lampe.masks#

Masking helpers.

Functions#

`mask2str`	Represents a binary mask as a string.
`str2mask`	Parses the string representation of a binary mask into a tensor.

Classes#

`BernoulliMask`	Creates a distribution \(P(b)\) over all binary masks \(b\) in the hypercube \(\{0, 1\}^D\) such that each bit \(b_i\) has a probability \(p_i\) of being positive.
`SelectionMask`	Creates a mask distribution \(P(b)\), uniform over a selection of binary masks \(\mathcal{B} \subseteq \{0, 1\}^D\).

Descriptions#

lampe.masks.mask2str(b)#

Represents a binary mask as a string.

Parameters: b (BoolTensor) – A binary mask \(b\), with shape \((D,)\).

Example

>>> b = torch.tensor([True, True, False, True, False])
>>> mask2str(b)
'11010'

lampe.masks.str2mask(string)#

Parses the string representation of a binary mask into a tensor.

Parameters: string (str) – A binary mask string representation.

Example

>>> str2mask('11010')
tensor([True, True, False, True, False])

class lampe.masks.BernoulliMask(p)#

Creates a distribution \(P(b)\) over all binary masks \(b\) in the hypercube \(\{0, 1\}^D\) such that each bit \(b_i\) has a probability \(p_i\) of being positive.

\[P(b) = \prod^D_{i = 1} p_i^{b_i} (1 - p_i)^{1 - b_i}\]

Parameters: p (Tensor) – The probability vector \(p\), with shape \((*, D)\).

Example

>>> d = BernoulliMask(torch.tensor([0.4, 0.1, 0.9]))
>>> d.sample()
tensor([True, False, True])

class lampe.masks.SelectionMask(selection)#

Creates a mask distribution \(P(b)\), uniform over a selection of binary masks \(\mathcal{B} \subseteq \{0, 1\}^D\).

\[\begin{split}P(b) = \begin{cases} \frac{1}{|\mathcal{B}|} & \text{if } b \in \mathcal{B} \\ 0 & \text{otherwise} \end{cases}\end{split}\]

Parameters: selection (BoolTensor) – A binary mask selection \(\mathcal{B}\), with shape \((N, D)\).

Example

>>> selection = torch.tensor([
...     [True, False, False],
...     [False, True, False],
...     [False, False, True],
... ])
>>> d = SelectionMask(selection)
>>> d.sample()
tensor([False, True, False])