Decouple parameter discovery from data extraction in _get_fit_args (facebook#5200)

Summary:

Adds a `data_parameters` argument to `TorchAdapter._get_fit_args` that decouples SSD construction (model params) from data column extraction (target params). This lets the TL adapter set `_model_space` to include source-only RangeParameters directly, so the SSD naturally covers the full joint feature space -- eliminating the need for the `_expand_ssd_to_joint_space` post-hoc expansion.

Overrides `_set_search_space` to add source-only RangeParameters from the joint search space to `_model_space` while preserving target bounds for shared params (Normalize stays anchored to target bounds). At gen time, `self.parameters` is temporarily swapped to target-only so `extract_search_space_digest` sees only params present in the gen-time search space.

Deletes `_expand_ssd_to_joint_space` (~90 lines).

Differential Revision: D104702983

Author: Carl Hvarfner

ax/adapter/torch.py

@@ -774,6 +774,7 @@ def _get_fit_args(
         search_space: SearchSpace,
         experiment_data: ExperimentData,
         update_outcomes_and_parameters: bool,
+        data_parameters: list[str] | None = None,
     ) -> tuple[
         list[SupervisedDataset],
         list[list[TCandidateMetadata]] | None,
@@ -791,6 +792,11 @@ def _get_fit_args(
             update_outcomes_and_parameters: Whether to update `self.outcomes` with
                 all outcomes found in the observations and `self.parameters` with
                 all parameters in the search space. Typically only used in `_fit`.
+            data_parameters: When provided, columns to extract from
+                ``experiment_data``. Defaults to ``self.parameters``. Useful when
+                the model space is larger than the data (e.g. transfer learning
+                with heterogeneous search spaces where the data only contains
+                target columns but the model operates in a joint feature space).
 
         Returns:
             The datasets & metadata, extracted from the ``experiment_data``, and the
@@ -818,12 +824,23 @@ def _get_fit_args(
         search_space_digest = extract_search_space_digest(
             search_space=search_space, param_names=self.parameters
         )
+        extract_params = (
+            data_parameters if data_parameters is not None else self.parameters
+        )
+        # When data_parameters differs from self.parameters, the SSD's
+        # task_feature indices don't match the data columns. Pass None
+        # to skip MultiTaskDataset wrapping (the caller handles it).
+        extraction_ssd = (
+            None
+            if data_parameters is not None and data_parameters != self.parameters
+            else search_space_digest
+        )
         # Convert observations to datasets
         datasets, ordered_outcomes, candidate_metadata = self._convert_experiment_data(
             experiment_data=experiment_data,
             outcomes=self.outcomes,
-            parameters=self.parameters,
-            search_space_digest=search_space_digest,
+            parameters=extract_params,
+            search_space_digest=extraction_ssd,
         )
         datasets = self._update_w_aux_exp_datasets(datasets=datasets)
 

ax/adapter/transfer_learning/adapter.py

@@ -7,7 +7,6 @@
 
 from __future__ import annotations
 
-import dataclasses
 import warnings
 from collections.abc import Mapping, Sequence
 from logging import Logger
@@ -39,7 +38,7 @@
 from ax.core.observation import ObservationData, ObservationFeatures
 from ax.core.optimization_config import OptimizationConfig
 from ax.core.parameter import FixedParameter, RangeParameter
-from ax.core.search_space import SearchSpace, SearchSpaceDigest
+from ax.core.search_space import SearchSpace
 from ax.exceptions.core import DataRequiredError, UnsupportedError, UserInputError
 from ax.generation_strategy.best_model_selector import (
     ReductionCriterion,
@@ -64,6 +63,8 @@
 
 
 class TransferLearningAdapter(TorchAdapter):
+    _source_only_params: set[str]
+
     def __init__(
         self,
         *,
@@ -150,22 +151,7 @@ def __init__(
             target_search_space=search_space,
         )
 
-        # Add source-only backfilled params as FixedParameter to the target
-        # search space so that the compatibility check passes and the model
-        # space includes these params (FixedToTunable will later convert them
-        # to RangeParameter using the joint space bounds).
         search_space = search_space.clone()  # avoid mutating caller's object
-        target_param_names = set(search_space.parameters.keys())
-        for name, param in self.joint_search_space.parameters.items():
-            if name not in target_param_names and param.backfill_value is not None:
-                search_space.add_parameter(
-                    FixedParameter(
-                        name=name,
-                        parameter_type=param.parameter_type,
-                        value=param.backfill_value,
-                    )
-                )
-
         # Include backfill param names in filled_params so Phase 1 of
         # check_search_space_compatibility passes for target-only params.
         filled_params.extend(self.joint_search_space.backfill_values().keys())
@@ -196,6 +182,28 @@ def __init__(
             default_model_gen_options=default_model_gen_options,
         )
 
