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
|
#include <catch2/catch_all.hpp>
#include <numeric>
#include <random>
#include <utility>
#include "sopt/imaging_padmm.h"
#include "sopt/logging.h"
#include "sopt/maths.h"
#include "sopt/mpi/communicator.h"
#include "sopt/mpi/utilities.h"
#include "sopt/relative_variation.h"
#include "sopt/sampling.h"
#include "sopt/types.h"
#include "sopt/utilities.h"
#include "sopt/wavelets.h"
// This header is not part of the installed sopt interface
// It is only present in tests
#include "tools_for_tests/directories.h"
#include "tools_for_tests/tiffwrappers.h"
using Catch::Approx;
TEST_CASE("Parallel vs serial inpainting") {
extern std::unique_ptr<std::mt19937_64> mersenne;
using namespace sopt;
auto const world = mpi::Communicator::World();
// split into serial and parallel
auto const split_comm = world.split(static_cast<t_int>(world.is_root()));
if (world.size() < 2) return;
// Some type aliases for simplicity
using Scalar = double;
using Vector = sopt::Vector<Scalar>;
using Image = sopt::Image<Scalar>;
std::string const input = "cameraman256";
// Read input file
Image const image = world.is_root()
? world.broadcast(sopt::tools::read_standard_tiff(input))
: world.broadcast<Image>();
// Initializing sensing operator
// The operator is obtained by world root proc and split across the procs in split_comm
sopt::t_uint const nmeasure = 0.33 * image.size();
auto indices = world.is_root()
? world.broadcast(sopt::Sampling(image.size(), nmeasure, *mersenne).indices())
: world.broadcast<std::vector<t_uint>>();
if (split_comm.size() > 1) {
auto const copy = indices;
auto const N = indices.size() / split_comm.size() +
(split_comm.rank() < indices.size() % split_comm.size() ? 1 : 0);
auto const start = split_comm.rank() * (indices.size() / split_comm.size()) +
std::min(indices.size() % split_comm.size(), split_comm.rank());
indices.resize(N);
std::copy(copy.begin() + start, copy.begin() + N + start, indices.begin());
}
auto const sampling = sopt::linear_transform<Scalar>(sopt::Sampling(image.size(), indices));
// Initializing wavelets
auto const wavelet = sopt::wavelets::factory("DB4", 4);
auto const psi = sopt::linear_transform<Scalar>(wavelet, image.rows(), image.cols());
// Computing proximal-ADMM parameters
Vector const y0 = sampling * Vector::Map(image.data(), image.size());
CHECK(y0.size() == indices.size());
auto constexpr snr = 30.0;
auto const sigma = y0.stableNorm() / std::sqrt(y0.size()) * std::pow(10.0, -(snr / 20.0));
auto const epsilon = world.broadcast(std::sqrt(nmeasure + 2 * std::sqrt(y0.size())) * sigma);
// Create dirty vector
std::normal_distribution<> gaussian_dist(0, sigma);
Vector y = world.is_root() ? y0 : world.broadcast<Vector>();
if (world.is_root()) {
for (sopt::t_int i = 0; i < y0.size(); i++) y(i) += gaussian_dist(*mersenne);
world.broadcast(y);
}
if (split_comm.size() > 1) {
auto const N =
y.size() / split_comm.size() + (split_comm.rank() < y.size() % split_comm.size() ? 1 : 0);
auto const start = split_comm.rank() * (y.size() / split_comm.size()) +
std::min(y.size() % split_comm.size(), split_comm.rank());
y = y.segment(start, N).eval();
}
CHECK(y.size() == indices.size());
// Creating proximal-ADMM Functor
auto padmm =
sopt::algorithm::ImagingProximalADMM<Scalar>(y)
.itermax(4)
.gamma(1e-1)
.relative_variation(5e-4)
.l2ball_proximal(sopt::proximal::WeightedL2Ball<Scalar>(epsilon).communicator(split_comm))
.tight_frame(false)
.l1_proximal_tolerance(1e-2)
.l1_proximal_nu(1)
.l1_proximal_itermax(50)
.l1_proximal_positivity_constraint(true)
.l1_proximal_real_constraint(true)
.lagrange_update_scale(0.9)
.nu(1e0)
.Psi(psi);
LinearTransform<Vector> const parallel_sampling(
[&sampling](Vector &out, Vector const &input) { out = sampling * input; }, sampling.sizes(),
[&sampling, split_comm](Vector &out, Vector const &input) {
out = sampling.adjoint() * input;
split_comm.all_sum_all(out);
},
sampling.adjoint().sizes());
padmm.Phi(parallel_sampling);
padmm.residual_convergence([&padmm, split_comm, world](Vector const &,
Vector const &residual) mutable -> bool {
auto const residual_norm =
sopt::mpi::l2_norm(residual, padmm.l2ball_proximal_weights(), split_comm);
SOPT_LOW_LOG(" - [PADMM] Residuals: {} <? {}", residual_norm, padmm.residual_tolerance());
CHECK(residual_norm == Approx(world.broadcast(residual_norm, world.root_id())));
return residual_norm < padmm.residual_tolerance();
});
sopt::ScalarRelativeVariation<Scalar> conv(padmm.relative_variation(), padmm.relative_variation(),
"Objective function");
padmm.objective_convergence(
[&padmm, split_comm, conv, world](Vector const &x, Vector const &) mutable -> bool {
auto const result = conv(
sopt::mpi::l1_norm(padmm.Psi().adjoint() * x, padmm.l1_proximal_weights(), split_comm));
CHECK(result == (world.broadcast<int>(result, world.root_id()) != 0));
return result;
});
padmm.is_converged([world](Vector const &image, Vector const &) -> bool {
auto const from_root = world.broadcast(image);
CHECK(from_root.isApprox(image, 1e-12));
return true;
});
auto const diagnostic = padmm();
CHECK(diagnostic.good == (world.broadcast<int>(diagnostic.good, world.root_id()) != 0));
CHECK(diagnostic.x.isApprox(world.broadcast(diagnostic.x)));
}
|
