introduce trial_indices argument to SupervisedDataset (facebook#2960)

Summary:

X-link: pytorch/botorch#2595

Adds optional `trial_indices` to SupervisedDataset, whose dimensionality should correspond 1:1 with the first few dimensions of X and Y tensors, as validated in `_validate` ([pointer](https://www.internalfb.com/diff/D64764019?permalink=1739375523489084)).

Reviewed By: Balandat

Differential Revision: D64764019

ax/core/observation.py

@@ -415,6 +415,15 @@ def _filter_data_on_status(
 
 
 def get_feature_cols(data: Data, is_map_data: bool = False) -> list[str]:
+    """Get the columns used to identify and group observations from a Data object.
+
+    Args:
+        data: the Data object from which to extract the feature columns.
+        is_map_data: If True, the Data object's map_keys will be included.
+
+    Returns:
+        A list of column names to be used to group observations.
+    """
     feature_cols = OBS_COLS.intersection(data.df.columns)
     # note we use this check, rather than isinstance, since
     # only some Modelbridges (e.g. MapTorchModelBridge)
@@ -432,8 +441,10 @@ def get_feature_cols(data: Data, is_map_data: bool = False) -> list[str]:
                 stacklevel=5,
             )
             feature_cols.discard(column)
-
-    return list(feature_cols)
+    # NOTE: This ensures the order of feature_cols is deterministic so that the order
+    # of lists of observations are deterministic, to avoid nondeterministic tests.
+    # Necessary for test_TorchModelBridge.
+    return sorted(feature_cols)
 
 
 def observations_from_data(

ax/modelbridge/map_torch.py

@@ -11,6 +11,7 @@
 
 import torch
 from ax.core.base_trial import TrialStatus
+from ax.core.batch_trial import BatchTrial
 from ax.core.data import Data
 from ax.core.experiment import Experiment
 from ax.core.map_data import MapData
@@ -118,6 +119,9 @@ def __init__(
             raise ValueError(
                 "`MapTorchModelBridge expects `MapData` instead of `Data`."
             )
+
+        if any(isinstance(t, BatchTrial) for t in experiment.trials.values()):
+            raise ValueError("MapTorchModelBridge does not support batch trials.")
         # pyre-fixme[4]: Attribute must be annotated.
         self._map_key_features = data.map_keys
         self._map_data_limit_rows_per_metric = map_data_limit_rows_per_metric
@@ -146,6 +150,12 @@ def statuses_to_fit_map_metric(self) -> set[TrialStatus]:
 
     @property
     def parameters_with_map_keys(self) -> list[str]:
+        """The parameters used for fitting the model, including map_keys."""
+        # NOTE: This list determines the order of feature columns in the training data.
+        # Learning-curve-based modeling methods assume that the last columns are
+        # map_keys, so we place self._map_key_features on the end.
+        # TODO: Plumb down the `map_key` feature indices to the model, so that we don't
+        # have to make the assumption in the above note.
         return self.parameters + self._map_key_features
 
     def _predict(

ax/modelbridge/pairwise.py

@@ -48,6 +48,7 @@ def _convert_observations(
             Yvars,
             candidate_metadata_dict,
             any_candidate_metadata_is_not_none,
+            trial_indices,
         ) = self._extract_observation_data(
             observation_data, observation_features, parameters
         )
@@ -69,6 +70,9 @@ def _convert_observations(
                     Y=Y,
                     feature_names=parameters,
                     outcome_names=[outcome],
+                    trial_indices=torch.tensor(trial_indices[outcome])
+                    if trial_indices
+                    else None,
                 )
 
             datasets.append(dataset)

ax/modelbridge/tests/test_map_torch_modelbridge.py

@@ -43,15 +43,24 @@ def test_TorchModelBridge(self) -> None:
             experiment.trials[i].mark_as(status=TrialStatus.COMPLETED)
 