+    def _set_search_space(self, search_space: SearchSpace) -> None:
+        """Set search space and model space for transfer learning.
+
+        Overrides the base class to add source-only params (as RangeParameters)
+        to ``_model_space`` while preserving target bounds for shared params.
+        This ensures the SSD naturally covers the full joint feature space
+        without needing post-hoc expansion, and Normalize is anchored to target
+        bounds so target data maps to [0, 1].
+        """
+        self._search_space = search_space.clone()
+        model_space = search_space.clone()
+        self._source_only_params: set[str] = set()
+        for name, param in self.joint_search_space.parameters.items():
+            if name not in model_space.parameters and isinstance(param, RangeParameter):
+                model_space.add_parameter(param.clone())
+                # Only mark as source-only if no backfill value exists.
+                # Backfilled params have known values (via FillMissingParameters)
+                # and should be included in data extraction.
+                if param.backfill_value is None:
+                    self._source_only_params.add(name)
+        self._model_space = model_space
+
     def _transform_data(
         self,
         experiment_data: ExperimentData,
@@ -505,94 +513,18 @@ def _get_task_datasets(
             )
         return task_datasets
 
-    def _expand_ssd_to_joint_space(
+    def _get_target_data_parameters(
         self,
-        search_space_digest: SearchSpaceDigest,
-    ) -> SearchSpaceDigest:
-        """Expand SSD bounds and feature_names to cover the joint search space.
-
-        The SSD produced by ``_get_fit_args`` reflects the target search space.
-        When source experiments have additional parameters, the model operates
-        in the full joint feature space. This method appends bounds and feature
-        names for source-only parameters so that input transforms receive
-        correct full-space bounds.
+        all_params: list[str],
+    ) -> list[str]:
+        """Filter a joint parameter list to target-only params + task feature.
+
+        Source-only params (those added by ``_set_search_space`` from the joint
+        space) are excluded because the target experiment data does not have
+        those columns. Uses untransformed names, which are stable across
+        transforms (Range params are never renamed by OneHot, IntToFloat, etc.).
         """
-        existing_names = set(search_space_digest.feature_names)
-        extra_names: list[str] = []
-        extra_bounds: list[tuple[int | float, int | float]] = []
-        # Only collect parameters absent from the target SSD. Shared
-        # parameters that appear in both target and source keep the target
-        # bounds -- source observations outside those bounds will normalize
-        # outside [0, 1]. This is intentional, as the GP hyperprior is calibrated
-        # for a __target__ task in [0, 1]^D.
-        for name, param in self.joint_search_space.parameters.items():
-            if name not in existing_names and isinstance(param, RangeParameter):
-                extra_names.append(name)
-                extra_bounds.append((param.lower, param.upper))
-        if not extra_names:
-            return search_space_digest
-        # Insert source-only params before the task feature
-        task_features = search_space_digest.task_features
-        if len(task_features) == 1:
-            tf_idx = task_features[0]
-            names = list(search_space_digest.feature_names)
-            bounds = list(search_space_digest.bounds)
-            # Raise if index-based fields (other than the task feature
-            # itself) reference indices at or above tf_idx, since we would
-            # need to shift them when inserting extra params.
-            for field_name in (
-                "ordinal_features",
-                "categorical_features",
-                "fidelity_features",
-            ):
-                indices = getattr(search_space_digest, field_name)
-                if any(i >= tf_idx for i in indices):
-                    raise UnsupportedError(
-                        f"Cannot expand SSD: {field_name} contains index >= {tf_idx}."
-                    )
-            if any(
-                i >= tf_idx and i not in task_features
-                for i in search_space_digest.discrete_choices
-            ):
-                raise UnsupportedError(
-                    f"Cannot expand SSD: discrete_choices contains index >= {tf_idx}."
-                )
-            if search_space_digest.hierarchical_dependencies is not None and any(
-                i >= tf_idx for i in search_space_digest.hierarchical_dependencies
-            ):
-                raise UnsupportedError(
-                    "Cannot expand SSD: hierarchical_dependencies contains "
-                    f"index >= {tf_idx}."
-                )
-            names[tf_idx:tf_idx] = extra_names
-            bounds[tf_idx:tf_idx] = extra_bounds
-            n_extra = len(extra_names)
-            new_task_features = [tf_idx + n_extra]
-            new_target_values = dict(search_space_digest.target_values)
-            if tf_idx in new_target_values:
-                new_target_values[new_task_features[0]] = new_target_values.pop(tf_idx)
-            new_discrete = dict(search_space_digest.discrete_choices)
-            if tf_idx in new_discrete:
-                new_discrete[new_task_features[0]] = new_discrete.pop(tf_idx)
-            return dataclasses.replace(
-                search_space_digest,
-                feature_names=names,
-                bounds=bounds,
-                task_features=new_task_features,
-                target_values=new_target_values,
-                discrete_choices=new_discrete,
-            )
-        elif len(task_features) == 0:
-            # No task feature -- just append.
-            return dataclasses.replace(
-                search_space_digest,
-                feature_names=search_space_digest.feature_names + extra_names,
-                bounds=search_space_digest.bounds + extra_bounds,
-            )
-        else:
-            raise UnsupportedError(
-                "Multiple task features are not supported in transfer learning."
-            )
+        return [p for p in all_params if p not in self._source_only_params]
 
