Actual source code: brgn.c

  1: #include <../src/tao/leastsquares/impls/brgn/brgn.h>

  3: #define BRGN_REGULARIZATION_USER    0
  4: #define BRGN_REGULARIZATION_L2PROX  1
  5: #define BRGN_REGULARIZATION_L2PURE  2
  6: #define BRGN_REGULARIZATION_L1DICT  3
  7: #define BRGN_REGULARIZATION_LM      4
  8: #define BRGN_REGULARIZATION_TYPES   5

 10: static const char *BRGN_REGULARIZATION_TABLE[64] = {"user","l2prox","l2pure","l1dict","lm"};

 12: static PetscErrorCode GNHessianProd(Mat H,Vec in,Vec out)
 13: {
 14:   TAO_BRGN              *gn;
 15:   PetscErrorCode        ierr;

 18:   MatShellGetContext(H,&gn);
 19:   MatMult(gn->subsolver->ls_jac,in,gn->r_work);
 20:   MatMultTranspose(gn->subsolver->ls_jac,gn->r_work,out);
 21:   switch (gn->reg_type) {
 22:   case BRGN_REGULARIZATION_USER:
 23:     MatMult(gn->Hreg,in,gn->x_work);
 24:     VecAXPY(out,gn->lambda,gn->x_work);
 25:     break;
 26:   case BRGN_REGULARIZATION_L2PURE:
 27:     VecAXPY(out,gn->lambda,in);
 28:     break;
 29:   case BRGN_REGULARIZATION_L2PROX:
 30:     VecAXPY(out,gn->lambda,in);
 31:     break;
 32:   case BRGN_REGULARIZATION_L1DICT:
 33:     /* out = out + lambda*D'*(diag.*(D*in)) */
 34:     if (gn->D) {
 35:       MatMult(gn->D,in,gn->y);/* y = D*in */
 36:     } else {
 37:       VecCopy(in,gn->y);
 38:     }
 39:     VecPointwiseMult(gn->y_work,gn->diag,gn->y);   /* y_work = diag.*(D*in), where diag = epsilon^2 ./ sqrt(x.^2+epsilon^2).^3 */
 40:     if (gn->D) {
 41:       MatMultTranspose(gn->D,gn->y_work,gn->x_work); /* x_work = D'*(diag.*(D*in)) */
 42:     } else {
 43:       VecCopy(gn->y_work,gn->x_work);
 44:     }
 45:     VecAXPY(out,gn->lambda,gn->x_work);
 46:     break;
 47:   case BRGN_REGULARIZATION_LM:
 48:     VecPointwiseMult(gn->x_work,gn->damping,in);
 49:     VecAXPY(out,1,gn->x_work);
 50:     break;
 51:   }
 52:   return(0);
 53: }
 54: static PetscErrorCode ComputeDamping(TAO_BRGN *gn)
 55: {
 56:   const PetscScalar *diag_ary;
 57:   PetscScalar       *damping_ary;
 58:   PetscInt          i,n;
 59:   PetscErrorCode    ierr;

 62:   /* update damping */
 63:   VecGetArray(gn->damping,&damping_ary);
 64:   VecGetArrayRead(gn->diag,&diag_ary);
 65:   VecGetLocalSize(gn->damping,&n);
 66:   for (i=0; i<n; i++) {
 67:     damping_ary[i] = PetscClipInterval(diag_ary[i],PETSC_SQRT_MACHINE_EPSILON,PetscSqrtReal(PETSC_MAX_REAL));
 68:   }
 69:   VecScale(gn->damping,gn->lambda);
 70:   VecRestoreArray(gn->damping,&damping_ary);
 71:   VecRestoreArrayRead(gn->diag,&diag_ary);
 72:   return(0);
 73: }

 75: PetscErrorCode TaoBRGNGetDampingVector(Tao tao,Vec *d)
 76: {
 77:   TAO_BRGN *gn = (TAO_BRGN *)tao->data;

