Teach DifferencingTransform to inverse_transform with NaNs

CHANGELOG.md

@@ -33,6 +33,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 - Fix `SklearnTransform` in per-segment mode to work on subset of segments and raise error on new segments ([#1107](https://github.com/tinkoff-ai/etna/pull/1107))
 - Fix `OutliersTransform` and its children to raise error on new segments ([#1139](https://github.com/tinkoff-ai/etna/pull/1139))
 - Fix `DifferencingTransform` to raise error on new segments during `transform` and `inverse_transform` in inplace mode ([#1141](https://github.com/tinkoff-ai/etna/pull/1141))
+- Teach `DifferencingTransform` to `inverse_transform` with NaNs ([#1155](https://github.com/tinkoff-ai/etna/pull/1155))
 ## [1.14.0] - 2022-12-16
 ### Added
 - Add python 3.10 support ([#1005](https://github.com/tinkoff-ai/etna/pull/1005))

etna/transforms/math/differencing.py

@@ -16,7 +16,8 @@
 class _SingleDifferencingTransform(Transform):
     """Calculate a time series differences of order 1.
 
-    This transform can work with NaNs at the beginning of the segment, but fails when meets NaN inside the segment.
+    During ``fit`` this transform can work with NaNs at the beginning of the segment, but fails when meets NaN inside the segment.
+    During ``transform`` and ``inverse_transform`` there is no special treatment of NaNs.
 
     Notes
     -----
@@ -86,6 +87,11 @@ def fit(self, df: pd.DataFrame) -> "_SingleDifferencingTransform":
         Returns
         -------
         result: _SingleDifferencingTransform
+
+        Raises
+        ------
+        ValueError:
+            if NaNs are present inside the segment
         """
         segments = sorted(set(df.columns.get_level_values("segment")))
         fit_df = df.loc[:, pd.IndexSlice[segments, self.in_column]].copy()
@@ -124,11 +130,8 @@ def transform(self, df: pd.DataFrame) -> pd.DataFrame:
         transformed = df.loc[:, pd.IndexSlice[segments, self.in_column]].copy()
         for current_segment in segments:
             to_transform = transformed.loc[:, pd.IndexSlice[current_segment, self.in_column]]
-            start_idx = to_transform.first_valid_index()
             # make a differentiation
-            transformed.loc[start_idx:, pd.IndexSlice[current_segment, self.in_column]] = to_transform.loc[
-                start_idx:
-            ].diff(periods=self.period)
+            transformed.loc[:, pd.IndexSlice[current_segment, self.in_column]] = to_transform.diff(periods=self.period)
 
         if self.inplace:
             result_df = df.copy()
@@ -188,10 +191,6 @@ def _reconstruct_test(self, df: pd.DataFrame, columns_to_inverse: Set[str]) -> p
             init_df = init_df[segments]
             to_transform = pd.concat([init_df, to_transform])
 
-            # validate values inside the series to transform
-            if to_transform.isna().sum().sum() > 0:
-                raise ValueError(f"There should be no NaNs inside the segments")
-
             # run reconstruction and save the result
             to_transform = self._make_inv_diff(to_transform)
             result_df.loc[:, pd.IndexSlice[segments, column]] = to_transform
@@ -210,6 +209,13 @@ def inverse_transform(self, df: pd.DataFrame) -> pd.DataFrame:
         -------
         result:
             transformed DataFrame.
+
+        Raises
+        ------
+        ValueError:
+            if inverse transform is applied not to full train nor to test that goes after train
+        ValueError:
+            if inverse transform is applied to test that goes after train with gap
         """
         # we assume this to be fitted
         self._train_timestamp = cast(pd.DatetimeIndex, self._train_timestamp)
@@ -241,7 +247,8 @@ def inverse_transform(self, df: pd.DataFrame) -> pd.DataFrame:
 class DifferencingTransform(Transform):
     """Calculate a time series differences.
 
-    This transform can work with NaNs at the beginning of the segment, but fails when meets NaN inside the segment.
+    During ``fit`` this transform can work with NaNs at the beginning of the segment, but fails when meets NaN inside the segment.
+    During ``transform`` and ``inverse_transform`` there is no special treatment of NaNs.
 
     Notes
     -----
@@ -334,6 +341,11 @@ def fit(self, df: pd.DataFrame) -> "DifferencingTransform":
         Returns
         -------
         result: DifferencingTransform
+
+        Raises
+        ------
+        ValueError:
+            if NaNs are present inside the segment
         """
         # this is made because transforms of high order may need some columns created by transforms of lower order
         result_df = df.copy()
@@ -395,6 +407,10 @@ def inverse_transform(self, df: pd.DataFrame) -> pd.DataFrame:
             if transform isn't fitted
         NotImplementedError:
             if there are segments that weren't present during training
+        ValueError:
+            if inverse transform is applied not to full train nor to test that goes after train
+        ValueError:
+            if inverse transform is applied to test that goes after train with gap
         """
         self._check_is_fitted()
         if not self.inplace:

tests/test_transforms/test_inference/test_inverse_transform.py

@@ -311,6 +311,7 @@ def _test_inverse_transform_future_subset_segments(self, ts, transform, segments
             (LogTransform(in_column="target", inplace=True), "positive_ts"),
             (LogTransform(in_column="positive", inplace=True), "ts_with_exog"),
             (DifferencingTransform(in_column="target", inplace=False), "regular_ts"),
+            (DifferencingTransform(in_column="target", inplace=True), "regular_ts"),
             (DifferencingTransform(in_column="positive", inplace=True), "ts_with_exog"),
             (MADTransform(in_column="target", window=14), "regular_ts"),
             (MaxTransform(in_column="target", window=14), "regular_ts"),
@@ -389,17 +390,6 @@ def test_inverse_transform_future_subset_segments(self, transform, dataset_name,
         ts = request.getfixturevalue(dataset_name)
         self._test_inverse_transform_future_subset_segments(ts, transform, segments=["segment_2"])
 
-    @to_be_fixed(ValueError, match="There should be no NaNs inside the segments")
-    @pytest.mark.parametrize(
-        "transform, dataset_name",
-        [
-            (DifferencingTransform(in_column="target", inplace=True), "regular_ts"),
-        ],
-    )
-    def test_inverse_transform_future_subset_difference_fail(self, transform, dataset_name, request):
-        ts = request.getfixturevalue(dataset_name)
-        self._test_inverse_transform_future_subset_segments(ts, transform, segments=["segment_2"])
-
 
 class TestInverseTransformTrainNewSegments:
     """Test inverse transform on train part of new segments.

tests/test_transforms/test_math/test_differencing_transform.py

@@ -40,6 +40,18 @@ def df_nans() -> pd.DataFrame:
     return df
 
 
+@pytest.fixture
+def df_nans_middle() -> pd.DataFrame:
+    """Create DataFrame with nans in the middle of the segment."""
+    timestamp = pd.date_range("2021-01-01", "2021-04-01")
+    df_1 = pd.DataFrame({"timestamp": timestamp, "target": np.arange(timestamp.shape[0]), "segment": "1"})
+    df_2 = pd.DataFrame({"timestamp": timestamp, "target": np.arange(timestamp.shape[0]) * 2, "segment": "2"})
+    df = pd.concat([df_1, df_2], ignore_index=True)
+    df = TSDataset.to_dataset(df)
+    df.iloc[5:10, 0] = np.NaN
+    return df
+
+
 @pytest.fixture
 def df_segments_split(df_nans) -> Tuple[pd.DataFrame, pd.DataFrame]:
     """Create a pair of DataFrames with different segments."""
@@ -106,15 +118,10 @@ def check_transform(
         series_init = df.loc[:, pd.IndexSlice[segment, "target"]]
         series_transformed = transformed_df.loc[:, pd.IndexSlice[segment, out_column]]
 
-        series_init = series_init.loc[series_init.first_valid_index() :]
-        series_transformed = series_transformed.loc[series_transformed.first_valid_index() :]
-
-        assert series_init.shape[0] == series_transformed.shape[0] + order * period
-
         for _ in range(order):
-            series_init = series_init.diff(periods=period).iloc[period:]
+            series_init = series_init.diff(periods=period)
 
-        assert np.all(series_init == series_transformed)
+        assert series_init.equals(series_transformed)
 
 
 def check_inverse_transform_not_inplace(
@@ -134,10 +141,10 @@ def check_inverse_transform_inplace_train(transform: GeneralDifferencingTransfor
     assert inverse_transformed_df.equals(df)
 
 
-def check_inverse_transform_inplace_test(
+def check_inverse_transform_inplace_filled_test(
     transform: GeneralDifferencingTransform, period: int, order: int, df: pd.DataFrame
 ):
-    """Check that differencing transform correctly makes inverse_transform on test data in inplace mode."""
+    """Check that differencing transform correctly makes inverse_transform on filled test data in inplace mode."""
     ts = TSDataset(df, freq="D")
     ts_train, ts_test = ts.train_test_split(test_size=20)
     ts_train.fit_transform(transforms=[transform])
@@ -158,6 +165,19 @@ def check_inverse_transform_inplace_test(
     assert np.all(future_ts.to_pandas() == ts_test.to_pandas())
 
 
+def check_inverse_transform_inplace_unfilled_test(transform: GeneralDifferencingTransform, df: pd.DataFrame):
+    """Check that differencing transform correctly makes inverse_transform on unfilled test data in inplace mode."""
+    ts = TSDataset(df, freq="D")
+    ts_train, ts_test = ts.train_test_split(test_size=20)
+    ts_train.fit_transform(transforms=[transform])
+
+    future_ts = ts_train.make_future(20)
+
+    # check values from inverse_transform
+    future_ts.inverse_transform()
+    assert future_ts.to_pandas().isna().all().all()
+
+
 def check_inverse_transform_inplace_test_quantiles(transform: GeneralDifferencingTransform, df: pd.DataFrame):
     """Check that differencing transform correctly makes inverse_transform on test data with quantiles."""
     ts = TSDataset(df, freq="D")
@@ -345,6 +365,25 @@ def test_full_transform(period, order, inplace, out_column, df_nans):
     check_transform(transform, period, order, out_column, df_nans, df_nans)
 
