feat: add bigframes.ml.compose.SQLScalarColumnTransformer to create custom SQL-based transformations

bigframes/ml/base.py

@@ -198,10 +198,6 @@ def _extract_output_names(self):
             # pass the columns that are not transformed
             if "transformSql" not in transform_col_dict:
                 continue
-            transform_sql: str = transform_col_dict["transformSql"]
-            if not transform_sql.startswith("ML."):
-                continue
-
             output_names.append(transform_col_dict["name"])
 
         self._output_names = output_names

bigframes/ml/compose.py

@@ -46,6 +46,101 @@
 )
 
 
+class SQLScalarColumnTransformer:
+    r"""
+    Wrapper for plain SQL code contained in a ColumnTransformer.
+
+    Create a single column transformer in plain sql.
+    This transformer can only be used inside ColumnTransformer.
+
+    When creating an instance '{0}' can be used as placeholder
+    for the column to transform:
+
+        SQLScalarColumnTransformer("{0}+1")
+
+    The default target column gets the prefix 'transformed\_'
+    but can also be changed when creating an instance:
+
+        SQLScalarColumnTransformer("{0}+1", "inc_{0}")
+
+    **Examples:**
+
+        >>> from bigframes.ml.compose import ColumnTransformer, SQLScalarColumnTransformer
+        >>> import bigframes.pandas as bpd
+        <BLANKLINE>
+        >>> df = bpd.DataFrame({'name': ["James", None, "Mary"], 'city': ["New York", "Boston", None]})
+        >>> col_trans = ColumnTransformer([
+        ...     ("strlen",
+        ...      SQLScalarColumnTransformer("CASE WHEN {0} IS NULL THEN 15 ELSE LENGTH({0}) END"),
+        ...      ['name', 'city']),
+        ... ])
+        >>> col_trans = col_trans.fit(df)
+        >>> df_transformed = col_trans.transform(df)
+        >>> df_transformed
+           transformed_name  transformed_city
+        0                 5                 8
+        1                15                 6
+        2                 4                15
+        <BLANKLINE>
+        [3 rows x 2 columns]
+
+    SQLScalarColumnTransformer can be combined with other transformers, like StandardScaler:
+
+        >>> col_trans = ColumnTransformer([
+        ...     ("identity", SQLScalarColumnTransformer("{0}", target_column="{0}"), ["col1", col5"]),
+        ...     ("increment", SQLScalarColumnTransformer("{0}+1", target_column="inc_{0}"), "col2"),
+        ...     ("stdscale", preprocessing.StandardScaler(), "col3"),
+        ...     ...
+        ... ])
+
+    """
+
+    def __init__(self, sql: str, target_column: str = "transformed_{0}"):
+        super().__init__()
+        self._sql = sql
+        self._target_column = target_column.replace("`", "")
+
+    PLAIN_COLNAME_RX = re.compile("^[a-z][a-z0-9_]*$", re.IGNORECASE)
+
+    def escape(self, colname: str):
+        colname = colname.replace("`", "")
+        if self.PLAIN_COLNAME_RX.match(colname):
+            return colname
+        return f"`{colname}`"
+
+    def _compile_to_sql(
+        self, X: bpd.DataFrame, columns: Optional[Iterable[str]] = None
+    ) -> List[str]:
+        if columns is None:
+            columns = X.columns
+        result = []
+        for column in columns:
+            current_sql = self._sql.format(self.escape(column))
+            current_target_column = self.escape(self._target_column.format(column))
+            result.append(f"{current_sql} AS {current_target_column}")
+        return result
+
+    def __repr__(self):
+        return f"SQLScalarColumnTransformer(sql='{self._sql}', target_column='{self._target_column}')"
+
+    def __eq__(self, other) -> bool:
+        return type(self) is type(other) and self._keys() == other._keys()
+
+    def __hash__(self) -> int:
+        return hash(self._keys())
+
+    def _keys(self):
+        return (self._sql, self._target_column)
+
+
+# Type hints for transformers contained in ColumnTransformer
+SingleColTransformer = Union[
+    preprocessing.PreprocessingType,
+    impute.SimpleImputer,
+    SQLScalarColumnTransformer,
+]
+
+
 @log_adapter.class_logger
 class ColumnTransformer(
     base.Transformer,
@@ -60,7 +155,7 @@ def __init__(
         transformers: Iterable[
             Tuple[
                 str,
-                Union[preprocessing.PreprocessingType, impute.SimpleImputer],
+                SingleColTransformer,
                 Union[str, Iterable[str]],
             ]
         ],
@@ -78,14 +173,12 @@ def _keys(self):
     @property
     def transformers_(
         self,
-    ) -> List[
-        Tuple[str, Union[preprocessing.PreprocessingType, impute.SimpleImputer], str]
-    ]:
+    ) -> List[Tuple[str, SingleColTransformer, str,]]:
         """The collection of transformers as tuples of (name, transformer, column)."""
         result: List[
             Tuple[
                 str,
-                Union[preprocessing.PreprocessingType, impute.SimpleImputer],
+                SingleColTransformer,
                 str,
             ]
         ] = []
@@ -103,6 +196,8 @@ def transformers_(
 
