altx.extract_methods

Statistical feature extraction methods for the altx algorithm.

type altx.extract_methods.ExtrMethod = list[str] | list[str | float]

type altx.extract_methods.ExtrMethods = list[ExtrMethod]

class altx.extract_methods.ExtractMethods

Bases: object

Provide statistical feature extraction methods for Altx.

Implement the most common statistical methods used for feature extraction in the Adaptive Law-Based Transformation.

extract(F, extr_methods, device)

Extract features from F with the given extraction methods.

nth_moment(percentiles, n)

Calculate the n-th moment of percentiles along dimension 1.

excess_kurtosis(percentiles)

Calculate the excess kurtosis along dimension 1.

static excess_kurtosis(percentiles)

Calculate the excess kurtosis of percentiles along dimension 1.

Parameters:

percentiles (Tensor) – Input tensor of pre-computed percentiles.

Returns:

The excess kurtosis of the computed percentiles.

Return type:

Tensor

Examples

>>> import torch
>>> from altx import ExtractMethods
>>> p = torch.tensor([[[1.], [2.], [3.], [4.], [5.]]])
>>> print(ExtractMethods.excess_kurtosis(p))
tensor([[-1.3000]])

static extract(F, extr_methods, device='cpu')

Extract features from F using the given extraction methods.

Parameters:

• F (Tensor) – Input tensor.

• extr_methods (list[list[str] | list[str | float]]) – Each element is either a one-element list [method] or a two-element list [method, percentile].

• device (device | str) – The device to calculate on. Default is CPU.

Returns:

The tensor of the collected features.

Return type:

Tensor

Raises:

ValueError – If the given extraction method is not implemented.

Notes

The return tensor has the shape (n, m), where n is the number of used extraction methods and m is the size of the input tensor along the third dimension.

Examples

>>> import torch
>>> from altx import ExtractMethods
>>> F = torch.ones(5, 20, 2)
>>> print(ExtractMethods.extract(F, [["mean", 0.5]]))
tensor([[1., 1.]])
>>> print(ExtractMethods.extract(F, [["var", 0.5]]))
tensor([[0., 0.]])
>>> # Note: Here None has to be given for the `mean_all`
>>> print(ExtractMethods.extract(F, [["mean_all", None]]))
tensor([[1., 1.]])
>>> _ = torch.manual_seed(0)
>>> F2 = torch.randn(5, 20, 1).abs() + 0.5
>>> methods = [["mean", 0.05], ["var", 0.1], ["mean_all", None]]
>>> print(ExtractMethods.extract(F2, methods))
tensor([[0.4577],
        [0.0499],
        [2.1054]])

static nth_moment(percentiles, n=4)

Calculate the n-th moment of percentiles along dimension 1.

Parameters:

• percentiles (Tensor) – Input tensor of pre-computed percentiles.

• n (int) – The order of the moment to compute. Default is 4.

Returns:

The n-th moment of the computed percentiles.

Return type:

Tensor

Examples

>>> import torch
>>> from altx import ExtractMethods
>>> p = torch.tensor([[[1.], [2.], [3.], [4.], [5.]]])
>>> print(ExtractMethods.nth_moment(p, n=2))
tensor([[2.]])
>>> print(ExtractMethods.nth_moment(p, n=4))
tensor([[6.8000]])