Allow fixed parameters and direct NLL.

albatross/core/parameter_handling_mixin.h

@@ -70,6 +70,8 @@ struct Parameter {
 
   bool is_valid() const { return within_bounds(); }
 
+  bool is_fixed() const { return has_prior() && prior->is_fixed(); }
+
   double prior_log_likelihood() const {
     if (has_prior()) {
       return prior->log_pdf(value);
@@ -219,7 +221,9 @@ class ParameterHandlingMixin {
     std::vector<ParameterValue> x;
     const ParameterStore params = get_params();
     for (const auto &pair : params) {
-      x.push_back(pair.second.value);
+      if (!pair.second.is_fixed()) {
+        x.push_back(pair.second.value);
+      }
     }
     return x;
   }
@@ -247,14 +251,16 @@ class ParameterHandlingMixin {
   }
 
   void set_params_from_vector(const std::vector<ParameterValue> &x) {
-    const std::size_t n = x.size();
     const ParameterStore params = get_params();
-    assert(n == static_cast<std::size_t>(params.size()));
-
-    const std::vector<ParameterKey> param_names = map_keys(params);
-    for (std::size_t i = 0; i < n; i++) {
-      unchecked_set_param(param_names[i], x[i]);
+    std::size_t i = 0;
+    for (const auto &pair : params) {
+      if (!pair.second.is_fixed()) {
+        unchecked_set_param(pair.first, x[i]);
+        i++;
+      }
     }
+    // Make sure we used all the parameters that were passed in.
+    assert(x.size() == i);
   }
 
   ParameterValue get_param_value(const ParameterKey &name) const {

albatross/core/priors.h

@@ -37,6 +37,7 @@ class Prior {
   virtual std::string get_name() const = 0;
   virtual double lower_bound() const { return -LARGE_VAL; }
   virtual double upper_bound() const { return LARGE_VAL; }
+  virtual bool is_fixed() const { return false; }
   virtual bool operator==(const Prior &other) const {
     return typeid(*this) == typeid(other);
   }
@@ -55,6 +56,18 @@ class UninformativePrior : public Prior {
   }
 };
 
+class FixedPrior : public Prior {
+public:
+  std::string get_name() const override { return "fixed"; };
+  double log_pdf(double x) const override { return 0.; }
+
+  bool is_fixed() const override { return true; }
+
+  template <typename Archive> void serialize(Archive &archive) {
+    archive(cereal::base_class<Prior>(this));
+  }
+};
+
 class PositivePrior : public Prior {
 public:
   double log_pdf(double x) const override { return x > 0. ? 0. : -LARGE_VAL; }
@@ -67,6 +80,18 @@ class PositivePrior : public Prior {
   }
 };
 
+class NonNegativePrior : public Prior {
+public:
+  double log_pdf(double x) const override { return x >= 0. ? 0. : -LARGE_VAL; }
+  std::string get_name() const override { return "non_negative"; };
+  double lower_bound() const override { return 0.; }
+  double upper_bound() const override { return LARGE_VAL; }
+
+  template <typename Archive> void serialize(Archive &archive) {
+    archive(cereal::base_class<Prior>(this));
+  }
+};
+
 class UniformPrior : public Prior {
 public:
   UniformPrior(double lower = 0., double upper = 1.)
@@ -168,6 +193,8 @@ class LogNormalPrior : public Prior {
 
 CEREAL_REGISTER_TYPE(albatross::UninformativePrior);
 CEREAL_REGISTER_TYPE(albatross::PositivePrior);
+CEREAL_REGISTER_TYPE(albatross::NonNegativePrior);
+CEREAL_REGISTER_TYPE(albatross::FixedPrior);
 CEREAL_REGISTER_TYPE(albatross::UniformPrior);
 CEREAL_REGISTER_TYPE(albatross::GaussianPrior);
 CEREAL_REGISTER_TYPE(albatross::LogNormalPrior);

albatross/evaluate.h

@@ -53,7 +53,6 @@ static inline double
 negative_log_likelihood(const Eigen::VectorXd &deviation,
                         const Eigen::LDLT<_MatrixType, _UpLo> &ldlt) {
   const auto diag = ldlt.vectorD();
-  const auto L = ldlt.matrixL();
   const double rank = static_cast<double>(diag.size());
   const double mahalanobis = deviation.dot(ldlt.solve(deviation));
   const double log_det = log_sum(diag);

albatross/tuning_metrics.h

@@ -50,6 +50,22 @@ inline double gp_fast_loo_nll(const RegressionDataset<FeatureType> &dataset,
          model->prior_log_likelihood();
 }
 
+template <typename FeatureType, typename SubFeatureType = FeatureType>
+inline double gp_nll(const RegressionDataset<FeatureType> &dataset,
+                     RegressionModel<FeatureType> *model) {
+  using SerializableGP =
+      SerializableRegressionModel<FeatureType,
+                                  GaussianProcessFit<SubFeatureType>>;
+  SerializableGP *gp_model = static_cast<SerializableGP *>(model);
+  gp_model->fit(dataset);
+  GaussianProcessFit<SubFeatureType> model_fit = gp_model->get_fit();
+
+  double nll =
+      negative_log_likelihood(dataset.targets.mean, model_fit.train_ldlt);
+  nll -= model->prior_log_likelihood();
+  return nll;
+}
+
 inline double loo_nll(const RegressionDataset<double> &dataset,
                       RegressionModel<double> *model) {
   auto loo_folds = leave_one_out(dataset);

tests/test_parameter_handling_mixin.cc

@@ -82,6 +82,37 @@ TEST(test_parameter_handler, test_get_set_from_vector) {
                                 original_handler.get_params_as_vector());
 }
 
+/*
+ * Test the helper functions that let you get and set parameters from
+ * a vector of values.
+ */
+TEST(test_parameter_handler, test_get_set_from_vector_with_fixed) {
+  const ParameterStore expected = {
+      {"1", 4.},
+      {"2", 5.},
+      {"foo", {1., std::make_shared<FixedPrior>()}},
+      {"3", 6.}};
+  const std::vector<ParameterValue> expected_param_vector = {4., 5., 6.};
+
+  ParameterStore original(expected);
+  original["1"] = 1.;
+  original["2"] = 2.;
+  original["foo"] = {-2., std::make_shared<FixedPrior>()};
+  original["3"] = 3.;
+  const std::vector<ParameterValue> original_param_vector = {1., 2., 3.};
+  MockParameterHandler original_handler(original);
+
+  // Make sure we start with the parameter vector we'd expect, even though
+  // it was initialized out of order.
+  expect_parameter_vector_equal(original_param_vector,
+                                original_handler.get_params_as_vector());
+
+  // Now set the parameters using a new vector and make sure they stick
+  original_handler.set_params_from_vector(expected_param_vector);
+  expect_parameter_vector_equal(expected_param_vector,
+                                original_handler.get_params_as_vector());
+}
+
 TEST(test_parameter_handler, test_prior_log_likelihood) {
   auto p = TestParameterHandler();