Update for op_norm inside Phi

Michael McLeod · Michael McLeod · commit ea60b789fe9d · 2025-02-24T16:28:53.000Z
diff --git a/cpp/benchmarks/algorithms.cc b/cpp/benchmarks/algorithms.cc
@@ -71,7 +71,7 @@ BENCHMARK_DEFINE_F(AlgoFixture, Padmm)(benchmark::State &state) {
   m_padmm = factory::padmm_factory<sopt::algorithm::ImagingProximalADMM<t_complex>>(
       factory::algo_distribution::serial, m_measurements_transform, wavelets, m_uv_data, m_sigma,
       m_imsizey, m_imsizex, m_sara.size(), state.range(3) + 1, true, true, false, 1e-3, 1e-2, 50,
-      1.0, 1.0);
+      1.0);
 
   while (state.KeepRunning()) {
     auto start = std::chrono::high_resolution_clock::now();
@@ -92,7 +92,7 @@ BENCHMARK_DEFINE_F(AlgoFixture, ForwardBackward)(benchmark::State &state) {
   m_fb = factory::fb_factory<sopt::algorithm::ImagingForwardBackward<t_complex>>(
       factory::algo_distribution::serial, m_measurements_transform, wavelets, m_uv_data, m_sigma,
       beta, gamma, m_imsizey, m_imsizex, m_sara.size(), state.range(3) + 1, true, true, false, 1e-3,
-      1e-2, 50, 1.0);
+      1e-2, 50);
 
   while (state.KeepRunning()) {
     auto start = std::chrono::high_resolution_clock::now();
diff --git a/cpp/benchmarks/algorithms_mpi.cc b/cpp/benchmarks/algorithms_mpi.cc
@@ -90,7 +90,7 @@ BENCHMARK_DEFINE_F(AlgoFixtureMPI, PadmmDistributeImage)(benchmark::State &state
   m_padmm = factory::padmm_factory<sopt::algorithm::ImagingProximalADMM<t_complex>>(
       factory::algo_distribution::mpi_distributed, m_measurements_distribute_image, wavelets,
       m_uv_data, m_sigma, m_imsizey, m_imsizex, m_sara.size(), state.range(3) + 1, true, true,
-      false, 1e-3, 1e-2, 50, 1.0, 1.0);
+      false, 1e-3, 1e-2, 50, 1.0);
 
   // Benchmark the application of the algorithm
   while (state.KeepRunning()) {
@@ -111,7 +111,7 @@ BENCHMARK_DEFINE_F(AlgoFixtureMPI, PadmmDistributeGrid)(benchmark::State &state)
   m_padmm = factory::padmm_factory<sopt::algorithm::ImagingProximalADMM<t_complex>>(
       factory::algo_distribution::mpi_distributed, m_measurements_distribute_grid, wavelets,
       m_uv_data, m_sigma, m_imsizey, m_imsizex, m_sara.size(), state.range(3) + 1, true, true,
-      false, 1e-3, 1e-2, 50, 1.0, 1.0);
+      false, 1e-3, 1e-2, 50, 1.0);
 
   // Benchmark the application of the algorithm
   while (state.KeepRunning()) {
diff --git a/cpp/example/padmm_mpi_random_coverage.cc b/cpp/example/padmm_mpi_random_coverage.cc
@@ -116,7 +116,6 @@ std::shared_ptr<sopt::algorithm::ImagingProximalADMM<t_complex>> padmm_factory(
       .l1_proximal_real_constraint(true)
       .residual_tolerance(epsilon)
       .lagrange_update_scale(0.9)
-      .sq_op_norm(1e0)
       .Psi(Psi)
       .Phi(*measurements);
   sopt::ScalarRelativeVariation<t_complex> conv(padmm->relative_variation(),
diff --git a/cpp/example/padmm_mpi_real_data.cc b/cpp/example/padmm_mpi_real_data.cc
@@ -89,7 +89,6 @@ std::shared_ptr<sopt::algorithm::ImagingProximalADMM<t_complex>> padmm_factory(
       .l1_proximal_real_constraint(true)
       .residual_tolerance(epsilon)
       .lagrange_update_scale(0.9)
-      .sq_op_norm(1e0)
       .Psi(Psi)
       .Phi(*measurements);
   sopt::ScalarRelativeVariation<t_complex> conv(padmm->relative_variation(),
diff --git a/cpp/example/padmm_random_coverage.cc b/cpp/example/padmm_random_coverage.cc
@@ -108,7 +108,6 @@ void padmm(const std::string &name, const Image<t_complex> &M31, const std::stri
 #ifdef PURIFY_CImg
           .is_converged(show_image)
 #endif
-          .sq_op_norm(1e0)
           .Psi(Psi)
           .Phi(*measurements_transform);
 
