perf: update df.corr, df.cov to be used with more than 30 columns case. (#1161)

Genesis929 · web-flow · commit 9dcf1aa91891 · 2024-12-09T10:36:05.000-08:00
* perf: update df.corr, df.cov to be used with more than 30 columns case.

* add large test

* remove print

* fix_index

* fix index

* test fix

* fix test

* fix test

* slightly improve multi_apply_unary_op to avoid RecursionError

* update recursion limit for nox session

* skip the test in e2e/python 3.12

* simplify code

* simplify code
diff --git a/bigframes/core/blocks.py b/bigframes/core/blocks.py
@@ -912,14 +912,17 @@ def multi_apply_unary_op(
             input_varname = input_varnames[0]
 
         block = self
+
+        result_ids = []
         for col_id in columns:
             label = self.col_id_to_label[col_id]
             block, result_id = block.project_expr(
                 expr.bind_variables({input_varname: ex.deref(col_id)}),
                 label=label,
             )
             block = block.copy_values(result_id, col_id)
-            block = block.drop_columns([result_id])
+            result_ids.append(result_id)
+        block = block.drop_columns(result_ids)
         # Special case, we can preserve transpose cache for full-frame unary ops
         if (self._transpose_cache is not None) and set(self.value_columns) == set(
             columns
@@ -1317,50 +1320,6 @@ def summarize(
             index_columns=index_cols,
         )
 
-    def calculate_pairwise_metric(self, op=agg_ops.CorrOp()):
-        """
-        Returns a block object to compute pairwise metrics among all value columns in this block.
-
-        The metric to be computed is specified by the `op` parameter, which can be either a
-        correlation operation (default) or a covariance operation.
-        """
-        if len(self.value_columns) > 30:
-            raise NotImplementedError(
-                "This function supports dataframes with 30 columns or fewer. "
-                f"Provided dataframe has {len(self.value_columns)} columns. {constants.FEEDBACK_LINK}"
-            )
-
-        aggregations = [
-            (
-                ex.BinaryAggregation(op, ex.deref(left_col), ex.deref(right_col)),
-                f"{left_col}-{right_col}",
-            )
-            for left_col in self.value_columns
-            for right_col in self.value_columns
-        ]
-        expr = self.expr.aggregate(aggregations)
-
-        input_count = len(self.value_columns)
-        unpivot_columns = tuple(
-            tuple(expr.column_ids[input_count * i : input_count * (i + 1)])
-            for i in range(input_count)
-        )
-        labels = self._get_labels_for_columns(self.value_columns)
-
-        # TODO(b/340896143): fix type error
-        expr, (index_col_ids, _, _) = unpivot(
-            expr,
-            row_labels=labels,
-            unpivot_columns=unpivot_columns,
-        )
-
-        return Block(
-            expr,
-            column_labels=self.column_labels,
-            index_columns=index_col_ids,
-            index_labels=self.column_labels.names,
-        )
-
     def explode(
         self,
         column_ids: typing.Sequence[str],
diff --git a/bigframes/dataframe.py b/bigframes/dataframe.py
@@ -1197,15 +1197,203 @@ def corr(self, method="pearson", min_periods=None, numeric_only=False) -> DataFr
         else:
             frame = self._drop_non_numeric()
 
