Implemented GSL optimizers

Signed-off-by: Daniel Claudino <[email protected]>

CMakeLists.txt

@@ -42,6 +42,8 @@ option(XACC_ENSMALLEN_INCLUDE_DIR "Path to ensmallen.hpp for mlpack optimizer" "
 option(XACC_ARMADILLO_INCLUDE_DIR "Path to armadillo header for mlpack optimizer" "")
 option(XACC_BUILD_SCIPY "Build Scipy optimizer plugin" OFF)
 option(XACC_BUILD_ANNEALING "Build annealing libraries" OFF)
+option(XACC_BUILD_GSL "Build GNU Scientific Library optimizer plugin" OFF)
+
 if(XACC_BUILD_ANNEALING)
   set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DANNEALING_ENABLED")
 else()

quantum/plugins/optimizers/CMakeLists.txt

@@ -1,5 +1,9 @@
 add_subdirectory(nlopt-optimizers)
 add_subdirectory(mlpack)
+
+if(XACC_BUILD_GSL)
+add_subdirectory(gsl)
+endif()
 if(XACC_BUILD_SCIPY)
   execute_process(COMMAND ${Python_EXECUTABLE} -c "import scipy" RESULT_VARIABLE SCIPY_EXISTS)
   if(SCIPY_EXISTS EQUAL "1")

quantum/plugins/optimizers/gsl/CMakeLists.txt

@@ -0,0 +1,53 @@
+message(STATUS "${BoldGreen}Building GSL Optimizer.${ColorReset}")
+set(LIBRARY_NAME xacc-gsl-optimizer)
+
+file(GLOB
+     SRC
+     gsl_optimizer.cpp)
+
+usfunctiongetresourcesource(TARGET
+                            ${LIBRARY_NAME}
+                            OUT
+                            SRC)
+usfunctiongeneratebundleinit(TARGET
+                             ${LIBRARY_NAME}
+                             OUT
+                             SRC)
+
+add_library(${LIBRARY_NAME} SHARED ${SRC})
+find_package(GSL REQUIRED)
+target_include_directories(${LIBRARY_NAME}
+                           PUBLIC . ${GSL_INCLUDE_DIR})
+
+target_link_libraries(${LIBRARY_NAME} PUBLIC xacc GSL::gsl GSL::gslcblas)
+
+set(_bundle_name xacc_optimizer_gsl)
+set_target_properties(${LIBRARY_NAME}
+                      PROPERTIES COMPILE_DEFINITIONS
+                                 US_BUNDLE_NAME=${_bundle_name}
+                                 US_BUNDLE_NAME
+                                 ${_bundle_name})
+
+usfunctionembedresources(TARGET
+                         ${LIBRARY_NAME}
+                         WORKING_DIRECTORY
+                         ${CMAKE_CURRENT_SOURCE_DIR}
+                         FILES
+                         manifest.json)
+
+if(APPLE)
+  set_target_properties(${LIBRARY_NAME}
+                        PROPERTIES INSTALL_RPATH "@loader_path/../lib")
+  set_target_properties(${LIBRARY_NAME}
+                        PROPERTIES LINK_FLAGS "-undefined dynamic_lookup")
+else()
+  set_target_properties(${LIBRARY_NAME}
+                        PROPERTIES INSTALL_RPATH "$ORIGIN/../lib")
+  set_target_properties(${LIBRARY_NAME} PROPERTIES LINK_FLAGS "-shared")
+endif()
+
+if(XACC_BUILD_TESTS)
+  add_subdirectory(tests)
+endif()
+
+install(TARGETS ${LIBRARY_NAME} DESTINATION ${CMAKE_INSTALL_PREFIX}/plugins)

