Optimize memory usage for TSDataset._merge_exog

CHANGELOG.md

@@ -45,7 +45,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 - **Breaking:** Bump minimum `scipy` version to 1.12 ([#599](https://github.com/etna-team/etna/pull/599))
 - **Breaking:** Bump minimum `optuna` version to 4.0 ([#599](https://github.com/etna-team/etna/pull/599))
 - **Breaking:** Bump minimum `statsforecast` version to 2.0 ([#599](https://github.com/etna-team/etna/pull/599))
--
+- Optimize performance of exogenous variables addition to the dataset ([#596](https://github.com/etna-team/etna/pull/596))
 -
 
 ### Fixed

etna/datasets/tsdataset.py

@@ -11,6 +11,7 @@
 from typing import List
 from typing import Optional
 from typing import Sequence
+from typing import Set
 from typing import Tuple
 from typing import Union
 
@@ -141,15 +142,15 @@ def __init__(
         """
         self.freq = freq
         self.df_exog = None
-        self.raw_df = self._prepare_df(df=df.copy(deep=True), freq=freq)
-        self.df = self.raw_df.copy(deep=True)
+        self.raw_df = self._prepare_df(df=df, freq=freq)
+        self.df = self.raw_df.copy(deep=None)
 
         self.hierarchical_structure = hierarchical_structure
         self.current_df_level: Optional[str] = self._get_dataframe_level(df=self.df)
         self.current_df_exog_level: Optional[str] = None
 
         if df_exog is not None:
-            self.df_exog = self._prepare_df_exog(df_exog=df_exog.copy(deep=True), freq=freq)
+            self.df_exog = self._prepare_df_exog(df_exog=df_exog, freq=freq)
 
             self.known_future = self._check_known_future(known_future, self.df_exog)
             self._regressors = copy(self.known_future)
@@ -244,15 +245,15 @@ def create_from_misaligned(
             freq=freq,
             timestamp_name=original_timestamp_name,
         )
-        timestamp_df = TSDataset.to_dataset(timestamp_df)
 
         if df_exog is not None:
             df_exog_realigned = apply_alignment(df=df_exog, alignment=alignment)
+
+            df_exog_realigned = pd.merge(df_exog_realigned, timestamp_df, how="outer", on=["timestamp", "segment"])
             df_exog_realigned = TSDataset.to_dataset(df_exog_realigned)
 
-            df_exog_realigned = df_exog_realigned.join(timestamp_df, how="outer")
         else:
-            df_exog_realigned = timestamp_df
+            df_exog_realigned = TSDataset.to_dataset(timestamp_df)
 
         known_future_realigned: Union[Literal["all"], Sequence]
         if known_future != "all":
@@ -328,6 +329,9 @@ def _prepare_df(cls, df: pd.DataFrame, freq: Optional[str]) -> pd.DataFrame:
         if df_format is DataFrameFormat.long:
             df = cls.to_dataset(df)
 
+        else:
+            df = df.copy(deep=True)
+
         # cast segment to str type
         cls._cast_segment_to_str(df)
 
@@ -338,7 +342,7 @@ def _prepare_df(cls, df: pd.DataFrame, freq: Optional[str]) -> pd.DataFrame:
 
             new_index = np.arange(df.index.min(), df.index.max() + 1)
             index_name = df.index.name
-            df = df.reindex(new_index)
+            df = df.reindex(new_index, copy=False)
             df.index.name = index_name
 
         else:
@@ -354,7 +358,9 @@ def _prepare_df(cls, df: pd.DataFrame, freq: Optional[str]) -> pd.DataFrame:
                     f"You probably set wrong freq. Discovered freq in you data is {inferred_freq}, you set {freq}"
                 )
 
-            df = df.asfreq(freq)
+            new_index = pd.date_range(start=df.index.min(), end=df.index.max(), freq=freq)
+            new_index.name = df.index.name  # type: ignore
+            df = df.reindex(new_index, copy=False)
 
         return df
 
@@ -364,6 +370,9 @@ def _prepare_df_exog(cls, df_exog: pd.DataFrame, freq: Optional[str]) -> pd.Data
         if df_format is DataFrameFormat.long:
             df_exog = cls.to_dataset(df_exog)
 
+        else:
+            df_exog = df_exog.copy(deep=True)
+
         df_exog = cls._cast_segment_to_str(df=df_exog)
         if freq is not None:
             cls._cast_index_to_datetime(df_exog, freq)
@@ -548,44 +557,73 @@ def _check_known_future(
                 return sorted(known_future_unique)
 
