination */
165     ierr = TaoGradientNorm(tao, tao->gradient,NORM_2,&gnorm);CHKERRQ(ierr);
166 
167     tao->niter++;
168     ierr = TaoMonitor(tao,tao->niter,f,gnorm,0.0,step,&reason);CHKERRQ(ierr);
169   }
170   PetscFunctionReturn(0);
171 }
172 
173 static PetscErrorCode TaoSetUp_LMVM(Tao tao)
174 {
175   TAO_LMVM       *lmP = (TAO_LMVM *)tao->data;
176   PetscInt       n,N;
177   PetscErrorCode ierr;
178   KSP            H0ksp;
179 
180   PetscFunctionBegin;
181   /* Existence of tao->solution checked in TaoSetUp() */
182   if (!tao->gradient) {ierr = VecDuplicate(tao->solution,&tao->gradient);CHKERRQ(ierr);  }
183   if (!tao->stepdirection) {ierr = VecDuplicate(tao->solution,&tao->stepdirection);CHKERRQ(ierr);  }
184   if (!lmP->D) {ierr = VecDuplicate(tao->solution,&lmP->D);CHKERRQ(ierr);  }
185   if (!lmP->Xold) {ierr = VecDuplicate(tao->solution,&lmP->Xold);CHKERRQ(ierr);  }
186   if (!lmP->Gold) {ierr = VecDuplicate(tao->solution,&lmP->Gold);CHKERRQ(ierr);  }
187 
188   /*  Create matrix for the limited memory approximation */
189   ierr = VecGetLocalSize(tao->solution,&n);CHKERRQ(ierr);
190   ierr = VecGetSize(tao->solution,&N);CHKERRQ(ierr);
191   ierr = MatCreateLMVM(((PetscObject)tao)->comm,n,N,&lmP->M);CHKERRQ(ierr);
192   ierr = MatLMVMAllocateVectors(lmP->M,tao->solution);CHKERRQ(ierr);
193 
194   /* If the user has set a matrix to solve as the initial H0, set the options prefix here, and set up the KSP */
195   if (lmP->H0) {
196     const char *prefix;
197     PC H0pc;
198 
199     ierr = MatLMVMSetH0(lmP->M, lmP->H0);CHKERRQ(ierr);
200     ierr = MatLMVMGetH0KSP(lmP->M, &H0ksp);CHKERRQ(ierr);
201 
202     ierr = TaoGetOptionsPrefix(tao, &prefix);CHKERRQ(ierr);
203     ierr = KSPSetOptionsPrefix(H0ksp, prefix);CHKERRQ(ierr);
204     ierr = PetscObjectAppendOptionsPrefix((PetscObject)H0ksp, "tao_h0_");CHKERRQ(ierr);
205     ierr = KSPGetPC(H0ksp, &H0pc);CHKERRQ(ierr);
206     ierr = PetscObjectAppendOptionsPrefix((PetscObject)H0pc,  "tao_h0_");CHKERRQ(ierr);
207 
208     ierr = KSPSetFromOptions(H0ksp);CHKERRQ(ierr);
209     ierr = KSPSetUp(H0ksp);CHKERRQ(ierr);
210   }
211 
212   PetscFunctionReturn(0);
213 }
214 
215 /* ---------------------------------------------------------- */
216 static PetscErrorCode TaoDestroy_LMVM(Tao tao)
217 {
218   TAO_LMVM       *lmP = (TAO_LMVM *)tao->data;
219   PetscErrorCode ierr;
220 
221   PetscFunctionBegin;
222   if (tao->setupcalled) {
223     ierr = VecDestroy(&lmP->Xold);CHKERRQ(ierr);
224     ierr = VecDestroy(&lmP->Gold);CHKERRQ(ierr);
225     ierr = VecDestroy(&lmP->D);CHKERRQ(ierr);
226     ierr = MatDestroy(&lmP->M);CHKERRQ(ierr);
227   }
228 
229   if (lmP->H0) {
230     ierr = PetscObjectDereference((PetscObject)lmP->H0);CHKERRQ(ierr);
231   }
232 
233   ierr = PetscFree(tao->data);CHKERRQ(ierr);
234 
235   PetscFunctionReturn(0);
236 }
237 
238 /*------------------------------------------------------------*/
239 static PetscErrorCode TaoSetFromOptions_LMVM(PetscOptionItems *PetscOptionsObject,Tao tao)
240 {
241   PetscErrorCode ierr;
242 
243   PetscFunctionBegin;
244   ierr = PetscOptionsHead(PetscOptionsObject,"Limited-memory variable-metric method for unconstrained optimization");CHKERRQ(ierr);
245   ierr = TaoLineSearchSetFromOptions(tao->linesearch);CHKERRQ(ierr);
246   ierr = PetscOptionsTail();CHKERRQ(ierr);
247   PetscFunctionReturn(0);
248 }
249 
250 /*------------------------------------------------------------*/
251 static PetscErrorCode TaoView_LMVM(Tao tao, PetscViewer viewer)
252 {
253   TAO_LMVM       *lm = (TAO_LMVM *)tao->data;
254   PetscBool      isascii, recycle;
255   PetscInt       recycled_its;
256   PetscErrorCode ierr;
257 
258   PetscFunctionBegin;