quantum/plugins/optimizers/gsl/gsl_optimizer.cpp

@@ -0,0 +1,245 @@
+#include "gsl_optimizer.hpp"
+#include "xacc_plugin.hpp"
+#include <gsl/gsl_multimin.h>
+
+struct GSLFunctionWrapper {
+  xacc::OptFunction &function;
+  std::vector<double> grad;
+
+  double operator()(const gsl_vector *v) {
+    std::vector<double> x(v->size);
+    for (size_t i = 0; i < v->size; ++i) {
+      x[i] = gsl_vector_get(v, i);
+    }
+    if (grad.empty()) {
+      return function(std::move(x));
+    } else {
+      return function(x, grad);
+    }
+  }
+
+  void getGradient(gsl_vector *df) {
+    for (size_t i = 0; i < grad.size(); ++i) {
+      gsl_vector_set(df, i, grad[i]);
+    }
+  }
+};
+
+namespace xacc {
+
+const std::string GSLOptimizer::get_algorithm() const {
+  std::string optimizerAlgo = "COBYLA";
+  if (options.stringExists("algorithm")) {
+    optimizerAlgo = options.getString("algorithm");
+  }
+  if (options.stringExists("scipy-optimizer")) {
+    optimizerAlgo = options.getString("scipy-optimizer");
+  }
+  return optimizerAlgo;
+}
+
+const bool GSLOptimizer::isGradientBased() const {
+
+  std::string optimizerAlgo = "cobyla";
+  if (options.stringExists("algorithm")) {
+    optimizerAlgo = options.getString("algorithm");
+  }
+  if (options.stringExists("scipy-optimizer")) {
+    optimizerAlgo = options.getString("scipy-optimizer");
+  }
+
+  if (options.stringExists("optimizer")) {
+    optimizerAlgo = options.getString("optimizer");
+  }
+
+  if (optimizerAlgo == "bfgs") {
+    return true;
+  } else {
+    return false;
+  }
+}
+
+OptResult GSLOptimizer::optimize(OptFunction &function) {
+
+  bool maximize = false;
+  if (options.keyExists<bool>("maximize")) {
+    xacc::info("Turning on maximize!");
+    maximize = options.get<bool>("maximize");
+  }
+
+  std::string algo = "COBYLA";
+  if (options.stringExists("algorithm")) {
+    algo = options.getString("algorithm");
+  }
+  if (options.stringExists("gsl-optimizer")) {
+    algo = options.getString("gsl-optimizer");
+  }
+  if (options.stringExists("optimizer")) {
+    algo = options.getString("optimizer");
+  }
+
+  if (algo == "cobyla" || algo == "COBYLA") {
+    algo = "COBYLA";
+  } else if (algo == "nelder-mead" || algo == "Nelder-Mead") {
+    algo = "Nelder-Mead";
+  } else if (algo == "bfgs" || algo == "BFGS" || algo == "l-bfgs") {
+    algo = "BFGS";
+  } else {
+    xacc::XACCLogger::instance()->error("Invalid optimizer at this time: " +
+                                        algo);
+  }
+
+  double tol = 1e-6;
+  if (options.keyExists<double>("ftol")) {
+    tol = options.get<double>("ftol");
+    xacc::info("[GSL] function tolerance set to " + std::to_string(tol));
+  }
+  if (options.keyExists<double>("gsl-ftol")) {
+    tol = options.get<double>("ftol");
+    xacc::info("[GSL] function tolerance set to " + std::to_string(tol));
+  }
+
+  int maxeval = 1000;
+  if (options.keyExists<int>("maxeval")) {
+    maxeval = options.get<int>("maxeval");
+    xacc::info("[GSL] max function evaluations set to " +
+               std::to_string(maxeval));
+  }
+  if (options.keyExists<int>("gsl-maxeval")) {
+    maxeval = options.get<int>("maxeval");
+    xacc::info("[GSL] max function evaluations set to " +
+               std::to_string(maxeval));
+  }
+
+  double step = 1.0e-4;
+  if (options.keyExists<double>("step-size")) {
+    step = options.get<double>("step-size");
+    xacc::info("[GSL] step size set to " + std::to_string(step));
+  }
+  if (options.keyExists<double>("gsl-step-size")) {
+    step = options.get<double>("gsl-step-size");
+    xacc::info("[GSL] step size set to " + std::to_string(step));
+  }
+
+  std::vector<double> x(function.dimensions());
+  if (options.keyExists<std::vector<double>>("initial-parameters")) {
+    x = options.get_with_throw<std::vector<double>>("initial-parameters");
+  } else if (options.keyExists<std::vector<int>>("initial-parameters")) {
+    auto tmpx = options.get<std::vector<int>>("initial-parameters");
+    x = std::vector<double>(tmpx.begin(), tmpx.end());
+  }
+
+  auto dim = function.dimensions();
+  gsl_vector *theta = gsl_vector_alloc(dim);
+  for (size_t i = 0; i < dim; ++i) {
+    gsl_vector_set(theta, i, x[i]);
+  }
+
+  GSLFunctionWrapper wrapper{function};
+  double minVal;
+  std::vector<double> result(dim);
+  if (isGradientBased()) {
+
+    wrapper.grad.resize(dim);
+    gsl_multimin_function_fdf gslFunction;
+    gslFunction.n = dim;
+
+    gslFunction.f = [](const gsl_vector *v, void *params) -> double {
+      GSLFunctionWrapper *w = static_cast<GSLFunctionWrapper *>(params);
+      return (*w)(v);
+    };
+
+    gslFunction.df = [](const gsl_vector *v, void *params, gsl_vector *df) {
+      GSLFunctionWrapper *w = static_cast<GSLFunctionWrapper *>(params);
+      w->getGradient(df);
+    };
+
+    // I don't understand why, but this seems kinda redundant
+    gslFunction.fdf = [](const gsl_vector *v, void *params, double *f,
+                         gsl_vector *df) {
+      GSLFunctionWrapper *w = static_cast<GSLFunctionWrapper *>(params);
+      (*f) = (*w)(v);
+      w->getGradient(df);
+    };
+
+    gslFunction.params = &wrapper;
+
+    const gsl_multimin_fdfminimizer_type *T = gsl_multimin_fdfminimizer_conjugate_fr;
+    gsl_multimin_fdfminimizer *s = gsl_multimin_fdfminimizer_alloc(T, dim);
+
+    int set_status = gsl_multimin_fdfminimizer_set(s, &gslFunction, theta, 0.01, 1e-4);
+    if (set_status) {
+      gsl_multimin_fdfminimizer_free(s);
+      gsl_vector_free(theta);
+      throw std::runtime_error("Failed to set the minimizer.");
+    }
+
+    int status, iter = 0;
+    do {
+      iter++;
+      status = gsl_multimin_fdfminimizer_iterate(s);
+
+      if (status)
+        break;
+
+      // Check for convergence using gradient norm
+      status = gsl_multimin_test_gradient(s->gradient, 1e-4);
+
+    } while (status == GSL_CONTINUE && iter < 100);
+
+    // std::vector<double> result(dim);
+    for (size_t i = 0; i < dim; ++i) {
+      result[i] = gsl_vector_get(s->x, i);
+    }
+
+    minVal = s->f;
+    gsl_vector_free(theta);
+    gsl_multimin_fdfminimizer_free(s);
+
+  } else {
+
+    gsl_multimin_function gslFunction;
+    gslFunction.n = dim;
+    gslFunction.f = [](const gsl_vector *v, void *params) -> double {
+      GSLFunctionWrapper *w = static_cast<GSLFunctionWrapper *>(params);
+      return (*w)(v);
+    };
+
+    gslFunction.params = &wrapper;
+
+    gsl_vector *steps = gsl_vector_alloc(gslFunction.n);
+    gsl_vector_set_all(steps, step);
+
+    const gsl_multimin_fminimizer_type *T = gsl_multimin_fminimizer_nmsimplex2;
+    gsl_multimin_fminimizer *s =
+        gsl_multimin_fminimizer_alloc(T, gslFunction.n);
+    gsl_multimin_fminimizer_set(s, &gslFunction, theta, steps);
+
+    int status, iter;
+    double size;
+    do {
+      iter++;
+      status = gsl_multimin_fminimizer_iterate(s);
+
+      if (status)
+        break;
+
+      size = gsl_multimin_fminimizer_size(s);
+      status = gsl_multimin_test_size(size, 1e-4);
+    } while (status == GSL_CONTINUE && iter < 100);
+
+    std::vector<double> result(gslFunction.n);
+    for (size_t i = 0; i < gslFunction.n; ++i) {
+      result[i] = gsl_vector_get(s->x, i);
+    }
+
+    minVal = s->fval;
+
+    gsl_multimin_fminimizer_free(s);
+    gsl_vector_free(theta);
+  }
+
+  return {minVal, result};
+}
+} // namespace xacc
+REGISTER_OPTIMIZER(xacc::GSLOptimizer)