diff --git a/cpp/example/padmm_real_data.cc b/cpp/example/padmm_real_data.cc
@@ -91,7 +91,6 @@ void padmm(const std::string &name, const t_uint &imsizex, const t_uint &imsizey
       .l1_proximal_real_constraint(true)
       .residual_convergence(epsilon)
       .lagrange_update_scale(0.9)
-      .sq_op_norm(1e0)
       .Psi(Psi)
       .Phi(*measurements_transform);
 
diff --git a/cpp/example/padmm_reweighted_simulation.cc b/cpp/example/padmm_reweighted_simulation.cc
@@ -102,7 +102,6 @@ int main(int nargs, char const **args) {
                          .l1_proximal_real_constraint(true)
                          .residual_convergence(epsilon * 1.001)
                          .lagrange_update_scale(0.9)
-                         .sq_op_norm(1e0)
                          .Psi(Psi)
                          .Phi(measurements_transform);
   // Timing reconstruction
diff --git a/cpp/example/padmm_simulation.cc b/cpp/example/padmm_simulation.cc
@@ -117,7 +117,6 @@ int main(int nargs, char const **args) {
                          .l1_proximal_real_constraint(true)
                          .residual_convergence(epsilon * 1.001)
                          .lagrange_update_scale(0.9)
-                         .sq_op_norm(1e0)
                          .Psi(Psi)
                          .itermax(100)
                          .is_converged(convergence_function)
diff --git a/cpp/example/sara_padmm_random_coverage.cc b/cpp/example/sara_padmm_random_coverage.cc
@@ -97,7 +97,6 @@ int main(int, char **) {
                          .l1_proximal_real_constraint(true)
                          .residual_convergence(epsilon * 1.001)
                          .lagrange_update_scale(0.9)
-                         .sq_op_norm(1e0)
                          .Psi(Psi)
                          .Phi(measurements_transform);
 
diff --git a/cpp/main.cc b/cpp/main.cc
@@ -61,14 +61,14 @@ int main(int argc, const char **argv) {
       getInputData(params, mop_algo, wop_algo, using_mpi);
 
   // create measurement operator
-  auto [measurements_transform, operator_norm] =
+  auto measurements_transform =
       createMeasurementOperator(params, mop_algo, wop_algo, using_mpi, image_index, w_stacks,
                                 uv_data, measurement_op_eigen_vector);
 
   // create wavelet operator
   const waveletInfo wavelets = createWaveletOperator(params, wop_algo);
 
-  PURIFY_LOW_LOG("Value of operator norm is {}", operator_norm);
+  PURIFY_LOW_LOG("Value of operator norm is {}", measurements_transform->norm());
   t_real const flux_scale = 1.;
   uv_data.vis = uv_data.vis.array() * uv_data.weights.array() / flux_scale;
 
@@ -95,7 +95,7 @@ int main(int argc, const char **argv) {
     beam_units = uv_data.size() / flux_scale / flux_scale;
   }
 
-  savePSF(params, def_header, measurements_transform, uv_data, flux_scale, sigma, operator_norm,
+  savePSF(params, def_header, measurements_transform, uv_data, flux_scale, sigma,
           beam_units);
 
