refactor(SearchSpace): removes a lot of methods from SearchSpace (#150

)

neps/optimizers/bayesian_optimization/acquisition_functions/ei.py

@@ -69,7 +69,7 @@ def eval(
         if len(space.fidelities) > 0 and self.optimize_on_max_fidelity:
             assert len(space.fidelities) == 1
             fid_name, fid = next(iter(space.fidelities.items()))
-            _x = [space.from_dict({**e.hp_values(), fid_name: fid.upper}) for e in x]
+            _x = [space.from_dict({**e._values, fid_name: fid.upper}) for e in x]
         else:
             _x = list(x)
 

neps/optimizers/grid_search/optimizer.py

@@ -9,12 +9,9 @@
 import torch
 
 from neps.optimizers.base_optimizer import BaseOptimizer, SampledConfig
+from neps.search_spaces import Categorical, Constant, Float, Integer
 from neps.search_spaces.architecture.graph_grammar import GraphParameter
 from neps.search_spaces.domain import UNIT_FLOAT_DOMAIN
-from neps.search_spaces.hyperparameters.categorical import Categorical
-from neps.search_spaces.hyperparameters.constant import Constant
-from neps.search_spaces.hyperparameters.float import Float
-from neps.search_spaces.hyperparameters.integer import Integer
 
 if TYPE_CHECKING:
     from neps.search_spaces.search_space import SearchSpace

neps/optimizers/multi_fidelity/mf_bo.py

@@ -4,10 +4,16 @@
 from copy import deepcopy
 from typing import TYPE_CHECKING, Any, Literal
 
+from neps.search_spaces.functions import sample_one_old
+
 
 def update_fidelity(config: SearchSpace, fidelity: int | float) -> SearchSpace:
     assert config.fidelity is not None
     config.fidelity.set_value(fidelity)
+    # TODO: Place holder until we can get rid of passing around search spaces
+    # as configurations
+    assert config.fidelity_name is not None
+    config._values[config.fidelity_name] = fidelity
     return config
 
 
@@ -93,7 +99,7 @@ def _fit_models(self) -> None:
             train_y = deepcopy(self.rung_histories[rung]["perf"])
             # extract only the pending configurations that are at `rung`
             pending_df = pending_df[pending_df.rung == rung]
-            pending_x = deepcopy(pending_df.config.values.tolist())
+            pending_x = deepcopy(pending_df["config"].values.tolist())
             # update fidelity
             fidelities = [self.rung_map[rung]] * len(pending_x)
             pending_x = list(map(update_fidelity, pending_x, fidelities))
@@ -196,7 +202,8 @@ def sample_new_config(
         elif self.sampling_policy is not None:
             config = self.sampling_policy.sample(**self.sampling_args)
         else:
-            config = self.pipeline_space.sample(
+            config = sample_one_old(
+                self.pipeline_space,
                 patience=self.patience,
                 user_priors=self.use_priors,
                 ignore_fidelity=True,

neps/optimizers/multi_fidelity/sampling_policy.py

@@ -22,6 +22,7 @@
 from neps.sampling.priors import Prior
 from neps.sampling.samplers import Sampler
 from neps.search_spaces.encoding import ConfigEncoder
+from neps.search_spaces.functions import sample_one_old
 
 if TYPE_CHECKING:
     from botorch.acquisition.analytic import SingleTaskGP
@@ -64,8 +65,11 @@ def __init__(self, pipeline_space: SearchSpace):
         super().__init__(pipeline_space=pipeline_space)
 
     def sample(self, *args: Any, **kwargs: Any) -> SearchSpace:
-        return self.pipeline_space.sample(
-            patience=self.patience, user_priors=False, ignore_fidelity=True
+        return sample_one_old(
+            self.pipeline_space,
+            patience=self.patience,
+            user_priors=False,
+            ignore_fidelity=True,
         )
 
 
@@ -88,8 +92,12 @@ def sample(self, *args: Any, **kwargs: Any) -> SearchSpace:
         user_priors = False
         if np.random.uniform() < self.fraction_from_prior:
             user_priors = True
-        return self.pipeline_space.sample(
-            patience=self.patience, user_priors=user_priors, ignore_fidelity=True
+
+        return sample_one_old(
+            self.pipeline_space,
+            patience=self.patience,
+            user_priors=user_priors,
+            ignore_fidelity=True,
         )
 
