diff --git a/optimization/optimizers/gradient_descent/gradient_descent.h b/optimization/optimizers/gradient_descent/gradient_descent.h
index 46dd99ee254c1f36287355935f228bff1262c87a..1a920961b5cec3069c54755972cb2d67b72c4e7c 100644
--- a/optimization/optimizers/gradient_descent/gradient_descent.h
+++ b/optimization/optimizers/gradient_descent/gradient_descent.h
@@ -7,54 +7,89 @@
namespace optimization {
//--------------------------------------------------------------------------------------------------
-template <int32_t N, typename Log = GradientDescentLogNothing>
-class GradientDescent : public Log {
+template <int32_t N>
+class GradientDescent {
public:
GradientDescent() {}
GradientDescent(Scalar learning_rate, uint32_t me, Scalar gt)
: learning_rate_(learning_rate), max_evaluations_(me), gradient_threshold_(gt)
{}
+ Scalar& learning_rate() { return learning_rate_; }
+ Scalar learning_rate() const { return learning_rate_; }
+
uint32_t n_evaluations() const { return n_evaluations_; }
// We evaluate the objective function exactly once per iteration.
uint32_t n_iterations() const { return n_evaluations_; }
-
- Scalar& learning_rate() { return learning_rate_; }
- Scalar learning_rate() const { return learning_rate_; }
+ VectorNs<N> const& point() const { return point_; }
+ VectorNs<N> const& gradient() const { return gradient_; }
+ Scalar value() const { return value_; }
template <typename Objective>
- VectorNs<N> optimize(Objective& objective, VectorNs<N> const& initial_point) {
- VectorNs<N> point = initial_point;
- Scalar value;
- VectorNs<N> gradient;
- gradient.resize(point.size());
- objective.eval(point, value, gradient);
- Log::initialize(objective);
- Log::push_back(point, value, gradient);
-
- for (n_evaluations_ = 1; n_evaluations_ < max_evaluations_; ++n_evaluations_) {
- if (gradient.norm() <= gradient_threshold_) {
- std::cout << "Gradient norm below threshold: " << gradient.norm() << '\n';
- break;
- }
-
- point -= learning_rate_ * gradient;
- objective.eval(point, value, gradient);
-
- Log::push_back(point, value, gradient);
- }
+ VectorNs<N> const& optimize(Objective& objective, VectorNs<N> const& initial_point);
- return point;
- }
+ template <typename Objective, typename Log>
+ VectorNs<N> const& optimize(Objective& objective, VectorNs<N> const& initial_point, Log& log);
private:
- uint32_t n_evaluations_;
-
+ // hyperparameters
Scalar learning_rate_;
uint32_t max_evaluations_;
Scalar gradient_threshold_;
+
+ // algorithm state
+ uint32_t n_evaluations_;
+ VectorNs<N> point_;
+ VectorNs<N> gradient_;
+ Scalar value_;
};
+//..................................................................................................
+template <int32_t N>
+template <typename Objective>
+VectorNs<N> const& GradientDescent<N>::optimize(
+ Objective& objective,
+ VectorNs<N> const& initial_point
+) {
+ GradientDescentLogNothing log;
+ return optimize(objective, initial_point, log);
+}
+
+//..................................................................................................
