示例#1
0
    def transform_batch(
        self, func: Callable[..., Union[pd.DataFrame, pd.Series]], *args: Any, **kwargs: Any
    ) -> DataFrameOrSeries:
        """
        Transform chunks with a function that takes pandas DataFrame and outputs pandas DataFrame.
        The pandas DataFrame given to the function is of a batch used internally. The length of
        each input and output should be the same.

        See also `Transform and apply a function
        <https://koalas.readthedocs.io/en/latest/user_guide/transform_apply.html>`_.

        .. note:: the `func` is unable to access to the whole input frame. pandas-on-Spark
            internally splits the input series into multiple batches and calls `func` with each
            batch multiple times. Therefore, operations such as global aggregations are impossible.
            See the example below.

            >>> # This case does not return the length of whole frame but of the batch internally
            ... # used.
            ... def length(pdf) -> ps.DataFrame[int]:
            ...     return pd.DataFrame([len(pdf)] * len(pdf))
            ...
            >>> df = ps.DataFrame({'A': range(1000)})
            >>> df.pandas_on_spark.transform_batch(length)  # doctest: +SKIP
                c0
            0   83
            1   83
            2   83
            ...

        .. note:: this API executes the function once to infer the type which is
            potentially expensive, for instance, when the dataset is created after
            aggregations or sorting.

            To avoid this, specify return type in ``func``, for instance, as below:

            >>> def plus_one(x) -> ps.DataFrame[int, [float, float]]:
            ...     return x + 1

            If the return type is specified, the output column names become
            `c0, c1, c2 ... cn`. These names are positionally mapped to the returned
            DataFrame in ``func``.

            To specify the column names, you can assign them in a NumPy compound type style
            as below:

            >>> def plus_one(x) -> ps.DataFrame[("index", int), [("a", float), ("b", float)]]:
            ...     return x + 1

            >>> pdf = pd.DataFrame({'a': [1, 2, 3], 'b': [3, 4, 5]})
            >>> def plus_one(x) -> ps.DataFrame[
            ...         (pdf.index.name, pdf.index.dtype), zip(pdf.dtypes, pdf.columns)]:
            ...     return x + 1

        Parameters
        ----------
        func : function
            Function to transform each pandas frame.
        *args
            Positional arguments to pass to func.
        **kwargs
            Keyword arguments to pass to func.

        Returns
        -------
        DataFrame or Series

        See Also
        --------
        DataFrame.pandas_on_spark.apply_batch: For row/columnwise operations.
        Series.pandas_on_spark.transform_batch: transform the search as each pandas chunks.

        Examples
        --------
        >>> df = ps.DataFrame([(1, 2), (3, 4), (5, 6)], columns=['A', 'B'])
        >>> df
           A  B
        0  1  2
        1  3  4
        2  5  6

        >>> def plus_one_func(pdf) -> ps.DataFrame[int, [int, int]]:
        ...     return pdf + 1
        >>> df.pandas_on_spark.transform_batch(plus_one_func)
           c0  c1
        0   2   3
        1   4   5
        2   6   7

        >>> def plus_one_func(pdf) -> ps.DataFrame[("index", int), [('A', int), ('B', int)]]:
        ...     return pdf + 1
        >>> df.pandas_on_spark.transform_batch(plus_one_func)  # doctest: +NORMALIZE_WHITESPACE
               A  B
        index
        0      2  3
        1      4  5
        2      6  7

        >>> def plus_one_func(pdf) -> ps.Series[int]:
        ...     return pdf.B + 1
        >>> df.pandas_on_spark.transform_batch(plus_one_func)
        0    3
        1    5
        2    7
        dtype: int64

        You can also omit the type hints so pandas-on-Spark infers the return schema as below:

        >>> df.pandas_on_spark.transform_batch(lambda pdf: pdf + 1)
           A  B
        0  2  3
        1  4  5
        2  6  7

        >>> (df * -1).pandas_on_spark.transform_batch(abs)
           A  B
        0  1  2
        1  3  4
        2  5  6

        Note that you should not transform the index. The index information will not change.

