Fix PriorAcquisitionFunction (#1185)

* Dummy commit to create Draft MR

* First preliminary fix for the iteration_number.

* Update iteration_number counter

* Update priors and examples

* Fix typo

* Remove doubled priro floor

* Update tests

* Add changelog.md

* Update format

CHANGELOG.md

@@ -7,6 +7,7 @@
 
 ## Bugfixes
 - Fix kwargs for DifferentialEvolution (#1187)
+- Fix PiBo implementation (#1076)
 
 # 2.2.1
 

examples/1_basics/6_priors.py

@@ -20,6 +20,7 @@
     ConfigurationSpace,
     NormalFloatHyperparameter,
     UniformIntegerHyperparameter,
+    UniformFloatHyperparameter,
 )
 from sklearn.datasets import load_digits
 from sklearn.exceptions import ConvergenceWarning
@@ -38,7 +39,7 @@
 
 class MLP:
     @property
-    def configspace(self) -> ConfigurationSpace:
+    def prior_configspace(self) -> ConfigurationSpace:
         # Build Configuration Space which defines all parameters and their ranges.
         # To illustrate different parameter types,
         # we use continuous, integer and categorical parameters.
@@ -100,7 +101,67 @@ def configspace(self) -> ConfigurationSpace:
         )
 
         # Add all hyperparameters at once:
-        cs.add([n_layer, n_neurons, activation, optimizer, batch_size, learning_rate_init])
+        cs.add(
+            [n_layer, n_neurons, activation, optimizer, batch_size, learning_rate_init]
+        )
+
+        return cs
+
+    @property
+    def configspace(self) -> ConfigurationSpace:
+        # Build Configuration Space which defines all parameters and their ranges.
+        # To illustrate different parameter types,
+        # we use continuous, integer and categorical parameters.
+        cs = ConfigurationSpace()
+
+        # We do not have an educated belief on the number of layers beforehand
+        n_layer = UniformIntegerHyperparameter(
+            "n_layer",
+            lower=1,
+            upper=5,
+        )
+
+        # Define network width without a specific prior
+        n_neurons = UniformIntegerHyperparameter(
+            "n_neurons",
+            lower=8,
+            upper=256,
+        )
+
+        # Define activation functions without specific weights
+        activation = CategoricalHyperparameter(
+            "activation",
+            ["logistic", "tanh", "relu"],
+            default_value="relu",
+        )
+
+        # Define optimizer without specific weights
+        optimizer = CategoricalHyperparameter(
+            "optimizer",
+            ["sgd", "adam"],
+            default_value="adam",
+        )
+
+        # Define batch size without specific distribution
+        batch_size = UniformIntegerHyperparameter(
+            "batch_size",
+            16,
+            512,
+            default_value=128,
+        )
+
+        # Define learning rate range without log-normal prior
+        learning_rate_init = UniformFloatHyperparameter(
+            "learning_rate_init",
+            lower=1e-5,
+            upper=1.0,
+            default_value=1e-3,
+        )
+
+        # Add all hyperparameters at once:
+        cs.add(
+            [n_layer, n_neurons, activation, optimizer, batch_size, learning_rate_init]
+        )
 
         return cs
 
@@ -119,8 +180,12 @@ def train(self, config: Configuration, seed: int = 0) -> float:
             )
 
             # Returns the 5-fold cross validation accuracy
-            cv = StratifiedKFold(n_splits=5, random_state=seed, shuffle=True)  # to make CV splits consistent
-            score = cross_val_score(classifier, digits.data, digits.target, cv=cv, error_score="raise")
+            cv = StratifiedKFold(
+                n_splits=5, random_state=seed, shuffle=True
+            )  # to make CV splits consistent
+            score = cross_val_score(
+                classifier, digits.data, digits.target, cv=cv, error_score="raise"
+            )
 
         return 1 - np.mean(score)
 
@@ -140,7 +205,9 @@ def train(self, config: Configuration, seed: int = 0) -> float:
 