         return result
 
+    AS_FLEXNAME_SUFFIX_RX = re.compile("^(.*)\\bAS\\s*`[^`]+`\\s*$", re.IGNORECASE)
+
     @classmethod
     def _extract_from_bq_model(
         cls,
@@ -114,7 +209,7 @@ def _extract_from_bq_model(
         transformers_set: Set[
             Tuple[
                 str,
-                Union[preprocessing.PreprocessingType, impute.SimpleImputer],
+                SingleColTransformer,
                 Union[str, List[str]],
             ]
         ] = set()
@@ -130,8 +225,10 @@ def camel_to_snake(name):
             if "transformSql" not in transform_col_dict:
                 continue
             transform_sql: str = transform_col_dict["transformSql"]
-            if not transform_sql.startswith("ML."):
-                continue
+
+            # workaround for bug in bq_model returning " AS `...`" suffix for flexible names
+            if cls.AS_FLEXNAME_SUFFIX_RX.match(transform_sql):
+                transform_sql = cls.AS_FLEXNAME_SUFFIX_RX.match(transform_sql).group(1)
 
             output_names.append(transform_col_dict["name"])
             found_transformer = False
@@ -148,8 +245,22 @@ def camel_to_snake(name):
                     found_transformer = True
                     break
             if not found_transformer:
-                raise NotImplementedError(
-                    f"Unsupported transformer type. {constants.FEEDBACK_LINK}"
+                if transform_sql.startswith("ML."):
+                    raise NotImplementedError(
+                        f"Unsupported transformer type. {constants.FEEDBACK_LINK}"
+                    )
+
+                target_column = transform_col_dict["name"]
+                sql_transformer = SQLScalarColumnTransformer(
+                    transform_sql, target_column=target_column
+                )
+                input_column_name = f"?{target_column}"
+                transformers_set.add(
+                    (
+                        camel_to_snake(sql_transformer.__class__.__name__),
+                        sql_transformer,
+                        input_column_name,
+                    )
                 )
 
         transformer = cls(transformers=list(transformers_set))
@@ -167,6 +278,8 @@ def _merge(
 
         assert len(transformers) > 0
         _, transformer_0, column_0 = transformers[0]
+        if isinstance(transformer_0, SQLScalarColumnTransformer):
+            return self  # SQLScalarColumnTransformer only work inside ColumnTransformer
         feature_columns_sorted = sorted(
             [
                 cast(str, feature_column.name)

tests/system/large/ml/test_compose.py

@@ -36,6 +36,32 @@ def test_columntransformer_standalone_fit_and_transform(
                 preprocessing.MinMaxScaler(),
                 ["culmen_length_mm"],
             ),
+            (
+                "increment",
+                compose.SQLScalarColumnTransformer("{0}+1"),
+                ["culmen_length_mm", "flipper_length_mm"],
+            ),
+            (
+                "length",
+                compose.SQLScalarColumnTransformer(
+                    "CASE WHEN {0} IS NULL THEN -1 ELSE LENGTH({0}) END",
+                    target_column="len_{0}",
+                ),
+                "species",
+            ),
+            (
+                "ohe",
+                compose.SQLScalarColumnTransformer(
+                    "CASE WHEN {0}='Adelie Penguin (Pygoscelis adeliae)' THEN 1 ELSE 0 END",
+                    target_column="ohe_adelie",
+                ),
+                "species",
+            ),
+            (
+                "identity",
+                compose.SQLScalarColumnTransformer("{0}", target_column="{0}"),
+                ["culmen_length_mm", "flipper_length_mm"],
+            ),
         ]
     )
 
