xref: /petsc/src/tao/unconstrained/impls/owlqn/owlqn.c (revision 3fc23ce8277e00a65c4b1be6e2b4535b56ba7978)
1 #include <petsctaolinesearch.h>
2 #include <../src/tao/unconstrained/impls/owlqn/owlqn.h>
3 
4 #define OWLQN_BFGS                0
5 #define OWLQN_SCALED_GRADIENT     1
6 #define OWLQN_GRADIENT            2
7 
8 static PetscErrorCode ProjDirect_OWLQN(Vec d, Vec g)
9 {
10   const PetscReal *gptr;
11   PetscReal       *dptr;
12   PetscInt        low,high,low1,high1,i;
13 
14   PetscFunctionBegin;
15   PetscCall(VecGetOwnershipRange(d,&low,&high));
16   PetscCall(VecGetOwnershipRange(g,&low1,&high1));
17 
18   PetscCall(VecGetArrayRead(g,&gptr));
19   PetscCall(VecGetArray(d,&dptr));
20   for (i = 0; i < high-low; i++) {
21     if (dptr[i] * gptr[i] <= 0.0) {
22       dptr[i] = 0.0;
23     }
24   }
25   PetscCall(VecRestoreArray(d,&dptr));
26   PetscCall(VecRestoreArrayRead(g,&gptr));
27   PetscFunctionReturn(0);
28 }
29 
30 static PetscErrorCode ComputePseudoGrad_OWLQN(Vec x, Vec gv, PetscReal lambda)
31 {
32   const PetscReal *xptr;
33   PetscReal       *gptr;
34   PetscInt        low,high,low1,high1,i;
35 
36   PetscFunctionBegin;
37   PetscCall(VecGetOwnershipRange(x,&low,&high));
38   PetscCall(VecGetOwnershipRange(gv,&low1,&high1));
39 
40   PetscCall(VecGetArrayRead(x,&xptr));
41   PetscCall(VecGetArray(gv,&gptr));
42   for (i = 0; i < high-low; i++) {
43     if (xptr[i] < 0.0)               gptr[i] = gptr[i] - lambda;
44     else if (xptr[i] > 0.0)          gptr[i] = gptr[i] + lambda;
45     else if (gptr[i] + lambda < 0.0) gptr[i] = gptr[i] + lambda;
46     else if (gptr[i] - lambda > 0.0) gptr[i] = gptr[i] - lambda;
47     else                             gptr[i] = 0.0;
48   }
49   PetscCall(VecRestoreArray(gv,&gptr));
50   PetscCall(VecRestoreArrayRead(x,&xptr));
51   PetscFunctionReturn(0);
52 }
53 
54 static PetscErrorCode TaoSolve_OWLQN(Tao tao)
55 {
56   TAO_OWLQN                    *lmP = (TAO_OWLQN *)tao->data;
57   PetscReal                    f, fold, gdx, gnorm;
58   PetscReal                    step = 1.0;
59   PetscReal                    delta;
60   PetscInt                     stepType;
61   PetscInt                     iter = 0;
62   TaoLineSearchConvergedReason ls_status = TAOLINESEARCH_CONTINUE_ITERATING;
63 
64   PetscFunctionBegin;
65   if (tao->XL || tao->XU || tao->ops->computebounds) {
66     PetscCall(PetscInfo(tao,"WARNING: Variable bounds have been set but will be ignored by owlqn algorithm\n"));
67   }
68 
69   /* Check convergence criteria */
70   PetscCall(TaoComputeObjectiveAndGradient(tao, tao->solution, &f, tao->gradient));
71   PetscCall(VecCopy(tao->gradient, lmP->GV));
72   PetscCall(ComputePseudoGrad_OWLQN(tao->solution,lmP->GV,lmP->lambda));
73   PetscCall(VecNorm(lmP->GV,NORM_2,&gnorm));
74   PetscCheck(!PetscIsInfOrNanReal(f) && !PetscIsInfOrNanReal(gnorm),PetscObjectComm((PetscObject)tao),PETSC_ERR_USER, "User provided compute function generated Inf or NaN");
75 
76   tao->reason = TAO_CONTINUE_ITERATING;
77   PetscCall(TaoLogConvergenceHistory(tao,f,gnorm,0.0,tao->ksp_its));
78   PetscCall(TaoMonitor(tao,iter,f,gnorm,0.0,step));
79   PetscCall((*tao->ops->convergencetest)(tao,tao->cnvP));
80   if (tao->reason != TAO_CONTINUE_ITERATING) PetscFunctionReturn(0);
81 
82   /* Set initial scaling for the function */
83   delta = 2.0 * PetscMax(1.0, PetscAbsScalar(f)) / (gnorm*gnorm);
84   PetscCall(MatLMVMSetJ0Scale(lmP->M, delta));
85 
86   /* Set counter for gradient/reset steps */
87   lmP->bfgs = 0;