259   ierr = PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERASCII, &isascii);CHKERRQ(ierr);
260   if (isascii) {
261     ierr = PetscViewerASCIIPushTab(viewer);CHKERRQ(ierr);
262     ierr = PetscViewerASCIIPrintf(viewer, "BFGS steps: %D\n", lm->bfgs);CHKERRQ(ierr);
263     ierr = PetscViewerASCIIPrintf(viewer, "Scaled gradient steps: %D\n", lm->sgrad);CHKERRQ(ierr);
264     ierr = PetscViewerASCIIPrintf(viewer, "Gradient steps: %D\n", lm->grad);CHKERRQ(ierr);
265     ierr = MatLMVMGetRecycleFlag(lm->M, &recycle);CHKERRQ(ierr);
266     if (recycle) {
267       ierr = PetscViewerASCIIPrintf(viewer, "Recycle: on\n");CHKERRQ(ierr);
268       recycled_its = lm->bfgs + lm->sgrad + lm->grad;
269       ierr = PetscViewerASCIIPrintf(viewer, "Total recycled iterations: %D\n", recycled_its);CHKERRQ(ierr);
270     }
271     ierr = PetscViewerASCIIPopTab(viewer);CHKERRQ(ierr);
272   }
273   PetscFunctionReturn(0);
274 }
275 
276 /* ---------------------------------------------------------- */
277 
278 /*MC
279      TAOLMVM - Limited Memory Variable Metric method is a quasi-Newton
280      optimization solver for unconstrained minimization. It solves
281      the Newton step
282               Hkdk = - gk
283 
284      using an approximation Bk in place of Hk, where Bk is composed using
285      the BFGS update formula. A More-Thuente line search is then used
286      to computed the steplength in the dk direction
287   Options Database Keys:
288 +     -tao_lmm_vectors - number of vectors to use for approximation
289 .     -tao_lmm_scale_type - "none","scalar","broyden"
290 .     -tao_lmm_limit_type - "none","average","relative","absolute"
291 .     -tao_lmm_rescale_type - "none","scalar","gl"
292 .     -tao_lmm_limit_mu - mu limiting factor
293 .     -tao_lmm_limit_nu - nu limiting factor
294 .     -tao_lmm_delta_min - minimum delta value
295 .     -tao_lmm_delta_max - maximum delta value
296 .     -tao_lmm_broyden_phi - phi factor for Broyden scaling
297 .     -tao_lmm_scalar_alpha - alpha factor for scalar scaling
298 .     -tao_lmm_rescale_alpha - alpha factor for rescaling diagonal
299 .     -tao_lmm_rescale_beta - beta factor for rescaling diagonal
300 .     -tao_lmm_scalar_history - amount of history for scalar scaling
301 .     -tao_lmm_rescale_history - amount of history for rescaling diagonal
302 -     -tao_lmm_eps - rejection tolerance
303 
304   Level: beginner
305 M*/
306 
307 PETSC_EXTERN PetscErrorCode TaoCreate_LMVM(Tao tao)
308 {
309   TAO_LMVM       *lmP;
310   const char     *morethuente_type = TAOLINESEARCHMT;
311   PetscErrorCode ierr;
312 
313   PetscFunctionBegin;
314   tao->ops->setup = TaoSetUp_LMVM;
315   tao->ops->solve = TaoSolve_LMVM;
316   tao->ops->view = TaoView_LMVM;
317   tao->ops->setfromoptions = TaoSetFromOptions_LMVM;
318   tao->ops->destroy = TaoDestroy_LMVM;
319 
320   ierr = PetscNewLog(tao,&lmP);CHKERRQ(ierr);
321   lmP->D = 0;
322   lmP->M = 0;
323   lmP->Xold = 0;
324   lmP->Gold = 0;
325   lmP->H0   = NULL;
326 
327   tao->data = (void*)lmP;
328   /* Override default settings (unless already changed) */
329   if (!tao->max_it_changed) tao->max_it = 2000;
330   if (!tao->max_funcs_changed) tao->max_funcs = 4000;
331 
332   ierr = TaoLineSearchCreate(((PetscObject)tao)->comm,&tao->linesearch);CHKERRQ(ierr);
333   ierr = PetscObjectIncrementTabLevel((PetscObject)tao->linesearch, (PetscObject)tao, 1);CHKERRQ(ierr);
334   ierr = TaoLineSearchSetType(tao->linesearch,morethuente_type);CHKERRQ(ierr);
335   ierr = TaoLineSearchUseTaoRoutines(tao->linesearch,tao);CHKERRQ(ierr);
336   ierr = TaoLineSearchSetOptionsPrefix(tao->linesearch,tao->hdr.prefix);CHKERRQ(ierr);
337   PetscFunctionReturn(0);
338 }
339