     @staticmethod
-    def _check_regressors(df: pd.DataFrame, df_regressors: pd.DataFrame):
+    def _get_min_max_valid_timestamp(
+        df: pd.DataFrame, segments: Set[str], regressors: Optional[Sequence[str]] = None
+    ) -> Tuple[Sequence[pd.Timestamp], Sequence[pd.Timestamp]]:
+        """Estimate first and last valid indices for the dataframe."""
+        # shape: (num_samples, num_segments, num_features)
+        df_values = df.values.reshape((len(df), len(segments), -1))
+
+        if regressors is not None:
+            # expected equal features for all segments and sorted column index
+            features = df.columns.get_level_values("feature")
+            segment_features = features[: len(features) // len(segments)]
+            regressors_mask = segment_features.isin(set(regressors))
+            # shape: (num_samples, num_segments, num_regressors)
+            df_values = df_values[..., regressors_mask]
+
+        # shape: (num_samples, num_segments)
+        df_mask = ~np.any(pd.isna(df_values), axis=-1)
+
+        # shape: (num_segments,)
+        min_ids = np.argmax(df_mask, axis=0)
+        max_ids = len(df_mask) - np.argmax(df_mask[::-1], axis=0) - 1
+
+        min_index = df.index.values[min_ids]
+        max_index = df.index.values[max_ids]
+
+        none_segments = ~np.any(df_mask, axis=0)
+        min_index[none_segments] = np.datetime64("NaT")
+        max_index[none_segments] = np.datetime64("NaT")
+
+        return min_index, max_index
+
+    def _check_regressors(self, df: pd.DataFrame):
         """Check that regressors begin not later than in ``df`` and end later than in ``df``."""
-        if df_regressors.shape[1] == 0:
+        if len(self.known_future) == 0:
             return
-        # TODO: check performance
-        df_segments = df.columns.get_level_values("segment")
-        for segment in df_segments:
-            target_min = df[segment]["target"].first_valid_index()
-            target_min = pd.NaT if target_min is None else target_min
-            target_max = df[segment]["target"].last_valid_index()
-            target_max = pd.NaT if target_max is None else target_max
-
-            exog_series_min = df_regressors[segment].first_valid_index()
-            exog_series_min = pd.NaT if exog_series_min is None else exog_series_min
-            exog_series_max = df_regressors[segment].last_valid_index()
-            exog_series_max = pd.NaT if exog_series_max is None else exog_series_max
-            if target_min < exog_series_min:
+
+        segments = set(df.columns.get_level_values("segment"))
+
+        target_min, target_max = self._get_min_max_valid_timestamp(df=df, segments=segments)
+        exog_series_min, exog_series_max = self._get_min_max_valid_timestamp(
+            df=self.df_exog, segments=segments, regressors=self.known_future
+        )
+
+        for i, segment in enumerate(segments):
+            if target_min[i] < exog_series_min[i]:
                 raise ValueError(
                     f"All the regressor series should start not later than corresponding 'target'."
                     f"Series of segment {segment} have not enough history: "
-                    f"{target_min} < {exog_series_min}."
+                    f"{target_min[i]} < {exog_series_min[i]}."
                 )
-            if target_max >= exog_series_max:
+            if target_max[i] >= exog_series_max[i]:
                 raise ValueError(
                     f"All the regressor series should finish later than corresponding 'target'."
                     f"Series of segment {segment} have not enough history: "
-                    f"{target_max} >= {exog_series_max}."
+                    f"{target_max[i]} >= {exog_series_max[i]}."
                 )
 
     def _merge_exog(self, df: pd.DataFrame) -> pd.DataFrame:
         if self.df_exog is None:
             raise ValueError("Something went wrong, Trying to merge df_exog which is None!")
 
         # TODO: this check could probably be skipped at make_future
-        df_regressors = self.df_exog.loc[:, pd.IndexSlice[:, self.known_future]]
-        self._check_regressors(df=df, df_regressors=df_regressors)
+        self._check_regressors(df=df)
+
+        df = df.merge(self.df_exog, how="left", left_index=True, right_index=True)
 
-        df = pd.concat((df, self.df_exog), axis=1).loc[df.index].sort_index(axis=1, level=(0, 1))
+        df.sort_index(axis=1, level=(0, 1), inplace=True)
 
         _check_features_in_segments(columns=df.columns)
 
@@ -964,22 +1002,23 @@ def to_dataset(df: pd.DataFrame) -> pd.DataFrame:
         2021-01-04           3           8
         2021-01-05           4           9
         """
-        df_copy = df.copy(deep=True)
+        df = df.set_index(["timestamp", "segment"])
 
-        if not pd.api.types.is_integer_dtype(df_copy["timestamp"]):
-            df_copy["timestamp"] = pd.to_datetime(df_copy["timestamp"])
+        df = df.unstack(level=-1)
+        if not pd.api.types.is_integer_dtype(df.index):
+            df.index = pd.to_datetime(df.index)
 
-        df_copy["segment"] = df_copy["segment"].astype(str)
+        if not pd.api.types.is_string_dtype(df.columns.levels[1]):
+            df.columns = df.columns.set_levels(df.columns.levels[1].astype(str), level=1)
 
-        feature_columns = df_copy.columns.tolist()
-        feature_columns.remove("timestamp")
-        feature_columns.remove("segment")
+        df.columns = df.columns.reorder_levels([1, 0])
+        df.columns.names = ["segment", "feature"]
+        df.sort_index(axis=1, level=(0, 1), inplace=True)
 
-        df_copy = df_copy.pivot(index="timestamp", columns="segment")
-        df_copy = df_copy.reorder_levels([1, 0], axis=1)
-        df_copy.columns.names = ["segment", "feature"]
-        df_copy = df_copy.sort_index(axis=1, level=(0, 1))
-        return df_copy
+        if df._is_view or df._is_copy is None:
+            df = df.copy(deep=None)
+
+        return df
 
