Functional utilities

Helpers for constructing attribution baselines and normalizing scores. For practical baseline guidance, see Choosing a baseline.

tslens.get_baseline

get_baseline(inputs: Union[Tensor, Tuple[Tensor, ...]], mode: str = 'zero') -> Union[torch.Tensor, Tuple[torch.Tensor, ...]]

Build a baseline tensor (or tuple of them) matching inputs.

Parameters:

Name	Type	Description	Default
inputs	Union[Tensor, Tuple[Tensor, ...]]	tensor or tuple of tensors shaped (batch, seq_len, features).	required
mode	str	zero, random, normal or mean.	'zero'

tslens.normalize_scale

normalize_scale(data: Tensor, dim: int = 1, norm_type: str = 'standard', legacy: bool = False) -> torch.Tensor

Normalize data along dim.

Parameters:

Name	Type	Description	Default
data	Tensor	tensor to normalize.	required
dim	int	dimension to reduce over.	1
norm_type	str	one of standard, minmax or l1.	'standard'
legacy	bool	reproduce the exact minmax behaviour of the original WinTSR research code, which normalizes the whole tensor by the first slice's min/max instead of per-slice. See :class:tslens.WinTSR for why this switch exists. Ignored for other norm_type values.	False