 
+@pytest.mark.parametrize("period", [1, 7])
+@pytest.mark.parametrize("inplace, out_column", [(False, "diff"), (True, "target")])
+def test_single_transform_nans_middle(period, inplace, out_column, df_nans, df_nans_middle):
+    """Test that _SingleDifferencingTransform generates correct values in transform with NaNs in the middle."""
+    transform = _SingleDifferencingTransform(in_column="target", period=period, inplace=inplace, out_column=out_column)
+    check_transform(transform, period, 1, out_column, df_nans, df_nans_middle)
+
+
+@pytest.mark.parametrize("period", [1, 7])
+@pytest.mark.parametrize("order", [1, 2])
+@pytest.mark.parametrize("inplace, out_column", [(False, "diff"), (True, "target")])
+def test_full_transform_nans_middle(period, order, inplace, out_column, df_nans, df_nans_middle):
+    """Test that DifferencingTransform generates correct values in transform with NaNs in the middle."""
+    transform = DifferencingTransform(
+        in_column="target", period=period, order=order, inplace=inplace, out_column=out_column
+    )
+    check_transform(transform, period, order, out_column, df_nans, df_nans_middle)
+
+
 @pytest.mark.parametrize("period", [1, 7])
 def test_single_transform_not_inplace_new_segments(period, df_segments_split):
     """Test that _SingleDifferencingTransform generates correct values in transform on new segments in non-inplace mode."""
@@ -466,43 +505,34 @@ def test_full_inverse_transform_inplace_train(period, order, df_nans):
     check_inverse_transform_inplace_train(transform, df_nans)
 
 
-@pytest.mark.parametrize(
-    "transform",
-    [
-        _SingleDifferencingTransform(in_column="target", period=1, inplace=True),
-        DifferencingTransform(in_column="target", period=1, order=1, inplace=True),
-    ],
-)
-def test_general_inverse_transform_inplace_test_fail_nans(transform, df_nans):
-    """Test that differencing transform fails to make inverse_transform on test data if there are NaNs."""
-    ts = TSDataset(df_nans, freq="D")
-    ts_train, ts_test = ts.train_test_split(test_size=20)
-
-    ts_train.fit_transform(transforms=[transform])
+@pytest.mark.parametrize("period", [1, 7])
+def test_single_inverse_transform_inplace_filled_test(period, df_nans):
+    """Test that _SingleDifferencingTransform correctly makes inverse_transform on filled test data in inplace mode."""
+    transform = _SingleDifferencingTransform(in_column="target", period=period, inplace=True)
+    check_inverse_transform_inplace_filled_test(transform, period, 1, df_nans)
 
-    # make predictions by hand only on one segment
-    future_ts = ts_train.make_future(20)
-    future_ts.df.loc[:, pd.IndexSlice["1", "target"]] = np.NaN
-    future_ts.df.loc[:, pd.IndexSlice["2", "target"]] = 2
 
-    # check fail on inverse_transform
-    with pytest.raises(ValueError, match="There should be no NaNs inside the segments"):
-        future_ts.inverse_transform()
+@pytest.mark.parametrize("period", [1, 7])
+@pytest.mark.parametrize("order", [1, 2])
+def test_full_inverse_transform_inplace_test(period, order, df_nans):
+    """Test that DifferencingTransform correctly makes inverse_transform on filled test data in inplace mode."""
+    transform = DifferencingTransform(in_column="target", period=period, order=order, inplace=True)
+    check_inverse_transform_inplace_filled_test(transform, period, order, df_nans)
 
 
 @pytest.mark.parametrize("period", [1, 7])
 def test_single_inverse_transform_inplace_test(period, df_nans):
-    """Test that _SingleDifferencingTransform correctly makes inverse_transform on test data in inplace mode."""
+    """Test that _SingleDifferencingTransform correctly makes inverse_transform on unfilled test data in inplace mode."""
     transform = _SingleDifferencingTransform(in_column="target", period=period, inplace=True)
-    check_inverse_transform_inplace_test(transform, period, 1, df_nans)
+    check_inverse_transform_inplace_unfilled_test(transform, df_nans)
 
 
 @pytest.mark.parametrize("period", [1, 7])
 @pytest.mark.parametrize("order", [1, 2])
 def test_full_inverse_transform_inplace_test(period, order, df_nans):
-    """Test that DifferencingTransform correctly makes inverse_transform on test data in inplace mode."""
+    """Test that DifferencingTransform correctly makes inverse_transform on unfilled test data in inplace mode."""
     transform = DifferencingTransform(in_column="target", period=period, order=order, inplace=True)
-    check_inverse_transform_inplace_test(transform, period, order, df_nans)
+    check_inverse_transform_inplace_unfilled_test(transform, df_nans)
 
 
 @pytest.mark.parametrize("period", [1, 7])