        >>> df.pandas_on_spark.transform_batch(lambda pdf: pdf.B + 1)
        0    3
        1    5
        2    7
        Name: B, dtype: int64

        You can also specify extra arguments as below.

        >>> df.pandas_on_spark.transform_batch(lambda pdf, a, b, c: pdf.B + a + b + c, 1, 2, c=3)
        0     8
        1    10
        2    12
        Name: B, dtype: int64
        """
        from pyspark.pandas.groupby import GroupBy
        from pyspark.pandas.frame import DataFrame
        from pyspark.pandas.series import first_series
        from pyspark import pandas as ps

        assert callable(func), "the first argument should be a callable function."
        spec = inspect.getfullargspec(func)
        return_sig = spec.annotations.get("return", None)
        should_infer_schema = return_sig is None
        should_retain_index = should_infer_schema
        original_func = func
        func = lambda o: original_func(o, *args, **kwargs)

        def apply_func(pdf: pd.DataFrame) -> pd.DataFrame:
            return func(pdf).to_frame()

        def pandas_series_func(
            f: Callable[[pd.DataFrame], pd.DataFrame], return_type: DataType
        ) -> "UserDefinedFunctionLike":
            ff = f

            @pandas_udf(returnType=return_type)  # type: ignore[call-overload]
            def udf(pdf: pd.DataFrame) -> pd.Series:
                return first_series(ff(pdf))

            return udf

        if should_infer_schema:
            # Here we execute with the first 1000 to get the return type.
            # If the records were less than 1000, it uses pandas API directly for a shortcut.
            log_advice(
                "If the type hints is not specified for `transform_batch`, "
                "it is expensive to infer the data type internally."
            )
            limit = ps.get_option("compute.shortcut_limit")
            pdf = self._psdf.head(limit + 1)._to_internal_pandas()
            transformed = func(pdf)
            if not isinstance(transformed, (pd.DataFrame, pd.Series)):
                raise ValueError(
                    "The given function should return a frame; however, "
                    "the return type was %s." % type(transformed)
                )
            if len(transformed) != len(pdf):
                raise ValueError("transform_batch cannot produce aggregated results")
            psdf_or_psser = ps.from_pandas(transformed)

            if isinstance(psdf_or_psser, ps.Series):
                psser = cast(ps.Series, psdf_or_psser)

                field = psser._internal.data_fields[0].normalize_spark_type()

                return_schema = StructType([field.struct_field])
                output_func = GroupBy._make_pandas_df_builder_func(
                    self._psdf, apply_func, return_schema, retain_index=False
                )

                pudf = pandas_series_func(output_func, return_type=field.spark_type)
                columns = self._psdf._internal.spark_columns
                # TODO: Index will be lost in this case.
                internal = self._psdf._internal.copy(
                    column_labels=psser._internal.column_labels,
                    data_spark_columns=[pudf(F.struct(*columns)).alias(field.name)],
                    data_fields=[field],
                    column_label_names=psser._internal.column_label_names,
                )
                return first_series(DataFrame(internal))
            else:
                psdf = cast(DataFrame, psdf_or_psser)
                if len(pdf) <= limit:
                    # only do the short cut when it returns a frame to avoid
                    # operations on different dataframes in case of series.
                    return psdf

                index_fields = [
                    field.normalize_spark_type() for field in psdf._internal.index_fields
                ]
                data_fields = [field.normalize_spark_type() for field in psdf._internal.data_fields]

                return_schema = StructType(
                    [field.struct_field for field in index_fields + data_fields]
                )

                self_applied: DataFrame = DataFrame(self._psdf._internal.resolved_copy)

                output_func = GroupBy._make_pandas_df_builder_func(
                    self_applied, func, return_schema, retain_index=True  # type: ignore[arg-type]
                )
                columns = self_applied._internal.spark_columns

