xref: /petsc/src/tao/tutorials/ex4.c (revision 0619917b5a674bb687c64e7daba2ab22be99af31)
1 static char help[] = "Simple example to test separable objective optimizers.\n";
2 
3 #include <petsc.h>
4 #include <petsctao.h>
5 #include <petscvec.h>
6 #include <petscmath.h>
7 
8 #define NWORKLEFT  4
9 #define NWORKRIGHT 12
10 
11 typedef struct _UserCtx {
12   PetscInt    m;       /* The row dimension of F */
13   PetscInt    n;       /* The column dimension of F */
14   PetscInt    matops;  /* Matrix format. 0 for stencil, 1 for random */
15   PetscInt    iter;    /* Number of iterations for ADMM */
16   PetscReal   hStart;  /* Starting point for Taylor test */
17   PetscReal   hFactor; /* Taylor test step factor */
18   PetscReal   hMin;    /* Taylor test end goal */
19   PetscReal   alpha;   /* regularization constant applied to || x ||_p */
20   PetscReal   eps;     /* small constant for approximating gradient of || x ||_1 */
21   PetscReal   mu;      /* the augmented Lagrangian term in ADMM */
22   PetscReal   abstol;
23   PetscReal   reltol;
24   Mat         F;                     /* matrix in least squares component $(1/2) * || F x - d ||_2^2$ */
25   Mat         W;                     /* Workspace matrix. ATA */
26   Mat         Hm;                    /* Hessian Misfit*/
27   Mat         Hr;                    /* Hessian Reg*/
28   Vec         d;                     /* RHS in least squares component $(1/2) * || F x - d ||_2^2$ */
29   Vec         workLeft[NWORKLEFT];   /* Workspace for temporary vec */
30   Vec         workRight[NWORKRIGHT]; /* Workspace for temporary vec */
31   NormType    p;
32   PetscRandom rctx;
33   PetscBool   taylor;   /* Flag to determine whether to run Taylor test or not */
34   PetscBool   use_admm; /* Flag to determine whether to run Taylor test or not */
35 } *UserCtx;
36 
37 static PetscErrorCode CreateRHS(UserCtx ctx)
38 {
39   PetscFunctionBegin;
40   /* build the rhs d in ctx */
41   PetscCall(VecCreate(PETSC_COMM_WORLD, &(ctx->d)));
42   PetscCall(VecSetSizes(ctx->d, PETSC_DECIDE, ctx->m));
43   PetscCall(VecSetFromOptions(ctx->d));
44   PetscCall(VecSetRandom(ctx->d, ctx->rctx));
45   PetscFunctionReturn(PETSC_SUCCESS);
46 }
47 
48 static PetscErrorCode CreateMatrix(UserCtx ctx)
49 {
50   PetscInt      Istart, Iend, i, j, Ii, gridN, I_n, I_s, I_e, I_w;
51   PetscLogStage stage;
52 
53   PetscFunctionBegin;
54   /* build the matrix F in ctx */
55   PetscCall(MatCreate(PETSC_COMM_WORLD, &(ctx->F)));
56   PetscCall(MatSetSizes(ctx->F, PETSC_DECIDE, PETSC_DECIDE, ctx->m, ctx->n));
57   PetscCall(MatSetType(ctx->F, MATAIJ));                          /* TODO: Decide specific SetType other than dummy*/
58   PetscCall(MatMPIAIJSetPreallocation(ctx->F, 5, NULL, 5, NULL)); /*TODO: some number other than 5?*/
59   PetscCall(MatSeqAIJSetPreallocation(ctx->F, 5, NULL));
60   PetscCall(MatSetUp(ctx->F));
61   PetscCall(MatGetOwnershipRange(ctx->F, &Istart, &Iend));
62   PetscCall(PetscLogStageRegister("Assembly", &stage));
63   PetscCall(PetscLogStagePush(stage));
64 
65   /* Set matrix elements in  2-D five point stencil format. */
66   if (!(ctx->matops)) {
67     PetscCheck(ctx->m == ctx->n, PETSC_COMM_WORLD, PETSC_ERR_ARG_SIZ, "Stencil matrix must be square");
68     gridN = (PetscInt)PetscSqrtReal((PetscReal)ctx->m);
69     PetscCheck(gridN * gridN == ctx->m, PETSC_COMM_WORLD, PETSC_ERR_ARG_SIZ, "Number of rows must be square");
70     for (Ii = Istart; Ii < Iend; Ii++) {
71       i   = Ii / gridN;
72       j   = Ii % gridN;
73       I_n = i * gridN + j + 1;
74       if (j + 1 >= gridN) I_n = -1;
75       I_s = i * gridN + j - 1;