     def _fit(
         self,
@@ -610,15 +542,20 @@ def _fit(
         if experiment_data.arm_data.empty:
             # Temporarily unset self.outcomes to avoid an error in _get_fit_args.
             self.outcomes = []
+        # Pre-compute the joint param ordering (mirrors _get_fit_args logic)
+        # so we can derive the target-only subset for data extraction.
+        all_params = list(search_space.parameters.keys())
+        task_name = Keys.TASK_FEATURE_NAME.value
+        if task_name in all_params:
+            all_params.remove(task_name)
+            all_params.append(task_name)
+        target_data_params = self._get_target_data_parameters(all_params)
         datasets, candidate_metadata, search_space_digest = self._get_fit_args(
             search_space=search_space,
             experiment_data=experiment_data,
             update_outcomes_and_parameters=True,
+            data_parameters=target_data_params,
         )
-        # Expand SSD bounds to cover source-only params from the joint search
-        # space. This ensures Normalize (and other input transforms) get bounds
-        # for the full feature space, not just the target dims.
-        search_space_digest = self._expand_ssd_to_joint_space(search_space_digest)
         if experiment_data.arm_data.empty:
             self.outcomes = outcomes
             # Temporarily set datasets to None. We will construct empty datasets
@@ -656,12 +593,13 @@ def _cross_validate(
     ) -> list[ObservationData]:
         if self.parameters is None:
             raise ValueError(FIT_MODEL_ERROR.format(action="_cross_validate"))
+        target_data_params = self._get_target_data_parameters(self.parameters)
         datasets, _, search_space_digest = self._get_fit_args(
             search_space=search_space,
             experiment_data=cv_training_data,
             update_outcomes_and_parameters=False,
+            data_parameters=target_data_params,
         )
-        search_space_digest = self._expand_ssd_to_joint_space(search_space_digest)
         # Add the task feature to SSD, to ensure that a multi-task model is selected.
         if len(search_space_digest.task_features) > 1:
             raise UnsupportedError(
@@ -714,23 +652,22 @@ def gen(
         to ``RemoveFixed.untransform_observation_features``, which requires updating
         the signature of all transforms.
         """
-        # If a fixed parameter in the target search space, is
-        # a range parameter in the joint search space, then we
-        # should set it as a fixed feature here.
+        # If a fixed parameter in the target search space is a range
+        # parameter in the joint search space, pin it as a fixed feature.
         search_space = search_space or self._search_space
         for name, target_p in search_space.parameters.items():
             if (
                 isinstance(target_p, FixedParameter)
                 and (p := self.joint_search_space.parameters.get(name)) is not None
                 and isinstance(p, RangeParameter)
             ):
-                # add to fixed features
                 if fixed_features is None:
                     fixed_features = ObservationFeatures(parameters={})
                 fixed_features.parameters.setdefault(name, target_p.value)
-        # Fix source-only params so the optimizer doesn't search over them.
-        # Center is a reasonable default; LearnedFeatureImputation overwrites
-        # these with learned values when configured.
+        # Fix source-only params that ARE in the search space (e.g. injected
+        # as FixedParam with a backfill value) so the optimizer doesn't search
+        # over them. Params NOT in the search space are handled by the model
+        # internally (HeterogeneousMTGP natively, LFI for MultiTaskGP).
         joint_center = self.joint_search_space.compute_naive_center()
         for name, param in self.joint_search_space.parameters.items():
             if name not in search_space.parameters and isinstance(
@@ -739,6 +676,20 @@ def gen(
                 if fixed_features is None:
                     fixed_features = ObservationFeatures(parameters={})
                 fixed_features.parameters.setdefault(name, joint_center[name])
+        # At gen time, optimize over the target search space only.
+        # _source_only_params covers params without backfill (stable
+        # untransformed names). Backfilled source-only params are identified
+        # by checking joint SS backfill_values — their names are also stable
+        # (RangeParameters are never renamed by transforms).
+        saved_parameters = self.parameters
+        backfilled_source_only = {
+            name
+            for name, param in self.joint_search_space.parameters.items()
+            if name not in self._experiment.search_space.parameters
+            and param.backfill_value is not None
+        }
+        exclude = self._source_only_params | backfilled_source_only
+        self.parameters = [p for p in self.parameters if p not in exclude]
         generator_run = super().gen(
             n=n,
             search_space=search_space,
@@ -747,6 +698,7 @@ def gen(
             fixed_features=fixed_features,
             model_gen_options=model_gen_options,
         )
+        self.parameters = saved_parameters
         # Remove the parameters that are not in the target experiment's search
         # space, and update candidate_metadata_by_arm_signature to reflect the
         # new arm. We use the experiment's search space rather than