xref: /petsc/src/tao/unconstrained/impls/lmvm/lmvm.c (revision 503c0ea9b45bcfbcebbb1ea5341243bbc69f0bea)
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   PetscInt                     stepType = LMVM_STEP_GRAD, nupdates;
13   TaoLineSearchConvergedReason ls_status = TAOLINESEARCH_CONTINUE_ITERATING;
14 
15   PetscFunctionBegin;
16 
17   if (tao->XL || tao->XU || tao->ops->computebounds) {
18     PetscCall(PetscInfo(tao,"WARNING: Variable bounds have been set but will be ignored by lmvm algorithm\n"));
19   }
20 
21   /*  Check convergence criteria */
22   PetscCall(TaoComputeObjectiveAndGradient(tao, tao->solution, &f, tao->gradient));
23   PetscCall(TaoGradientNorm(tao, tao->gradient,NORM_2,&gnorm));
24 
25   PetscCheck(!PetscIsInfOrNanReal(f) && !PetscIsInfOrNanReal(gnorm),PetscObjectComm((PetscObject)tao),PETSC_ERR_USER, "User provided compute function generated Inf or NaN");
26 
27   tao->reason = TAO_CONTINUE_ITERATING;
28   PetscCall(TaoLogConvergenceHistory(tao,f,gnorm,0.0,tao->ksp_its));
29   PetscCall(TaoMonitor(tao,tao->niter,f,gnorm,0.0,step));
30   PetscCall((*tao->ops->convergencetest)(tao,tao->cnvP));
31   if (tao->reason != TAO_CONTINUE_ITERATING) PetscFunctionReturn(0);
32 
33   /*  Set counter for gradient/reset steps */
34   if (!lmP->recycle) {
35     lmP->bfgs = 0;
36     lmP->grad = 0;
37     PetscCall(MatLMVMReset(lmP->M, PETSC_FALSE));
38   }
39 
40   /*  Have not converged; continue with Newton method */
41   while (tao->reason == TAO_CONTINUE_ITERATING) {
42     /* Call general purpose update function */
43     if (tao->ops->update) {
44       PetscCall((*tao->ops->update)(tao, tao->niter, tao->user_update));
45     }
46 
47     /*  Compute direction */
48     if (lmP->H0) {
49       PetscCall(MatLMVMSetJ0(lmP->M, lmP->H0));
50       stepType = LMVM_STEP_BFGS;
51     }
52     PetscCall(MatLMVMUpdate(lmP->M,tao->solution,tao->gradient));
53     PetscCall(MatSolve(lmP->M, tao->gradient, lmP->D));
54     PetscCall(MatLMVMGetUpdateCount(lmP->M, &nupdates));
55     if (nupdates > 0) stepType = LMVM_STEP_BFGS;
56 
57     /*  Check for success (descent direction) */
58     PetscCall(VecDot(lmP->D, tao->gradient, &gdx));
59     if ((gdx <= 0.0) || PetscIsInfOrNanReal(gdx)) {
60       /* Step is not descent or direction produced not a number
61          We can assert bfgsUpdates > 1 in this case because
62          the first solve produces the scaled gradient direction,
63          which is guaranteed to be descent
64 
65          Use steepest descent direction (scaled)
66       */
67 
68       PetscCall(MatLMVMReset(lmP->M, PETSC_FALSE));
69       PetscCall(MatLMVMClearJ0(lmP->M));
70       PetscCall(MatLMVMUpdate(lmP->M, tao->solution, tao->gradient));
71       PetscCall(MatSolve(lmP->M,tao->gradient, lmP->D));
72 
73       /* On a reset, the direction cannot be not a number; it is a
74          scaled gradient step.  No need to check for this condition. */
75       stepType = LMVM_STEP_GRAD;
76     }
77     PetscCall(VecScale(lmP->D, -1.0));
78 
79     /*  Perform the linesearch */
80     fold = f;
81     PetscCall(VecCopy(tao->solution, lmP->Xold));
82     PetscCall(VecCopy(tao->gradient, lmP->Gold));
83 
84     PetscCall(TaoLineSearchApply(tao->linesearch, tao->solution, &f, tao->gradient, lmP->D, &step,&ls_status));
85     PetscCall(TaoAddLineSearchCounts(tao));
86 
87     if (ls_status != TAOLINESEARCH_SUCCESS && ls_status != TAOLINESEARCH_SUCCESS_USER && (stepType != LMVM_STEP_GRAD)) {
88       /*  Reset factors and use scaled gradient step */
89       f = fold;
90       PetscCall(VecCopy(lmP->Xold, tao->solution));
91       PetscCall(VecCopy(lmP->Gold, tao->gradient));
92 
93       /*  Failed to obtain acceptable iterate with BFGS step */
94       /*  Attempt to use the scaled gradient direction */