76       if (j - 1 < 0) I_s = -1;
77       I_e = (i + 1) * gridN + j;
78       if (i + 1 >= gridN) I_e = -1;
79       I_w = (i - 1) * gridN + j;
80       if (i - 1 < 0) I_w = -1;
81       PetscCall(MatSetValue(ctx->F, Ii, Ii, 4., INSERT_VALUES));
82       PetscCall(MatSetValue(ctx->F, Ii, I_n, -1., INSERT_VALUES));
83       PetscCall(MatSetValue(ctx->F, Ii, I_s, -1., INSERT_VALUES));
84       PetscCall(MatSetValue(ctx->F, Ii, I_e, -1., INSERT_VALUES));
85       PetscCall(MatSetValue(ctx->F, Ii, I_w, -1., INSERT_VALUES));
86     }
87   } else PetscCall(MatSetRandom(ctx->F, ctx->rctx));
88   PetscCall(MatAssemblyBegin(ctx->F, MAT_FINAL_ASSEMBLY));
89   PetscCall(MatAssemblyEnd(ctx->F, MAT_FINAL_ASSEMBLY));
90   PetscCall(PetscLogStagePop());
91   /* Stencil matrix is symmetric. Setting symmetric flag for ICC/Cholesky preconditioner */
92   if (!(ctx->matops)) PetscCall(MatSetOption(ctx->F, MAT_SYMMETRIC, PETSC_TRUE));
93   PetscCall(MatTransposeMatMult(ctx->F, ctx->F, MAT_INITIAL_MATRIX, PETSC_DEFAULT, &(ctx->W)));
94   /* Setup Hessian Workspace in same shape as W */
95   PetscCall(MatDuplicate(ctx->W, MAT_DO_NOT_COPY_VALUES, &(ctx->Hm)));
96   PetscCall(MatDuplicate(ctx->W, MAT_DO_NOT_COPY_VALUES, &(ctx->Hr)));
97   PetscFunctionReturn(PETSC_SUCCESS);
98 }
99 
100 static PetscErrorCode SetupWorkspace(UserCtx ctx)
101 {
102   PetscInt i;
103 
104   PetscFunctionBegin;
105   PetscCall(MatCreateVecs(ctx->F, &ctx->workLeft[0], &ctx->workRight[0]));
106   for (i = 1; i < NWORKLEFT; i++) PetscCall(VecDuplicate(ctx->workLeft[0], &(ctx->workLeft[i])));
107   for (i = 1; i < NWORKRIGHT; i++) PetscCall(VecDuplicate(ctx->workRight[0], &(ctx->workRight[i])));
108   PetscFunctionReturn(PETSC_SUCCESS);
109 }
110 
111 static PetscErrorCode ConfigureContext(UserCtx ctx)
112 {
113   PetscFunctionBegin;
114   ctx->m        = 16;
115   ctx->n        = 16;
116   ctx->eps      = 1.e-3;
117   ctx->abstol   = 1.e-4;
118   ctx->reltol   = 1.e-2;
119   ctx->hStart   = 1.;
120   ctx->hMin     = 1.e-3;
121   ctx->hFactor  = 0.5;
122   ctx->alpha    = 1.;
123   ctx->mu       = 1.0;
124   ctx->matops   = 0;
125   ctx->iter     = 10;
126   ctx->p        = NORM_2;
127   ctx->taylor   = PETSC_TRUE;
128   ctx->use_admm = PETSC_FALSE;
129   PetscOptionsBegin(PETSC_COMM_WORLD, NULL, "Configure separable objection example", "ex4.c");
130   PetscCall(PetscOptionsInt("-m", "The row dimension of matrix F", "ex4.c", ctx->m, &(ctx->m), NULL));
131   PetscCall(PetscOptionsInt("-n", "The column dimension of matrix F", "ex4.c", ctx->n, &(ctx->n), NULL));
132   PetscCall(PetscOptionsInt("-matrix_format", "Decide format of F matrix. 0 for stencil, 1 for random", "ex4.c", ctx->matops, &(ctx->matops), NULL));
133   PetscCall(PetscOptionsInt("-iter", "Iteration number ADMM", "ex4.c", ctx->iter, &(ctx->iter), NULL));
134   PetscCall(PetscOptionsReal("-alpha", "The regularization multiplier. 1 default", "ex4.c", ctx->alpha, &(ctx->alpha), NULL));
135   PetscCall(PetscOptionsReal("-epsilon", "The small constant added to |x_i| in the denominator to approximate the gradient of ||x||_1", "ex4.c", ctx->eps, &(ctx->eps), NULL));
136   PetscCall(PetscOptionsReal("-mu", "The augmented lagrangian multiplier in ADMM", "ex4.c", ctx->mu, &(ctx->mu), NULL));
137   PetscCall(PetscOptionsReal("-hStart", "Taylor test starting point. 1 default.", "ex4.c", ctx->hStart, &(ctx->hStart), NULL));
138   PetscCall(PetscOptionsReal("-hFactor", "Taylor test multiplier factor. 0.5 default", "ex4.c", ctx->hFactor, &(ctx->hFactor), NULL));
139   PetscCall(PetscOptionsReal("-hMin", "Taylor test ending condition. 1.e-3 default", "ex4.c", ctx->hMin, &(ctx->hMin), NULL));