88   lmP->sgrad = 0;
89   lmP->grad = 0;
90 
91   /* Have not converged; continue with Newton method */
92   while (tao->reason == TAO_CONTINUE_ITERATING) {
93     /* Call general purpose update function */
94     if (tao->ops->update) {
95       PetscCall((*tao->ops->update)(tao, tao->niter, tao->user_update));
96     }
97 
98     /* Compute direction */
99     PetscCall(MatLMVMUpdate(lmP->M,tao->solution,tao->gradient));
100     PetscCall(MatSolve(lmP->M, lmP->GV, lmP->D));
101 
102     PetscCall(ProjDirect_OWLQN(lmP->D,lmP->GV));
103 
104     ++lmP->bfgs;
105 
106     /* Check for success (descent direction) */
107     PetscCall(VecDot(lmP->D, lmP->GV , &gdx));
108     if ((gdx <= 0.0) || PetscIsInfOrNanReal(gdx)) {
109 
110       /* Step is not descent or direction produced not a number
111          We can assert bfgsUpdates > 1 in this case because
112          the first solve produces the scaled gradient direction,
113          which is guaranteed to be descent
114 
115          Use steepest descent direction (scaled) */
116       ++lmP->grad;
117 
118       delta = 2.0 * PetscMax(1.0, PetscAbsScalar(f)) / (gnorm*gnorm);
119       PetscCall(MatLMVMSetJ0Scale(lmP->M, delta));
120       PetscCall(MatLMVMReset(lmP->M, PETSC_FALSE));
121       PetscCall(MatLMVMUpdate(lmP->M, tao->solution, tao->gradient));
122       PetscCall(MatSolve(lmP->M,lmP->GV, lmP->D));
123 
124       PetscCall(ProjDirect_OWLQN(lmP->D,lmP->GV));
125 
126       lmP->bfgs = 1;
127       ++lmP->sgrad;
128       stepType = OWLQN_SCALED_GRADIENT;
129     } else {
130       if (1 == lmP->bfgs) {
131         /* The first BFGS direction is always the scaled gradient */
132         ++lmP->sgrad;
133         stepType = OWLQN_SCALED_GRADIENT;
134       } else {
135         ++lmP->bfgs;
136         stepType = OWLQN_BFGS;
137       }
138     }
139 
140     PetscCall(VecScale(lmP->D, -1.0));
141 
142     /* Perform the linesearch */
143     fold = f;
144     PetscCall(VecCopy(tao->solution, lmP->Xold));
145     PetscCall(VecCopy(tao->gradient, lmP->Gold));
146 
147     PetscCall(TaoLineSearchApply(tao->linesearch, tao->solution, &f, lmP->GV, lmP->D, &step,&ls_status));
148     PetscCall(TaoAddLineSearchCounts(tao));
149 
150     while (((int)ls_status < 0) && (stepType != OWLQN_GRADIENT)) {
151 
152       /* Reset factors and use scaled gradient step */
153       f = fold;
154       PetscCall(VecCopy(lmP->Xold, tao->solution));
155       PetscCall(VecCopy(lmP->Gold, tao->gradient));
156       PetscCall(VecCopy(tao->gradient, lmP->GV));
157 
158       PetscCall(ComputePseudoGrad_OWLQN(tao->solution,lmP->GV,lmP->lambda));
159 
160       switch(stepType) {
161       case OWLQN_BFGS:
162         /* Failed to obtain acceptable iterate with BFGS step
163            Attempt to use the scaled gradient direction */
164 
165         delta = 2.0 * PetscMax(1.0, PetscAbsScalar(f)) / (gnorm*gnorm);
166         PetscCall(MatLMVMSetJ0Scale(lmP->M, delta));
167         PetscCall(MatLMVMReset(lmP->M, PETSC_FALSE));
168         PetscCall(MatLMVMUpdate(lmP->M, tao->solution, tao->gradient));
169         PetscCall(MatSolve(lmP->M, lmP->GV, lmP->D));
170 
171         PetscCall(ProjDirect_OWLQN(lmP->D,lmP->GV));
172 
173         lmP->bfgs = 1;
174         ++lmP->sgrad;
175         stepType = OWLQN_SCALED_GRADIENT;
176         break;
177 
178       case OWLQN_SCALED_GRADIENT:
179         /* The scaled gradient step did not produce a new iterate;
180            attempt to use the gradient direction.
