1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081 | #include "AverageAngleCShellOptimization.hh"
template<template<typename> class Object>
AverageAngleCShellOptimization<Object>::AverageAngleCShellOptimization(Object<Real> &flat, Object<Real> &deployed, const NewtonOptimizerOptions &eopts, std::unique_ptr<LOMinAngleConstraint<Object>> &&minAngleConstraint, int pinJoint,
bool allowFlatActuation, bool optimizeTargetAngle, bool fixDeployedVars, const std::vector<size_t> &additionalFixedFlatVars, const std::vector<size_t> &additionalFixedDeployedVars)
: LinkageOptimization<Object>(flat, eopts, deployed.energy(), BBox<Point3D>(deployed.deformedPoints()).dimensions().norm(), deployed.totalRestLength(), deployed.averageAbsRestKappaVars() < 1.0e-10 ? 1.0 : deployed.averageAbsRestKappaVars()),
m_deployed(deployed), m_linesearch_deployed(deployed),
m_minAngleConstraint(std::move(minAngleConstraint)),
m_fixDeployedVars(fixDeployedVars),
m_allowFlatActuation(allowFlatActuation)
{
std::runtime_error mismatch("Linkage mismatch");
if (m_numParams != deployed.numDesignParams()) throw mismatch;
if ((deployed.getDesignParameters() - flat.getDesignParameters()).norm() > 1.0e-16) throw mismatch;
m_alpha_tgt = deployed.getAverageActuatedJointsAngle();
m_Linesearch_alpha_tgt = deployed.getAverageActuatedJointsAngle();
m_numFullParams = flat.numDesignParams() + int(m_optimizeTargetAngle);
// Create the objective terms
using OET = OptEnergyType;
using EEO = ElasticEnergyObjective<Object>;
using TSF = TargetFittingDOOT<Object>;
using RLM = RegularizationTermDOOWrapper<Object, RestLengthMinimization>;
using RCS = RegularizationTermDOOWrapper<Object, RestCurvatureSmoothing>;
auto &tsf = target_surface_fitter;
objective.add("ElasticEnergyFlat", OET::ElasticBase, std::make_shared<EEO>(m_linesearch_base), gamma / m_E0);
objective.add("ElasticEnergyDeployed", OET::ElasticDeployed, std::make_shared<EEO>(m_linesearch_deployed), (1.0 - gamma) / m_E0);
objective.add("TargetFitting", OET::Target, std::make_shared<TSF>(m_linesearch_deployed, tsf), beta / (m_l0 * m_l0));
objective.add("RestLengthMinimization", OET::Regularization, std::make_shared<RLM>(m_linesearch_deployed), 1.0 / m_rl0);
objective.add("RestCurvatureSmoothing", OET::Smoothing, std::make_shared<RCS>(m_linesearch_deployed), 1.0 / (m_rk0 * m_rk0));
// Cannot use enveloppe theorem if we choose to optimize alpha
setOptimizeTargetAngle(optimizeTargetAngle);
// Unless the user specifies otherwise, use the current deployed linkage joint positions as the target
target_surface_fitter.joint_pos_tgt = deployed.jointPositions();
constructTargetSurface(2, 1);
// Set to true for testing the derivatives
target_surface_fitter.holdClosestPointsFixed = false;
// Trade off between fitting to the individual joint targets and the target surface.
target_surface_fitter.setTargetJointPosVsTargetSurfaceTradeoff(deployed, 0.1);
// Constrain the position and orientation of the centermost joint to prevent global rigid motion.
if (pinJoint != -1) {
m_rm_constrained_joint = pinJoint;
if (m_rm_constrained_joint >= flat.numJoints()) throw std::runtime_error("Manually specified pinJoint is out of bounds");
}
else {
m_rm_constrained_joint = flat.centralJoint();
}
const size_t jdo = flat.dofOffsetForJoint(m_rm_constrained_joint);
for (size_t i = 0; i < 6; ++i) m_rigidMotionFixedVars.push_back(jdo + i);
const size_t idxAlphaBarDeployed = m_deployed.getAverageAngleIndex();
std::vector<size_t> fixedFlatVars = m_rigidMotionFixedVars;
fixedFlatVars.insert(fixedFlatVars.end(), additionalFixedFlatVars.begin(), additionalFixedFlatVars.end());
m_flat_optimizer = get_equilibrium_optimizer(m_linesearch_base, TARGET_ANGLE_NONE, fixedFlatVars);
std::vector<size_t> fixedDeployedVars;
if (m_fixDeployedVars) { fixedDeployedVars = m_rigidMotionFixedVars; }
fixedDeployedVars.push_back(idxAlphaBarDeployed);
fixedDeployedVars.insert(fixedDeployedVars.end(), additionalFixedDeployedVars.begin(), additionalFixedDeployedVars.end());
// We use TARGET_ANGLE_NONE since this constraint is taken care of in fixedDeployedVars
m_deployed_optimizer = get_equilibrium_optimizer(m_linesearch_deployed, TARGET_ANGLE_NONE, fixedDeployedVars);
m_flat_optimizer ->options = m_equilibrium_options;
m_deployed_optimizer->options = m_equilibrium_options;
// Ensure we start at an equilibrium (using the passed equilibrium solver options)
m_forceEquilibriumUpdate();
m_updateMinAngleConstraintActuation();
commitLinesearchLinkage();
}
template<template<typename> class Object>
void AverageAngleCShellOptimization<Object>::m_forceEquilibriumUpdate() {
// Update the flat/deployed equilibria
m_equilibriumSolveSuccessful = true;
try {
if (m_equilibrium_options.verbose)
std::cout << "Flat equilibrium solve" << std::endl;
auto cr_flat = getFlatOptimizer().optimize();
// A backtracking failure will happen if the gradient tolerance is set too low
// and generally does not indicate a complete failure/bad estimate of the equilibrium.
// We therefore accept such equilibria with a warning.
// (We would prefer to reject saddle points, but the surface-attracted structures
// appear to // sometimes have backtracking failures in saddle points
// close to reasonably stable equilibria...)
bool acceptable_failed_flat_equilibrium = cr_flat.backtracking_failure;
if (!cr_flat.success && !acceptable_failed_flat_equilibrium) throw std::runtime_error("Flat equilibrium solve did not converge");
if (acceptable_failed_flat_equilibrium) { std::cout << "WARNING: Flat equillibrium solve backtracking failure." << std::endl; }
if (m_equilibrium_options.verbose)
std::cout << "Deployed equilibrium solve" << std::endl;
std::cout << "Initial energy (deployed) " << getDeployedOptimizer().get_problem().energy() << std::endl;
std::cout << "Has constraints: " << getDeployedOptimizer().get_problem().hasLEQConstraint() << std::endl;
// std::cout << "HDeployed optimizer: " << getDeployedOptimizer() << std::endl;
auto cr_deploy = getDeployedOptimizer().optimize();
std::cout << "Final energy (deployed) " << getDeployedOptimizer().get_problem().energy() << std::endl;
bool acceptable_failed_deployed_equilibrium = cr_deploy.backtracking_failure;
if (!cr_deploy.success && !acceptable_failed_deployed_equilibrium) throw std::runtime_error("Deployed equilibrium solve did not converge");
if (acceptable_failed_deployed_equilibrium) { std::cout << "WARNING: Deployed equillibrium solve backtracking failure." << std::endl; }
}
catch (const std::runtime_error &e) {
std::cout << "Equilibrium solve failed: " << e.what() << std::endl;
m_equilibriumSolveSuccessful = false;
return; // subsequent update_factorizations will fail if we caught a Tau runaway...