     # We define the prior acquisition function, which conduct the optimization using priors over the optimum
     acquisition_function = PriorAcquisitionFunction(
-        acquisition_function=HyperparameterOptimizationFacade.get_acquisition_function(scenario),
+        acquisition_function=HyperparameterOptimizationFacade.get_acquisition_function(
+            scenario
+        ),
         decay_beta=scenario.n_trials / 10,  # Proven solid value
     )
 

smac/acquisition/function/prior_acquisition_function.py

@@ -72,6 +72,9 @@ def __init__(
             acquisition_type = self._acquisition_function
 
         self._rescale = isinstance(acquisition_type, (LCB, TS))
+
+        # Variables needed to adapt the weighting of the prior
+        self._initial_design_size = None
         self._iteration_number = 0
 
     @property
@@ -116,7 +119,12 @@ def _update(self, **kwargs: Any) -> None:
             Current incumbent value.
         """
         assert "eta" in kwargs
-        self._iteration_number += 1
+
+        # Compute intiial design size
+        if self._initial_design_size is None:
+            self._initial_design_size = kwargs["num_data"]
+
+        self._iteration_number = kwargs["num_data"] - self._initial_design_size
         self._eta = kwargs["eta"]
 
         assert self.model is not None
@@ -146,8 +154,10 @@ def _compute_prior(self, X: np.ndarray) -> np.ndarray:
         for parameter, X_col in zip(self._hyperparameters.values(), X.T):
             if self._discretize and isinstance(parameter, FloatHyperparameter):
                 assert self._discrete_bins_factor is not None
-                number_of_bins = int(np.ceil(self._discrete_bins_factor * self._decay_beta / self._iteration_number))
-                prior_values *= self._compute_discretized_pdf(parameter, X_col, number_of_bins) + self._prior_floor
+                number_of_bins = int(
+                    np.ceil(self._discrete_bins_factor * self._decay_beta / (self._iteration_number + 1))
+                )
+                prior_values *= self._compute_discretized_pdf(parameter, X_col, number_of_bins)
             else:
                 prior_values *= parameter._pdf(X_col[:, np.newaxis])
 
@@ -178,6 +188,7 @@ def _compute_discretized_pdf(
             The user prior over the optimum for the parameter at hand.
         """
         # Evaluates the actual pdf on all the relevant points
+        # Replace deprecated method
         pdf_values = hyperparameter._pdf(X_col[:, np.newaxis])
 
         # Retrieves the largest value of the pdf in the domain
@@ -221,6 +232,6 @@ def _compute(self, X: np.ndarray) -> np.ndarray:
             acq_values = self._acquisition_function._compute(X)
 
         prior_values = self._compute_prior(X) + self._prior_floor
-        decayed_prior_values = np.power(prior_values, self._decay_beta / self._iteration_number)
+        decayed_prior_values = np.power(prior_values, self._decay_beta / (self._iteration_number + 1))
 
         return acq_values * decayed_prior_values

tests/test_acquisition/test_functions.py

@@ -298,16 +298,16 @@ def test_prior_init_ts(prior_model, acq_ts, beta):
 
 def test_prior_update(prior_model, acquisition_function, beta):
     paf = PriorAcquisitionFunction(acquisition_function=acquisition_function, decay_beta=beta)
-    paf.update(model=prior_model, eta=2)
+    paf.update(model=prior_model, eta=2, num_data=10)
     assert paf._eta == 2
     assert paf._acquisition_function._eta == 2
-    assert paf._iteration_number == 1
+    assert paf._iteration_number == 0
 
 
 def test_prior_compute_prior_Nx1(prior_model, hyperparameter_dict, acquisition_function, beta):
     prior_model.update_prior(hyperparameter_dict)
     paf = PriorAcquisitionFunction(acquisition_function=acquisition_function, decay_beta=beta)
-    paf.update(model=prior_model, eta=1)
+    paf.update(model=prior_model, eta=1, num_data=1)
 
     X = np.array([0, 0.5, 1]).reshape(3, 1)
     prior_values = paf._compute_prior(X)
@@ -321,7 +321,7 @@ def test_prior_compute_prior_Nx1(prior_model, hyperparameter_dict, acquisition_f
 def test_prior_compute_prior_NxD(prior_model, hyperparameter_dict, acquisition_function, beta):
     prior_model.update_prior(hyperparameter_dict)
     paf = PriorAcquisitionFunction(acquisition_function=acquisition_function, decay_beta=beta)
-    paf.update(model=prior_model, eta=1)
+    paf.update(model=prior_model, eta=1, num_data=1)
 