 80:   if (gn->reg_type != BRGN_REGULARIZATION_LM) SETERRQ(PetscObjectComm((PetscObject)tao),PETSC_ERR_SUP,"Damping vector is only available if regularization type is lm.");
 81:   *d = gn->damping;
 82:   return(0);
 83: }

 85: static PetscErrorCode GNObjectiveGradientEval(Tao tao,Vec X,PetscReal *fcn,Vec G,void *ptr)
 86: {
 87:   TAO_BRGN              *gn = (TAO_BRGN *)ptr;
 88:   PetscInt              K;                    /* dimension of D*X */
 89:   PetscScalar           yESum;
 90:   PetscErrorCode        ierr;
 91:   PetscReal             f_reg;

 94:   /* compute objective *fcn*/
 95:   /* compute first term 0.5*||ls_res||_2^2 */
 96:   TaoComputeResidual(tao,X,tao->ls_res);
 97:   VecDot(tao->ls_res,tao->ls_res,fcn);
 98:   *fcn *= 0.5;
 99:   /* compute gradient G */
100:   TaoComputeResidualJacobian(tao,X,tao->ls_jac,tao->ls_jac_pre);
101:   MatMultTranspose(tao->ls_jac,tao->ls_res,G);
102:   /* add the regularization contribution */
103:   switch (gn->reg_type) {
104:   case BRGN_REGULARIZATION_USER:
105:     (*gn->regularizerobjandgrad)(tao,X,&f_reg,gn->x_work,gn->reg_obj_ctx);
106:     *fcn += gn->lambda*f_reg;
107:     VecAXPY(G,gn->lambda,gn->x_work);
108:     break;
109:   case BRGN_REGULARIZATION_L2PURE:
110:     /* compute f = f + lambda*0.5*xk'*xk */
111:     VecDot(X,X,&f_reg);
112:     *fcn += gn->lambda*0.5*f_reg;
113:     /* compute G = G + lambda*xk */
114:     VecAXPY(G,gn->lambda,X);
115:     break;
116:   case BRGN_REGULARIZATION_L2PROX:
117:     /* compute f = f + lambda*0.5*(xk - xkm1)'*(xk - xkm1) */
118:     VecAXPBYPCZ(gn->x_work,1.0,-1.0,0.0,X,gn->x_old);
119:     VecDot(gn->x_work,gn->x_work,&f_reg);
120:     *fcn += gn->lambda*0.5*f_reg;
121:     /* compute G = G + lambda*(xk - xkm1) */
122:     VecAXPBYPCZ(G,gn->lambda,-gn->lambda,1.0,X,gn->x_old);
123:     break;
124:   case BRGN_REGULARIZATION_L1DICT:
125:     /* compute f = f + lambda*sum(sqrt(y.^2+epsilon^2) - epsilon), where y = D*x*/
126:     if (gn->D) {
127:       MatMult(gn->D,X,gn->y);/* y = D*x */
128:     } else {
129:       VecCopy(X,gn->y);
130:     }
131:     VecPointwiseMult(gn->y_work,gn->y,gn->y);
132:     VecShift(gn->y_work,gn->epsilon*gn->epsilon);
133:     VecSqrtAbs(gn->y_work);  /* gn->y_work = sqrt(y.^2+epsilon^2) */
134:     VecSum(gn->y_work,&yESum);
135:     VecGetSize(gn->y,&K);
136:     *fcn += gn->lambda*(yESum - K*gn->epsilon);
137:     /* compute G = G + lambda*D'*(y./sqrt(y.^2+epsilon^2)),where y = D*x */
138:     VecPointwiseDivide(gn->y_work,gn->y,gn->y_work); /* reuse y_work = y./sqrt(y.^2+epsilon^2) */
139:     if (gn->D) {
140:       MatMultTranspose(gn->D,gn->y_work,gn->x_work);
141:     } else {
142:       VecCopy(gn->y_work,gn->x_work);
143:     }
144:     VecAXPY(G,gn->lambda,gn->x_work);
145:     break;
146:   }
147:   return(0);
148: }