}
// We will be evaluating the Hessian/using the simplified gradient expressions:
m_linesearch_base .updateSourceFrame();
m_linesearch_base .updateRotationParametrizations();
m_linesearch_deployed.updateSourceFrame();
m_linesearch_deployed.updateRotationParametrizations();
// Use the final equilibria's Hessians for sensitivity analysis, not the second-to-last iterates'
try {
getFlatOptimizer() .update_factorizations();
getDeployedOptimizer().update_factorizations();
}
catch (const std::runtime_error &e) {
std::cout << "Equilibrium solve failed: " << e.what() << std::endl;
m_equilibriumSolveSuccessful = false;
return;
}
// The cached adjoint state is invalidated whenever the equilibrium is updated...
m_adjointStateIsCurrent = false;
m_autodiffLinkagesAreCurrent = false;
objective.update();
}
template<template<typename> class Object>
bool AverageAngleCShellOptimization<Object>::m_updateEquilibria(const Eigen::Ref<const Eigen::VectorXd> &newParams) {
const size_t np = numParams(), nfp = numFullParams();
const size_t idxAlphaBarDeployed = m_linesearch_deployed.getAverageAngleIndex();
std::runtime_error mismatch("Dimension mismatch for the design parameters in update equilibria");
if (newParams.size() != int(nfp)) throw mismatch;
// The linesearch linkage is already up to date
if ((getLinesearchDesignParameters() - newParams).norm() < 1.0e-14) { return false; }
m_linesearch_deployed.set(m_deployed);
m_linesearch_base. set(m_base);
const Eigen::VectorXd currParams = getFullDesignParameters();
Eigen::VectorXd delta_p = newParams - currParams;
if (delta_p.norm() == 0) { // returning to linesearch start; no prediction/Hessian factorization necessary
std::cout << "Nothing has changed compared to last commit" << std::endl;
m_Linesearch_alpha_tgt = m_alpha_tgt;
Eigen::VectorXd curr_x3d = m_deployed.getDoFs();
curr_x3d[idxAlphaBarDeployed] = m_alpha_tgt;
m_linesearch_deployed.setDoFs(curr_x3d);
m_forceEquilibriumUpdate();
return true;
}
// Apply the new design parameters and measure the energy with the 0^th order prediction
// (i.e. at the current equilibrium).
// We will only replace this equilibrium if the higher order predictions achieve a lower energy.
m_linesearch_deployed.setDesignParameters(newParams.head(np));
m_linesearch_base .setDesignParameters(newParams.head(np));
Real bestEnergy3d = m_linesearch_deployed.energy(),
bestEnergy2d = m_linesearch_base .energy();
Eigen::VectorXd curr_x3d = m_deployed.getDoFs(),
curr_x2d = m_base .getDoFs();
Eigen::VectorXd best_x3d = curr_x3d,
best_x2d = curr_x2d;
if (m_optimizeTargetAngle) {
m_Linesearch_alpha_tgt = newParams[np];
}
// In case we decide not to use first order prediction,
// we should still update the average opening angle
if (prediction_order == PredictionOrder::Zero) {
curr_x3d[idxAlphaBarDeployed] = m_Linesearch_alpha_tgt;
m_linesearch_deployed.setDoFs(curr_x3d);
} else {
BENCHMARK_SCOPED_TIMER_SECTION timer("Predict equilibrium");
// Return to using the Hessian for the last committed linkage
// (i.e. for the equilibrium stored in m_flat and m_deployed).
auto &opt_2D = getFlatOptimizer();
auto &opt_3D = getDeployedOptimizer();
if (!(opt_2D.solver.hasStashedFactorization() && opt_3D.solver.hasStashedFactorization()))
throw std::runtime_error("Factorization was not stashed... was commitLinesearchLinkage() called?");
opt_2D.solver.swapStashedFactorization();
opt_3D.solver.swapStashedFactorization();
{
// Former:
// Solve for equilibrium perturbation corresponding to delta_p:
// [H_3D a][delta x ] = [-d2E/dxdp delta_p]
// [a^T 0][delta lambda] [ 0 ]
// \_________________/
// b
// In case alpha_t is also optimized, this transforms into
// [H_3D a][delta x ] = [-d2E/dxdp delta_p]
// [a^T 0][delta lambda] [ delta_alpha_t ]
// \_________________/
// b
//
// Now:
// Solve for equilibrium perturbation corresponding to [delta_p delta_alpha_p]:
// H_3D delta x = - d2E/dxdp delta p - d2E/dxdalpha_x delta alpha_p
// and delta alpha_x = delta alpha_p (if alpha bar is also optimized)
const size_t np = numParams(), nd = m_base.numDoF();
VecX_T<Real> neg_deltap_padded(nd + np);
neg_deltap_padded.setZero();
neg_deltap_padded.tail(np) = -delta_p.head(np);
if (m_optimizeTargetAngle){ neg_deltap_padded[idxAlphaBarDeployed] = -delta_p[np]; }
// Computing -d2E/dxdp delta_p can skip the *-x and designParameter-* blocks
HessianComputationMask mask_dxdp;
mask_dxdp.dof_in = m_optimizeTargetAngle;
mask_dxdp.designParameter_out = false;
Eigen::VectorXd b_reduced_3D = opt_3D.removeFixedEntries(m_deployed.applyHessianPerSegmentRestlen(neg_deltap_padded, mask_dxdp).head(nd));
m_delta_x3d = opt_3D.extractFullSolution(opt_3D.solver.solve(b_reduced_3D));
if (m_optimizeTargetAngle){
// Should we instead do the following?
// m_delta_x3d[idxAlphaBarDeployed] = newParams[np] - curr_x3d[idxAlphaBarDeployed];
m_delta_x3d[idxAlphaBarDeployed] = delta_p[np];
neg_deltap_padded[idxAlphaBarDeployed] = 0.; // Reset to 0 for the 2D prediction
mask_dxdp.dof_in = false; // No need to compute the input DoF block for 2D
}
Eigen::VectorXd b_reduced_2D = opt_2D.removeFixedEntries(m_base.applyHessianPerSegmentRestlen(neg_deltap_padded, mask_dxdp).head(nd));
m_delta_x2d = opt_2D.extractFullSolution(opt_2D.solver.solve(b_reduced_2D));
// Evaluate the energy at the 1st order-predicted equilibrium
{
auto first_order_x3d = (curr_x3d + m_delta_x3d).eval(),
first_order_x2d = (curr_x2d + m_delta_x2d).eval();
m_linesearch_deployed.setDoFs(first_order_x3d);
m_linesearch_base .setDoFs(first_order_x2d);
Real energy1stOrder3d = m_linesearch_deployed.energy(),
energy1stOrder2d = m_linesearch_base .energy();
if (energy1stOrder3d < bestEnergy3d) { std::cout << " used first order prediction, energy reduction " << bestEnergy3d - energy1stOrder3d << std::endl; bestEnergy3d = energy1stOrder3d; best_x3d = first_order_x3d; } else { m_linesearch_deployed.setDoFs(best_x3d); }
if (energy1stOrder2d < bestEnergy2d) { std::cout << " used first order prediction, energy reduction " << bestEnergy2d - energy1stOrder2d << std::endl; bestEnergy2d = energy1stOrder2d; best_x2d = first_order_x2d; } else { m_linesearch_base .setDoFs(best_x2d); }
}
if (prediction_order > PredictionOrder::One) {
// TODO: also stash autodiff linkages for committed linkages?
