Actual source code: owlqn.c
petsc-3.13.6 2020-09-29
1: #include <petsctaolinesearch.h>
2: #include <../src/tao/unconstrained/impls/owlqn/owlqn.h>
4: #define OWLQN_BFGS 0
5: #define OWLQN_SCALED_GRADIENT 1
6: #define OWLQN_GRADIENT 2
8: static PetscErrorCode ProjDirect_OWLQN(Vec d, Vec g)
9: {
10: PetscErrorCode ierr;
11: const PetscReal *gptr;
12: PetscReal *dptr;
13: PetscInt low,high,low1,high1,i;
16: ierr=VecGetOwnershipRange(d,&low,&high);
17: ierr=VecGetOwnershipRange(g,&low1,&high1);
19: VecGetArrayRead(g,&gptr);
20: VecGetArray(d,&dptr);
21: for (i = 0; i < high-low; i++) {
22: if (dptr[i] * gptr[i] <= 0.0 ) {
23: dptr[i] = 0.0;
24: }
25: }
26: VecRestoreArray(d,&dptr);
27: VecRestoreArrayRead(g,&gptr);
28: return(0);
29: }
31: static PetscErrorCode ComputePseudoGrad_OWLQN(Vec x, Vec gv, PetscReal lambda)
32: {
33: PetscErrorCode ierr;
34: const PetscReal *xptr;
35: PetscReal *gptr;
36: PetscInt low,high,low1,high1,i;
39: ierr=VecGetOwnershipRange(x,&low,&high);
40: ierr=VecGetOwnershipRange(gv,&low1,&high1);
42: VecGetArrayRead(x,&xptr);
43: VecGetArray(gv,&gptr);
44: for (i = 0; i < high-low; i++) {
45: if (xptr[i] < 0.0) gptr[i] = gptr[i] - lambda;
46: else if (xptr[i] > 0.0) gptr[i] = gptr[i] + lambda;
47: else if (gptr[i] + lambda < 0.0) gptr[i] = gptr[i] + lambda;
48: else if (gptr[i] - lambda > 0.0) gptr[i] = gptr[i] - lambda;
49: else gptr[i] = 0.0;
50: }
51: VecRestoreArray(gv,&gptr);
52: VecRestoreArrayRead(x,&xptr);
53: return(0);
54: }
56: static PetscErrorCode TaoSolve_OWLQN(Tao tao)
57: {
58: TAO_OWLQN *lmP = (TAO_OWLQN *)tao->data;
59: PetscReal f, fold, gdx, gnorm;
60: PetscReal step = 1.0;
61: PetscReal delta;
62: PetscErrorCode ierr;
63: PetscInt stepType;
64: PetscInt iter = 0;
65: TaoLineSearchConvergedReason ls_status = TAOLINESEARCH_CONTINUE_ITERATING;
68: if (tao->XL || tao->XU || tao->ops->computebounds) {
69: PetscInfo(tao,"WARNING: Variable bounds have been set but will be ignored by owlqn algorithm\n");
70: }
72: /* Check convergence criteria */
73: TaoComputeObjectiveAndGradient(tao, tao->solution, &f, tao->gradient);
74: VecCopy(tao->gradient, lmP->GV);
75: ComputePseudoGrad_OWLQN(tao->solution,lmP->GV,lmP->lambda);
76: VecNorm(lmP->GV,NORM_2,&gnorm);
77: if (PetscIsInfOrNanReal(f) || PetscIsInfOrNanReal(gnorm)) SETERRQ(PetscObjectComm((PetscObject)tao),PETSC_ERR_USER, "User provided compute function generated Inf or NaN");
79: tao->reason = TAO_CONTINUE_ITERATING;
80: TaoLogConvergenceHistory(tao,f,gnorm,0.0,tao->ksp_its);
81: TaoMonitor(tao,iter,f,gnorm,0.0,step);
82: (*tao->ops->convergencetest)(tao,tao->cnvP);
83: if (tao->reason != TAO_CONTINUE_ITERATING) return(0);
85: /* Set initial scaling for the function */
86: delta = 2.0 * PetscMax(1.0, PetscAbsScalar(f)) / (gnorm*gnorm);
87: MatLMVMSetJ0Scale(lmP->M, delta);
89: /* Set counter for gradient/reset steps */
90: lmP->bfgs = 0;
91: lmP->sgrad = 0;
92: lmP->grad = 0;
94: /* Have not converged; continue with Newton method */
95: while (tao->reason == TAO_CONTINUE_ITERATING) {
96: /* Call general purpose update function */
97: if (tao->ops->update) {
98: (*tao->ops->update)(tao, tao->niter, tao->user_update);
99: }
101: /* Compute direction */
102: MatLMVMUpdate(lmP->M,tao->solution,tao->gradient);
103: MatSolve(lmP->M, lmP->GV, lmP->D);
105: ProjDirect_OWLQN(lmP->D,lmP->GV);
107: ++lmP->bfgs;
109: /* Check for success (descent direction) */
110: VecDot(lmP->D, lmP->GV , &gdx);
111: if ((gdx <= 0.0) || PetscIsInfOrNanReal(gdx)) {
113: /* Step is not descent or direction produced not a number
