Skip to content

ratiopath.ray.aggregate.TensorStd

Bases: AggregateFnV2[dict, ndarray | float]

Calculates the standard deviation of a column containing Tensors.

This aggregator treats the data column as a high-dimensional array where axis 0 represents the batch dimension. To satisfy the requirements of a reduction and prevent memory growth proportional to the number of rows, axis 0 must be included in the aggregation.

It uses a parallel variance accumulation algorithm (Chan's method) to maintain numerical stability while processing data across multiple Ray blocks.

Parameters:

Name Type Description Default
on str

The name of the column containing tensors or numbers.

 required
axis int | tuple[int, ...] | None

The axis or axes along which the reduction is computed. - None: Global reduction. Collapses all dimensions (including batch) to a single scalar. - int: Aggregates over both the batch (axis 0) AND the specified tensor dimension. For example, axis=1 collapses the batch and the first dimension of the tensors. - tuple: A sequence of axes that must explicitly include 0.

 None
ddof float

Delta Degrees of Freedom. The divisor used in calculations is $N - ddof$, where $N$ represents the number of elements. Defaults to 1.0 (sample standard deviation).

 1.0
ignore_nulls bool

Whether to ignore null values. Defaults to True.

 True
alias_name str | None

Optional name for the resulting column. Defaults to "std()".

 None

Raises:

Type Description
ValueError

If axis is provided as a tuple but does not include 0.
Note

This aggregator is designed for "reduction" operations. If you wish to calculate statistics per-row without collapsing the batch dimension, use .map() instead.

Example

import ray import numpy as np from ratiopath.ray.aggregate import TensorStd ds = ray.data.from_items( ... [ ... {"m": np.array([[1, 2], [1, 2]])}, ... {"m": np.array([[5, 6], [5, 6]])}, ... ] ... )

1. Global Std (axis=None) -> All elements reduced to one scalar

ds.aggregate(TensorStd(on="m", axis=None))

2. Batch Std (axis=0) -> Result is a 2x2 matrix of std values

calculated across the dataset rows.

ds.aggregate(TensorStd(on="m", axis=0))

3. Int shorthand (axis=1) -> Internally uses axis=(0, 1)

Collapses batch and the first dimension of the tensor.

ds.aggregate(TensorStd(on="m", axis=1))

Source code in ratiopath/ray/aggregate/tensor_std.py 
  9
 10
 11
 12
 13
 14
 15
 16
 17
 18
 19
 20
 21
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
class TensorStd(AggregateFnV2[dict, np.ndarray | float]):
    """Calculates the standard deviation of a column containing Tensors.

    This aggregator treats the data column as a high-dimensional array where
    **axis 0 represents the batch dimension**. To satisfy the requirements
    of a reduction and prevent memory growth proportional to the number of rows,
    axis 0 must be included in the aggregation.

    It uses a parallel variance accumulation algorithm (Chan's method) to maintain
    numerical stability while processing data across multiple Ray blocks.

    Args:
        on: The name of the column containing tensors or numbers.
        axis: The axis or axes along which the reduction is computed.
            - `None`: Global reduction. Collapses all dimensions (including batch)
              to a single scalar.
            - `int`: Aggregates over both the batch (axis 0) AND the specified
              tensor dimension. For example, `axis=1` collapses the batch and
              the first dimension of the tensors.
            - `tuple`: A sequence of axes that **must** explicitly include `0`.
        ddof: Delta Degrees of Freedom. The divisor used in calculations
            is $N - ddof$, where $N$ represents the number of elements.
            Defaults to 1.0 (sample standard deviation).
        ignore_nulls: Whether to ignore null values. Defaults to True.
        alias_name: Optional name for the resulting column. Defaults to "std(<on>)".

    Raises:
        ValueError: If `axis` is provided as a tuple but does not include `0`.

    Note:
        This aggregator is designed for "reduction" operations. If you wish to
        calculate statistics per-row without collapsing the batch dimension,
        use `.map()` instead.