181            Need to make sure we are not using a different diagonal scaling */
182         PetscCall(MatLMVMSetJ0Scale(lmP->M, 1.0));
183         PetscCall(MatLMVMReset(lmP->M, PETSC_FALSE));
184         PetscCall(MatLMVMUpdate(lmP->M, tao->solution, tao->gradient));
185         PetscCall(MatSolve(lmP->M, lmP->GV, lmP->D));
186 
187         PetscCall(ProjDirect_OWLQN(lmP->D,lmP->GV));
188 
189         lmP->bfgs = 1;
190         ++lmP->grad;
191         stepType = OWLQN_GRADIENT;
192         break;
193       }
194       PetscCall(VecScale(lmP->D, -1.0));
195 
196       /* Perform the linesearch */
197       PetscCall(TaoLineSearchApply(tao->linesearch, tao->solution, &f, lmP->GV, lmP->D, &step, &ls_status));
198       PetscCall(TaoAddLineSearchCounts(tao));
199     }
200 
201     if ((int)ls_status < 0) {
202       /* Failed to find an improving point*/
203       f = fold;
204       PetscCall(VecCopy(lmP->Xold, tao->solution));
205       PetscCall(VecCopy(lmP->Gold, tao->gradient));
206       PetscCall(VecCopy(tao->gradient, lmP->GV));
207       step = 0.0;
208     } else {
209       /* a little hack here, because that gv is used to store g */
210       PetscCall(VecCopy(lmP->GV, tao->gradient));
211     }
212 
213     PetscCall(ComputePseudoGrad_OWLQN(tao->solution,lmP->GV,lmP->lambda));
214 
215     /* Check for termination */
216 
217     PetscCall(VecNorm(lmP->GV,NORM_2,&gnorm));
218 
219     iter++;
220     PetscCall(TaoLogConvergenceHistory(tao,f,gnorm,0.0,tao->ksp_its));
221     PetscCall(TaoMonitor(tao,iter,f,gnorm,0.0,step));
222     PetscCall((*tao->ops->convergencetest)(tao,tao->cnvP));
223 
224     if ((int)ls_status < 0) break;
225   }
226   PetscFunctionReturn(0);
227 }
228 
229 static PetscErrorCode TaoSetUp_OWLQN(Tao tao)
230 {
231   TAO_OWLQN      *lmP = (TAO_OWLQN *)tao->data;
232   PetscInt       n,N;
233 
234   PetscFunctionBegin;
235   /* Existence of tao->solution checked in TaoSetUp() */
236   if (!tao->gradient) PetscCall(VecDuplicate(tao->solution,&tao->gradient));
237   if (!tao->stepdirection) PetscCall(VecDuplicate(tao->solution,&tao->stepdirection));
238   if (!lmP->D) PetscCall(VecDuplicate(tao->solution,&lmP->D));
239   if (!lmP->GV) PetscCall(VecDuplicate(tao->solution,&lmP->GV));
240   if (!lmP->Xold) PetscCall(VecDuplicate(tao->solution,&lmP->Xold));
241   if (!lmP->Gold) PetscCall(VecDuplicate(tao->solution,&lmP->Gold));
242 
243   /* Create matrix for the limited memory approximation */
244   PetscCall(VecGetLocalSize(tao->solution,&n));
245   PetscCall(VecGetSize(tao->solution,&N));
246   PetscCall(MatCreateLMVMBFGS(((PetscObject)tao)->comm,n,N,&lmP->M));
247   PetscCall(MatLMVMAllocate(lmP->M,tao->solution,tao->gradient));
248   PetscFunctionReturn(0);
249 }
250 
251 /* ---------------------------------------------------------- */
252 static PetscErrorCode TaoDestroy_OWLQN(Tao tao)
253 {
254   TAO_OWLQN      *lmP = (TAO_OWLQN *)tao->data;
255 
256   PetscFunctionBegin;
257   if (tao->setupcalled) {
258     PetscCall(VecDestroy(&lmP->Xold));
259     PetscCall(VecDestroy(&lmP->Gold));