114: We can assert bfgsUpdates > 1 in this case because
115: the first solve produces the scaled gradient direction,
116: which is guaranteed to be descent
118: Use steepest descent direction (scaled) */
119: ++lmP->grad;
121: delta = 2.0 * PetscMax(1.0, PetscAbsScalar(f)) / (gnorm*gnorm);
122: MatLMVMSetJ0Scale(lmP->M, delta);
123: MatLMVMReset(lmP->M, PETSC_FALSE);
124: MatLMVMUpdate(lmP->M, tao->solution, tao->gradient);
125: MatSolve(lmP->M,lmP->GV, lmP->D);
127: ProjDirect_OWLQN(lmP->D,lmP->GV);
129: lmP->bfgs = 1;
130: ++lmP->sgrad;
131: stepType = OWLQN_SCALED_GRADIENT;
132: } else {
133: if (1 == lmP->bfgs) {
134: /* The first BFGS direction is always the scaled gradient */
135: ++lmP->sgrad;
136: stepType = OWLQN_SCALED_GRADIENT;
137: } else {
138: ++lmP->bfgs;
139: stepType = OWLQN_BFGS;
140: }
141: }
143: VecScale(lmP->D, -1.0);
145: /* Perform the linesearch */
146: fold = f;
147: VecCopy(tao->solution, lmP->Xold);
148: VecCopy(tao->gradient, lmP->Gold);
150: TaoLineSearchApply(tao->linesearch, tao->solution, &f, lmP->GV, lmP->D, &step,&ls_status);
151: TaoAddLineSearchCounts(tao);
153: while (((int)ls_status < 0) && (stepType != OWLQN_GRADIENT)) {
155: /* Reset factors and use scaled gradient step */
156: f = fold;
157: VecCopy(lmP->Xold, tao->solution);
158: VecCopy(lmP->Gold, tao->gradient);
159: VecCopy(tao->gradient, lmP->GV);
161: ComputePseudoGrad_OWLQN(tao->solution,lmP->GV,lmP->lambda);
163: switch(stepType) {
164: case OWLQN_BFGS:
165: /* Failed to obtain acceptable iterate with BFGS step
166: Attempt to use the scaled gradient direction */
168: delta = 2.0 * PetscMax(1.0, PetscAbsScalar(f)) / (gnorm*gnorm);
169: MatLMVMSetJ0Scale(lmP->M, delta);
170: MatLMVMReset(lmP->M, PETSC_FALSE);
171: MatLMVMUpdate(lmP->M, tao->solution, tao->gradient);
172: MatSolve(lmP->M, lmP->GV, lmP->D);
174: ProjDirect_OWLQN(lmP->D,lmP->GV);
176: lmP->bfgs = 1;
177: ++lmP->sgrad;
178: stepType = OWLQN_SCALED_GRADIENT;
179: break;
181: case OWLQN_SCALED_GRADIENT:
182: /* The scaled gradient step did not produce a new iterate;
183: attempt to use the gradient direction.
184: Need to make sure we are not using a different diagonal scaling */
185: MatLMVMSetJ0Scale(lmP->M, 1.0);
186: MatLMVMReset(lmP->M, PETSC_FALSE);
187: MatLMVMUpdate(lmP->M, tao->solution, tao->gradient);
188: MatSolve(lmP->M, lmP->GV, lmP->D);
190: ProjDirect_OWLQN(lmP->D,lmP->GV);
192: lmP->bfgs = 1;
193: ++lmP->grad;
194: stepType = OWLQN_GRADIENT;
195: break;
196: }
197: VecScale(lmP->D, -1.0);
199: /* Perform the linesearch */
200: TaoLineSearchApply(tao->linesearch, tao->solution, &f, lmP->GV, lmP->D, &step, &ls_status);
201: TaoAddLineSearchCounts(tao);
202: }
204: if ((int)ls_status < 0) {
205: /* Failed to find an improving point*/
206: f = fold;
207: VecCopy(lmP->Xold, tao->solution);
208: VecCopy(lmP->Gold, tao->gradient);
209: VecCopy(tao->gradient, lmP->GV);
210: step = 0.0;
211: } else {
212: /* a little hack here, because that gv is used to store g */
213: VecCopy(lmP->GV, tao->gradient);
214: }
216: ComputePseudoGrad_OWLQN(tao->solution,lmP->GV,lmP->lambda);
218: /* Check for termination */
220: VecNorm(lmP->GV,NORM_2,&gnorm);
222: iter++;
223: TaoLogConvergenceHistory(tao,f,gnorm,0.0,tao->ksp_its);
224: TaoMonitor(tao,iter,f,gnorm,0.0,step);
225: (*tao->ops->convergencetest)(tao,tao->cnvP);
227: if ((int)ls_status < 0) break;
228: }
229: return(0);
230: }
232: static PetscErrorCode TaoSetUp_OWLQN(Tao tao)
233: {
234: TAO_OWLQN *lmP = (TAO_OWLQN *)tao->data;
235: PetscInt n,N;
239: /* Existence of tao->solution checked in TaoSetUp() */