// Solve for perturbation of equilibrium perturbation corresponding to delta_p:
// H_3D delta_p^T d2x/dp^2 delta_p = - (d3E/dx3 delta_x + d3E/dx2dp delta_p) delta_x - (d3E/dxdpdx delta_x + d3E/dxdpdp delta_p) delta_p
// = - [d3E/dx3 delta_x + d3E/dx2dp delta_p d3E/dxdpdx delta_x + d3E/dxdpdp delta_p][delta_x]
// [delta_p]
// and delta_p^T d2x_alpha/dp^2 delta_p = 0 (if alpha bar is also optimized)
m_diff_linkage_deployed.set(m_deployed);
m_diff_linkage_flat .set(m_base);
Eigen::VectorXd neg_d3E_delta_x3d, neg_d3E_delta_x2d;
{
// inject design parameter perturbation.
VecX_T<ADReal> ad_p = currParams;
for (size_t i = 0; i < np; ++i) ad_p[i].derivatives()[0] = delta_p[i];
m_diff_linkage_deployed.setDesignParameters(ad_p);
m_diff_linkage_flat .setDesignParameters(ad_p);
// inject equilibrium perturbation
VecX_T<ADReal> ad_x_3d = curr_x3d;
VecX_T<ADReal> ad_x_2d = curr_x2d;
for (int i = 0; i < ad_x_3d.size(); ++i) ad_x_3d[i].derivatives()[0] = m_delta_x3d[i];
for (int i = 0; i < ad_x_2d.size(); ++i) ad_x_2d[i].derivatives()[0] = m_delta_x2d[i];
m_diff_linkage_deployed.setDoFs(ad_x_3d);
m_diff_linkage_flat .setDoFs(ad_x_2d);
VecX_T<Real> delta_edof_3d(nd + np);
VecX_T<Real> delta_edof_2d(nd + np);
delta_edof_3d.head(nd) = m_delta_x3d;
delta_edof_2d.head(nd) = m_delta_x2d;
delta_edof_3d.tail(np) = delta_p.head(np);
delta_edof_2d.tail(np) = delta_p.head(np);
neg_d3E_delta_x3d = -extractDirectionalDerivative(m_diff_linkage_deployed.applyHessianPerSegmentRestlen(delta_edof_3d)).head(nd);
neg_d3E_delta_x2d = -extractDirectionalDerivative(m_diff_linkage_flat .applyHessianPerSegmentRestlen(delta_edof_2d)).head(nd);
}
// The fixed entries are automatically zeroed out by extractFullSolution.
b_reduced_3D = opt_3D.removeFixedEntries(neg_d3E_delta_x3d);
m_delta_delta_x3d = opt_3D.extractFullSolution(opt_3D.solver.solve(b_reduced_3D));
b_reduced_2D = opt_2D.removeFixedEntries(neg_d3E_delta_x2d);
m_delta_delta_x2d = opt_2D.extractFullSolution(opt_2D.solver.solve(b_reduced_2D));
// Evaluate the energy at the 2nd order-predicted equilibrium, roll back to previous best if energy is higher.
{
m_second_order_x3d = (curr_x3d + m_delta_x3d + 0.5 * m_delta_delta_x3d).eval(),
m_second_order_x2d = (curr_x2d + m_delta_x2d + 0.5 * m_delta_delta_x2d).eval();
m_linesearch_deployed.setDoFs(m_second_order_x3d);
m_linesearch_base .setDoFs(m_second_order_x2d);
Real energy2ndOrder3d = m_linesearch_deployed.energy(),
energy2ndOrder2d = m_linesearch_base .energy();
if (energy2ndOrder3d < bestEnergy3d) { std::cout << " used second order prediction, energy reduction " << bestEnergy3d - energy2ndOrder3d << std::endl; bestEnergy3d = energy2ndOrder3d; best_x3d = m_second_order_x3d;} else { m_linesearch_deployed.setDoFs(best_x3d); }
if (energy2ndOrder2d < bestEnergy2d) { std::cout << " used second order prediction, energy reduction " << bestEnergy2d - energy2ndOrder2d << std::endl; bestEnergy2d = energy2ndOrder2d; best_x2d = m_second_order_x2d;} else { m_linesearch_base .setDoFs(best_x2d); }
}
}
}
// Return to using the primary factorization, storing the committed
// linkages' factorizations back in the stash for later use.
opt_2D.solver.swapStashedFactorization();
opt_3D.solver.swapStashedFactorization();
}
std::cout << "Will run equilibrium from equilibria predictions" << std::endl;
// Make sure the target angle is correct
if (m_optimizeTargetAngle) {
curr_x3d = m_linesearch_deployed.getDoFs();
if (std::abs(curr_x3d[idxAlphaBarDeployed] - m_Linesearch_alpha_tgt) > 1e-16) {
curr_x3d[idxAlphaBarDeployed] = m_Linesearch_alpha_tgt;
m_linesearch_deployed.setDoFs(curr_x3d);
}
}
m_forceEquilibriumUpdate();
return true;
}
// TODO: should this be updated?
template<template<typename> class Object>
void AverageAngleCShellOptimization<Object>::m_updateMinAngleConstraintActuation() {
if (!m_minAngleConstraint || !m_allowFlatActuation) return;
getFlatOptimizer().optimize(); // We need to update the flat equilibrium to determine if the minimum angle constraint is in the working set
// Add/remove the minimum angle constraint to the working set.
if (m_minAngleConstraint->shouldRelease(m_linesearch_base, getFlatOptimizer())) {
m_minAngleConstraint->inWorkingSet = false;
}
else if (m_minAngleConstraint->violated(m_linesearch_base)) {
m_minAngleConstraint->inWorkingSet = true;
Real alpha_bar_0 = m_linesearch_base.getAverageActuatedJointsAngle();
m_minAngleConstraint->actuationAngle = alpha_bar_0;
// Construct the actuated flat equilibrium solver if it hasn't been.
if (!m_flat_optimizer_actuated) {
m_flat_optimizer_actuated = get_equilibrium_optimizer(m_linesearch_base, TARGET_ANGLE_NONE, m_rigidMotionFixedVars);
// m_flat_optimizer_actuated = get_equilibrium_optimizer(m_linesearch_base, alpha_bar_0, m_rigidMotionFixedVars);
m_flat_optimizer_actuated->options = m_equilibrium_options;
m_flat_optimizer_actuated->optimize();
m_linesearch_base.updateSourceFrame();
m_linesearch_base.updateRotationParametrizations();
m_flat_optimizer_actuated->update_factorizations();
}
}
// If the minimum angle is in the working set, solve for the actuation angle such
// that the bound is satisifed.
m_minAngleConstraint->enforce(m_linesearch_base, getFlatOptimizer());
}
// Update the adjoint state vectors "w", "y", and "s"
template<template<typename> class Object>
bool AverageAngleCShellOptimization<Object>::m_updateAdjointState(const Eigen::Ref<const Eigen::VectorXd> ¶ms, const OptEnergyType opt_eType) {
const size_t nd = m_linesearch_deployed.numDoF(), nfp = numFullParams();
std::runtime_error mismatch("Dimension mismatch for the design parameters in update adjoint state");
if (params.size() != int(nfp)) throw mismatch;
m_updateEquilibria(params);
if (m_adjointStateIsCurrent) return false;
std::cout << "Updating adjoint state" << std::endl;
// Solve the adjoint problems needed to efficiently evaluate the gradient.
// Note: if the Hessian modification failed (tau runaway), the adjoint state
// solves will fail. To keep the solver from giving up entirely, we simply
// set the adjoint state to 0 in these cases. Presumably this only happens
// at bad iterates that will be discarded anyway.
try {
// Adjoint solve for the flatness constraint on the closed linkage:
// H_2D y = 2 S_z^T S_z x_2D or [H_2D a][y_x ] = [2 S_z^T S_z x_2D]
// [a^T 0][y_lambda] [ 0 ]
// Depending on whether the closed linkage is actuated.