-        return DataFrame(frame._block.calculate_pairwise_metric(op=agg_ops.CorrOp()))
+        orig_columns = frame.columns
+        # Replace column names with 0 to n - 1 to keep order
+        # and avoid the influence of duplicated column name
+        frame.columns = pandas.Index(range(len(orig_columns)))
+        frame = frame.astype(bigframes.dtypes.FLOAT_DTYPE)
+        block = frame._block
+
+        # A new column that uniquely identifies each row
+        block, ordering_col = frame._block.promote_offsets(label="_bigframes_idx")
+
+        val_col_ids = [
+            col_id for col_id in block.value_columns if col_id != ordering_col
+        ]
+
+        block = block.melt(
+            [ordering_col], val_col_ids, ["_bigframes_variable"], "_bigframes_value"
+        )
+
+        block = block.merge(
+            block,
+            left_join_ids=[ordering_col],
+            right_join_ids=[ordering_col],
+            how="inner",
+            sort=False,
+        )
+
+        frame = DataFrame(block).dropna(
+            subset=["_bigframes_value_x", "_bigframes_value_y"]
+        )
+
+        paired_mean_frame = (
+            frame.groupby(["_bigframes_variable_x", "_bigframes_variable_y"])
+            .agg(
+                _bigframes_paired_mean_x=bigframes.pandas.NamedAgg(
+                    column="_bigframes_value_x", aggfunc="mean"
+                ),
+                _bigframes_paired_mean_y=bigframes.pandas.NamedAgg(
+                    column="_bigframes_value_y", aggfunc="mean"
+                ),
+            )
+            .reset_index()
+        )
+
+        frame = frame.merge(
+            paired_mean_frame, on=["_bigframes_variable_x", "_bigframes_variable_y"]
+        )
+        frame["_bigframes_value_x"] -= frame["_bigframes_paired_mean_x"]
+        frame["_bigframes_value_y"] -= frame["_bigframes_paired_mean_y"]
+
+        frame["_bigframes_dividend"] = (
+            frame["_bigframes_value_x"] * frame["_bigframes_value_y"]
+        )
+        frame["_bigframes_x_square"] = (
+            frame["_bigframes_value_x"] * frame["_bigframes_value_x"]
+        )
+        frame["_bigframes_y_square"] = (
+            frame["_bigframes_value_y"] * frame["_bigframes_value_y"]
+        )
+
+        result = (
+            frame.groupby(["_bigframes_variable_x", "_bigframes_variable_y"])
+            .agg(
+                _bigframes_dividend_sum=bigframes.pandas.NamedAgg(
+                    column="_bigframes_dividend", aggfunc="sum"
+                ),
+                _bigframes_x_square_sum=bigframes.pandas.NamedAgg(
+                    column="_bigframes_x_square", aggfunc="sum"
+                ),
+                _bigframes_y_square_sum=bigframes.pandas.NamedAgg(
+                    column="_bigframes_y_square", aggfunc="sum"
+                ),
+            )
+            .reset_index()
+        )
+        result["_bigframes_corr"] = result["_bigframes_dividend_sum"] / (
+            (
+                result["_bigframes_x_square_sum"] * result["_bigframes_y_square_sum"]
+            )._apply_unary_op(ops.sqrt_op)
+        )
+        result = result._pivot(
+            index="_bigframes_variable_x",
+            columns="_bigframes_variable_y",
+            values="_bigframes_corr",
+        )
+
+        map_data = {
+            f"_bigframes_level_{i}": orig_columns.get_level_values(i)
+            for i in range(orig_columns.nlevels)
+        }
+        map_data["_bigframes_keys"] = range(len(orig_columns))
+        map_df = bigframes.dataframe.DataFrame(
+            map_data,
+            session=self._get_block().expr.session,
+        ).set_index("_bigframes_keys")
+        result = result.join(map_df).sort_index()
+        index_columns = [f"_bigframes_level_{i}" for i in range(orig_columns.nlevels)]
+        result = result.set_index(index_columns)
+        result.index.names = orig_columns.names
+        result.columns = orig_columns
+
+        return result
 
     def cov(self, *, numeric_only: bool = False) -> DataFrame:
         if not numeric_only:
             frame = self._raise_on_non_numeric("corr")
         else:
             frame = self._drop_non_numeric()
 