     @staticmethod
     def _hierarchical_structure_from_level_columns(

tests/test_datasets/test_dataset.py

@@ -1213,19 +1213,20 @@ def test_check_regressors_error(exog_starts_later: bool, exog_ends_earlier: bool
     df_regressors = TSDataset.to_dataset(df_regressors)
 
     with pytest.raises(ValueError):
-        TSDataset._check_regressors(df=df, df_regressors=df_regressors)
+        _ = TSDataset(df=df, df_exog=df_regressors, known_future="all", freq="D")
 
 
 def test_check_regressors_pass(df_and_regressors):
     """Check that regressors check on creation passes with correct regressors."""
     df, df_exog, _ = df_and_regressors
-    _ = TSDataset._check_regressors(df=df, df_regressors=df_exog)
+    _ = TSDataset(df=df, df_exog=df_exog, known_future="all", freq="D")
 
 
 def test_check_regressors_pass_empty(df_and_regressors):
     """Check that regressors check on creation passes with no regressors."""
     df, _, _ = df_and_regressors
-    _ = TSDataset._check_regressors(df=df, df_regressors=pd.DataFrame())
+    df_exog = pd.DataFrame(columns=["timestamp", "segment", "exog"])
+    _ = TSDataset(df=df, df_exog=df_exog, known_future="all", freq="D")
 
 
 def test_getitem_only_date(tsdf_with_exog):
@@ -1968,15 +1969,17 @@ def test_create_from_misaligned_with_exog(
     expected_raw_df = TSDataset.to_dataset(apply_alignment(df=df, alignment=alignment))
     pd.testing.assert_frame_equal(ts.raw_df, expected_raw_df)
 
-    expected_df_exog = TSDataset.to_dataset(apply_alignment(df=df_exog, alignment=alignment))
     timestamp_df = make_timestamp_df_from_alignment(
         alignment=alignment,
         start=expected_raw_df.index[0],
         periods=len(expected_raw_df) + future_steps,
         freq=freq,
         timestamp_name=original_timestamp_name,
     )
-    expected_df_exog = expected_df_exog.join(TSDataset.to_dataset(timestamp_df), how="outer")
+    expected_df_exog = apply_alignment(df=df_exog, alignment=alignment)
+    expected_df_exog = pd.merge(expected_df_exog, timestamp_df, how="outer", on=["timestamp", "segment"])
+    expected_df_exog = TSDataset.to_dataset(expected_df_exog)
+
     pd.testing.assert_frame_equal(ts.df_exog, expected_df_exog)
 
     expected_known_future = sorted(set(known_future).union([original_timestamp_name]))
@@ -2016,15 +2019,18 @@ def test_create_from_misaligned_with_exog_all(
     expected_raw_df = TSDataset.to_dataset(apply_alignment(df=df, alignment=alignment))
     pd.testing.assert_frame_equal(ts.raw_df, expected_raw_df)
 
-    expected_df_exog = TSDataset.to_dataset(apply_alignment(df=df_exog, alignment=alignment))
     timestamp_df = make_timestamp_df_from_alignment(
         alignment=alignment,
         start=expected_raw_df.index[0],
         periods=len(expected_raw_df) + future_steps,
         freq=freq,
         timestamp_name=original_timestamp_name,
     )
-    expected_df_exog = expected_df_exog.join(TSDataset.to_dataset(timestamp_df), how="outer")
+
+    expected_df_exog = apply_alignment(df=df_exog, alignment=alignment)
+    expected_df_exog = pd.merge(expected_df_exog, timestamp_df, how="outer", on=["timestamp", "segment"])
+
+    expected_df_exog = TSDataset.to_dataset(expected_df_exog)
     pd.testing.assert_frame_equal(ts.df_exog, expected_df_exog)
 
     assert ts.known_future == expected_known_future

tests/test_transforms/test_math/test_exog_shift_transform.py

@@ -183,9 +183,9 @@ def test_transformed_names(ts_name, lag, horizon, expected, request):
 @pytest.mark.parametrize(
     "lag,horizon,expected_types",
     (
-        (1, None, {"feat1_shift_1": "float", "feat2_shift_1": "float", "feat3_shift_1": "float", "target": "float"}),
-        ("auto", 1, {"feat1_shift_1": "float", "feat2_shift_2": "float", "feat3": "int", "target": "float"}),
-        ("auto", 2, {"feat1_shift_2": "float", "feat2_shift_3": "float", "feat3_shift_1": "float", "target": "float"}),
+        (1, None, {"feat1_shift_1": "float", "feat2_shift_1": "float", "feat3_shift_1": "float", "target": "int"}),
+        ("auto", 1, {"feat1_shift_1": "float", "feat2_shift_2": "float", "feat3": "int", "target": "int"}),
+        ("auto", 2, {"feat1_shift_2": "float", "feat2_shift_3": "float", "feat3_shift_1": "float", "target": "int"}),
     ),
 )
 @pytest.mark.parametrize(