150: static PetscErrorCode GNComputeHessian(Tao tao,Vec X,Mat H,Mat Hpre,void *ptr)
151: {
152:   TAO_BRGN       *gn = (TAO_BRGN *)ptr;
153:   PetscInt       i,n,cstart,cend;
154:   PetscScalar    *cnorms,*diag_ary;

158:   TaoComputeResidualJacobian(tao,X,tao->ls_jac,tao->ls_jac_pre);
159:   if (gn->mat_explicit) {
160:     MatTransposeMatMult(tao->ls_jac, tao->ls_jac, MAT_REUSE_MATRIX, PETSC_DEFAULT, &gn->H);
161:   }

163:   switch (gn->reg_type) {
164:   case BRGN_REGULARIZATION_USER:
165:     (*gn->regularizerhessian)(tao,X,gn->Hreg,gn->reg_hess_ctx);
166:     if (gn->mat_explicit) {
167:       MatAXPY(gn->H, 1.0, gn->Hreg, DIFFERENT_NONZERO_PATTERN);
168:     }
169:     break;
170:   case BRGN_REGULARIZATION_L2PURE:
171:     if (gn->mat_explicit) {
172:       MatShift(gn->H, gn->lambda);
173:     }
174:     break;
175:   case BRGN_REGULARIZATION_L2PROX:
176:     if (gn->mat_explicit) {
177:       MatShift(gn->H, gn->lambda);
178:     }
179:     break;
180:   case BRGN_REGULARIZATION_L1DICT:
181:     /* calculate and store diagonal matrix as a vector: diag = epsilon^2 ./ sqrt(x.^2+epsilon^2).^3* --> diag = epsilon^2 ./ sqrt(y.^2+epsilon^2).^3,where y = D*x */
182:     if (gn->D) {
183:       MatMult(gn->D,X,gn->y);/* y = D*x */
184:     } else {
185:       VecCopy(X,gn->y);
186:     }
187:     VecPointwiseMult(gn->y_work,gn->y,gn->y);
188:     VecShift(gn->y_work,gn->epsilon*gn->epsilon);
189:     VecCopy(gn->y_work,gn->diag);                  /* gn->diag = y.^2+epsilon^2 */
190:     VecSqrtAbs(gn->y_work);                        /* gn->y_work = sqrt(y.^2+epsilon^2) */
191:     VecPointwiseMult(gn->diag,gn->y_work,gn->diag);/* gn->diag = sqrt(y.^2+epsilon^2).^3 */
192:     VecReciprocal(gn->diag);
193:     VecScale(gn->diag,gn->epsilon*gn->epsilon);
194:     if (gn->mat_explicit) {
195:       MatDiagonalSet(gn->H, gn->diag, ADD_VALUES);
196:     }
197:     break;
198:   case BRGN_REGULARIZATION_LM:
199:     /* compute diagonal of J^T J */
200:     MatGetSize(gn->parent->ls_jac,NULL,&n);
201:     PetscMalloc1(n,&cnorms);
202:     MatGetColumnNorms(gn->parent->ls_jac,NORM_2,cnorms);
203:     MatGetOwnershipRangeColumn(gn->parent->ls_jac,&cstart,&cend);
204:     VecGetArray(gn->diag,&diag_ary);
205:     for (i = 0; i < cend-cstart; i++) {
206:       diag_ary[i] = cnorms[cstart+i] * cnorms[cstart+i];
207:     }
208:     VecRestoreArray(gn->diag,&diag_ary);
209:     PetscFree(cnorms);
210:     ComputeDamping(gn);
211:     if (gn->mat_explicit) {
212:       MatDiagonalSet(gn->H, gn->damping, ADD_VALUES);
213:     }
214:     break;
215:   }
216:   return(0);
217: }