240: if (!tao->gradient) {VecDuplicate(tao->solution,&tao->gradient); }
241: if (!tao->stepdirection) {VecDuplicate(tao->solution,&tao->stepdirection); }
242: if (!lmP->D) {VecDuplicate(tao->solution,&lmP->D); }
243: if (!lmP->GV) {VecDuplicate(tao->solution,&lmP->GV); }
244: if (!lmP->Xold) {VecDuplicate(tao->solution,&lmP->Xold); }
245: if (!lmP->Gold) {VecDuplicate(tao->solution,&lmP->Gold); }
247: /* Create matrix for the limited memory approximation */
248: VecGetLocalSize(tao->solution,&n);
249: VecGetSize(tao->solution,&N);
250: MatCreateLMVMBFGS(((PetscObject)tao)->comm,n,N,&lmP->M);
251: MatLMVMAllocate(lmP->M,tao->solution,tao->gradient);
252: return(0);
253: }
255: /* ---------------------------------------------------------- */
256: static PetscErrorCode TaoDestroy_OWLQN(Tao tao)
257: {
258: TAO_OWLQN *lmP = (TAO_OWLQN *)tao->data;
262: if (tao->setupcalled) {
263: VecDestroy(&lmP->Xold);
264: VecDestroy(&lmP->Gold);
265: VecDestroy(&lmP->D);
266: MatDestroy(&lmP->M);
267: VecDestroy(&lmP->GV);
268: }
269: PetscFree(tao->data);
270: return(0);
271: }
273: /*------------------------------------------------------------*/
274: static PetscErrorCode TaoSetFromOptions_OWLQN(PetscOptionItems *PetscOptionsObject,Tao tao)
275: {
276: TAO_OWLQN *lmP = (TAO_OWLQN *)tao->data;
280: PetscOptionsHead(PetscOptionsObject,"Orthant-Wise Limited-memory method for Quasi-Newton unconstrained optimization");
281: PetscOptionsReal("-tao_owlqn_lambda", "regulariser weight","", 100,&lmP->lambda,NULL);
282: PetscOptionsTail();
283: TaoLineSearchSetFromOptions(tao->linesearch);
284: return(0);
285: }
287: /*------------------------------------------------------------*/
288: static PetscErrorCode TaoView_OWLQN(Tao tao, PetscViewer viewer)
289: {
290: TAO_OWLQN *lm = (TAO_OWLQN *)tao->data;
291: PetscBool isascii;
295: PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERASCII, &isascii);
296: if (isascii) {
297: PetscViewerASCIIPushTab(viewer);
298: PetscViewerASCIIPrintf(viewer, "BFGS steps: %D\n", lm->bfgs);
299: PetscViewerASCIIPrintf(viewer, "Scaled gradient steps: %D\n", lm->sgrad);
300: PetscViewerASCIIPrintf(viewer, "Gradient steps: %D\n", lm->grad);
301: PetscViewerASCIIPopTab(viewer);
302: }
303: return(0);
304: }
306: /* ---------------------------------------------------------- */
307: /*MC
308: TAOOWLQN - orthant-wise limited memory quasi-newton algorithm
310: . - tao_owlqn_lambda - regulariser weight
312: Level: beginner
313: M*/
316: PETSC_EXTERN PetscErrorCode TaoCreate_OWLQN(Tao tao)
317: {
318: TAO_OWLQN *lmP;
319: const char *owarmijo_type = TAOLINESEARCHOWARMIJO;
323: tao->ops->setup = TaoSetUp_OWLQN;
324: tao->ops->solve = TaoSolve_OWLQN;
325: tao->ops->view = TaoView_OWLQN;
326: tao->ops->setfromoptions = TaoSetFromOptions_OWLQN;
327: tao->ops->destroy = TaoDestroy_OWLQN;
329: PetscNewLog(tao,&lmP);
330: lmP->D = 0;
331: lmP->M = 0;
332: lmP->GV = 0;
333: lmP->Xold = 0;
334: lmP->Gold = 0;
335: lmP->lambda = 1.0;
337: tao->data = (void*)lmP;
338: /* Override default settings (unless already changed) */
339: if (!tao->max_it_changed) tao->max_it = 2000;
340: if (!tao->max_funcs_changed) tao->max_funcs = 4000;
342: TaoLineSearchCreate(((PetscObject)tao)->comm,&tao->linesearch);
343: PetscObjectIncrementTabLevel((PetscObject)tao->linesearch, (PetscObject)tao, 1);
344: TaoLineSearchSetType(tao->linesearch,owarmijo_type);
345: TaoLineSearchUseTaoRoutines(tao->linesearch,tao);
346: TaoLineSearchSetOptionsPrefix(tao->linesearch,tao->hdr.prefix);
347: return(0);
348: }