 
@@ -140,9 +148,11 @@ def sample_neighbour(
         )
 
         while True:
-            # sampling a config
-            config = self.pipeline_space.sample(
-                patience=self.patience, user_priors=False, ignore_fidelity=False
+            config = sample_one_old(
+                self.pipeline_space,
+                patience=self.patience,
+                user_priors=False,
+                ignore_fidelity=False,
             )
             # computing distance from incumbent
             d = compute_config_dist(config, incumbent)
@@ -188,8 +198,11 @@ def sample(  # noqa: PLR0912, C901, PLR0915
         logger.info(f"Sampling from {policy} with weights (i, p, r)={prob_weights}")
 
         if policy == "prior":
-            config = self.pipeline_space.sample(
-                patience=self.patience, user_priors=True, ignore_fidelity=True
+            config = sample_one_old(
+                self.pipeline_space,
+                patience=self.patience,
+                user_priors=True,
+                ignore_fidelity=True,
             )
         elif policy == "inc":
             if (
@@ -201,7 +214,7 @@ def sample(  # noqa: PLR0912, C901, PLR0915
                 user_priors = False
 
             if inc is None:
-                inc = self.pipeline_space.sample_default_configuration().clone()
+                inc = self.pipeline_space.from_dict(self.pipeline_space.default_config)
                 logger.warning(
                     "No incumbent config found, using default as the incumbent."
                 )
@@ -251,7 +264,8 @@ def sample(  # noqa: PLR0912, C901, PLR0915
                     f"Crossing over with user_priors={user_priors} with p={probs}"
                 )
                 # sampling a configuration either randomly or from a prior
-                _config = self.pipeline_space.sample(
+                _config = sample_one_old(
+                    self.pipeline_space,
                     patience=self.patience,
                     user_priors=user_priors,
                     ignore_fidelity=True,
@@ -274,9 +288,11 @@ def sample(  # noqa: PLR0912, C901, PLR0915
                     f"{{'mutation', 'crossover', 'hypersphere', 'gaussian'}}"
                 )
         else:
-            # random
-            config = self.pipeline_space.sample(
-                patience=self.patience, user_priors=False, ignore_fidelity=True
+            config = sample_one_old(
+                self.pipeline_space,
+                patience=self.patience,
+                user_priors=False,
+                ignore_fidelity=True,
             )
         return config
 
@@ -316,8 +332,8 @@ def update_model(
         pending_x: list[SearchSpace],
         decay_t: float | None = None,
     ) -> None:
-        x_train = self._encoder.encode([config.hp_values() for config in train_x])
-        x_pending = self._encoder.encode([config.hp_values() for config in pending_x])
+        x_train = self._encoder.encode([config._values for config in train_x])
+        x_pending = self._encoder.encode([config._values for config in pending_x])
         y_train = torch.tensor(train_y, dtype=torch.float64, device=self.device)
 
         # TODO: Most of this just copies BO and the duplication can be replaced
@@ -377,7 +393,7 @@ def sample(
         """
         # sampling random configurations
         samples = [
-            self.pipeline_space.sample(user_priors=False, ignore_fidelity=True)
+            sample_one_old(self.pipeline_space, user_priors=False, ignore_fidelity=True)
             for _ in range(SAMPLE_THRESHOLD)
         ]
 

neps/optimizers/multi_fidelity/successive_halving.py

@@ -26,6 +26,7 @@
     Integer,
     SearchSpace,
 )
+from neps.search_spaces.functions import sample_one_old
 
 if TYPE_CHECKING:
     from neps.state.optimizer import BudgetInfo
@@ -371,14 +372,14 @@ def sample_new_config(
     ) -> SearchSpace:
         # Samples configuration from policy or random
         if self.sampling_policy is None:
-            config = self.pipeline_space.sample(
+            return sample_one_old(
+                self.pipeline_space,
                 patience=self.patience,
                 user_priors=self.use_priors,
                 ignore_fidelity=True,
             )
-        else:
-            config = self.sampling_policy.sample(**self.sampling_args)
-        return config
+
+        return self.sampling_policy.sample(**self.sampling_args)
 
     def _generate_new_config_id(self) -> int:
         if len(self.observed_configs) == 0:
@@ -405,22 +406,27 @@ def get_config_and_ids(self) -> tuple[RawConfig, str, str | None]:
         Returns:
             [type]: [description]
         """
+        fidelity_name = self.pipeline_space.fidelity_name
+        assert fidelity_name is not None
+
         rung_to_promote = self.is_promotable()
         if rung_to_promote is not None:
             # promotes the first recorded promotable config in the argsort-ed rung
             row = self.observed_configs.iloc[self.rung_promotions[rung_to_promote][0]]
             config = row["config"].clone()
             rung = rung_to_promote + 1
             # assigning the fidelity to evaluate the config at
-            config.fidelity.set_value(self.rung_map[rung])
+
+            config_values = config._values
+            config_values[fidelity_name] = self.rung_map[rung]
+
             # updating config IDs
             previous_config_id = f"{row.name}_{rung_to_promote}"
             config_id = f"{row.name}_{rung}"
         else:
             rung_id = self.min_rung
             # using random instead of np.random to be consistent with NePS BO
             rng = random.Random(None)  # TODO: Seeding
-
             if (
                 self.use_priors
                 and self.sample_default_first
@@ -431,10 +437,10 @@ def get_config_and_ids(self) -> tuple[RawConfig, str, str | None]:
                     rung_id = self.max_rung
                     logger.info("Next config will be evaluated at target fidelity.")
                 logger.info("Sampling the default configuration...")
-                config = self.pipeline_space.sample_default_configuration()
-
+                config = self.pipeline_space.from_dict(self.pipeline_space.default_config)
             elif rng.random() < self.random_interleave_prob:
-                config = self.pipeline_space.sample(
+                config = sample_one_old(
+                    self.pipeline_space,
                     patience=self.patience,
                     user_priors=False,  # sample uniformly random
                     ignore_fidelity=True,
@@ -443,13 +449,13 @@ def get_config_and_ids(self) -> tuple[RawConfig, str, str | None]:
                 config = self.sample_new_config(rung=rung_id)
 