219: static PetscErrorCode GNHookFunction(Tao tao,PetscInt iter, void *ctx)
220: {
221:   TAO_BRGN              *gn = (TAO_BRGN *)ctx;
222:   PetscErrorCode        ierr;

225:   /* Update basic tao information from the subsolver */
226:   gn->parent->nfuncs = tao->nfuncs;
227:   gn->parent->ngrads = tao->ngrads;
228:   gn->parent->nfuncgrads = tao->nfuncgrads;
229:   gn->parent->nhess = tao->nhess;
230:   gn->parent->niter = tao->niter;
231:   gn->parent->ksp_its = tao->ksp_its;
232:   gn->parent->ksp_tot_its = tao->ksp_tot_its;
233:   gn->parent->fc = tao->fc;
234:   TaoGetConvergedReason(tao,&gn->parent->reason);
235:   /* Update the solution vectors */
236:   if (iter == 0) {
237:     VecSet(gn->x_old,0.0);
238:   } else {
239:     VecCopy(tao->solution,gn->x_old);
240:     VecCopy(tao->solution,gn->parent->solution);
241:   }
242:   /* Update the gradient */
243:   VecCopy(tao->gradient,gn->parent->gradient);

245:   /* Update damping parameter for LM */
246:   if (gn->reg_type == BRGN_REGULARIZATION_LM) {
247:     if (iter > 0) {
248:       if (gn->fc_old > tao->fc) {
249:         gn->lambda = gn->lambda * gn->downhill_lambda_change;
250:       } else {
251:         /* uphill step */
252:         gn->lambda = gn->lambda * gn->uphill_lambda_change;
253:       }
254:     }
255:     gn->fc_old = tao->fc;
256:   }

258:   /* Call general purpose update function */
259:   if (gn->parent->ops->update) {
260:     (*gn->parent->ops->update)(gn->parent,gn->parent->niter,gn->parent->user_update);
261:   }
262:   return(0);
263: }

265: static PetscErrorCode TaoSolve_BRGN(Tao tao)
266: {
267:   TAO_BRGN              *gn = (TAO_BRGN *)tao->data;
268:   PetscErrorCode        ierr;

271:   TaoSolve(gn->subsolver);
272:   /* Update basic tao information from the subsolver */
273:   tao->nfuncs = gn->subsolver->nfuncs;
274:   tao->ngrads = gn->subsolver->ngrads;
275:   tao->nfuncgrads = gn->subsolver->nfuncgrads;
276:   tao->nhess = gn->subsolver->nhess;
277:   tao->niter = gn->subsolver->niter;
278:   tao->ksp_its = gn->subsolver->ksp_its;
279:   tao->ksp_tot_its = gn->subsolver->ksp_tot_its;
280:   TaoGetConvergedReason(gn->subsolver,&tao->reason);
281:   /* Update vectors */
282:   VecCopy(gn->subsolver->solution,tao->solution);
283:   VecCopy(gn->subsolver->gradient,tao->gradient);
284:   return(0);
285: }

287: static PetscErrorCode TaoSetFromOptions_BRGN(PetscOptionItems *PetscOptionsObject,Tao tao)
288: {
289:   TAO_BRGN              *gn = (TAO_BRGN *)tao->data;
290:   TaoLineSearch         ls;
291:   PetscErrorCode        ierr;