//
// Now:
// H_2D y = 2 S_z^T S_z x_2D
{
auto &opt_2D = getFlatOptimizer();
Eigen::VectorXd b_reduced = opt_2D.removeFixedEntries(m_apply_S_z_transpose(2 * m_apply_S_z(m_linesearch_base.getDoFs())));
m_y = opt_2D.extractFullSolution(opt_2D.solver.solve(b_reduced));
}
// Adjoint solve for the target fitting objective on the deployed linkage
// H_3D w = ∂J/∂x
// Note that these gradients should be expressed using the change of coordinates
{
// if (!prob3D.hasLEQConstraint()) throw std::runtime_error("The deployed linkage must have a linear equality constraint applied!");
auto &opt_3D = getDeployedOptimizer();
Eigen::VectorXd grad_x;
grad_x.setZero(nd);
for (const auto &t : objective.terms) {
if (t.term->getWeight() == 0.0) continue;
if ((opt_eType == OptEnergyType::Full) || (opt_eType == t.type)) {
if (t.type == OptEnergyType::ElasticBase){
continue;
} else if (t.type != OptEnergyType::ElasticDeployed) {
// We should change variables before inverting the system
grad_x += m_linesearch_deployed.applyTransformationTransposeDoFSize(t.term->grad_x());
// grad_x += m_linesearch_deployed.applyTransformationDoFSize(t.term->grad_x());
} else {
// OptEnergyType::ElasticDeployed already expresses the gradient using the change of variables
grad_x += t.term->grad_x();
}
}
}
Eigen::VectorXd b_reduced_3D = opt_3D.removeFixedEntries(grad_x);
m_w_x = opt_3D.extractFullSolution(opt_3D.solver.solve(b_reduced_3D));
// m_w_lambda = w_lambda;
}
// Adjoint solve for the minimum opening angle constraint
// [H_2D a][s_x ] = [d alpha_min / d_x]
// [a^T 0][s_lambda] [ 0 ]
//
// Now:
// H_2D s_x = ∂ alpha_min / ∂x
if (m_minAngleConstraint) {
auto &opt_2D = getFlatOptimizer();
Eigen::VectorXd Hinv_b_reduced;
Eigen::VectorXd grad_x = m_linesearch_base.applyTransformationTransposeDoFSize(m_minAngleConstraint->grad(m_linesearch_base));
opt_2D.solver.solve(opt_2D.removeFixedEntries(grad_x), Hinv_b_reduced);
m_s_x = opt_2D.extractFullSolution(Hinv_b_reduced);
}
}
catch (...) {
std::cout << "WARNING: Adjoint state solve failed" << std::endl;
m_y.setZero();
m_w_x.setZero();
m_s_x.setZero();
m_w_lambda = 0.;
}
m_adjointStateIsCurrent = true;
return true;
}
template<template<typename> class Object>
Eigen::VectorXd AverageAngleCShellOptimization<Object>::gradp_J(const Eigen::Ref<const Eigen::VectorXd> ¶ms, OptEnergyType opt_eType) {
const size_t np = numParams(), nd = m_linesearch_deployed.numDoF(), nfp = numFullParams();
const size_t idxAlphaBarDeployed = m_linesearch_deployed.getAverageAngleIndex();
std::runtime_error mismatch("Dimension mismatch for the design parameters in gradp_J");
if (params.size() != int(nfp)) throw mismatch;
// Then force equilibria and update adjoint state
m_updateAdjointState(params, opt_eType);
Eigen::VectorXd gradp_tot;
gradp_tot.setZero(nfp);
// Compute terms for the adjoint
HessianComputationMask mask;
mask.dof_out = m_optimizeTargetAngle; // Actually we need ∂^2E/∂α_x∂x
mask.designParameter_in = false;
Eigen::VectorXd w_padded(nd + np);
w_padded.head(nd) = m_w_x;
w_padded.tail(np).setZero();
// gradxp_adj = - [∂^2E/∂x∂x] w_x
// [∂^2E/∂p∂x]
Eigen::VectorXd gradxp_adj = - m_linesearch_deployed.applyHessianPerSegmentRestlen(w_padded, mask);
// gradp_obj = ∂J/∂p
Eigen::VectorXd gradp_obj = objective.grad_p(opt_eType);
// Collect all the gradients
gradp_tot.head(np) = gradp_obj + gradxp_adj.tail(np);
// Add gradient with respect to the target angle
// This should be - ∂^2E/∂α_x∂x w_x + ∂J/∂α_x + ∂J/∂α_p
// The last term is assumed to be equal to zero.
if (m_optimizeTargetAngle) {
Eigen::VectorXd gradx_obj;
gradx_obj.setZero(nd);
for (const auto &t : objective.terms) {
if (t.term->getWeight() == 0.0) continue;
if ((opt_eType == OptEnergyType::Full) || (opt_eType == t.type)) {
if (t.type == OptEnergyType::ElasticBase){
continue;
} else if (t.type != OptEnergyType::ElasticDeployed) {
// We should change variables before inverting the system
gradx_obj += m_linesearch_deployed.applyTransformationTransposeDoFSize(t.term->grad_x());
} else {
// OptEnergyType::ElasticDeployed already expresses the gradient using the change of variables
gradx_obj += t.term->grad_x();
}
}
}
gradp_tot[np] = gradxp_adj[idxAlphaBarDeployed] + gradx_obj[idxAlphaBarDeployed];
}
return gradp_tot;
}
template<template<typename> class Object>
Eigen::VectorXd AverageAngleCShellOptimization<Object>::gradp_J_target(const Eigen::Ref<const Eigen::VectorXd> ¶ms) {
const size_t np = numParams(), nd = m_linesearch_deployed.numDoF(), nfp = numFullParams();
std::runtime_error mismatch("Dimension mismatch for the design parameters in gradp_J_target");
if (params.size() != int(nfp)) throw mismatch;
m_updateAdjointState(params, OptEnergyType::Target);
Eigen::VectorXd gradp_tot;
gradp_tot.setZero(nfp);
HessianComputationMask mask;
mask.dof_out = false;
mask.designParameter_in = false;
Eigen::VectorXd w_padded(nd + np);
w_padded.head(nd) = m_w_x;
w_padded.tail(np).setZero();
return -m_linesearch_deployed.applyHessianPerSegmentRestlen(w_padded, mask).tail(np);
}
template<template<typename> class Object>
Eigen::VectorXd AverageAngleCShellOptimization<Object>::gradp_c(const Eigen::Ref<const Eigen::VectorXd> ¶ms) {
const size_t np = numParams(), nd = m_linesearch_base.numDoF(), nfp = numFullParams();
std::runtime_error mismatch("Dimension mismatch for the design parameters in gradp_c");
if (params.size() != int(nfp)) throw mismatch;
m_updateAdjointState(params);
Eigen::VectorXd gradp_tot;
gradp_tot.setZero(nfp);
HessianComputationMask mask;
mask.dof_out = false;
mask.designParameter_in = false;
Eigen::VectorXd y_padded(nd + np);
y_padded.head(m_y.size()) = m_y;
y_padded.tail(numParams()).setZero();