-        return DataFrame(frame._block.calculate_pairwise_metric(agg_ops.CovOp()))
+        orig_columns = frame.columns
+        # Replace column names with 0 to n - 1 to keep order
+        # and avoid the influence of duplicated column name
+        frame.columns = pandas.Index(range(len(orig_columns)))
+        frame = frame.astype(bigframes.dtypes.FLOAT_DTYPE)
+        block = frame._block
+
+        # A new column that uniquely identifies each row
+        block, ordering_col = frame._block.promote_offsets(label="_bigframes_idx")
+
+        val_col_ids = [
+            col_id for col_id in block.value_columns if col_id != ordering_col
+        ]
+
+        block = block.melt(
+            [ordering_col], val_col_ids, ["_bigframes_variable"], "_bigframes_value"
+        )
+        block = block.merge(
+            block,
+            left_join_ids=[ordering_col],
+            right_join_ids=[ordering_col],
+            how="inner",
+            sort=False,
+        )
+
+        frame = DataFrame(block).dropna(
+            subset=["_bigframes_value_x", "_bigframes_value_y"]
+        )
+
+        paired_mean_frame = (
+            frame.groupby(["_bigframes_variable_x", "_bigframes_variable_y"])
+            .agg(
+                _bigframes_paired_mean_x=bigframes.pandas.NamedAgg(
+                    column="_bigframes_value_x", aggfunc="mean"
+                ),
+                _bigframes_paired_mean_y=bigframes.pandas.NamedAgg(
+                    column="_bigframes_value_y", aggfunc="mean"
+                ),
+            )
+            .reset_index()
+        )
+
+        frame = frame.merge(
+            paired_mean_frame, on=["_bigframes_variable_x", "_bigframes_variable_y"]
+        )
+        frame["_bigframes_value_x"] -= frame["_bigframes_paired_mean_x"]
+        frame["_bigframes_value_y"] -= frame["_bigframes_paired_mean_y"]
+
+        frame["_bigframes_dividend"] = (
+            frame["_bigframes_value_x"] * frame["_bigframes_value_y"]
+        )
+
+        result = (
+            frame.groupby(["_bigframes_variable_x", "_bigframes_variable_y"])
+            .agg(
+                _bigframes_dividend_sum=bigframes.pandas.NamedAgg(
+                    column="_bigframes_dividend", aggfunc="sum"
+                ),
+                _bigframes_dividend_count=bigframes.pandas.NamedAgg(
+                    column="_bigframes_dividend", aggfunc="count"
+                ),
+            )
+            .reset_index()
+        )
+        result["_bigframes_cov"] = result["_bigframes_dividend_sum"] / (
+            result["_bigframes_dividend_count"] - 1
+        )
+        result = result._pivot(
+            index="_bigframes_variable_x",
+            columns="_bigframes_variable_y",
+            values="_bigframes_cov",
+        )
+
+        map_data = {
+            f"_bigframes_level_{i}": orig_columns.get_level_values(i)
+            for i in range(orig_columns.nlevels)
+        }
+        map_data["_bigframes_keys"] = range(len(orig_columns))
+        map_df = bigframes.dataframe.DataFrame(
+            map_data,
+            session=self._get_block().expr.session,
+        ).set_index("_bigframes_keys")
+        result = result.join(map_df).sort_index()
+        index_columns = [f"_bigframes_level_{i}" for i in range(orig_columns.nlevels)]
+        result = result.set_index(index_columns)
+        result.index.names = orig_columns.names
+        result.columns = orig_columns
+
+        return result
 
     def to_arrow(
         self,
diff --git a/tests/system/conftest.py b/tests/system/conftest.py
@@ -611,6 +611,28 @@ def scalars_dfs_maybe_ordered(
     )
 