294:   PetscOptionsHead(PetscOptionsObject,"least-squares problems with regularizer: ||f(x)||^2 + lambda*g(x), g(x) = ||xk-xkm1||^2 or ||Dx||_1 or user defined function.");
295:   PetscOptionsBool("-tao_brgn_mat_explicit","switches the Hessian construction to be an explicit matrix rather than MATSHELL","",gn->mat_explicit,&gn->mat_explicit,NULL);
296:   PetscOptionsReal("-tao_brgn_regularizer_weight","regularizer weight (default 1e-4)","",gn->lambda,&gn->lambda,NULL);
297:   PetscOptionsReal("-tao_brgn_l1_smooth_epsilon","L1-norm smooth approximation parameter: ||x||_1 = sum(sqrt(x.^2+epsilon^2)-epsilon) (default 1e-6)","",gn->epsilon,&gn->epsilon,NULL);
298:   PetscOptionsReal("-tao_brgn_lm_downhill_lambda_change","Factor to decrease trust region by on downhill steps","",gn->downhill_lambda_change,&gn->downhill_lambda_change,NULL);
299:   PetscOptionsReal("-tao_brgn_lm_uphill_lambda_change","Factor to increase trust region by on uphill steps","",gn->uphill_lambda_change,&gn->uphill_lambda_change,NULL);
300:   PetscOptionsEList("-tao_brgn_regularization_type","regularization type", "",BRGN_REGULARIZATION_TABLE,BRGN_REGULARIZATION_TYPES,BRGN_REGULARIZATION_TABLE[gn->reg_type],&gn->reg_type,NULL);
301:   PetscOptionsTail();
302:   /* set unit line search direction as the default when using the lm regularizer */
303:   if (gn->reg_type == BRGN_REGULARIZATION_LM) {
304:     TaoGetLineSearch(gn->subsolver,&ls);
305:     TaoLineSearchSetType(ls,TAOLINESEARCHUNIT);
306:   }
307:   TaoSetFromOptions(gn->subsolver);
308:   return(0);
309: }

311: static PetscErrorCode TaoView_BRGN(Tao tao,PetscViewer viewer)
312: {
313:   TAO_BRGN              *gn = (TAO_BRGN *)tao->data;
314:   PetscErrorCode        ierr;

317:   PetscViewerASCIIPushTab(viewer);
318:   TaoView(gn->subsolver,viewer);
319:   PetscViewerASCIIPopTab(viewer);
320:   return(0);
321: }

323: static PetscErrorCode TaoSetUp_BRGN(Tao tao)
324: {
325:   TAO_BRGN              *gn = (TAO_BRGN *)tao->data;
326:   PetscErrorCode        ierr;
327:   PetscBool             is_bnls,is_bntr,is_bntl;
328:   PetscInt              i,n,N,K; /* dict has size K*N*/

331:   if (!tao->ls_res) SETERRQ(PetscObjectComm((PetscObject)tao),PETSC_ERR_ORDER,"TaoSetResidualRoutine() must be called before setup!");
332:   PetscObjectTypeCompare((PetscObject)gn->subsolver,TAOBNLS,&is_bnls);
333:   PetscObjectTypeCompare((PetscObject)gn->subsolver,TAOBNTR,&is_bntr);
334:   PetscObjectTypeCompare((PetscObject)gn->subsolver,TAOBNTL,&is_bntl);
335:   if ((is_bnls || is_bntr || is_bntl) && !tao->ls_jac) SETERRQ(PetscObjectComm((PetscObject)tao),PETSC_ERR_ORDER,"TaoSetResidualJacobianRoutine() must be called before setup!");
336:   if (!tao->gradient) {
337:     VecDuplicate(tao->solution,&tao->gradient);
338:   }
339:   if (!gn->x_work) {
340:     VecDuplicate(tao->solution,&gn->x_work);
341:   }
342:   if (!gn->r_work) {
343:     VecDuplicate(tao->ls_res,&gn->r_work);
344:   }
345:   if (!gn->x_old) {
346:     VecDuplicate(tao->solution,&gn->x_old);
347:     VecSet(gn->x_old,0.0);
348:   }