gradp_tot.head(np) = - m_linesearch_base.applyHessianPerSegmentRestlen(y_padded, mask).tail(np);
return gradp_tot;
}
template<template<typename> class Object>
Eigen::VectorXd AverageAngleCShellOptimization<Object>::gradp_angle_constraint(const Eigen::Ref<const Eigen::VectorXd> ¶ms) {
if (!m_minAngleConstraint) throw std::runtime_error("No minimum angle constraint is applied.");
const size_t np = numParams(), nd = m_linesearch_base.numDoF(), nfp = numFullParams();
const size_t idxAlphaBarBase = m_linesearch_base.getAverageAngleIndex();
std::runtime_error mismatch("Dimension mismatch for the design parameters in gradp_c");
if (params.size() != int(nfp)) throw mismatch;
m_updateAdjointState(params);
Eigen::VectorXd gradp_tot;
gradp_tot.setZero(nfp);
HessianComputationMask mask;
mask.dof_out = m_optimizeTargetAngle;
mask.designParameter_in = false;
Eigen::VectorXd s_padded(nd + np);
s_padded.head(m_s_x.size()) = m_s_x;
s_padded.tail(numParams()).setZero();
// gradxp_adj = - [∂^2E/∂x∂x] s_x
// [∂^2E/∂p∂x]
Eigen::VectorXd gradxp_adj = - m_linesearch_base.applyHessianPerSegmentRestlen(s_padded, mask);
// gradp_adj = - ∂^2E/∂p∂x s_x
gradp_tot.head(np) = gradxp_adj.tail(np);
// Add gradient with respect to the target angle
// This should be - ∂^2E/∂α_x∂x s_x + ∂ alpha_min/∂α_x
if (m_optimizeTargetAngle) {
Eigen::VectorXd gradx_obj = m_linesearch_base.applyTransformationTransposeDoFSize(m_minAngleConstraint->grad(m_linesearch_base));
gradp_tot[np] = gradxp_adj[idxAlphaBarBase] + gradx_obj[idxAlphaBarBase];
}
return gradp_tot;
}
template<template<typename> class Object>
Eigen::VectorXd AverageAngleCShellOptimization<Object>::pushforward(const Eigen::Ref<const Eigen::VectorXd> ¶ms, const Eigen::Ref<const Eigen::VectorXd> &delta_p) {
Eigen::VectorXd delta_x2d,
delta_x3d;
const size_t np = numParams(), nd = m_linesearch_base.numDoF();
const size_t idxAlphaBarDeployed = m_linesearch_deployed.getAverageAngleIndex();
m_updateAdjointState(params);
std::cout << "Done updating adjoint in pushforward" << std::endl;
// Use current factorization
auto &opt_2D = getFlatOptimizer();
auto &opt_3D = getDeployedOptimizer();
{
// Formerly:
// Solve for equilibrium perturbation corresponding to delta_p:
// [H_3D a][delta x ] = [-d2E/dxdp delta_p]
// [a^T 0][delta lambda] [ 0 ]
// \_________________/
// b
// In case alpha_t is also optimized, this transforms into
// [H_3D a][delta x ] = [-d2E/dxdp delta_p]
// [a^T 0][delta lambda] [ delta_alpha_t ]
// \_________________/
// b
//
// Now:
// Solve for equilibrium perturbation corresponding to [delta_p delta_alpha_p]:
// H_3D delta x = - d2E/dxdp delta p - d2E/dxdalpha_x delta alpha_p
// and delta alpha_x = delta alpha_p (if alpha bar is also optimized)
VecX_T<Real> neg_deltap_padded(nd + np);
neg_deltap_padded.setZero();
neg_deltap_padded.tail(np) = -delta_p.head(np);
if (m_optimizeTargetAngle){ neg_deltap_padded[idxAlphaBarDeployed] = -delta_p[np]; }
// Computing -d2E/dxdp delta_p can skip the *-x and designParameter-* blocks
HessianComputationMask mask_dxdp;
mask_dxdp.dof_in = m_optimizeTargetAngle;
mask_dxdp.designParameter_out = false;
Eigen::VectorXd b_reduced_3D = opt_3D.removeFixedEntries(m_linesearch_deployed.applyHessianPerSegmentRestlen(neg_deltap_padded, mask_dxdp).head(nd));
delta_x3d = opt_3D.extractFullSolution(opt_3D.solver.solve(b_reduced_3D));
if (m_optimizeTargetAngle){
delta_x3d[idxAlphaBarDeployed] = delta_p[np];
neg_deltap_padded[idxAlphaBarDeployed] = 0.; // Reset to 0 for the 2D prediction
mask_dxdp.dof_in = false; // No need to compute the input DoF block for 2D
}
Eigen::VectorXd b_reduced_2D = opt_2D.removeFixedEntries(m_linesearch_base.applyHessianPerSegmentRestlen(neg_deltap_padded, mask_dxdp).head(nd));
delta_x2d = opt_2D.extractFullSolution(opt_2D.solver.solve(b_reduced_2D));
}
Eigen::VectorXd delta_x(2 * nd);
delta_x.head(nd) = delta_x2d;
delta_x.tail(nd) = delta_x3d;
return delta_x;
}
template<template<typename> class Object>
Eigen::VectorXd AverageAngleCShellOptimization<Object>::apply_hess(const Eigen::Ref<const Eigen::VectorXd> ¶ms,
const Eigen::Ref<const Eigen::VectorXd> &delta_p,
Real coeff_J, Real coeff_c, Real coeff_angle_constraint, OptEnergyType opt_eType) {
BENCHMARK_SCOPED_TIMER_SECTION timer("apply_hess_J");
BENCHMARK_START_TIMER_SECTION("Preamble");
const size_t np = numParams(), nd = m_linesearch_base.numDoF(), nfp = numFullParams();
const size_t idxAlphaBarDeployed = m_linesearch_deployed.getAverageAngleIndex();
if (params.size() != int(nfp)) throw std::runtime_error("Incorrect parameter vector size");
if (delta_p.size() != int(nfp)) throw std::runtime_error("Incorrect delta parameter vector size");
m_updateAdjointState(params, opt_eType);
if (!m_autodiffLinkagesAreCurrent) {
BENCHMARK_SCOPED_TIMER_SECTION timer2("Update autodiff linkages");
m_diff_linkage_deployed.set(m_linesearch_deployed);
m_diff_linkage_flat .set(m_linesearch_base);
m_autodiffLinkagesAreCurrent = true;
}
auto &opt_3D = getDeployedOptimizer();
BENCHMARK_STOP_TIMER_SECTION("Preamble");
VecX_T<Real> neg_deltap_padded(nd + np);
neg_deltap_padded.head(nd).setZero();
neg_deltap_padded.tail(np) = - delta_p.head(np);
// Computing -d2E/dxdp delta_p can skip the *-x and designParameter-* blocks
HessianComputationMask mask_dxdp, mask_dxpdx;
mask_dxdp.dof_in = false;
mask_dxdp.designParameter_out = false;
mask_dxpdx.designParameter_in = false;
// VecX_T<Real> delta_dJ_dx3dp;
// Note: if the Hessian modification failed (tau runaway), the delta forward/adjoint state
// solves will fail. To keep the solver from giving up entirely, we simply
// set the failed quantities to 0 in these cases. Presumably this only happens
// at bad iterates that will be discarded anyway.