140   PetscCall(PetscOptionsReal("-abstol", "Absolute stopping criterion for ADMM", "ex4.c", ctx->abstol, &(ctx->abstol), NULL));
141   PetscCall(PetscOptionsReal("-reltol", "Relative stopping criterion for ADMM", "ex4.c", ctx->reltol, &(ctx->reltol), NULL));
142   PetscCall(PetscOptionsBool("-taylor", "Flag for Taylor test. Default is true.", "ex4.c", ctx->taylor, &(ctx->taylor), NULL));
143   PetscCall(PetscOptionsBool("-use_admm", "Use the ADMM solver in this example.", "ex4.c", ctx->use_admm, &(ctx->use_admm), NULL));
144   PetscCall(PetscOptionsEnum("-p", "Norm type.", "ex4.c", NormTypes, (PetscEnum)ctx->p, (PetscEnum *)&(ctx->p), NULL));
145   PetscOptionsEnd();
146   /* Creating random ctx */
147   PetscCall(PetscRandomCreate(PETSC_COMM_WORLD, &(ctx->rctx)));
148   PetscCall(PetscRandomSetFromOptions(ctx->rctx));
149   PetscCall(CreateMatrix(ctx));
150   PetscCall(CreateRHS(ctx));
151   PetscCall(SetupWorkspace(ctx));
152   PetscFunctionReturn(PETSC_SUCCESS);
153 }
154 
155 static PetscErrorCode DestroyContext(UserCtx *ctx)
156 {
157   PetscInt i;
158 
159   PetscFunctionBegin;
160   PetscCall(MatDestroy(&((*ctx)->F)));
161   PetscCall(MatDestroy(&((*ctx)->W)));
162   PetscCall(MatDestroy(&((*ctx)->Hm)));
163   PetscCall(MatDestroy(&((*ctx)->Hr)));
164   PetscCall(VecDestroy(&((*ctx)->d)));
165   for (i = 0; i < NWORKLEFT; i++) PetscCall(VecDestroy(&((*ctx)->workLeft[i])));
166   for (i = 0; i < NWORKRIGHT; i++) PetscCall(VecDestroy(&((*ctx)->workRight[i])));
167   PetscCall(PetscRandomDestroy(&((*ctx)->rctx)));
168   PetscCall(PetscFree(*ctx));
169   PetscFunctionReturn(PETSC_SUCCESS);
170 }
171 
172 /* compute (1/2) * ||F x - d||^2 */
173 static PetscErrorCode ObjectiveMisfit(Tao tao, Vec x, PetscReal *J, void *_ctx)
174 {
175   UserCtx ctx = (UserCtx)_ctx;
176   Vec     y;
177 
178   PetscFunctionBegin;
179   y = ctx->workLeft[0];
180   PetscCall(MatMult(ctx->F, x, y));
181   PetscCall(VecAXPY(y, -1., ctx->d));
182   PetscCall(VecDot(y, y, J));
183   *J *= 0.5;
184   PetscFunctionReturn(PETSC_SUCCESS);
185 }
186 
187 /* compute V = FTFx - FTd */
188 static PetscErrorCode GradientMisfit(Tao tao, Vec x, Vec V, void *_ctx)
189 {
190   UserCtx ctx = (UserCtx)_ctx;
191   Vec     FTFx, FTd;
192 
193   PetscFunctionBegin;
194   /* work1 is A^T Ax, work2 is Ab, W is A^T A*/
195   FTFx = ctx->workRight[0];
196   FTd  = ctx->workRight[1];
197   PetscCall(MatMult(ctx->W, x, FTFx));
198   PetscCall(MatMultTranspose(ctx->F, ctx->d, FTd));
199   PetscCall(VecWAXPY(V, -1., FTd, FTFx));
200   PetscFunctionReturn(PETSC_SUCCESS);
201 }
202 
203 /* returns FTF */
204 static PetscErrorCode HessianMisfit(Tao tao, Vec x, Mat H, Mat Hpre, void *_ctx)
205 {
206   UserCtx ctx = (UserCtx)_ctx;
207 
208   PetscFunctionBegin;
209   if (H != ctx->W) PetscCall(MatCopy(ctx->W, H, DIFFERENT_NONZERO_PATTERN));
210   if (Hpre != ctx->W) PetscCall(MatCopy(ctx->W, Hpre, DIFFERENT_NONZERO_PATTERN));
211   PetscFunctionReturn(PETSC_SUCCESS);
212 }
213 
214 /* computes augment Lagrangian objective (with scaled dual):
215  * 0.5 * ||F x - d||^2  + 0.5 * mu ||x - z + u||^2 */
216 static PetscErrorCode ObjectiveMisfitADMM(Tao tao, Vec x, PetscReal *J, void *_ctx)
217 {
218   UserCtx   ctx = (UserCtx)_ctx;
219   PetscReal mu, workNorm, misfit;
220   Vec       z, u, temp;
221 
222   PetscFunctionBegin;
223   mu   = ctx->mu;
224   z    = ctx->workRight[5];
225   u    = ctx->workRight[6];
226   temp = ctx->workRight[10];
227   /* misfit = f(x) */
228   PetscCall(ObjectiveMisfit(tao, x, &misfit, _ctx));
229   PetscCall(VecCopy(x, temp));
230   /* temp = x - z + u */