                pudf = pandas_udf(  # type: ignore[call-overload]
                    output_func, returnType=return_schema
                )
                temp_struct_column = verify_temp_column_name(
                    self_applied._internal.spark_frame, "__temp_struct__"
                )
                applied = pudf(F.struct(*columns)).alias(temp_struct_column)
                sdf = self_applied._internal.spark_frame.select(applied)
                sdf = sdf.selectExpr("%s.*" % temp_struct_column)

                return DataFrame(
                    psdf._internal.with_new_sdf(
                        spark_frame=sdf, index_fields=index_fields, data_fields=data_fields
                    )
                )
        else:
            return_type = infer_return_type(original_func)
            is_return_series = isinstance(return_type, SeriesType)
            is_return_dataframe = isinstance(return_type, DataFrameType)
            if not is_return_dataframe and not is_return_series:
                raise TypeError(
                    "The given function should specify a frame or series as its type "
                    "hints; however, the return type was %s." % return_sig
                )
            if is_return_series:
                field = InternalField(
                    dtype=cast(SeriesType, return_type).dtype,
                    struct_field=StructField(
                        name=SPARK_DEFAULT_SERIES_NAME,
                        dataType=cast(SeriesType, return_type).spark_type,
                    ),
                ).normalize_spark_type()

                return_schema = StructType([field.struct_field])
                output_func = GroupBy._make_pandas_df_builder_func(
                    self._psdf, apply_func, return_schema, retain_index=False
                )

                pudf = pandas_series_func(output_func, return_type=field.spark_type)
                columns = self._psdf._internal.spark_columns
                internal = self._psdf._internal.copy(
                    column_labels=[None],
                    data_spark_columns=[pudf(F.struct(*columns)).alias(field.name)],
                    data_fields=[field],
                    column_label_names=None,
                )
                return first_series(DataFrame(internal))
            else:
                index_fields = cast(DataFrameType, return_type).index_fields
                index_fields = [index_field.normalize_spark_type() for index_field in index_fields]
                data_fields = [
                    field.normalize_spark_type()
                    for field in cast(DataFrameType, return_type).data_fields
                ]
                normalized_fields = index_fields + data_fields
                return_schema = StructType([field.struct_field for field in normalized_fields])
                should_retain_index = len(index_fields) > 0

                self_applied = DataFrame(self._psdf._internal.resolved_copy)

                output_func = GroupBy._make_pandas_df_builder_func(
                    self_applied, func, return_schema, retain_index=should_retain_index  # type: ignore[arg-type]
                )
                columns = self_applied._internal.spark_columns

                pudf = pandas_udf(  # type: ignore[call-overload]
                    output_func, returnType=return_schema
                )
                temp_struct_column = verify_temp_column_name(
                    self_applied._internal.spark_frame, "__temp_struct__"
                )
                applied = pudf(F.struct(*columns)).alias(temp_struct_column)
                sdf = self_applied._internal.spark_frame.select(applied)
                sdf = sdf.selectExpr("%s.*" % temp_struct_column)

                index_spark_columns = None
                index_names: Optional[List[Optional[Tuple[Any, ...]]]] = None

                if should_retain_index:
                    index_spark_columns = [
                        scol_for(sdf, index_field.struct_field.name) for index_field in index_fields
                    ]

                    if not any(
                        [
                            SPARK_INDEX_NAME_PATTERN.match(index_field.struct_field.name)
                            for index_field in index_fields
                        ]
                    ):
                        index_names = [
                            (index_field.struct_field.name,) for index_field in index_fields
                        ]
                internal = InternalFrame(
                    spark_frame=sdf,
                    index_names=index_names,
                    index_spark_columns=index_spark_columns,
                    index_fields=index_fields,
                    data_fields=data_fields,
                )
                return DataFrame(internal)
示例#2
0
    def apply_batch(
        self, func: Callable[..., pd.DataFrame], args: Tuple = (), **kwds: Any
    ) -> "DataFrame":
        """
        Apply a function that takes pandas DataFrame and outputs pandas DataFrame. The pandas
        DataFrame given to the function is of a batch used internally.