@@ -51,6 +77,12 @@ def test_columntransformer_standalone_fit_and_transform(
             "standard_scaled_culmen_length_mm",
             "min_max_scaled_culmen_length_mm",
             "standard_scaled_flipper_length_mm",
+            "transformed_culmen_length_mm",
+            "transformed_flipper_length_mm",
+            "len_species",
+            "ohe_adelie",
+            "culmen_length_mm",
+            "flipper_length_mm",
         ],
         index=[1633, 1672, 1690],
         col_exact=False,
@@ -70,6 +102,19 @@ def test_columntransformer_standalone_fit_transform(new_penguins_df):
                 preprocessing.StandardScaler(),
                 ["culmen_length_mm", "flipper_length_mm"],
             ),
+            (
+                "length",
+                compose.SQLScalarColumnTransformer(
+                    "CASE WHEN {0} IS NULL THEN -1 ELSE LENGTH({0}) END",
+                    target_column="len_{0}",
+                ),
+                "species",
+            ),
+            (
+                "identity",
+                compose.SQLScalarColumnTransformer("{0}", target_column="{0}"),
+                ["culmen_length_mm", "flipper_length_mm"],
+            ),
         ]
     )
 
@@ -83,6 +128,9 @@ def test_columntransformer_standalone_fit_transform(new_penguins_df):
             "onehotencoded_species",
             "standard_scaled_culmen_length_mm",
             "standard_scaled_flipper_length_mm",
+            "len_species",
+            "culmen_length_mm",
+            "flipper_length_mm",
         ],
         index=[1633, 1672, 1690],
         col_exact=False,
@@ -102,6 +150,27 @@ def test_columntransformer_save_load(new_penguins_df, dataset_id):
                 preprocessing.StandardScaler(),
                 ["culmen_length_mm", "flipper_length_mm"],
             ),
+            (
+                "length",
+                compose.SQLScalarColumnTransformer(
+                    "CASE WHEN {0} IS NULL THEN -1 ELSE LENGTH({0}) END",
+                    target_column="len_{0}",
+                ),
+                "species",
+            ),
+            (
+                "identity",
+                compose.SQLScalarColumnTransformer("{0}", target_column="{0}"),
+                ["culmen_length_mm", "flipper_length_mm"],
+            ),
+            (
+                "flexname",
+                compose.SQLScalarColumnTransformer(
+                    "CASE WHEN {0} IS NULL THEN -1 ELSE LENGTH({0}) END",
+                    target_column="Flex {0} Name",
+                ),
+                "species",
+            ),
         ]
     )
     transformer.fit(
@@ -122,6 +191,36 @@ def test_columntransformer_save_load(new_penguins_df, dataset_id):
         ),
         ("standard_scaler", preprocessing.StandardScaler(), "culmen_length_mm"),
         ("standard_scaler", preprocessing.StandardScaler(), "flipper_length_mm"),
+        (
+            "sql_scalar_column_transformer",
+            compose.SQLScalarColumnTransformer(
+                "CASE WHEN species IS NULL THEN -1 ELSE LENGTH(species) END",
+                target_column="len_species",
+            ),
+            "?len_species",
+        ),
+        (
+            "sql_scalar_column_transformer",
+            compose.SQLScalarColumnTransformer(
+                "flipper_length_mm", target_column="flipper_length_mm"
+            ),
+            "?flipper_length_mm",
+        ),
+        (
+            "sql_scalar_column_transformer",
+            compose.SQLScalarColumnTransformer(
+                "culmen_length_mm", target_column="culmen_length_mm"
+            ),
+            "?culmen_length_mm",
+        ),
+        (
+            "sql_scalar_column_transformer",
+            compose.SQLScalarColumnTransformer(
+                "CASE WHEN species IS NULL THEN -1 ELSE LENGTH(species) END ",
+                target_column="Flex species Name",
+            ),
+            "?Flex species Name",
+        ),
     ]
     assert set(reloaded_transformer.transformers) == set(expected)
     assert reloaded_transformer._bqml_model is not None
@@ -136,6 +235,10 @@ def test_columntransformer_save_load(new_penguins_df, dataset_id):
             "onehotencoded_species",
             "standard_scaled_culmen_length_mm",
             "standard_scaled_flipper_length_mm",
+            "len_species",
+            "culmen_length_mm",
+            "flipper_length_mm",
+            "Flex species Name",
         ],
         index=[1633, 1672, 1690],
         col_exact=False,