350:   if (BRGN_REGULARIZATION_L1DICT == gn->reg_type) {
351:     if (!gn->y) {
352:       if (gn->D) {
353:         MatGetSize(gn->D,&K,&N); /* Shell matrices still must have sizes defined. K = N for identity matrix, K=N-1 or N for gradient matrix */
354:         MatCreateVecs(gn->D,NULL,&gn->y);
355:       } else {
356:         VecDuplicate(tao->solution,&gn->y); /* If user does not setup dict matrix, use identiy matrix, K=N */
357:       }
358:       VecSet(gn->y,0.0);
359:     }
360:     if (!gn->y_work) {
361:       VecDuplicate(gn->y,&gn->y_work);
362:     }
363:     if (!gn->diag) {
364:       VecDuplicate(gn->y,&gn->diag);
365:       VecSet(gn->diag,0.0);
366:     }
367:   }
368:   if (BRGN_REGULARIZATION_LM == gn->reg_type) {
369:     if (!gn->diag) {
370:       MatCreateVecs(tao->ls_jac,&gn->diag,NULL);
371:     }
372:     if (!gn->damping) {
373:       MatCreateVecs(tao->ls_jac,&gn->damping,NULL);
374:     }
375:   }

377:   if (!tao->setupcalled) {
378:     /* Hessian setup */
379:     if (gn->mat_explicit) {
380:       TaoComputeResidualJacobian(tao,tao->solution,tao->ls_jac,tao->ls_jac_pre);
381:       MatTransposeMatMult(tao->ls_jac, tao->ls_jac, MAT_INITIAL_MATRIX, PETSC_DEFAULT, &gn->H);
382:     } else {
383:       VecGetLocalSize(tao->solution,&n);
384:       VecGetSize(tao->solution,&N);
385:       MatCreate(PetscObjectComm((PetscObject)tao),&gn->H);
386:       MatSetSizes(gn->H,n,n,N,N);
387:       MatSetType(gn->H,MATSHELL);
388:       MatSetOption(gn->H, MAT_SYMMETRIC, PETSC_TRUE);
389:       MatShellSetOperation(gn->H,MATOP_MULT,(void (*)(void))GNHessianProd);
390:       MatShellSetContext(gn->H,(void*)gn);
391:     }
392:     MatSetUp(gn->H);
393:     /* Subsolver setup,include initial vector and dicttionary D */
394:     TaoSetUpdate(gn->subsolver,GNHookFunction,(void*)gn);
395:     TaoSetInitialVector(gn->subsolver,tao->solution);
396:     if (tao->bounded) {
397:       TaoSetVariableBounds(gn->subsolver,tao->XL,tao->XU);
398:     }
399:     TaoSetResidualRoutine(gn->subsolver,tao->ls_res,tao->ops->computeresidual,tao->user_lsresP);
400:     TaoSetJacobianResidualRoutine(gn->subsolver,tao->ls_jac,tao->ls_jac,tao->ops->computeresidualjacobian,tao->user_lsjacP);
401:     TaoSetObjectiveAndGradientRoutine(gn->subsolver,GNObjectiveGradientEval,(void*)gn);
402:     TaoSetHessianRoutine(gn->subsolver,gn->H,gn->H,GNComputeHessian,(void*)gn);
403:     /* Propagate some options down */
404:     TaoSetTolerances(gn->subsolver,tao->gatol,tao->grtol,tao->gttol);
405:     TaoSetMaximumIterations(gn->subsolver,tao->max_it);
406:     TaoSetMaximumFunctionEvaluations(gn->subsolver,tao->max_funcs);
407:     for (i=0; i<tao->numbermonitors; ++i) {
408:       TaoSetMonitor(gn->subsolver,tao->monitor[i],tao->monitorcontext[i],tao->monitordestroy[i]);
409:       PetscObjectReference((PetscObject)(tao->monitorcontext[i]));
410:     }
411:     TaoSetUp(gn->subsolver);
412:   }
413:   return(0);
414: }

416: static PetscErrorCode TaoDestroy_BRGN(Tao tao)
417: {
418:   TAO_BRGN              *gn = (TAO_BRGN *)tao->data;
419:   PetscErrorCode        ierr;