        See also `Transform and apply a function
        <https://koalas.readthedocs.io/en/latest/user_guide/transform_apply.html>`_.

        .. note:: the `func` is unable to access to the whole input frame. pandas-on-Spark
            internally splits the input series into multiple batches and calls `func` with each
            batch multiple times. Therefore, operations such as global aggregations are impossible.
            See the example below.

            >>> # This case does not return the length of whole frame but of the batch internally
            ... # used.
            ... def length(pdf) -> ps.DataFrame[int, [int]]:
            ...     return pd.DataFrame([len(pdf)])
            ...
            >>> df = ps.DataFrame({'A': range(1000)})
            >>> df.pandas_on_spark.apply_batch(length)  # doctest: +SKIP
                c0
            0   83
            1   83
            2   83
            ...
            10  83
            11  83

        .. note:: this API executes the function once to infer the type which is
            potentially expensive, for instance, when the dataset is created after
            aggregations or sorting.

            To avoid this, specify return type in ``func``, for instance, as below:

            >>> def plus_one(x) -> ps.DataFrame[int, [float, float]]:
            ...     return x + 1

            If the return type is specified, the output column names become
            `c0, c1, c2 ... cn`. These names are positionally mapped to the returned
            DataFrame in ``func``.

            To specify the column names, you can assign them in a NumPy compound type style
            as below:

            >>> def plus_one(x) -> ps.DataFrame[("index", int), [("a", float), ("b", float)]]:
            ...     return x + 1

            >>> pdf = pd.DataFrame({'a': [1, 2, 3], 'b': [3, 4, 5]})
            >>> def plus_one(x) -> ps.DataFrame[
            ...         (pdf.index.name, pdf.index.dtype), zip(pdf.dtypes, pdf.columns)]:
            ...     return x + 1

        Parameters
        ----------
        func : function
            Function to apply to each pandas frame.
        args : tuple
            Positional arguments to pass to `func` in addition to the
            array/series.
        **kwds
            Additional keyword arguments to pass as keywords arguments to
            `func`.

        Returns
        -------
        DataFrame

        See Also
        --------
        DataFrame.apply: For row/columnwise operations.
        DataFrame.applymap: For elementwise operations.
        DataFrame.aggregate: Only perform aggregating type operations.
        DataFrame.transform: Only perform transforming type operations.
        Series.pandas_on_spark.transform_batch: transform the search as each pandas chunks.

        Examples
        --------
        >>> df = ps.DataFrame([(1, 2), (3, 4), (5, 6)], columns=['A', 'B'])
        >>> df
           A  B
        0  1  2
        1  3  4
        2  5  6

        >>> def query_func(pdf) -> ps.DataFrame[int, [int, int]]:
        ...     return pdf.query('A == 1')
        >>> df.pandas_on_spark.apply_batch(query_func)
           c0  c1
        0   1   2

        >>> def query_func(pdf) -> ps.DataFrame[("idx", int), [("A", int), ("B", int)]]:
        ...     return pdf.query('A == 1')
        >>> df.pandas_on_spark.apply_batch(query_func)  # doctest: +NORMALIZE_WHITESPACE
             A  B
        idx
        0    1  2

        You can also omit the type hints so pandas-on-Spark infers the return schema as below:

        >>> df.pandas_on_spark.apply_batch(lambda pdf: pdf.query('A == 1'))
           A  B
        0  1  2

        You can also specify extra arguments.

        >>> def calculation(pdf, y, z) -> ps.DataFrame[int, [int, int]]:
        ...     return pdf ** y + z
        >>> df.pandas_on_spark.apply_batch(calculation, args=(10,), z=20)
                c0        c1
        0       21      1044
        1    59069   1048596
        2  9765645  60466196

        You can also use ``np.ufunc`` and built-in functions as input.