260     PetscCall(VecDestroy(&lmP->D));
261     PetscCall(MatDestroy(&lmP->M));
262     PetscCall(VecDestroy(&lmP->GV));
263   }
264   PetscCall(PetscFree(tao->data));
265   PetscFunctionReturn(0);
266 }
267 
268 /*------------------------------------------------------------*/
269 static PetscErrorCode TaoSetFromOptions_OWLQN(PetscOptionItems *PetscOptionsObject,Tao tao)
270 {
271   TAO_OWLQN      *lmP = (TAO_OWLQN *)tao->data;
272 
273   PetscFunctionBegin;
274   PetscOptionsHeadBegin(PetscOptionsObject,"Orthant-Wise Limited-memory method for Quasi-Newton unconstrained optimization");
275   PetscCall(PetscOptionsReal("-tao_owlqn_lambda", "regulariser weight","", 100,&lmP->lambda,NULL));
276   PetscOptionsHeadEnd();
277   PetscCall(TaoLineSearchSetFromOptions(tao->linesearch));
278   PetscFunctionReturn(0);
279 }
280 
281 /*------------------------------------------------------------*/
282 static PetscErrorCode TaoView_OWLQN(Tao tao, PetscViewer viewer)
283 {
284   TAO_OWLQN      *lm = (TAO_OWLQN *)tao->data;
285   PetscBool      isascii;
286 
287   PetscFunctionBegin;
288   PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERASCII, &isascii));
289   if (isascii) {
290     PetscCall(PetscViewerASCIIPushTab(viewer));
291     PetscCall(PetscViewerASCIIPrintf(viewer, "BFGS steps: %" PetscInt_FMT "\n", lm->bfgs));
292     PetscCall(PetscViewerASCIIPrintf(viewer, "Scaled gradient steps: %" PetscInt_FMT "\n", lm->sgrad));
293     PetscCall(PetscViewerASCIIPrintf(viewer, "Gradient steps: %" PetscInt_FMT "\n", lm->grad));
294     PetscCall(PetscViewerASCIIPopTab(viewer));
295   }
296   PetscFunctionReturn(0);
297 }
298 
299 /* ---------------------------------------------------------- */
300 /*MC
301   TAOOWLQN - orthant-wise limited memory quasi-newton algorithm
302 
303 . - tao_owlqn_lambda - regulariser weight
304 
305   Level: beginner
306 M*/
307 
308 PETSC_EXTERN PetscErrorCode TaoCreate_OWLQN(Tao tao)
309 {
310   TAO_OWLQN      *lmP;
311   const char     *owarmijo_type = TAOLINESEARCHOWARMIJO;
312 
313   PetscFunctionBegin;
314   tao->ops->setup = TaoSetUp_OWLQN;
315   tao->ops->solve = TaoSolve_OWLQN;
316   tao->ops->view = TaoView_OWLQN;
317   tao->ops->setfromoptions = TaoSetFromOptions_OWLQN;
318   tao->ops->destroy = TaoDestroy_OWLQN;
319 
320   PetscCall(PetscNewLog(tao,&lmP));
321   lmP->D = NULL;
322   lmP->M = NULL;
323   lmP->GV = NULL;
324   lmP->Xold = NULL;
325   lmP->Gold = NULL;
326   lmP->lambda = 1.0;
327 
328   tao->data = (void*)lmP;
329   /* Override default settings (unless already changed) */
330   if (!tao->max_it_changed) tao->max_it = 2000;
331   if (!tao->max_funcs_changed) tao->max_funcs = 4000;
332 
333   PetscCall(TaoLineSearchCreate(((PetscObject)tao)->comm,&tao->linesearch));
334   PetscCall(PetscObjectIncrementTabLevel((PetscObject)tao->linesearch, (PetscObject)tao, 1));
335   PetscCall(TaoLineSearchSetType(tao->linesearch,owarmijo_type));
336   PetscCall(TaoLineSearchUseTaoRoutines(tao->linesearch,tao));
337   PetscCall(TaoLineSearchSetOptionsPrefix(tao->linesearch,tao->hdr.prefix));
338   PetscFunctionReturn(0);
339 }
340