231   PetscCall(VecAXPBYPCZ(temp, -1., 1., 1., z, u));
232   /* workNorm = ||x - z + u||^2 */
233   PetscCall(VecDot(temp, temp, &workNorm));
234   /* augment Lagrangian objective (with scaled dual): f(x) + 0.5 * mu ||x -z + u||^2 */
235   *J = misfit + 0.5 * mu * workNorm;
236   PetscFunctionReturn(PETSC_SUCCESS);
237 }
238 
239 /* computes FTFx - FTd  mu*(x - z + u) */
240 static PetscErrorCode GradientMisfitADMM(Tao tao, Vec x, Vec V, void *_ctx)
241 {
242   UserCtx   ctx = (UserCtx)_ctx;
243   PetscReal mu;
244   Vec       z, u, temp;
245 
246   PetscFunctionBegin;
247   mu   = ctx->mu;
248   z    = ctx->workRight[5];
249   u    = ctx->workRight[6];
250   temp = ctx->workRight[10];
251   PetscCall(GradientMisfit(tao, x, V, _ctx));
252   PetscCall(VecCopy(x, temp));
253   /* temp = x - z + u */
254   PetscCall(VecAXPBYPCZ(temp, -1., 1., 1., z, u));
255   /* V =  FTFx - FTd  mu*(x - z + u) */
256   PetscCall(VecAXPY(V, mu, temp));
257   PetscFunctionReturn(PETSC_SUCCESS);
258 }
259 
260 /* returns FTF + diag(mu) */
261 static PetscErrorCode HessianMisfitADMM(Tao tao, Vec x, Mat H, Mat Hpre, void *_ctx)
262 {
263   UserCtx ctx = (UserCtx)_ctx;
264 
265   PetscFunctionBegin;
266   PetscCall(MatCopy(ctx->W, H, DIFFERENT_NONZERO_PATTERN));
267   PetscCall(MatShift(H, ctx->mu));
268   if (Hpre != H) PetscCall(MatCopy(H, Hpre, DIFFERENT_NONZERO_PATTERN));
269   PetscFunctionReturn(PETSC_SUCCESS);
270 }
271 
272 /* computes || x ||_p (mult by 0.5 in case of NORM_2) */
273 static PetscErrorCode ObjectiveRegularization(Tao tao, Vec x, PetscReal *J, void *_ctx)
274 {
275   UserCtx   ctx = (UserCtx)_ctx;
276   PetscReal norm;
277 
278   PetscFunctionBegin;
279   *J = 0;
280   PetscCall(VecNorm(x, ctx->p, &norm));
281   if (ctx->p == NORM_2) norm = 0.5 * norm * norm;
282   *J = ctx->alpha * norm;
283   PetscFunctionReturn(PETSC_SUCCESS);
284 }
285 
286 /* NORM_2 Case: return x
287  * NORM_1 Case: x/(|x| + eps)
288  * Else: TODO */
289 static PetscErrorCode GradientRegularization(Tao tao, Vec x, Vec V, void *_ctx)
290 {
291   UserCtx   ctx = (UserCtx)_ctx;
292   PetscReal eps = ctx->eps;
293 
294   PetscFunctionBegin;
295   if (ctx->p == NORM_2) {
296     PetscCall(VecCopy(x, V));
297   } else if (ctx->p == NORM_1) {
298     PetscCall(VecCopy(x, ctx->workRight[1]));
299     PetscCall(VecAbs(ctx->workRight[1]));
300     PetscCall(VecShift(ctx->workRight[1], eps));
301     PetscCall(VecPointwiseDivide(V, x, ctx->workRight[1]));
302   } else SETERRQ(PetscObjectComm((PetscObject)tao), PETSC_ERR_ARG_OUTOFRANGE, "Example only works for NORM_1 and NORM_2");
303   PetscFunctionReturn(PETSC_SUCCESS);
304 }
305 
306 /* NORM_2 Case: returns diag(mu)
307  * NORM_1 Case: diag(mu* 1/sqrt(x_i^2 + eps) * (1 - x_i^2/ABS(x_i^2+eps)))  */
308 static PetscErrorCode HessianRegularization(Tao tao, Vec x, Mat H, Mat Hpre, void *_ctx)
309 {
310   UserCtx   ctx = (UserCtx)_ctx;
311   PetscReal eps = ctx->eps;
312   Vec       copy1, copy2, copy3;
313 
314   PetscFunctionBegin;
315   if (ctx->p == NORM_2) {
316     /* Identity matrix scaled by mu */
317     PetscCall(MatZeroEntries(H));
318     PetscCall(MatShift(H, ctx->mu));
319     if (Hpre != H) {
320       PetscCall(MatZeroEntries(Hpre));
321       PetscCall(MatShift(Hpre, ctx->mu));
322     }
323   } else if (ctx->p == NORM_1) {
324     /* 1/sqrt(x_i^2 + eps) * (1 - x_i^2/ABS(x_i^2+eps)) */
325     copy1 = ctx->workRight[1];
326     copy2 = ctx->workRight[2];
327     copy3 = ctx->workRight[3];
328     /* copy1 : 1/sqrt(x_i^2 + eps) */
329     PetscCall(VecCopy(x, copy1));
330     PetscCall(VecPow(copy1, 2));