422:   if (tao->setupcalled) {
423:     VecDestroy(&tao->gradient);
424:     VecDestroy(&gn->x_work);
425:     VecDestroy(&gn->r_work);
426:     VecDestroy(&gn->x_old);
427:     VecDestroy(&gn->diag);
428:     VecDestroy(&gn->y);
429:     VecDestroy(&gn->y_work);
430:   }
431:   VecDestroy(&gn->damping);
432:   VecDestroy(&gn->diag);
433:   MatDestroy(&gn->H);
434:   MatDestroy(&gn->D);
435:   MatDestroy(&gn->Hreg);
436:   TaoDestroy(&gn->subsolver);
437:   gn->parent = NULL;
438:   PetscFree(tao->data);
439:   return(0);
440: }

442: /*MC
443:   TAOBRGN - Bounded Regularized Gauss-Newton method for solving nonlinear least-squares
444:             problems with bound constraints. This algorithm is a thin wrapper around TAOBNTL
445:             that constructs the Gauss-Newton problem with the user-provided least-squares
446:             residual and Jacobian. The algorithm offers an L2-norm ("l2pure"), L2-norm proximal point ("l2prox")
447:             regularizer, and L1-norm dictionary regularizer ("l1dict"), where we approximate the
448:             L1-norm ||x||_1 by sum_i(sqrt(x_i^2+epsilon^2)-epsilon) with a small positive number epsilon.
449:             Also offered is the "lm" regularizer which uses a scaled diagonal of J^T J.
450:             With the "lm" regularizer, BRGN is a Levenberg-Marquardt optimizer.
451:             The user can also provide own regularization function.

453:   Options Database Keys:
454: + -tao_brgn_regularization_type - regularization type ("user", "l2prox", "l2pure", "l1dict", "lm") (default "l2prox")
455: . -tao_brgn_regularizer_weight  - regularizer weight (default 1e-4)
456: - -tao_brgn_l1_smooth_epsilon   - L1-norm smooth approximation parameter: ||x||_1 = sum(sqrt(x.^2+epsilon^2)-epsilon) (default 1e-6)

458:   Level: beginner
459: M*/
460: PETSC_EXTERN PetscErrorCode TaoCreate_BRGN(Tao tao)
461: {
462:   TAO_BRGN       *gn;

466:   PetscNewLog(tao,&gn);

468:   tao->ops->destroy = TaoDestroy_BRGN;
469:   tao->ops->setup = TaoSetUp_BRGN;
470:   tao->ops->setfromoptions = TaoSetFromOptions_BRGN;
471:   tao->ops->view = TaoView_BRGN;
472:   tao->ops->solve = TaoSolve_BRGN;

474:   tao->data = (void*)gn;
475:   gn->reg_type = BRGN_REGULARIZATION_L2PROX;
476:   gn->lambda = 1e-4;
477:   gn->epsilon = 1e-6;
478:   gn->downhill_lambda_change = 1./5.;
479:   gn->uphill_lambda_change = 1.5;
480:   gn->parent = tao;

482:   TaoCreate(PetscObjectComm((PetscObject)tao),&gn->subsolver);
483:   TaoSetType(gn->subsolver,TAOBNLS);
484:   TaoSetOptionsPrefix(gn->subsolver,"tao_brgn_subsolver_");
485:   return(0);
486: }

488: /*@
489:   TaoBRGNGetSubsolver - Get the pointer to the subsolver inside BRGN

491:   Collective on Tao

493:   Level: advanced

495:   Input Parameters:
496: +  tao - the Tao solver context
497: -  subsolver - the Tao sub-solver context
498: @*/
499: PetscErrorCode TaoBRGNGetSubsolver(Tao tao,Tao *subsolver)
500: {
501:   TAO_BRGN       *gn = (TAO_BRGN *)tao->data;

504:   *subsolver = gn->subsolver;
505:   return(0);
506: }

508: /*@
509:   TaoBRGNSetRegularizerWeight - Set the regularizer weight for the Gauss-Newton least-squares algorithm

511:   Collective on Tao

513:   Input Parameters:
514: +  tao - the Tao solver context
515: -  lambda - L1-norm regularizer weight