    Example:
        >>> import ray
        >>> import numpy as np
        >>> from ratiopath.ray.aggregate import TensorStd
        >>> ds = ray.data.from_items(
        ...     [
        ...         {"m": np.array([[1, 2], [1, 2]])},
        ...         {"m": np.array([[5, 6], [5, 6]])},
        ...     ]
        ... )
        >>> # 1. Global Std (axis=None) -> All elements reduced to one scalar
        >>> ds.aggregate(TensorStd(on="m", axis=None))
        >>> # 2. Batch Std (axis=0) -> Result is a 2x2 matrix of std values
        >>> # calculated across the dataset rows.
        >>> ds.aggregate(TensorStd(on="m", axis=0))
        >>> # 3. Int shorthand (axis=1) -> Internally uses axis=(0, 1)
        >>> # Collapses batch and the first dimension of the tensor.
        >>> ds.aggregate(TensorStd(on="m", axis=1))
    """

    _aggregate_axis: tuple[int, ...] | None = None

    def __init__(
        self,
        on: str,
        axis: int | tuple[int, ...] | None = None,
        ddof: float = 1.0,
        ignore_nulls: bool = True,
        alias_name: str | None = None,
    ):
        super().__init__(
            name=alias_name if alias_name else f"std({on})",
            on=on,
            ignore_nulls=ignore_nulls,
            zero_factory=self.zero_factory,
        )

        self._ddof = ddof

        if axis is not None:
            axes = {0, axis} if isinstance(axis, int) else set(axis)

            if 0 not in axes:
                raise ValueError(
                    f"Invalid axis configuration: {axis}. Axis 0 (the batch dimension) "
                    "must be included to perform a reduction. To process rows "
                    "independently without collapsing the batch, use .map() instead."
                )

            self._aggregate_axis = tuple(axes)

    @staticmethod
    def zero_factory() -> dict:
        return {"mean": 0, "ssd": 0, "shape": None, "count": 0}

    def aggregate_block(self, block: Block) -> dict:
        block_acc = BlockAccessor.for_block(block)

        # Get exact counts before any NumPy conversion obscures the nulls
        valid_count = cast(
            "int",
            block_acc.count(self._target_col_name, ignore_nulls=True),  # type: ignore [arg-type]
        )
        total_count = cast(
            "int",
            block_acc.count(self._target_col_name, ignore_nulls=False),  # type: ignore [arg-type]
        )

        # Catch nulls immediately if strict mode is on
        if valid_count < total_count and not self._ignore_nulls:
            raise ValueError(
                f"Column '{self._target_col_name}' contains null values, but "
                "ignore_nulls is False."
            )

        if valid_count == 0:
            return self.zero_factory()

        col_np = cast("np.ndarray", block_acc.to_numpy(self._target_col_name))

        # Filter out nulls if necessary
        if valid_count < total_count:
            valid_tensors = [x for x in col_np if x is not None]
            col_np = np.stack(valid_tensors)

        # Partial sum and element count
        block_sum = np.sum(col_np, axis=self._aggregate_axis, keepdims=True)
        block_count = col_np.size // block_sum.size

        mean = block_sum / block_count
        block_ssd = np.sum((col_np - mean) ** 2, axis=self._aggregate_axis)

        return {
            "mean": mean.ravel(),
            "ssd": block_ssd.ravel(),
            "shape": block_ssd.shape,
            "count": block_count,
        }

    def combine(self, current_accumulator: dict, new: dict) -> dict:
        if new["count"] == 0:
            return current_accumulator

        if current_accumulator["count"] == 0:
            return new

        n_current = current_accumulator["count"]
        n_new = new["count"]
        combined_count = n_current + n_new

        mean_current = np.asarray(current_accumulator["mean"])
        mean_new = np.asarray(new["mean"])

        delta = mean_new - mean_current

        # Chan's formula for the combined true mean
        combined_mean = (mean_current * n_current + mean_new * n_new) / combined_count

        combined_ssd = (
            np.asarray(current_accumulator["ssd"])
            + np.asarray(new["ssd"])
            + (delta**2 * n_current * n_new / combined_count)
        )

        return {
            "mean": combined_mean,
            "ssd": combined_ssd,
            "shape": new["shape"],
            "count": combined_count,
        }

    def finalize(self, accumulator: dict) -> np.ndarray | float:  # type: ignore [override]
        count = accumulator["count"]

        if count - self._ddof <= 0:
            return np.nan

        # np.maximum added as a safety net. Floating point jitter can occasionally
        # result in trivially negative numbers (e.g., -1e-16), which crashes np.sqrt
        variance = np.maximum(
            0.0,
            np.asarray(accumulator["ssd"]).reshape(accumulator["shape"])
            / (count - self._ddof),
        )

        return np.sqrt(variance)