// Solve for closed state perturbation
try {
// H_2D delta_x = [-d2E/dxdp delta_p]
// \_________________/
// b
// Note that delta_x is expressed using the change of angle variables
BENCHMARK_SCOPED_TIMER_SECTION timer2("solve delta x2d");
auto &opt_2D = getFlatOptimizer();
Eigen::VectorXd b_reduced = opt_2D.removeFixedEntries(m_linesearch_base.applyHessianPerSegmentRestlen(neg_deltap_padded, mask_dxdp).head(nd));
m_delta_x2d = opt_2D.extractFullSolution(opt_2D.solver.solve(b_reduced));
}
catch (...) {
m_delta_x2d.setZero();
}
VecX_T<Real> d3E_s, d3E_y;
VecX_T<Real> delta_grad_xp;
delta_grad_xp.setZero(nd + np);
VecX_T<Real> delta_grad_minAngleConstraint_xp;
delta_grad_minAngleConstraint_xp.setZero(nd + np);
try {
// Solve for deployed state perturbation
// [H_3D a][delta x ] = [-d2E/dxdp delta_p]
// [a^T 0][delta lambda] [ 0 ]
// \_________________/
// b
//
// In case alpha_t is also optimized, this transforms into
// [H_3D a][delta x ] = [-d2E/dxdp delta_p]
// [a^T 0][delta lambda] [ delta_alpha_t ]
// \_________________/
// b
//
// Now:
// Solve for equilibrium perturbation corresponding to [delta_p delta_alpha_p]:
// H_3D delta x = - d2E/dxdp delta p - d2E/dxdalpha_x delta alpha_p
// and delta alpha_x = delta alpha_p (if alpha bar is also optimized)
{
BENCHMARK_SCOPED_TIMER_SECTION timer2("solve delta x3d");
// Computing -d2E/dxdp delta_p can skip the *-x and designParameter-* blocks
// The above only holds if we don't optimize the target angle
mask_dxdp.dof_in = m_optimizeTargetAngle;
if (m_optimizeTargetAngle){ neg_deltap_padded[idxAlphaBarDeployed] = -delta_p[np]; }
// Expressed using change of variables
VecX_T<Real> b_reduced_3D = opt_3D.removeFixedEntries(m_linesearch_deployed.applyHessianPerSegmentRestlen(neg_deltap_padded, mask_dxdp).head(nd));
m_delta_x3d = opt_3D.extractFullSolution(opt_3D.solver.solve(b_reduced_3D)); // Expressed using change of variables
if (m_optimizeTargetAngle){
m_delta_x3d[idxAlphaBarDeployed] = delta_p[np];
neg_deltap_padded[idxAlphaBarDeployed] = 0.; // Reset to 0
mask_dxdp.dof_in = false; // No need to compute the input DoF block
}
}
VecX_T<Real> delta_x2dp(nd + np);
delta_x2dp.head(nd) = m_delta_x2d;
delta_x2dp.tail(np) = delta_p.head(np);
VecX_T<Real> delta_x3dp(nd + np);
delta_x3dp.head(nd) = m_delta_x3d;
delta_x3dp.tail(np) = delta_p.head(np);
// Solve for deployed adjoint state perturbation
BENCHMARK_START_TIMER_SECTION("getDoFs and inject state");
bool need_2d_autodiff = (coeff_c != 0.0) || (coeff_angle_constraint != 0.0);
auto ad_xp_3d = m_diff_linkage_deployed.getExtendedDoFsPSRL();
auto ad_xp_2d = m_diff_linkage_flat .getExtendedDoFsPSRL();
auto inject_delta_state_3d = [&](VecX_T<Real> delta) {
for (int i = 0; i < ad_xp_3d.size(); ++i) ad_xp_3d[i].derivatives()[0] = delta[i];
m_diff_linkage_deployed.setExtendedDoFsPSRL(ad_xp_3d);
};
auto inject_delta_state_2d = [&](VecX_T<Real> delta) {
for (int i = 0; i < ad_xp_2d.size(); ++i) ad_xp_2d[i].derivatives()[0] = delta[i];
m_diff_linkage_flat.setExtendedDoFsPSRL(ad_xp_2d);
};
inject_delta_state_3d(delta_x3dp);
if (need_2d_autodiff)
inject_delta_state_2d(delta_x2dp);
BENCHMARK_STOP_TIMER_SECTION("getDoFs and inject state");
// H_3D delta w_x = [∂2J/∂x∂x delta_x + ∂2J/∂x∂p delta_p ] - [∂3E/∂x∂x∂x delta_x + ∂3E/∂x∂x∂p delta_p] w_x
// \____________________________________________________________________________________/
// b
if (coeff_J != 0.0) {
BENCHMARK_SCOPED_TIMER_SECTION timer2("solve delta w x");
BENCHMARK_START_TIMER_SECTION("Hw");
VecX_T<ADReal> w_padded(nd + np);
w_padded.head(nd) = m_w_x;
w_padded.tail(np).setZero();
// Note: we need the "p" rows of d3E_w for evaluating the full Hessian matvec expressions below...
// m_d3E_w = [∂3E/∂x∂x∂x delta_x + ∂3E/∂x∂x∂p delta_p] w_x
// [∂3E/∂p∂x∂x delta_x + ∂3E/∂p∂x∂p delta_p]
// This is all expressed using the change of variables
m_d3E_w = extractDirectionalDerivative(m_diff_linkage_deployed.applyHessianPerSegmentRestlen(w_padded, mask_dxpdx));
BENCHMARK_STOP_TIMER_SECTION("Hw");
BENCHMARK_START_TIMER_SECTION("KKT_solve");
// Sensitivity of adjoint solve for the target fitting objective on the deployed linkage
// if (!prob3D.hasLEQConstraint()) throw std::runtime_error("The deployed linkage must have a linear equality constraint applied!");
auto &opt_3D = getDeployedOptimizer();
// Note that delta_x3dp is already expressed using the change of angle variables
// We simply need to apply the transformation at the output, and transform delta_x3dp whenever
// we look at a term that does not involve any kind of energy.
for (const auto &t : objective.terms) {
if (t.term->getWeight() == 0.0) continue;
if ((opt_eType == OptEnergyType::Full) || (opt_eType == t.type)) {
if (t.type == OptEnergyType::ElasticBase){
continue;
} else if (t.type != OptEnergyType::ElasticDeployed) {
delta_grad_xp += m_linesearch_deployed.applyTransformationTransposeDoFSize(t.term->delta_grad(m_linesearch_deployed.applyTransformationDoFSize(delta_x3dp), m_linesearch_deployed));
} else {
delta_grad_xp += t.term->delta_grad(delta_x3dp, m_linesearch_deployed);
}
}
}
// delta_grad_xp = [∂2J/∂x∂x delta_x + ∂2J/∂x∂p delta_p ]
// [∂2J/∂p∂x delta_x + ∂2J/∂p∂p delta_p ]
//
// Note that extractFullSolution automatically zeroes out the component corresponding to the average opening angle
Eigen::VectorXd b_reduced_3D = opt_3D.removeFixedEntries((delta_grad_xp.head(nd) - m_d3E_w.head(nd)).eval());
m_delta_w_x = opt_3D.extractFullSolution(opt_3D.solver.solve(b_reduced_3D));
BENCHMARK_STOP_TIMER_SECTION("KKT_solve");
}
// H_2D delta s_x = [∂2 alpha_min/∂x2 delta_x] - [∂3E/∂x∂x∂x delta_x + ∂3E/∂x∂x∂p delta_p] s
// \______________________________________________________________________/
// b
if (m_minAngleConstraint && (coeff_angle_constraint != 0.0)) {
auto &opt_2D = getFlatOptimizer();
BENCHMARK_SCOPED_TIMER_SECTION timer2("solve delta s x");
BENCHMARK_START_TIMER_SECTION("Hs");
VecX_T<ADReal> s_padded(nd + np);
s_padded.head(nd) = m_s_x;
s_padded.tail(np).setZero();
// d3E_s = [∂3E/∂x∂x∂x delta_x + ∂3E/∂x∂x∂p delta_p] s_x
// [∂3E/∂p∂x∂x delta_x + ∂3E/∂p∂x∂p delta_p]
// Note: we need the "p" rows of d3E_s for evaluating the full angle constraint Hessian matvec expression below...