517:   Level: beginner
518: @*/
519: PetscErrorCode TaoBRGNSetRegularizerWeight(Tao tao,PetscReal lambda)
520: {
521:   TAO_BRGN       *gn = (TAO_BRGN *)tao->data;

523:   /* Initialize lambda here */

526:   gn->lambda = lambda;
527:   return(0);
528: }

530: /*@
531:   TaoBRGNSetL1SmoothEpsilon - Set the L1-norm smooth approximation parameter for L1-regularized least-squares algorithm

533:   Collective on Tao

535:   Input Parameters:
536: +  tao - the Tao solver context
537: -  epsilon - L1-norm smooth approximation parameter

539:   Level: advanced
540: @*/
541: PetscErrorCode TaoBRGNSetL1SmoothEpsilon(Tao tao,PetscReal epsilon)
542: {
543:   TAO_BRGN       *gn = (TAO_BRGN *)tao->data;

545:   /* Initialize epsilon here */

548:   gn->epsilon = epsilon;
549:   return(0);
550: }

552: /*@
553:    TaoBRGNSetDictionaryMatrix - bind the dictionary matrix from user application context to gn->D, for compressed sensing (with least-squares problem)

555:    Input Parameters:
556: +  tao  - the Tao context
557: -  dict - the user specified dictionary matrix.  We allow to set a null dictionary, which means identity matrix by default

559:     Level: advanced
560: @*/
561: PetscErrorCode TaoBRGNSetDictionaryMatrix(Tao tao,Mat dict)
562: {
563:   TAO_BRGN       *gn = (TAO_BRGN *)tao->data;
567:   if (dict) {
570:     PetscObjectReference((PetscObject)dict);
571:   }
572:   MatDestroy(&gn->D);
573:   gn->D = dict;
574:   return(0);
575: }

577: /*@C
578:    TaoBRGNSetRegularizerObjectiveAndGradientRoutine - Sets the user-defined regularizer call-back
579:    function into the algorithm.

581:    Input Parameters:
582:    + tao - the Tao context
583:    . func - function pointer for the regularizer value and gradient evaluation
584:    - ctx - user context for the regularizer

586:    Level: advanced
587: @*/
588: PetscErrorCode TaoBRGNSetRegularizerObjectiveAndGradientRoutine(Tao tao,PetscErrorCode (*func)(Tao,Vec,PetscReal *,Vec,void*),void *ctx)
589: {
590:   TAO_BRGN       *gn = (TAO_BRGN *)tao->data;

594:   if (ctx) {
595:     gn->reg_obj_ctx = ctx;
596:   }
597:   if (func) {
598:     gn->regularizerobjandgrad = func;
599:   }
600:   return(0);
601: }

603: /*@C
604:    TaoBRGNSetRegularizerHessianRoutine - Sets the user-defined regularizer call-back
605:    function into the algorithm.

607:    Input Parameters:
608:    + tao - the Tao context
609:    . Hreg - user-created matrix for the Hessian of the regularization term
610:    . func - function pointer for the regularizer Hessian evaluation
611:    - ctx - user context for the regularizer Hessian

613:    Level: advanced
614: @*/
615: PetscErrorCode TaoBRGNSetRegularizerHessianRoutine(Tao tao,Mat Hreg,PetscErrorCode (*func)(Tao,Vec,Mat,void*),void *ctx)
616: {
617:   TAO_BRGN       *gn = (TAO_BRGN *)tao->data;

622:   if (Hreg) {
625:   } else SETERRQ(PetscObjectComm((PetscObject)tao),PETSC_ERR_ARG_WRONG,"NULL Hessian detected! User must provide valid Hessian for the regularizer.");
626:   if (ctx) {
627:     gn->reg_hess_ctx = ctx;
628:   }
629:   if (func) {
630:     gn->regularizerhessian = func;
631:   }
632:   if (Hreg) {
633:     PetscObjectReference((PetscObject)Hreg);
634:     MatDestroy(&gn->Hreg);
635:     gn->Hreg = Hreg;
636:   }
637:   return(0);
638: }