quantum/plugins/optimizers/gsl/gsl_optimizer.hpp

@@ -0,0 +1,24 @@
+#ifndef GSL_OPTIMIZER_HPP
+#define GSL_OPTIMIZER_HPP
+
+#include <xacc.hpp>
+#include <xacc_service.hpp>
+#include <Optimizer.hpp>
+
+namespace xacc {
+
+class GSLOptimizer : public xacc::Optimizer {
+public:
+  GSLOptimizer() = default;
+  ~GSLOptimizer() = default;
+
+  const std::string name() const override { return "gsl"; }
+  const std::string description() const override { return ""; }
+
+  OptResult optimize(OptFunction &function) override;
+  const bool isGradientBased() const override;
+  virtual const std::string get_algorithm() const override;
+};
+
+} // namespace xacc
+#endif

quantum/plugins/optimizers/gsl/manifest.json

@@ -0,0 +1,6 @@
+{
+  "bundle.symbolic_name" : "xacc_optimizer_gsl",
+  "bundle.activator" : true,
+  "bundle.name" : "XACC GNU Scientific Library Optimizers",
+  "bundle.description" : ""
+}

quantum/plugins/optimizers/gsl/tests/CMakeLists.txt

@@ -0,0 +1,2 @@
+add_xacc_test(GSLOptimizer)
+target_link_libraries(GSLOptimizerTester xacc)