+template <int32_t N>
+template <typename Objective, typename Log>
+VectorNs<N> const& GradientDescent<N>::optimize(
+ Objective& objective,
+ VectorNs<N> const& initial_point,
+ Log& log
+) {
+ n_evaluations_ = 0;
+ point_ = initial_point;
+ gradient_.resize(point_.size());
+ log.initialize(objective);
+
+ while (true) {
+ objective.eval(point_, value_, gradient_);
+ ++n_evaluations_;
+ log.push_back(point_, value_, gradient_);
+
+ // termination conditions
+ if (n_evaluations_ >= max_evaluations_) {
+ std::cout << "Gradient descent reached max evaluations: " << n_evaluations_ << '\n';
+ break;
+ }
+ if (gradient_.norm() <= gradient_threshold_) {
+ std::cout << "Gradient descent reached gradient threshold: " << gradient_.norm();
+ std::cout << '\n';
+ break;
+ }
+
+ point_ -= learning_rate_ * gradient_;
+ }
+
+ return point_;
+}
+
}
#endif
diff --git a/optimization/optimizers/gradient_descent/logs/everything.h b/optimization/optimizers/gradient_descent/logs/everything.h
index 8eaf55772a7043ca1f1834162003d8ec28c8c5de..53ced123e29d1a2619805208f411e7b6c028b274 100644
--- a/optimization/optimizers/gradient_descent/logs/everything.h
+++ b/optimization/optimizers/gradient_descent/logs/everything.h
@@ -5,7 +5,7 @@
#include "objectives/samples_vis.h"
#include "utils/eigen_json.h"
#include "utils/vector.h"
-#include "utils/vis_only.h"
+#include <string>
#include <vector>
namespace optimization {
diff --git a/optimization/optimizers/gradient_descent/main.cpp b/optimization/optimizers/gradient_descent/main.cpp
index b71f58896ae4bdbe003c49c7587eac5cdf1a2059..3ab9d34129d6cd234c22ab19d0e8e1ce09ae6e07 100644
--- a/optimization/optimizers/gradient_descent/main.cpp
+++ b/optimization/optimizers/gradient_descent/main.cpp
@@ -8,9 +8,8 @@
#include <fstream>
#include <iostream>
-using json = nlohmann::json;
-
using namespace optimization;
+using json = nlohmann::json;
//--------------------------------------------------------------------------------------------------
int main(int const argc, char const** argv) {
@@ -52,31 +51,29 @@ int main(int const argc, char const** argv) {
Objective objective;
objective.dim() = dim;
+ GradientDescent<-1> optimizer(learning_rate, max_evaluations, gradient_threshold);
+ GradientDescentLogEverything<-1> log;
+
+ // Only log stuff if we're going to use it.
if (log_file_path.empty() && vis_file_path.empty()) {
- // If we're not saving data, use a lean optimizer.
- // TODO: Find a way to deduplicate code between these branches.
- GradientDescent<-1> optimizer(learning_rate, max_evaluations, gradient_threshold);
- VectorXs minimum = optimizer.optimize(objective, initial_point);
- std::cout << "n evaluations: " << optimizer.n_evaluations() << '\n';
- std::cout << "final point: " << minimum << '\n';
+ optimizer.optimize(objective, initial_point);
} else {
- using Log = GradientDescentLogEverything<-1>;
- GradientDescent<-1, Log> optimizer(learning_rate, max_evaluations, gradient_threshold);
- VectorXs minimum = optimizer.optimize(objective, initial_point);
- std::cout << "n evaluations: " << optimizer.n_evaluations() << '\n';
- std::cout << "final point: " << minimum << '\n';
+ optimizer.optimize(objective, initial_point, log);
+ }
+
+ std::cout << "n evaluations: " << optimizer.n_evaluations() << '\n';
+ std::cout << "final point: " << optimizer.point() << '\n';
- if (!log_file_path.empty()) {
- json data = optimizer;
- std::ofstream log_file(log_file_path);
- log_file << data.dump(4) << '\n';
- }
+ if (!log_file_path.empty()) {
+ json data = log;
+ std::ofstream log_file(log_file_path);
+ log_file << data.dump(4) << '\n';
+ }
- if (!vis_file_path.empty()) {
- json data = GradientDescentVis<-1>{optimizer};
- std::ofstream vis_file(vis_file_path);
- vis_file << data.dump(4) << '\n';
- }
+ if (!vis_file_path.empty()) {
+ json data = GradientDescentVis<-1>{log};
+ std::ofstream vis_file(vis_file_path);
+ vis_file << data.dump(4) << '\n';
}
return 0;