   // the dirty image
@@ -114,7 +114,7 @@ int main(int argc, const char **argv) {
         (params.wavelet_basis().size() < 2) and (not params.realValueConstraint()) and
             (not params.positiveValueConstraint()),
         params.relVarianceConvergence(), params.dualFBVarianceConvergence(), 50,
-        params.epsilonConvergenceScaling(), operator_norm);
+        params.epsilonConvergenceScaling());
   if (params.algorithm() == "fb") {
     std::shared_ptr<DifferentiableFunc<t_complex>> f;
     if (params.diffFuncType() == diff_func_type::L2Norm_with_CRR) {
@@ -135,7 +135,7 @@ int main(int argc, const char **argv) {
         params.iterations(), params.realValueConstraint(), params.positiveValueConstraint(),
         (params.wavelet_basis().size() < 2) and (not params.realValueConstraint()) and
             (not params.positiveValueConstraint()),
-        params.relVarianceConvergence(), params.dualFBVarianceConvergence(), 50, operator_norm,
+        params.relVarianceConvergence(), params.dualFBVarianceConvergence(), 50,
         params.model_path(), params.nondiffFuncType(), f);
   }
   if (params.algorithm() == "primaldual")
@@ -144,7 +144,7 @@ int main(int argc, const char **argv) {
         sigma * params.epsilonScaling() / flux_scale, params.height(), params.width(),
         wavelets.sara_size, params.iterations(), params.realValueConstraint(),
         params.positiveValueConstraint(), params.relVarianceConvergence(),
-        params.epsilonConvergenceScaling(), operator_norm);
+        params.epsilonConvergenceScaling());
   // Add primal dual preconditioning
   if (params.algorithm() == "primaldual" and params.precondition_iters() > 0) {
     PURIFY_HIGH_LOG(
diff --git a/cpp/purify/algorithm_factory.h b/cpp/purify/algorithm_factory.h
@@ -60,7 +60,7 @@ padmm_factory(const algo_distribution dist,
               const bool tight_frame = false, const t_real relative_variation = 1e-3,
               const t_real l1_proximal_tolerance = 1e-2,
               const t_uint maximum_proximal_iterations = 50,
-              const t_real residual_tolerance_scaling = 1, const t_real op_norm = 1) {
+              const t_real residual_tolerance_scaling = 1) {
   typedef typename Algorithm::Scalar t_scalar;
   if (sara_size > 1 and tight_frame)
     throw std::runtime_error(
@@ -78,7 +78,6 @@ padmm_factory(const algo_distribution dist,
       .l1_proximal_positivity_constraint(positive_constraint)
       .l1_proximal_real_constraint(real_constraint)
       .lagrange_update_scale(0.9)
-      .sq_op_norm(op_norm * op_norm)
       .Psi(*wavelets)
       .Phi(*measurements);
 #ifdef PURIFY_MPI
@@ -162,7 +161,7 @@ fb_factory(const algo_distribution dist,
            const bool real_constraint = true, const bool positive_constraint = true,
            const bool tight_frame = false, const t_real relative_variation = 1e-3,
            const t_real l1_proximal_tolerance = 1e-2, const t_uint maximum_proximal_iterations = 50,
-           const t_real op_norm = 1, const std::string model_path = "",
+           const std::string model_path = "",
            const nondiff_func_type g_proximal = nondiff_func_type::L1Norm,
            std::shared_ptr<DifferentiableFunc<typename Algorithm::Scalar>> f_function = nullptr) {
   typedef typename Algorithm::Scalar t_scalar;
@@ -178,7 +177,6 @@ fb_factory(const algo_distribution dist,
       .step_size(step_size * std::sqrt(2))
       .relative_variation(relative_variation)
       .tight_frame(tight_frame)
-      .sq_op_norm(op_norm * op_norm)
       .Phi(*measurements);
 
   if (f_function) fb->f_function(f_function);  // only override f_function default if non-null
@@ -262,8 +260,7 @@ primaldual_factory(
     const utilities::vis_params &uv_data, const t_real sigma, const t_uint imsizey,
     const t_uint imsizex, const t_uint sara_size, const t_uint max_iterations = 500,
     const bool real_constraint = true, const bool positive_constraint = true,
-    const t_real relative_variation = 1e-3, const t_real residual_tolerance_scaling = 1,
-    const t_real op_norm = 1) {
+    const t_real relative_variation = 1e-3, const t_real residual_tolerance_scaling = 1) {
   typedef typename Algorithm::Scalar t_scalar;
   PURIFY_INFO("Constructing Primal Dual algorithm");
   auto epsilon = std::sqrt(2 * uv_data.size() + 2 * std::sqrt(4 * uv_data.size())) * sigma;
@@ -274,7 +271,7 @@ primaldual_factory(
       .positivity_constraint(positive_constraint)
       .Psi(*wavelets)
       .Phi(*measurements)
-      .tau(0.5 / (op_norm * op_norm + 1))
+      .tau(0.5 / (measurements->sq_norm() + 1))
       .xi(1.)
       .sigma(1.);
 #ifdef PURIFY_MPI
diff --git a/cpp/purify/setup_utils.cc b/cpp/purify/setup_utils.cc
@@ -247,7 +247,7 @@ inputData getInputData(const YamlParser &params,
   return {uv_data, sigma, measurement_op_eigen_vector, image_index, w_stacks};
 }
 