             fidelity_value = self.rung_map[rung_id]
-            assert config.fidelity is not None
-            config.fidelity.set_value(fidelity_value)
+            config_values = config._values
+            config_values[fidelity_name] = fidelity_value
 
             previous_config_id = None
             config_id = f"{self._generate_new_config_id()}_{rung_id}"
 
-        return config.hp_values(), config_id, previous_config_id
+        return config_values, config_id, previous_config_id
 
     def _enhance_priors(self, confidence_score: dict[str, float] | None = None) -> None:
         """Only applicable when priors are given along with a confidence.

neps/optimizers/multi_fidelity_prior/priorband.py

@@ -91,7 +91,7 @@ def find_incumbent(self, rung: int | None = None) -> SearchSpace:
         else:
             # THIS block should not ever execute, but for runtime anomalies, if no
             # incumbent can be extracted, the prior is treated as the incumbent
-            inc = self.pipeline_space.sample_default_configuration()
+            inc = self.pipeline_space.from_dict(self.pipeline_space.default_config)
             logger.warning(
                 "Treating the prior as the incumbent. "
                 "Please check if this should not happen."
@@ -259,7 +259,7 @@ def _prior_to_incumbent_ratio_dynamic(self, rung: int) -> tuple[float, float]:
         # requires at least eta completed configurations to begin computing scores
         if len(self.rung_histories[rung]["config"]) >= self.eta:
             # retrieve the prior
-            prior = self.pipeline_space.sample_default_configuration()
+            prior = self.pipeline_space.from_dict(self.pipeline_space.default_config)
             # retrieve the global incumbent
             inc = self.find_incumbent()
             # subsetting the top 1/eta configs from the rung

neps/optimizers/multi_fidelity_prior/utils.py

@@ -1,6 +1,6 @@
 from __future__ import annotations
 
-from typing import TYPE_CHECKING, Any
+from typing import Any
 
 import numpy as np
 import torch
@@ -10,14 +10,12 @@
     Categorical,
     Constant,
     GraphParameter,
+    Float,
+    Integer,
     SearchSpace,
 )
 from neps.search_spaces.encoding import ConfigEncoder
-from neps.search_spaces.hyperparameters.float import Float
-from neps.search_spaces.hyperparameters.integer import Integer
-
-if TYPE_CHECKING:
-    import pandas as pd
+from neps.search_spaces.functions import sample_one_old, pairwise_dist
 
 
 def update_fidelity(config: SearchSpace, fidelity: int | float) -> SearchSpace:
@@ -98,7 +96,7 @@ def custom_crossover(
     Returns a configuration where each HP in config1 has `crossover_prob`% chance of
     getting config2's value of the corresponding HP. By default, crossover rate is 50%.
     """
-    _existing = config1.hp_values()
+    _existing = config1._values
 
     for _ in range(patience):
         child_config = {}
@@ -114,7 +112,8 @@ def custom_crossover(
     # fail safe check to handle edge cases where config1=config2 or
     # config1 extremely local to config2 such that crossover fails to
     # generate new config in a discrete (sub-)space
-    return config1.sample(
+    return sample_one_old(
+        config1,
         patience=patience,
         user_priors=False,
         ignore_fidelity=True,
@@ -130,8 +129,8 @@ def compute_config_dist(config1: SearchSpace, config2: SearchSpace) -> float:
     the Hamming distance of the categorical subspace.
     """
     encoder = ConfigEncoder.from_parameters({**config1.numerical, **config1.categoricals})
-    configs = encoder.encode([config1.hp_values(), config2.hp_values()])
-    dist = encoder.pdist(configs, square_form=False)
+    configs = encoder.encode([config1._values, config2._values])
+    dist = pairwise_dist(configs, encoder, square_form=False)
     return float(dist.item())
 
 
@@ -146,16 +145,16 @@ def compute_scores(
     # TODO: This could lifted up and just done in the class itself
     # in a vectorized form.
     encoder = ConfigEncoder.from_space(config, include_fidelity=include_fidelity)
-    encoded_config = encoder.encode([config.hp_values()])
+    encoded_config = encoder.encode([config._values])
 
     prior_dist = Prior.from_space(
         prior,
-        center_values=prior.hp_values(),
+        center_values=prior._values,
         include_fidelity=include_fidelity,
     )
     inc_dist = Prior.from_space(
         inc,
-        center_values=inc.hp_values(),
+        center_values=inc._values,
         include_fidelity=include_fidelity,
     )