331     PetscCall(VecShift(copy1, eps));
332     PetscCall(VecSqrtAbs(copy1));
333     PetscCall(VecReciprocal(copy1));
334     /* copy2:  x_i^2.*/
335     PetscCall(VecCopy(x, copy2));
336     PetscCall(VecPow(copy2, 2));
337     /* copy3: abs(x_i^2 + eps) */
338     PetscCall(VecCopy(x, copy3));
339     PetscCall(VecPow(copy3, 2));
340     PetscCall(VecShift(copy3, eps));
341     PetscCall(VecAbs(copy3));
342     /* copy2: 1 - x_i^2/abs(x_i^2 + eps) */
343     PetscCall(VecPointwiseDivide(copy2, copy2, copy3));
344     PetscCall(VecScale(copy2, -1.));
345     PetscCall(VecShift(copy2, 1.));
346     PetscCall(VecAXPY(copy1, 1., copy2));
347     PetscCall(VecScale(copy1, ctx->mu));
348     PetscCall(MatZeroEntries(H));
349     PetscCall(MatDiagonalSet(H, copy1, INSERT_VALUES));
350     if (Hpre != H) {
351       PetscCall(MatZeroEntries(Hpre));
352       PetscCall(MatDiagonalSet(Hpre, copy1, INSERT_VALUES));
353     }
354   } else SETERRQ(PetscObjectComm((PetscObject)tao), PETSC_ERR_ARG_OUTOFRANGE, "Example only works for NORM_1 and NORM_2");
355   PetscFunctionReturn(PETSC_SUCCESS);
356 }
357 
358 /* NORM_2 Case: 0.5 || x ||_2 + 0.5 * mu * ||x + u - z||^2
359  * Else : || x ||_2 + 0.5 * mu * ||x + u - z||^2 */
360 static PetscErrorCode ObjectiveRegularizationADMM(Tao tao, Vec z, PetscReal *J, void *_ctx)
361 {
362   UserCtx   ctx = (UserCtx)_ctx;
363   PetscReal mu, workNorm, reg;
364   Vec       x, u, temp;
365 
366   PetscFunctionBegin;
367   mu   = ctx->mu;
368   x    = ctx->workRight[4];
369   u    = ctx->workRight[6];
370   temp = ctx->workRight[10];
371   PetscCall(ObjectiveRegularization(tao, z, &reg, _ctx));
372   PetscCall(VecCopy(z, temp));
373   /* temp = x + u -z */
374   PetscCall(VecAXPBYPCZ(temp, 1., 1., -1., x, u));
375   /* workNorm = ||x + u - z ||^2 */
376   PetscCall(VecDot(temp, temp, &workNorm));
377   *J = reg + 0.5 * mu * workNorm;
378   PetscFunctionReturn(PETSC_SUCCESS);
379 }
380 
381 /* NORM_2 Case: x - mu*(x + u - z)
382  * NORM_1 Case: x/(|x| + eps) - mu*(x + u - z)
383  * Else: TODO */
384 static PetscErrorCode GradientRegularizationADMM(Tao tao, Vec z, Vec V, void *_ctx)
385 {
386   UserCtx   ctx = (UserCtx)_ctx;
387   PetscReal mu;
388   Vec       x, u, temp;
389 
390   PetscFunctionBegin;
391   mu   = ctx->mu;
392   x    = ctx->workRight[4];
393   u    = ctx->workRight[6];
394   temp = ctx->workRight[10];
395   PetscCall(GradientRegularization(tao, z, V, _ctx));
396   PetscCall(VecCopy(z, temp));
397   /* temp = x + u -z */
398   PetscCall(VecAXPBYPCZ(temp, 1., 1., -1., x, u));
399   PetscCall(VecAXPY(V, -mu, temp));
400   PetscFunctionReturn(PETSC_SUCCESS);
401 }
402 
403 /* NORM_2 Case: returns diag(mu)
404  * NORM_1 Case: FTF + diag(mu) */
405 static PetscErrorCode HessianRegularizationADMM(Tao tao, Vec x, Mat H, Mat Hpre, void *_ctx)
406 {
407   UserCtx ctx = (UserCtx)_ctx;
408 
409   PetscFunctionBegin;
410   if (ctx->p == NORM_2) {
411     /* Identity matrix scaled by mu */
412     PetscCall(MatZeroEntries(H));
413     PetscCall(MatShift(H, ctx->mu));
414     if (Hpre != H) {
415       PetscCall(MatZeroEntries(Hpre));
416       PetscCall(MatShift(Hpre, ctx->mu));
417     }
418   } else if (ctx->p == NORM_1) {
419     PetscCall(HessianMisfit(tao, x, H, Hpre, (void *)ctx));
420     PetscCall(MatShift(H, ctx->mu));
421     if (Hpre != H) PetscCall(MatShift(Hpre, ctx->mu));
422   } else SETERRQ(PetscObjectComm((PetscObject)tao), PETSC_ERR_ARG_OUTOFRANGE, "Example only works for NORM_1 and NORM_2");
423   PetscFunctionReturn(PETSC_SUCCESS);
424 }
425 
426 /* NORM_2 Case : (1/2) * ||F x - d||^2 + 0.5 * || x ||_p