-measurementOpInfo createMeasurementOperator(
+std::shared_ptr<sopt::LinearTransform<Vector<t_complex>>> createMeasurementOperator(
     const YamlParser &params, const factory::distributed_measurement_operator mop_algo,
     const factory::distributed_wavelet_operator wop_algo, const bool using_mpi,
     const std::vector<t_int> &image_index, const std::vector<t_real> &w_stacks,
@@ -294,6 +294,7 @@ measurementOpInfo createMeasurementOperator(
                   params.powMethod_tolerance(), comm.broadcast(measurement_op_eigen_vector).eval());
     measurement_op_eigen_vector = std::get<1>(power_method_result);
     operator_norm = std::get<0>(power_method_result);
+    measurements_transform->set_norm(operator_norm);
   } else
 #endif
   {
@@ -302,9 +303,10 @@ measurementOpInfo createMeasurementOperator(
         measurement_op_eigen_vector);
     measurement_op_eigen_vector = std::get<1>(power_method_result);
     operator_norm = std::get<0>(power_method_result);
+    measurements_transform->set_norm(operator_norm);
   }
 
-  return {measurements_transform, operator_norm};
+  return measurements_transform;
 }
 
 void setupCostFunctions(const YamlParser &params, std::unique_ptr<DifferentiableFunc<t_complex>> &f,
@@ -401,8 +403,7 @@ void saveMeasurementEigenVector(const YamlParser &params,
 void savePSF(
     const YamlParser &params, const pfitsio::header_params &def_header,
     const std::shared_ptr<sopt::LinearTransform<Vector<t_complex>>> &measurements_transform,
-    const utilities::vis_params &uv_data, const t_real flux_scale, const t_real sigma,
-    const t_real operator_norm, const t_real beam_units) {
+    const utilities::vis_params &uv_data, const t_real flux_scale, const t_real sigma, const t_real beam_units) {
   pfitsio::header_params psf_header = def_header;
   psf_header.fits_name = params.output_path() + "/psf.fits";
   psf_header.pix_units = "Jy/Pixel";
@@ -417,7 +418,7 @@ void savePSF(
     auto const world = sopt::mpi::Communicator::World();
     PURIFY_LOW_LOG(
         "Expected image domain residual RMS is {} jy/beam",
-        sigma * params.epsilonScaling() * operator_norm /
+        sigma * params.epsilonScaling() * measurements_transform->norm() /
             (std::sqrt(params.width() * params.height()) * world.all_sum_all(uv_data.size())));
     if (world.is_root())
 #else
@@ -426,7 +427,7 @@ void savePSF(
       pfitsio::write2d(psf_image, psf_header, true);
   } else {
     PURIFY_LOW_LOG("Expected image domain residual RMS is {} jy/beam",
-                   sigma * params.epsilonScaling() * operator_norm /
+                   sigma * params.epsilonScaling() * measurements_transform->norm() /
                        (std::sqrt(params.width() * params.height()) * uv_data.size()));
     pfitsio::write2d(psf_image, psf_header, true);
   }
diff --git a/cpp/purify/setup_utils.h b/cpp/purify/setup_utils.h
@@ -41,12 +41,7 @@ inputData getInputData(const YamlParser &params,
                        const factory::distributed_measurement_operator mop_algo,
                        const factory::distributed_wavelet_operator wop_algo, const bool using_mpi);
 
-struct measurementOpInfo {
-  std::shared_ptr<sopt::LinearTransform<Vector<t_complex>>> measurement_transform;
-  t_real operator_norm;
-};
-
-measurementOpInfo createMeasurementOperator(
+std::shared_ptr<sopt::LinearTransform<Vector<t_complex>>> createMeasurementOperator(
     const YamlParser &params, const factory::distributed_measurement_operator mop_algo,
     const factory::distributed_wavelet_operator wop_algo, const bool using_mpi,
     const std::vector<t_int> &image_index, const std::vector<t_real> &w_stacks,
@@ -72,8 +67,7 @@ void saveMeasurementEigenVector(const YamlParser &params,
 void savePSF(
     const YamlParser &params, const pfitsio::header_params &def_header,
     const std::shared_ptr<sopt::LinearTransform<Vector<t_complex>>> &measurements_transform,
-    const utilities::vis_params &uv_data, const t_real flux_scale, const t_real sigma,
-    const t_real operator_norm, const t_real beam_units);
+    const utilities::vis_params &uv_data, const t_real flux_scale, const t_real sigma, const t_real beam_units);
 
 void saveDirtyImage(
     const YamlParser &params, const pfitsio::header_params &def_header,
diff --git a/cpp/tests/algo_factory.cc b/cpp/tests/algo_factory.cc
diff --git a/cpp/tests/mpi_algo_factory.cc b/cpp/tests/mpi_algo_factory.cc
diff --git a/cpp/uncertainty_quantification/uq_main.cc b/cpp/uncertainty_quantification/uq_main.cc