        >>> df.pandas_on_spark.apply_batch(np.add, args=(10,))
            A   B
        0  11  12
        1  13  14
        2  15  16

        >>> (df * -1).pandas_on_spark.apply_batch(abs)
           A  B
        0  1  2
        1  3  4
        2  5  6

        """
        # TODO: codes here partially duplicate `DataFrame.apply`. Can we deduplicate?

        from pyspark.pandas.groupby import GroupBy
        from pyspark.pandas.frame import DataFrame
        from pyspark import pandas as ps

        if not isinstance(func, FunctionType):
            assert callable(func), "the first argument should be a callable function."
            f = func
            func = lambda *args, **kwargs: f(*args, **kwargs)

        spec = inspect.getfullargspec(func)
        return_sig = spec.annotations.get("return", None)
        should_infer_schema = return_sig is None

        original_func = func
        func = lambda o: original_func(o, *args, **kwds)

        self_applied: DataFrame = DataFrame(self._psdf._internal.resolved_copy)

        if should_infer_schema:
            # Here we execute with the first 1000 to get the return type.
            # If the records were less than 1000, it uses pandas API directly for a shortcut.
            log_advice(
                "If the type hints is not specified for `apply_batch`, "
                "it is expensive to infer the data type internally."
            )
            limit = ps.get_option("compute.shortcut_limit")
            pdf = self_applied.head(limit + 1)._to_internal_pandas()
            applied = func(pdf)
            if not isinstance(applied, pd.DataFrame):
                raise ValueError(
                    "The given function should return a frame; however, "
                    "the return type was %s." % type(applied)
                )
            psdf: DataFrame = DataFrame(applied)
            if len(pdf) <= limit:
                return psdf

            index_fields = [field.normalize_spark_type() for field in psdf._internal.index_fields]
            data_fields = [field.normalize_spark_type() for field in psdf._internal.data_fields]

            return_schema = StructType([field.struct_field for field in index_fields + data_fields])

            output_func = GroupBy._make_pandas_df_builder_func(
                self_applied, func, return_schema, retain_index=True
            )
            sdf = self_applied._internal.spark_frame.mapInPandas(
                lambda iterator: map(output_func, iterator), schema=return_schema
            )

            # If schema is inferred, we can restore indexes too.
            internal = psdf._internal.with_new_sdf(
                spark_frame=sdf, index_fields=index_fields, data_fields=data_fields
            )
        else:
            return_type = infer_return_type(original_func)
            is_return_dataframe = isinstance(return_type, DataFrameType)
            if not is_return_dataframe:
                raise TypeError(
                    "The given function should specify a frame as its type "
                    "hints; however, the return type was %s." % return_sig
                )
            index_fields = cast(DataFrameType, return_type).index_fields
            should_retain_index = len(index_fields) > 0
            return_schema = cast(DataFrameType, return_type).spark_type

            output_func = GroupBy._make_pandas_df_builder_func(
                self_applied, func, return_schema, retain_index=should_retain_index
            )
            sdf = self_applied._internal.to_internal_spark_frame.mapInPandas(
                lambda iterator: map(output_func, iterator), schema=return_schema
            )

            index_spark_columns = None
            index_names: Optional[List[Optional[Tuple[Any, ...]]]] = None

            if should_retain_index:
                index_spark_columns = [
                    scol_for(sdf, index_field.struct_field.name) for index_field in index_fields
                ]

                if not any(
                    [
                        SPARK_INDEX_NAME_PATTERN.match(index_field.struct_field.name)
                        for index_field in index_fields
                    ]
                ):
                    index_names = [(index_field.struct_field.name,) for index_field in index_fields]
            internal = InternalFrame(
                spark_frame=sdf,
                index_names=index_names,
                index_spark_columns=index_spark_columns,
                index_fields=index_fields,
                data_fields=cast(DataFrameType, return_type).data_fields,
            )
        return DataFrame(internal)