427 *  NORM_1 Case : (1/2) * ||F x - d||^2 + || x ||_p */
428 static PetscErrorCode ObjectiveComplete(Tao tao, Vec x, PetscReal *J, void *ctx)
429 {
430   PetscReal Jm, Jr;
431 
432   PetscFunctionBegin;
433   PetscCall(ObjectiveMisfit(tao, x, &Jm, ctx));
434   PetscCall(ObjectiveRegularization(tao, x, &Jr, ctx));
435   *J = Jm + Jr;
436   PetscFunctionReturn(PETSC_SUCCESS);
437 }
438 
439 /* NORM_2 Case: FTFx - FTd + x
440  * NORM_1 Case: FTFx - FTd + x/(|x| + eps) */
441 static PetscErrorCode GradientComplete(Tao tao, Vec x, Vec V, void *ctx)
442 {
443   UserCtx cntx = (UserCtx)ctx;
444 
445   PetscFunctionBegin;
446   PetscCall(GradientMisfit(tao, x, cntx->workRight[2], ctx));
447   PetscCall(GradientRegularization(tao, x, cntx->workRight[3], ctx));
448   PetscCall(VecWAXPY(V, 1, cntx->workRight[2], cntx->workRight[3]));
449   PetscFunctionReturn(PETSC_SUCCESS);
450 }
451 
452 /* NORM_2 Case: diag(mu) + FTF
453  * NORM_1 Case: diag(mu* 1/sqrt(x_i^2 + eps) * (1 - x_i^2/ABS(x_i^2+eps))) + FTF  */
454 static PetscErrorCode HessianComplete(Tao tao, Vec x, Mat H, Mat Hpre, void *ctx)
455 {
456   Mat tempH;
457 
458   PetscFunctionBegin;
459   PetscCall(MatDuplicate(H, MAT_SHARE_NONZERO_PATTERN, &tempH));
460   PetscCall(HessianMisfit(tao, x, H, H, ctx));
461   PetscCall(HessianRegularization(tao, x, tempH, tempH, ctx));
462   PetscCall(MatAXPY(H, 1., tempH, DIFFERENT_NONZERO_PATTERN));
463   if (Hpre != H) PetscCall(MatCopy(H, Hpre, DIFFERENT_NONZERO_PATTERN));
464   PetscCall(MatDestroy(&tempH));
465   PetscFunctionReturn(PETSC_SUCCESS);
466 }
467 
468 static PetscErrorCode TaoSolveADMM(UserCtx ctx, Vec x)
469 {
470   PetscInt  i;
471   PetscReal u_norm, r_norm, s_norm, primal, dual, x_norm, z_norm;
472   Tao       tao1, tao2;
473   Vec       xk, z, u, diff, zold, zdiff, temp;
474   PetscReal mu;
475 
476   PetscFunctionBegin;
477   xk    = ctx->workRight[4];
478   z     = ctx->workRight[5];
479   u     = ctx->workRight[6];
480   diff  = ctx->workRight[7];
481   zold  = ctx->workRight[8];
482   zdiff = ctx->workRight[9];
483   temp  = ctx->workRight[11];
484   mu    = ctx->mu;
485   PetscCall(VecSet(u, 0.));
486   PetscCall(TaoCreate(PETSC_COMM_WORLD, &tao1));
487   PetscCall(TaoSetType(tao1, TAONLS));
488   PetscCall(TaoSetObjective(tao1, ObjectiveMisfitADMM, (void *)ctx));
489   PetscCall(TaoSetGradient(tao1, NULL, GradientMisfitADMM, (void *)ctx));
490   PetscCall(TaoSetHessian(tao1, ctx->Hm, ctx->Hm, HessianMisfitADMM, (void *)ctx));
491   PetscCall(VecSet(xk, 0.));
492   PetscCall(TaoSetSolution(tao1, xk));
493   PetscCall(TaoSetOptionsPrefix(tao1, "misfit_"));
494   PetscCall(TaoSetFromOptions(tao1));
495   PetscCall(TaoCreate(PETSC_COMM_WORLD, &tao2));
496   if (ctx->p == NORM_2) {
497     PetscCall(TaoSetType(tao2, TAONLS));
498     PetscCall(TaoSetObjective(tao2, ObjectiveRegularizationADMM, (void *)ctx));
499     PetscCall(TaoSetGradient(tao2, NULL, GradientRegularizationADMM, (void *)ctx));
500     PetscCall(TaoSetHessian(tao2, ctx->Hr, ctx->Hr, HessianRegularizationADMM, (void *)ctx));
501   }
502   PetscCall(VecSet(z, 0.));
503   PetscCall(TaoSetSolution(tao2, z));
504   PetscCall(TaoSetOptionsPrefix(tao2, "reg_"));
505   PetscCall(TaoSetFromOptions(tao2));
506 
507   for (i = 0; i < ctx->iter; i++) {
508     PetscCall(VecCopy(z, zold));
509     PetscCall(TaoSolve(tao1)); /* Updates xk */
510     if (ctx->p == NORM_1) {
511       PetscCall(VecWAXPY(temp, 1., xk, u));
512       PetscCall(TaoSoftThreshold(temp, -ctx->alpha / mu, ctx->alpha / mu, z));
513     } else {
514       PetscCall(TaoSolve(tao2)); /* Update zk */
515     }
516     /* u = u + xk -z */