d3E_s = extractDirectionalDerivative(m_diff_linkage_flat.applyHessianPerSegmentRestlen(s_padded, mask_dxpdx));
BENCHMARK_STOP_TIMER_SECTION("Hs");
BENCHMARK_START_TIMER_SECTION("KKT_solve");
// Not the same as DOT->delta_grad, only outputs the derivative against the DoFs
// m_minAngleConstraint->delta_grad = ∂2alpha_min/∂x∂x delta_x
// Still need to apply the transformation so that delta_grad_minAngleConstraint_xp uses the change of variables,
// and since delta_grad is only implemented using the initial angle parameterization, we should first transform m_delta_x2d
delta_grad_minAngleConstraint_xp = m_linesearch_base.applyTransformationTransposeDoFSize(m_minAngleConstraint->delta_grad(m_linesearch_base, m_linesearch_base.applyTransformationDoFSize(m_delta_x2d)));
auto b = (delta_grad_minAngleConstraint_xp - d3E_s.head(nd)).eval();
m_delta_s_x = opt_2D.extractFullSolution(opt_2D.solver.solve(opt_2D.removeFixedEntries(b)));
BENCHMARK_STOP_TIMER_SECTION("KKT_solve");
}
// H_2D delta y = 2 S_z^T S_z delta x_2D - delta H_2D y
//
// depending on whether the closed linkage is actuated,
// where delta H_2D = d3E/dx dx dx delta_x + d3E/dx dx dp delta_p.
if (coeff_c != 0.0) {
auto &opt_2D = getFlatOptimizer();
BENCHMARK_SCOPED_TIMER_SECTION timer2("solve delta y");
BENCHMARK_START_TIMER_SECTION("Hy");
VecX_T<ADReal> y_padded(nd + np);
y_padded.head(nd) = m_y;
y_padded.tail(np).setZero();
// Note: we need the "p" rows of d3E_y for evaluating the full Hessian matvec expressions below...
d3E_y = extractDirectionalDerivative(m_diff_linkage_flat.applyHessianPerSegmentRestlen(y_padded, mask_dxpdx));
BENCHMARK_STOP_TIMER_SECTION("Hy");
BENCHMARK_START_TIMER_SECTION("KKT_solve");
// Note that this remains the same under the angle change of variables
auto b = (m_apply_S_z_transpose(2 * m_apply_S_z(m_delta_x2d)) - d3E_y.head(nd)).eval();
m_delta_y = opt_2D.extractFullSolution(opt_2D.solver.solve(opt_2D.removeFixedEntries(b)));
BENCHMARK_STOP_TIMER_SECTION("KKT_solve");
}
}
catch (...) {
m_delta_x3d = VecX_T<Real>::Zero(nd );
m_delta_w_x = VecX_T<Real>::Zero(nd );
m_delta_w_lambda = 0.;
m_delta_s_x = VecX_T<Real>::Zero(nd );
m_delta_y = VecX_T<Real>::Zero(nd );
m_d3E_w = VecX_T<Real>::Zero(nd + np);
d3E_s = VecX_T<Real>::Zero(nd + np);
delta_grad_xp.setZero(nd + np);
}
VecX_T<Real> result;
result.setZero(nfp);
// Accumulate the J hessian matvec
{
BENCHMARK_SCOPED_TIMER_SECTION timer3("evaluate hessian matvec");
if (coeff_J != 0.0) {
if (objective.terms.empty()) throw std::runtime_error("no terms present");
VecX_T<Real> delta_edofs_2d(nd + np);
delta_edofs_2d.head(nd) = m_delta_x2d;
delta_edofs_2d.tail(np) = delta_p.head(np);
VecX_T<Real> delta_edofs_adj3d(nd + np);
delta_edofs_adj3d.head(nd) = m_delta_w_x;
delta_edofs_adj3d.tail(np).setZero();
HessianComputationMask mask;
mask.dof_out = m_optimizeTargetAngle; // Actually we need ∂^2E/∂α_x∂x
mask.designParameter_in = false;
// First accumulate the delta grad we previously computed
// result.head(np) += ∂2J/∂p∂x delta_x + ∂2J/∂p∂p delta_p
result.head(np) += delta_grad_xp.tail(np);
// Then add the remaining terms
// H_delta_w = [∂2E/∂x ∂x] delta_w_x
// [∂2E/∂p ∂x]
Eigen::VectorXd H_delta_w = m_linesearch_deployed.applyHessianPerSegmentRestlen(delta_edofs_adj3d, mask);
// result.head(np) += - ∂2E/∂p ∂x delta_w_x
result.head(np) += - H_delta_w.tail(np);
// result.head(np) += - (∂3E/∂p ∂x ∂x delta_x + ∂3E/∂p ∂x ∂p delta_p) w_x
result.head(np) += - m_d3E_w.tail(np);
for (const auto &t : objective.terms) {
if (t.term->getWeight() == 0.0) continue;
if ((opt_eType == OptEnergyType::Full) || (opt_eType == t.type)) {
if (t.type == OptEnergyType::ElasticBase){
result.head(np) += t.term->delta_grad(delta_edofs_2d, m_linesearch_base).tail(np);
}
}
}
// Handle the last component
// result[np] = - ∂2E/∂α_x∂x delta_w_x - (∂3E/∂α_x∂x∂x delta_x + ∂3E/∂α_x∂x∂p delta_p) w_x + ∂2J/∂α_x∂x delta_x + ∂2J/∂α_x∂p delta_p
if (m_optimizeTargetAngle) { result[np] = - H_delta_w[idxAlphaBarDeployed] - m_d3E_w[idxAlphaBarDeployed] + delta_grad_xp[idxAlphaBarDeployed]; }
result *= coeff_J;
}
if (coeff_c != 0) {
HessianComputationMask mask;
mask.dof_out = false;
VecX_T<Real> delta_edofs(nd + np);
delta_edofs.head(nd) = -m_delta_y;
delta_edofs.tail(np).setZero();
result.head(np) += coeff_c * (m_linesearch_base.applyHessianPerSegmentRestlen(delta_edofs, mask).tail(np) - d3E_y.tail(np));
}
if (coeff_angle_constraint != 0) {
HessianComputationMask mask;
mask.dof_out = m_optimizeTargetAngle;
mask.designParameter_in = false;
VecX_T<Real> delta_edofs(nd + np);
delta_edofs.head(nd) = m_delta_s_x;
delta_edofs.tail(np).setZero();
// Then add the remaining terms
// H_delta_s = [∂2E/∂x ∂x] delta_s_x
// [∂2E/∂p ∂x]
Eigen::VectorXd H_delta_s = m_linesearch_base.applyHessianPerSegmentRestlen(delta_edofs, mask);
// result.head(np) += c_alpha_min * ( - ∂E/∂p∂x delta_s_x - (∂3E/∂p∂x∂x delta_x + ∂3E/∂p∂x∂p delta_p) s_x )
result.head(np) += coeff_angle_constraint * ( - H_delta_s.tail(np) - d3E_s.tail(np));
if (m_optimizeTargetAngle) {
// result[np] = c_alpha_min * ( - ∂2E/∂α_x∂x delta_s_x - (∂3E/∂α_x∂x∂x delta_x + ∂3E/∂α_x∂x∂p delta_p) s_x + ∂2 alpha_min/∂α_x∂x delta_x )
result[np] += coeff_angle_constraint * ( - H_delta_s[idxAlphaBarDeployed] - d3E_s[idxAlphaBarDeployed] + delta_grad_minAngleConstraint_xp[idxAlphaBarDeployed]);
}
}
}
return result;
}
template<template<typename> class Object>
void AverageAngleCShellOptimization<Object>::dumpHessians(const std::string &hess_J_path, const std::string &hess_ac_path, Real fd_eps) {