     X = np.array([[0, 0], [0, 1], [1, 1]])
     prior_values = paf._compute_prior(X)
@@ -339,7 +339,7 @@ def test_prior_compute_prior_1xD(prior_model, acquisition_function, beta):
 
     prior_model.update_prior(hyperparameter_dict)
     paf = PriorAcquisitionFunction(acquisition_function=acquisition_function, decay_beta=beta)
-    paf.update(model=prior_model, eta=1)
+    paf.update(model=prior_model, eta=1, num_data=1)
 
     X = np.array([[0.5, 0.5]])
     prior_values = paf._compute_prior(X)
@@ -351,7 +351,7 @@ def test_prior_compute_prior_1xD(prior_model, acquisition_function, beta):
 def test_prior_compute_prior_1x1(prior_model, hyperparameter_dict, acquisition_function, beta):
     prior_model.update_prior(hyperparameter_dict)
     paf = PriorAcquisitionFunction(acquisition_function=acquisition_function, decay_beta=beta)
-    paf.update(model=prior_model, eta=1)
+    paf.update(model=prior_model, eta=1, num_data=1)
 
     X = np.array([0.5]).reshape(1, 1)
     prior_values = paf._compute_prior(X)
@@ -378,7 +378,7 @@ def hp_dict3(x0_prior, x1_prior, x2_prior):
 def test_prior_1xD(hp_dict3, prior_model, acquisition_function, beta, prior_floor):
     prior_model.update_prior(hp_dict3)
     paf = PriorAcquisitionFunction(acquisition_function=acquisition_function, decay_beta=beta, prior_floor=prior_floor)
-    paf.update(model=prior_model, eta=1.0)
+    paf.update(model=prior_model, eta=1.0, num_data=1)
     configurations = [ConfigurationMock([1.0, 1.0, 1.0])]
     acq = paf(configurations)
     assert acq.shape == (1, 1)
@@ -391,7 +391,7 @@ def test_prior_1xD(hp_dict3, prior_model, acquisition_function, beta, prior_floo
 def test_prior_NxD(hp_dict3, prior_model, acquisition_function, beta, prior_floor):
     prior_model.update_prior(hp_dict3)
     paf = PriorAcquisitionFunction(acquisition_function=acquisition_function, decay_beta=beta, prior_floor=prior_floor)
-    paf.update(model=prior_model, eta=1.0)
+    paf.update(model=prior_model, eta=1.0, num_data=1)
 
     # These are the exact same numbers as in the EI tests below
     configurations = [
@@ -449,15 +449,15 @@ def test_prior_NxD_TS(prior_model, hp_dict3, acq_ts, beta, prior_floor):
 def test_prior_decay(hp_dict3, prior_model, acquisition_function, beta, prior_floor):
     prior_model.update_prior(hp_dict3)
     paf = PriorAcquisitionFunction(acquisition_function=acquisition_function, decay_beta=beta, prior_floor=prior_floor)
-    paf.update(model=prior_model, eta=1.0)
+    paf.update(model=prior_model, eta=1.0, num_data=0)
     configurations = [ConfigurationMock([0.1, 0.1, 0.1])]
 
     for i in range(1, 6):
         prior_factor = np.power(0.2 * 1.0 * 1.8 + paf._prior_floor, beta / i)
         acq = paf(configurations)
         print(acq, 0.90020601136712231 * prior_factor)
         assert np.isclose(acq[0][0], 0.90020601136712231 * prior_factor)
-        paf.update(model=prior_model, eta=1.0)  # increase iteration number
+        paf.update(model=prior_model, eta=1.0, num_data = i)  # increase iteration number
 
 
 def test_prior_discretize_pdf(prior_model, acquisition_function, beta, prior_floor):
@@ -467,7 +467,7 @@ def test_prior_discretize_pdf(prior_model, acquisition_function, beta, prior_flo
     paf = PriorAcquisitionFunction(
         acquisition_function=acquisition_function, decay_beta=beta, prior_floor=prior_floor, discretize=True
     )
-    paf.update(model=prior_model, eta=1)
+    paf.update(model=prior_model, eta=1, num_data=1)
 
     number_of_bins_1 = 13
     number_of_bins_2 = 27521