517     PetscCall(VecAXPBYPCZ(u, 1., -1., 1., xk, z));
518     /* r_norm : norm(x-z) */
519     PetscCall(VecWAXPY(diff, -1., z, xk));
520     PetscCall(VecNorm(diff, NORM_2, &r_norm));
521     /* s_norm : norm(-mu(z-zold)) */
522     PetscCall(VecWAXPY(zdiff, -1., zold, z));
523     PetscCall(VecNorm(zdiff, NORM_2, &s_norm));
524     s_norm = s_norm * mu;
525     /* primal : sqrt(n)*ABSTOL + RELTOL*max(norm(x), norm(-z))*/
526     PetscCall(VecNorm(xk, NORM_2, &x_norm));
527     PetscCall(VecNorm(z, NORM_2, &z_norm));
528     primal = PetscSqrtReal(ctx->n) * ctx->abstol + ctx->reltol * PetscMax(x_norm, z_norm);
529     /* Duality : sqrt(n)*ABSTOL + RELTOL*norm(mu*u)*/
530     PetscCall(VecNorm(u, NORM_2, &u_norm));
531     dual = PetscSqrtReal(ctx->n) * ctx->abstol + ctx->reltol * u_norm * mu;
532     PetscCall(PetscPrintf(PetscObjectComm((PetscObject)tao1), "Iter %" PetscInt_FMT " : ||x-z||: %g, mu*||z-zold||: %g\n", i, (double)r_norm, (double)s_norm));
533     if (r_norm < primal && s_norm < dual) break;
534   }
535   PetscCall(VecCopy(xk, x));
536   PetscCall(TaoDestroy(&tao1));
537   PetscCall(TaoDestroy(&tao2));
538   PetscFunctionReturn(PETSC_SUCCESS);
539 }
540 
541 /* Second order Taylor remainder convergence test */
542 static PetscErrorCode TaylorTest(UserCtx ctx, Tao tao, Vec x, PetscReal *C)
543 {
544   PetscReal  h, J, temp;
545   PetscInt   i, j;
546   PetscInt   numValues;
547   PetscReal  Jx, Jxhat_comp, Jxhat_pred;
548   PetscReal *Js, *hs;
549   PetscReal  gdotdx;
550   PetscReal  minrate = PETSC_MAX_REAL;
551   MPI_Comm   comm    = PetscObjectComm((PetscObject)x);
552   Vec        g, dx, xhat;
553 
554   PetscFunctionBegin;
555   PetscCall(VecDuplicate(x, &g));
556   PetscCall(VecDuplicate(x, &xhat));
557   /* choose a perturbation direction */
558   PetscCall(VecDuplicate(x, &dx));
559   PetscCall(VecSetRandom(dx, ctx->rctx));
560   /* evaluate objective at x: J(x) */
561   PetscCall(TaoComputeObjective(tao, x, &Jx));
562   /* evaluate gradient at x, save in vector g */
563   PetscCall(TaoComputeGradient(tao, x, g));
564   PetscCall(VecDot(g, dx, &gdotdx));
565 
566   for (numValues = 0, h = ctx->hStart; h >= ctx->hMin; h *= ctx->hFactor) numValues++;
567   PetscCall(PetscCalloc2(numValues, &Js, numValues, &hs));
568   for (i = 0, h = ctx->hStart; h >= ctx->hMin; h *= ctx->hFactor, i++) {
569     PetscCall(VecWAXPY(xhat, h, dx, x));
570     PetscCall(TaoComputeObjective(tao, xhat, &Jxhat_comp));
571     /* J(\hat(x)) \approx J(x) + g^T (xhat - x) = J(x) + h * g^T dx */
572     Jxhat_pred = Jx + h * gdotdx;
573     /* Vector to dJdm scalar? Dot?*/
574     J = PetscAbsReal(Jxhat_comp - Jxhat_pred);
575     PetscCall(PetscPrintf(comm, "J(xhat): %g, predicted: %g, diff %g\n", (double)Jxhat_comp, (double)Jxhat_pred, (double)J));
576     Js[i] = J;
577     hs[i] = h;
578   }
579   for (j = 1; j < numValues; j++) {
580     temp = PetscLogReal(Js[j] / Js[j - 1]) / PetscLogReal(hs[j] / hs[j - 1]);
581     PetscCall(PetscPrintf(comm, "Convergence rate step %" PetscInt_FMT ": %g\n", j - 1, (double)temp));
582     minrate = PetscMin(minrate, temp);
583   }
584   /* If O is not ~2, then the test is wrong */
585   PetscCall(PetscFree2(Js, hs));
586   *C = minrate;
587   PetscCall(VecDestroy(&dx));
588   PetscCall(VecDestroy(&xhat));
589   PetscCall(VecDestroy(&g));
590   PetscFunctionReturn(PETSC_SUCCESS);
591 }
592 
593 int main(int argc, char **argv)
594 {
595   UserCtx ctx;
596   Tao     tao;
597   Vec     x;
598   Mat     H;
599 
600   PetscFunctionBeginUser;
601   PetscCall(PetscInitialize(&argc, &argv, NULL, help));
602   PetscCall(PetscNew(&ctx));