auto curr_params = getFullDesignParameters();
auto grad_J = gradp_J(curr_params);
size_t nperturbs = 3;
Eigen::VectorXd relerror_fd_diff_grad_p_J(nperturbs),
relerror_delta_Hw(nperturbs),
relerror_delta_w(nperturbs),
relerror_delta_w_rhs(nperturbs),
relerror_delta_x(nperturbs),
relerror_delta_J(nperturbs);
Eigen::VectorXd matvec_relerror_fd_diff_grad_p_J(nperturbs);
Eigen::VectorXd grad_J_relerror(nperturbs);
auto w = m_w_x;
auto H = m_linesearch_deployed.hessian();
auto Hw = H.apply(w);
for (size_t i = 0; i < nperturbs; ++i) {
Eigen::VectorXd perturb = Eigen::VectorXd::Random(m_numParams + int(m_optimizeTargetAngle));
apply_hess_J(curr_params, perturb);
auto delta_w = m_delta_w_x;
auto H_delta_w = m_linesearch_deployed.applyHessian(delta_w);
Real Jplus = J(curr_params + fd_eps * perturb);
auto gradp_J_plus = gradp_J(curr_params + fd_eps * perturb);
auto w_plus = m_w_x;
auto x_plus = m_linesearch_deployed.getDoFs();
auto Hw_plus = m_linesearch_deployed.applyHessian(w);
auto H_plus_w_plus = m_linesearch_deployed.applyHessian(w_plus);
auto w_rhs_plus = m_w_rhs;
auto H_plus = m_linesearch_deployed.hessian();
{
Eigen::VectorXd v = Eigen::VectorXd::Random(m_linesearch_deployed.numDoF());
auto my_H = m_linesearch_deployed.hessianSparsityPattern(false);
m_linesearch_deployed.hessian(my_H);
auto matvec_one = my_H.apply(v);
auto matvec_two = m_linesearch_deployed.applyHessian(v);
std::cout << "matvec error: " << (matvec_one - matvec_two).norm() / matvec_one.norm() << std::endl;
v = w_plus;
matvec_one = my_H.apply(v);
matvec_two = m_linesearch_deployed.applyHessian(v);
auto matvec_three = m_linesearch_deployed.applyHessian(v);
std::cout << "w_plus matec error: " << (matvec_one - matvec_two).norm() / matvec_one.norm() << std::endl;
std::cout << "w_plus.norm(): " << w_plus.norm() << std::endl;
std::cout << "H w_plus.norm() 1: " << matvec_one.norm() << std::endl;
std::cout << "H w_plus.norm() 2: " << matvec_two.norm() << std::endl;
std::cout << "H w_plus.norm() 3: " << matvec_three.norm() << std::endl;
std::ofstream out_file_w("w_plus.txt");
out_file_w.precision(16);
out_file_w << w_plus << std::endl;
std::ofstream out_file("matvec_one.txt");
out_file.precision(16);
out_file << matvec_one << std::endl;
std::ofstream out_file2("matvec_two.txt");
out_file2.precision(16);
out_file2 << matvec_two << std::endl;
std::ofstream out_file3("matvec_three.txt");
out_file3.precision(16);
out_file3 << matvec_two << std::endl;
}
Real Jminus = J(curr_params - fd_eps * perturb);
auto gradp_J_minus = gradp_J(curr_params - fd_eps * perturb);
auto w_minus = m_w_x;
auto x_minus = m_linesearch_deployed.getDoFs();
auto Hw_minus = m_linesearch_deployed.applyHessian(w);
auto H_minus_w_minus = m_linesearch_deployed.applyHessian(w_minus);
auto w_rhs_minus = m_w_rhs;
Real fd_J = (Jplus - Jminus) / (2 * fd_eps);
relerror_delta_J[i] = std::abs((grad_J.dot(perturb) - fd_J) / fd_J);
Eigen::VectorXd fd_diff_grad_p_J = (gradp_J_plus - gradp_J_minus) / (2 * fd_eps);
Eigen::VectorXd fd_delta_w = (w_plus - w_minus) / (2 * fd_eps);
Eigen::VectorXd fd_delta_x = (x_plus - x_minus) / (2 * fd_eps);
Eigen::VectorXd fd_delta_Hw = (Hw_plus - Hw_minus) / (2 * fd_eps);
#if 0
auto fd_H_delta_w = H.apply(fd_delta_w);
Eigen::VectorXd soln_error = ((Hw + fd_eps * fd_delta_Hw + fd_eps * fd_H_delta_w) - w_rhs_plus) + opt.extractFullSolution(opt.kkt_solver.a * (w_lambda + fd_eps * fd_delta_lambda));
std::cout << "||Hw + delta_H w + H delta w + a (lambda + delta lambda) - b||: " << opt.removeFixedEntries(soln_error).norm() << std::endl;
std::cout << soln_error.head(8).transpose() << std::endl;
std::cout << soln_error.segment<8>(m_linesearch_deployed.dofOffsetForJoint(0)).transpose() << std::endl;
Eigen::VectorXd soln_error2 = (H_plus.apply(w_plus) + w_lambda_plus * opt.extractFullSolution(opt.kkt_solver.a) - w_rhs_plus);
std::cout << "||Hw_plus + a lambda_plus - b_plus||: " << opt.removeFixedEntries(soln_error2).norm() << std::endl;
Eigen::VectorXd soln_error3 = ((Hw + fd_eps * fd_delta_Hw + fd_eps * H.apply(delta_w)) - w_rhs_plus) + opt.extractFullSolution(opt.kkt_solver.a * (w_lambda + fd_eps * fd_delta_lambda));
std::cout << "||Hw + delta_H w + H delta w + a (lambda + delta lambda) - b||: " << opt.removeFixedEntries(soln_error3).norm() << std::endl;
Eigen::VectorXd soln_error4 = H_plus_w_plus - w_rhs_plus + opt.extractFullSolution(opt.kkt_solver.a * w_lambda_plus);
std::cout << "||H_plus w_plus + a (lambda + delta lambda) - b||: " << opt.removeFixedEntries(soln_error4).norm() << std::endl;
#endif
matvec_relerror_fd_diff_grad_p_J[i] = (apply_hess_J(curr_params, perturb) - fd_diff_grad_p_J ).norm() / fd_diff_grad_p_J .norm();
relerror_delta_x[i] = (m_delta_x3d - fd_delta_x).norm() / fd_delta_x.norm();
relerror_delta_w[i] = (m_delta_w_x - fd_delta_w).norm() / fd_delta_w.norm();
relerror_delta_Hw[i] = (m_d3E_w.head(w.size()) - fd_delta_Hw).norm() / fd_delta_Hw.norm();
}
std::cout << "Wrote " << hess_J_path << ", " << hess_ac_path << std::endl;
std::cout << "Rel errors in delta J: " << relerror_delta_J .transpose() << std::endl;
std::cout << "Rel errors in matvec hessian-vec J: " << matvec_relerror_fd_diff_grad_p_J .transpose() << std::endl;
std::cout << "Rel errors in delta x: " << relerror_delta_x.transpose() << std::endl;
std::cout << "Rel errors in delta w: " << relerror_delta_w.transpose() << std::endl;
std::cout << "Rel errors in delta Hw: " << relerror_delta_Hw.transpose() << std::endl;
}
|