noether.data.stats¶

Classes¶

`RunningMoments`	Calculates running moments of data (mean, variance, min, max) for data normalization purposes.
`RunningStats`	Calculates statistics of data (min, max, mean and variance) for data normalization purposes.

Module Contents¶

class noether.data.stats.RunningMoments(log_scale=False)¶

Bases: torch.nn.Module

Calculates running moments of data (mean, variance, min, max) for data normalization purposes.

This class implements Welford’s online algorithm for computing mean and variance in a single pass, making it memory-efficient for large datasets.

Initialize a RunningMoments instance.

Parameters:: log_scale (bool) – Whether to apply log scaling to the data before calculating moments. This affects all operations unless overridden.

log_scale = False¶

push_tensor(x, dim=1, log_scale=None)¶

Add a tensor to the moment calculation. Calculations are carried out in float64 to avoid numerical imprecisions.

Parameters:

x (torch.Tensor) – Tensor to push. Shape should be compatible with dim parameter.
dim (int) – Which dimension contains the feature dimension. For example: - In a pointcloud (N, 3), dim=1 calculates statistics per coordinate (x, y, z) - In an RGB image (B, C, H, W), dim=1 calculates per-channel statistics
log_scale (bool | None) – Whether to apply log scaling to this tensor. If None, uses the class default.

Raises:

ValueError – If tensor has invalid dimensions or is incompatible with previous data

Return type:

None

push_scalar(x, log_scale=None)¶

Add a scalar to the moment calculation. Calculations are carried out in float64.

Parameters:

x (float) – Scalar value to add to the moment calculations.
log_scale (bool | None) – Whether to apply log scaling to this scalar. If None, uses the class default.

Raises:

TypeError – If x is not a Python float

Return type:

None

property mean: float | torch.Tensor¶

Return type:: Union[float, torch.Tensor]

property var: float | torch.Tensor¶

Return type:: Union[float, torch.Tensor]

property std: float | torch.Tensor¶

Return type:: Union[float, torch.Tensor]

property min: float | torch.Tensor¶

Return type:: Union[float, torch.Tensor]

property max: float | torch.Tensor¶

Return type:: Union[float, torch.Tensor]

property count: int¶

Return type:: int

print()¶

Print the calculated statistics in a readable format.

Return type:: None

dump()¶

Return type:: dict[str, Union[float, torch.Tensor]]

reset()¶

Reset the statistics, clearing all accumulated data.

Return type:: None

is_empty()¶

Return type:: bool

class noether.data.stats.RunningStats(name=None)¶

Calculates statistics of data (min, max, mean and variance) for data normalization purposes.

Parameters:: name (str | None)

name = None¶

push_tensor(x, dim=1)¶

Add a tensor to the statistics. Calculations are carried out in float64 to avoid numerical imprecision.

Parameters:

x (torch.Tensor)
dim (int)

Return type:

None

property min: torch.Tensor¶

Return type:: torch.Tensor

property max: torch.Tensor¶

Return type:: torch.Tensor

property mean: torch.Tensor¶

Return type:: torch.Tensor

property var: torch.Tensor¶

Return type:: torch.Tensor

property std: torch.Tensor¶

Return type:: torch.Tensor

property logmean: torch.Tensor¶

Return type:: torch.Tensor

property logvar: torch.Tensor¶

Return type:: torch.Tensor

property logstd: torch.Tensor¶

Return type:: torch.Tensor