603   PetscCall(ConfigureContext(ctx));
604   /* Define two functions that could pass as objectives to TaoSetObjective(): one
605    * for the misfit component, and one for the regularization component */
606   /* ObjectiveMisfit() and ObjectiveRegularization() */
607 
608   /* Define a single function that calls both components adds them together: the complete objective,
609    * in the absence of a Tao implementation that handles separability */
610   /* ObjectiveComplete() */
611   PetscCall(TaoCreate(PETSC_COMM_WORLD, &tao));
612   PetscCall(TaoSetType(tao, TAONM));
613   PetscCall(TaoSetObjective(tao, ObjectiveComplete, (void *)ctx));
614   PetscCall(TaoSetGradient(tao, NULL, GradientComplete, (void *)ctx));
615   PetscCall(MatDuplicate(ctx->W, MAT_SHARE_NONZERO_PATTERN, &H));
616   PetscCall(TaoSetHessian(tao, H, H, HessianComplete, (void *)ctx));
617   PetscCall(MatCreateVecs(ctx->F, NULL, &x));
618   PetscCall(VecSet(x, 0.));
619   PetscCall(TaoSetSolution(tao, x));
620   PetscCall(TaoSetFromOptions(tao));
621   if (ctx->use_admm) PetscCall(TaoSolveADMM(ctx, x));
622   else PetscCall(TaoSolve(tao));
623   /* examine solution */
624   PetscCall(VecViewFromOptions(x, NULL, "-view_sol"));
625   if (ctx->taylor) {
626     PetscReal rate;
627     PetscCall(TaylorTest(ctx, tao, x, &rate));
628   }
629   PetscCall(MatDestroy(&H));
630   PetscCall(TaoDestroy(&tao));
631   PetscCall(VecDestroy(&x));
632   PetscCall(DestroyContext(&ctx));
633   PetscCall(PetscFinalize());
634   return 0;
635 }
636 
637 /*TEST
638 
639   build:
640     requires: !complex
641 
642   test:
643     suffix: 0
644     args:
645 
646   test:
647     suffix: l1_1
648     args: -p 1 -tao_type lmvm -alpha 1. -epsilon 1.e-7 -m 64 -n 64 -view_sol -matrix_format 1
649 
650   test:
651     suffix: hessian_1
652     args: -matrix_format 1 -m 100 -n 100 -tao_monitor -p 1 -tao_type nls
653 
654   test:
655     suffix: hessian_2
656     args: -matrix_format 1 -m 100 -n 100 -tao_monitor -p 2 -tao_type nls
657 
658   test:
659     suffix: nm_1
660     args: -matrix_format 1 -m 100 -n 100 -tao_monitor -p 1 -tao_type nm -tao_max_it 50
661 
662   test:
663     suffix: nm_2
664     args: -matrix_format 1 -m 100 -n 100 -tao_monitor -p 2 -tao_type nm -tao_max_it 50
665 
666   test:
667     suffix: lmvm_1
668     args: -matrix_format 1 -m 100 -n 100 -tao_monitor -p 1 -tao_type lmvm -tao_max_it 40
669 
670   test:
671     suffix: lmvm_2
672     args: -matrix_format 1 -m 100 -n 100 -tao_monitor -p 2 -tao_type lmvm -tao_max_it 15
673 
674   test:
675     suffix: soft_threshold_admm_1
676     args: -matrix_format 1 -m 100 -n 100 -tao_monitor -p 1 -use_admm
677 
678   test:
679     suffix: hessian_admm_1
680     args: -matrix_format 1 -m 100 -n 100 -tao_monitor -p 1 -use_admm -reg_tao_type nls -misfit_tao_type nls
681 
682   test:
683     suffix: hessian_admm_2
684     args: -matrix_format 1 -m 100 -n 100 -tao_monitor -p 2 -use_admm -reg_tao_type nls -misfit_tao_type nls
685 
686   test:
687     suffix: nm_admm_1
688     args: -matrix_format 1 -m 100 -n 100 -tao_monitor -p 1 -use_admm -reg_tao_type nm -misfit_tao_type nm
689 
690   test:
691     suffix: nm_admm_2
692     args: -matrix_format 1 -m 100 -n 100 -tao_monitor -p 2 -use_admm -reg_tao_type nm -misfit_tao_type nm -iter 7
693 
694   test:
695     suffix: lmvm_admm_1
696     args: -matrix_format 1 -m 100 -n 100 -tao_monitor -p 1 -use_admm -reg_tao_type lmvm -misfit_tao_type lmvm
697 
698   test:
699     suffix: lmvm_admm_2
700     args: -matrix_format 1 -m 100 -n 100 -tao_monitor -p 2 -use_admm -reg_tao_type lmvm -misfit_tao_type lmvm
701 
702 TEST*/
703