         experiment.attach_data(data=experiment.fetch_data())
+        model = mock.MagicMock(TorchModel, autospec=True, instance=True)
         modelbridge = MapTorchModelBridge(
             experiment=experiment,
             search_space=experiment.search_space,
             data=experiment.lookup_data(),
-            model=TorchModel(),
+            model=model,
             transforms=[],
             fit_out_of_design=True,
             default_model_gen_options={"target_map_values": {"timestamp": 4.0}},
         )
+        # Check that indices are set correctly.
+        datasets_arg = model.fit.mock_calls[0][2]["datasets"]
+        t1 = datasets_arg[0].trial_indices
+        t2 = torch.tensor([0, 1, 2])
+        self.assertTrue(torch.equal(t1, t2), msg=f"{t1} != {t2}")
+        t1 = datasets_arg[1].trial_indices
+        t2 = torch.tensor([0, 0, 1, 1, 2, 2])
+        self.assertTrue(torch.equal(t1, t2), msg=f"{t1} != {t2}")
         # Check map data is converted to observations, that we get one Observation
         # per row of MapData
         # pyre-fixme[16]: `Data` has no attribute `map_df`.

ax/modelbridge/torch.py

@@ -343,6 +343,7 @@ def _convert_observations(
             Yvars,
             candidate_metadata_dict,
             any_candidate_metadata_is_not_none,
+            trial_indices,
         ) = self._extract_observation_data(
             observation_data, observation_features, parameters
         )
@@ -369,6 +370,9 @@ def _convert_observations(
                 Yvar=Yvar,
                 feature_names=parameters,
                 outcome_names=[outcome],
+                trial_indices=torch.tensor(trial_indices[outcome])
+                if trial_indices
+                else None,
             )
             datasets.append(dataset)
             candidate_metadata.append(candidate_metadata_dict[outcome])
@@ -656,7 +660,7 @@ def _fit(
         )
         # Fit
         self.model = model
-        self.model.fit(
+        not_none(self.model).fit(
             datasets=datasets,
             search_space_digest=search_space_digest,
             candidate_metadata=candidate_metadata,
@@ -970,6 +974,7 @@ def _extract_observation_data(
         dict[str, list[Tensor]],
         dict[str, list[TCandidateMetadata]],
         bool,
+        dict[str, list[int]] | None,
     ]:
         """Extract observation features & data into tensors and metadata.
 
@@ -991,14 +996,19 @@ def _extract_observation_data(
                 observation tensors `Yvar`.
             - A dictionary mapping metric names to lists of corresponding metadata.
             - A boolean denoting whether any candidate metadata is not none.
+            - A dictionary mapping metric names to lists of corresponding trial indices,
+                or None if trial indices are not provided. This is used to support
+                learning-curve-based modeling.
         """
         Xs: dict[str, list[Tensor]] = defaultdict(list)
         Ys: dict[str, list[Tensor]] = defaultdict(list)
         Yvars: dict[str, list[Tensor]] = defaultdict(list)
         candidate_metadata_dict: dict[str, list[TCandidateMetadata]] = defaultdict(list)
         any_candidate_metadata_is_not_none = False
+        trial_indices: dict[str, list[int]] = defaultdict(list)
 
-        for obsd, obsf in zip(observation_data, observation_features):
+        at_least_one_trial_index_is_none = False
+        for obsd, obsf in zip(observation_data, observation_features, strict=True):
             try:
                 x = torch.tensor(
                     [obsf.parameters[p] for p in parameters],
@@ -1016,13 +1026,24 @@ def _extract_observation_data(
                 if obsf.metadata is not None:
                     any_candidate_metadata_is_not_none = True
                 candidate_metadata_dict[metric_name].append(obsf.metadata)
+                trial_index = obsf.trial_index
+                if trial_index is not None:
+                    trial_indices[metric_name].append(trial_index)
+                else:
+                    at_least_one_trial_index_is_none = True
+        if len(trial_indices) > 0 and at_least_one_trial_index_is_none:
+            raise ValueError(
+                "Trial indices must be provided for all observation_features "
+                "or none of them."
+            )
 
         return (
             Xs,
             Ys,
             Yvars,
             candidate_metadata_dict,
             any_candidate_metadata_is_not_none,
+            trial_indices if len(trial_indices) > 0 else None,
         )