95 
96       PetscCall(MatLMVMReset(lmP->M, PETSC_FALSE));
97       PetscCall(MatLMVMClearJ0(lmP->M));
98       PetscCall(MatLMVMUpdate(lmP->M, tao->solution, tao->gradient));
99       PetscCall(MatSolve(lmP->M, tao->solution, tao->gradient));
100 
101       /* On a reset, the direction cannot be not a number; it is a
102           scaled gradient step.  No need to check for this condition. */
103       stepType = LMVM_STEP_GRAD;
104       PetscCall(VecScale(lmP->D, -1.0));
105 
106       /*  Perform the linesearch */
107       PetscCall(TaoLineSearchApply(tao->linesearch, tao->solution, &f, tao->gradient, lmP->D, &step, &ls_status));
108       PetscCall(TaoAddLineSearchCounts(tao));
109     }
110 
111     if (ls_status != TAOLINESEARCH_SUCCESS && ls_status != TAOLINESEARCH_SUCCESS_USER) {
112       /*  Failed to find an improving point */
113       f = fold;
114       PetscCall(VecCopy(lmP->Xold, tao->solution));
115       PetscCall(VecCopy(lmP->Gold, tao->gradient));
116       step = 0.0;
117       tao->reason = TAO_DIVERGED_LS_FAILURE;
118     } else {
119       /* LS found valid step, so tally up step type */
120       switch (stepType) {
121       case LMVM_STEP_BFGS:
122         ++lmP->bfgs;
123         break;
124       case LMVM_STEP_GRAD:
125         ++lmP->grad;
126         break;
127       default:
128         break;
129       }
130       /*  Compute new gradient norm */
131       PetscCall(TaoGradientNorm(tao, tao->gradient,NORM_2,&gnorm));
132     }
133 
134     /* Check convergence */
135     tao->niter++;
136     PetscCall(TaoLogConvergenceHistory(tao,f,gnorm,0.0,tao->ksp_its));
137     PetscCall(TaoMonitor(tao,tao->niter,f,gnorm,0.0,step));
138     PetscCall((*tao->ops->convergencetest)(tao,tao->cnvP));
139   }
140   PetscFunctionReturn(0);
141 }
142 
143 static PetscErrorCode TaoSetUp_LMVM(Tao tao)
144 {
145   TAO_LMVM       *lmP = (TAO_LMVM *)tao->data;
146   PetscInt       n,N;
147   PetscBool      is_spd;
148 
149   PetscFunctionBegin;
150   /* Existence of tao->solution checked in TaoSetUp() */
151   if (!tao->gradient) PetscCall(VecDuplicate(tao->solution,&tao->gradient));
152   if (!tao->stepdirection) PetscCall(VecDuplicate(tao->solution,&tao->stepdirection));
153   if (!lmP->D) PetscCall(VecDuplicate(tao->solution,&lmP->D));
154   if (!lmP->Xold) PetscCall(VecDuplicate(tao->solution,&lmP->Xold));
155   if (!lmP->Gold) PetscCall(VecDuplicate(tao->solution,&lmP->Gold));
156 
157   /*  Create matrix for the limited memory approximation */
158   PetscCall(VecGetLocalSize(tao->solution,&n));
159   PetscCall(VecGetSize(tao->solution,&N));
160   PetscCall(MatSetSizes(lmP->M, n, n, N, N));
161   PetscCall(MatLMVMAllocate(lmP->M,tao->solution,tao->gradient));
162   PetscCall(MatGetOption(lmP->M, MAT_SPD, &is_spd));
163   PetscCheck(is_spd,PetscObjectComm((PetscObject)tao), PETSC_ERR_ARG_INCOMP, "LMVM matrix is not symmetric positive-definite.");
164 
165   /* If the user has set a matrix to solve as the initial H0, set the options prefix here, and set up the KSP */
166   if (lmP->H0) {
167     PetscCall(MatLMVMSetJ0(lmP->M, lmP->H0));
168   }
169 
170   PetscFunctionReturn(0);
171 }
172 
173 /* ---------------------------------------------------------- */
174 static PetscErrorCode TaoDestroy_LMVM(Tao tao)
175 {
176   TAO_LMVM       *lmP = (TAO_LMVM *)tao->data;
177 
178   PetscFunctionBegin;
179   if (tao->setupcalled) {
180     PetscCall(VecDestroy(&lmP->Xold));
181     PetscCall(VecDestroy(&lmP->Gold));
182     PetscCall(VecDestroy(&lmP->D));
183   }
184   PetscCall(MatDestroy(&lmP->M));
185   if (lmP->H0) {
186     PetscCall(PetscObjectDereference((PetscObject)lmP->H0));
187   }
188   PetscCall(PetscFree(tao->data));
189 
190   PetscFunctionReturn(0);
191 }
192 