__init__(on, axis=None, ddof=1.0, ignore_nulls=True, alias_name=None)

Source code in ratiopath/ray/aggregate/tensor_std.py 
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
def __init__(
    self,
    on: str,
    axis: int | tuple[int, ...] | None = None,
    ddof: float = 1.0,
    ignore_nulls: bool = True,
    alias_name: str | None = None,
):
    super().__init__(
        name=alias_name if alias_name else f"std({on})",
        on=on,
        ignore_nulls=ignore_nulls,
        zero_factory=self.zero_factory,
    )

    self._ddof = ddof

    if axis is not None:
        axes = {0, axis} if isinstance(axis, int) else set(axis)

        if 0 not in axes:
            raise ValueError(
                f"Invalid axis configuration: {axis}. Axis 0 (the batch dimension) "
                "must be included to perform a reduction. To process rows "
                "independently without collapsing the batch, use .map() instead."
            )

        self._aggregate_axis = tuple(axes)

aggregate_block(block)

Source code in ratiopath/ray/aggregate/tensor_std.py 
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
def aggregate_block(self, block: Block) -> dict:
    block_acc = BlockAccessor.for_block(block)

    # Get exact counts before any NumPy conversion obscures the nulls
    valid_count = cast(
        "int",
        block_acc.count(self._target_col_name, ignore_nulls=True),  # type: ignore [arg-type]
    )
    total_count = cast(
        "int",
        block_acc.count(self._target_col_name, ignore_nulls=False),  # type: ignore [arg-type]
    )

    # Catch nulls immediately if strict mode is on
    if valid_count < total_count and not self._ignore_nulls:
        raise ValueError(
            f"Column '{self._target_col_name}' contains null values, but "
            "ignore_nulls is False."
        )

    if valid_count == 0:
        return self.zero_factory()

    col_np = cast("np.ndarray", block_acc.to_numpy(self._target_col_name))

    # Filter out nulls if necessary
    if valid_count < total_count:
        valid_tensors = [x for x in col_np if x is not None]
        col_np = np.stack(valid_tensors)

    # Partial sum and element count
    block_sum = np.sum(col_np, axis=self._aggregate_axis, keepdims=True)
    block_count = col_np.size // block_sum.size

    mean = block_sum / block_count
    block_ssd = np.sum((col_np - mean) ** 2, axis=self._aggregate_axis)

    return {
        "mean": mean.ravel(),
        "ssd": block_ssd.ravel(),
        "shape": block_ssd.shape,
        "count": block_count,
    }

combine(current_accumulator, new)

Source code in ratiopath/ray/aggregate/tensor_std.py 
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
def combine(self, current_accumulator: dict, new: dict) -> dict:
    if new["count"] == 0:
        return current_accumulator

    if current_accumulator["count"] == 0:
        return new

    n_current = current_accumulator["count"]
    n_new = new["count"]
    combined_count = n_current + n_new

    mean_current = np.asarray(current_accumulator["mean"])
    mean_new = np.asarray(new["mean"])

    delta = mean_new - mean_current

    # Chan's formula for the combined true mean
    combined_mean = (mean_current * n_current + mean_new * n_new) / combined_count

    combined_ssd = (
        np.asarray(current_accumulator["ssd"])
        + np.asarray(new["ssd"])
        + (delta**2 * n_current * n_new / combined_count)
    )

    return {
        "mean": combined_mean,
        "ssd": combined_ssd,
        "shape": new["shape"],
        "count": combined_count,
    }

finalize(accumulator)

Source code in ratiopath/ray/aggregate/tensor_std.py 
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
def finalize(self, accumulator: dict) -> np.ndarray | float:  # type: ignore [override]
    count = accumulator["count"]

    if count - self._ddof <= 0:
        return np.nan

    # np.maximum added as a safety net. Floating point jitter can occasionally
    # result in trivially negative numbers (e.g., -1e-16), which crashes np.sqrt
    variance = np.maximum(
        0.0,
        np.asarray(accumulator["ssd"]).reshape(accumulator["shape"])
        / (count - self._ddof),
    )

    return np.sqrt(variance)

zero_factory() staticmethod

Source code in ratiopath/ray/aggregate/tensor_std.py 
94
95
96
@staticmethod
def zero_factory() -> dict:
    return {"mean": 0, "ssd": 0, "shape": None, "count": 0}