 
+@pytest.fixture(scope="session")
+def scalars_df_numeric_150_columns_maybe_ordered(
+    maybe_ordered_session,
+    scalars_pandas_df_index,
+):
+    """DataFrame pointing at test data."""
+    # TODO(b/379911038): After the error fixed, add numeric type.
+    pandas_df = scalars_pandas_df_index.reset_index(drop=False)[
+        [
+            "rowindex",
+            "rowindex_2",
+            "float64_col",
+            "int64_col",
+            "int64_too",
+        ]
+        * 30
+    ]
+
+    df = maybe_ordered_session.read_pandas(pandas_df)
+    return (df, pandas_df)
+
+
 @pytest.fixture(scope="session")
 def hockey_df(
     hockey_table_id: str, session: bigframes.Session
diff --git a/tests/system/large/test_dataframe.py b/tests/system/large/test_dataframe.py
@@ -0,0 +1,42 @@
+import sys
+
+import pandas as pd
+import pytest
+
+
+@pytest.mark.skipif(
+    sys.version_info >= (3, 12),
+    # See: https://github.com/python/cpython/issues/112282
+    reason="setrecursionlimit has no effect on the Python C stack since Python 3.12.",
+)
+def test_corr_w_numeric_only(scalars_df_numeric_150_columns_maybe_ordered):
+    scalars_df, scalars_pandas_df = scalars_df_numeric_150_columns_maybe_ordered
+    bf_result = scalars_df.corr(numeric_only=True).to_pandas()
+    pd_result = scalars_pandas_df.corr(numeric_only=True)
+
+    pd.testing.assert_frame_equal(
+        bf_result,
+        pd_result,
+        check_dtype=False,
+        check_index_type=False,
+        check_column_type=False,
+    )
+
+
+@pytest.mark.skipif(
+    sys.version_info >= (3, 12),
+    # See: https://github.com/python/cpython/issues/112282
+    reason="setrecursionlimit has no effect on the Python C stack since Python 3.12.",
+)
+def test_cov_w_numeric_only(scalars_df_numeric_150_columns_maybe_ordered):
+    scalars_df, scalars_pandas_df = scalars_df_numeric_150_columns_maybe_ordered
+    bf_result = scalars_df.cov(numeric_only=True).to_pandas()
+    pd_result = scalars_pandas_df.cov(numeric_only=True)
+
+    pd.testing.assert_frame_equal(
+        bf_result,
+        pd_result,
+        check_dtype=False,
+        check_index_type=False,
+        check_column_type=False,
+    )
diff --git a/tests/system/small/test_dataframe.py b/tests/system/small/test_dataframe.py
@@ -2075,8 +2075,8 @@ def test_combine_first(
         ),
     ],
 )
-def test_corr_w_numeric_only(scalars_dfs, columns, numeric_only):
-    scalars_df, scalars_pandas_df = scalars_dfs
+def test_corr_w_numeric_only(scalars_dfs_maybe_ordered, columns, numeric_only):
+    scalars_df, scalars_pandas_df = scalars_dfs_maybe_ordered
 
     bf_result = scalars_df[columns].corr(numeric_only=numeric_only).to_pandas()
     pd_result = scalars_pandas_df[columns].corr(numeric_only=numeric_only)
@@ -2115,11 +2115,10 @@ def test_corr_w_invalid_parameters(scalars_dfs):
         ),
     ],
 )
-def test_cov_w_numeric_only(scalars_dfs, columns, numeric_only):
-    scalars_df, scalars_pandas_df = scalars_dfs
+def test_cov_w_numeric_only(scalars_dfs_maybe_ordered, columns, numeric_only):
+    scalars_df, scalars_pandas_df = scalars_dfs_maybe_ordered
     bf_result = scalars_df[columns].cov(numeric_only=numeric_only).to_pandas()
     pd_result = scalars_pandas_df[columns].cov(numeric_only=numeric_only)
-
     # BigFrames and Pandas differ in their data type handling:
     # - Column types: BigFrames uses Float64, Pandas uses float64.
     # - Index types: BigFrames uses strign, Pandas uses object.
diff --git a/tests/system/small/test_multiindex.py b/tests/system/small/test_multiindex.py
@@ -910,7 +910,9 @@ def test_column_multi_index_unstack(scalars_df_index, scalars_pandas_df_index):
 
 def test_corr_w_multi_index(scalars_df_index, scalars_pandas_df_index):
     columns = ["int64_too", "float64_col", "int64_col"]
-    multi_columns = pandas.MultiIndex.from_tuples(zip(["a", "b", "b"], [1, 2, 2]))
+    multi_columns = pandas.MultiIndex.from_tuples(
+        zip(["a", "b", "b"], [1, 2, 2]), names=[None, "level_2"]
+    )
 
     bf = scalars_df_index[columns].copy()
     bf.columns = multi_columns
@@ -931,7 +933,9 @@ def test_corr_w_multi_index(scalars_df_index, scalars_pandas_df_index):
 
 def test_cov_w_multi_index(scalars_df_index, scalars_pandas_df_index):
     columns = ["int64_too", "float64_col", "int64_col"]
-    multi_columns = pandas.MultiIndex.from_tuples(zip(["a", "b", "b"], [1, 2, 2]))
+    multi_columns = pandas.MultiIndex.from_tuples(
+        zip(["a", "b", "b"], [1, 2, 2]), names=["level_1", None]
+    )
 
     bf = scalars_df_index[columns].copy()
     bf.columns = multi_columns