193 /*------------------------------------------------------------*/
194 static PetscErrorCode TaoSetFromOptions_LMVM(PetscOptionItems *PetscOptionsObject,Tao tao)
195 {
196   TAO_LMVM       *lm = (TAO_LMVM *)tao->data;
197 
198   PetscFunctionBegin;
199   PetscCall(PetscOptionsHead(PetscOptionsObject,"Limited-memory variable-metric method for unconstrained optimization"));
200   PetscCall(PetscOptionsBool("-tao_lmvm_recycle","enable recycling of the BFGS matrix between subsequent TaoSolve() calls","",lm->recycle,&lm->recycle,NULL));
201   PetscCall(TaoLineSearchSetFromOptions(tao->linesearch));
202   PetscCall(MatSetFromOptions(lm->M));
203   PetscCall(PetscOptionsTail());
204   PetscFunctionReturn(0);
205 }
206 
207 /*------------------------------------------------------------*/
208 static PetscErrorCode TaoView_LMVM(Tao tao, PetscViewer viewer)
209 {
210   TAO_LMVM       *lm = (TAO_LMVM *)tao->data;
211   PetscBool      isascii;
212   PetscInt       recycled_its;
213 
214   PetscFunctionBegin;
215   PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERASCII, &isascii));
216   if (isascii) {
217     PetscCall(PetscViewerASCIIPrintf(viewer, "  Gradient steps: %D\n", lm->grad));
218     if (lm->recycle) {
219       PetscCall(PetscViewerASCIIPrintf(viewer, "  Recycle: on\n"));
220       recycled_its = lm->bfgs + lm->grad;
221       PetscCall(PetscViewerASCIIPrintf(viewer, "  Total recycled iterations: %D\n", recycled_its));
222     }
223   }
224   PetscFunctionReturn(0);
225 }
226 
227 /* ---------------------------------------------------------- */
228 
229 /*MC
230   TAOLMVM - Limited Memory Variable Metric method is a quasi-Newton
231   optimization solver for unconstrained minimization. It solves
232   the Newton step
233           Hkdk = - gk
234 
235   using an approximation Bk in place of Hk, where Bk is composed using
236   the BFGS update formula. A More-Thuente line search is then used
237   to computed the steplength in the dk direction
238 
239   Options Database Keys:
240 +   -tao_lmvm_recycle - enable recycling LMVM updates between TaoSolve() calls
241 -   -tao_lmvm_no_scale - (developer) disables diagonal Broyden scaling on the LMVM approximation
242 
243   Level: beginner
244 M*/
245 
246 PETSC_EXTERN PetscErrorCode TaoCreate_LMVM(Tao tao)
247 {
248   TAO_LMVM       *lmP;
249   const char     *morethuente_type = TAOLINESEARCHMT;
250 
251   PetscFunctionBegin;
252   tao->ops->setup = TaoSetUp_LMVM;
253   tao->ops->solve = TaoSolve_LMVM;
254   tao->ops->view = TaoView_LMVM;
255   tao->ops->setfromoptions = TaoSetFromOptions_LMVM;
256   tao->ops->destroy = TaoDestroy_LMVM;
257 
258   PetscCall(PetscNewLog(tao,&lmP));
259   lmP->D = NULL;
260   lmP->M = NULL;
261   lmP->Xold = NULL;
262   lmP->Gold = NULL;
263   lmP->H0   = NULL;
264   lmP->recycle = PETSC_FALSE;
265 
266   tao->data = (void*)lmP;
267   /* Override default settings (unless already changed) */
268   if (!tao->max_it_changed) tao->max_it = 2000;
269   if (!tao->max_funcs_changed) tao->max_funcs = 4000;
270 
271   PetscCall(TaoLineSearchCreate(((PetscObject)tao)->comm,&tao->linesearch));
272   PetscCall(PetscObjectIncrementTabLevel((PetscObject)tao->linesearch, (PetscObject)tao, 1));
273   PetscCall(TaoLineSearchSetType(tao->linesearch,morethuente_type));
274   PetscCall(TaoLineSearchUseTaoRoutines(tao->linesearch,tao));
275   PetscCall(TaoLineSearchSetOptionsPrefix(tao->linesearch,tao->hdr.prefix));
276 
277   PetscCall(KSPInitializePackage());
278   PetscCall(MatCreate(((PetscObject)tao)->comm, &lmP->M));
279   PetscCall(PetscObjectIncrementTabLevel((PetscObject)lmP->M, (PetscObject)tao, 1));
280   PetscCall(MatSetType(lmP->M, MATLMVMBFGS));
281   PetscCall(MatSetOptionsPrefix(lmP->M, "tao_lmvm_"));
282   PetscFunctionReturn(0);
283 }
284