List of Probabilistic Dropout Layers

All dropout layers are nearly identical implementations to their counterparts in Pytorch, based on the work by Gal et al [41]. The difference is these class have a dropping attribute that controls whether or not they are active, rather than relying on the training attribute as Pytorch’s implementations do.

This allows us to more conveniently call the dropout methods on forward calls of a neural network, which is helpful for various computations in uncertainty quantification.

Probabilistic Dropout

Randomly zero out elements.

The ProbabilisticDropout class is imported using the following command:

>>> from UQpy.scientific_machine_learning import ProbabilisticDropout
class ProbabilisticDropout(p=0.5, inplace=False, dropping=True, **kwargs)[source]

Randomly zero out some elements of the input tensor with probability \(p\)

Parameters:

  • p (float) – Probability of an element to be zeroed. Default: 0.5

  • inplace (bool) – If True, will do this operation in-place. Default: False

  • dropping (bool) – If True, will perform dropout, otherwise acts as identity function. Default: True

Shape:

  • Input: Any shape

  • Output: Any shape (same shape as input)

Example:

>>> dropout = sml.ProbabilisticDropout(p=0.75)
>>> input = torch.rand(12, 100)
>>> output = dropout(input)
forward(x)[source]

Calls torch.nn.functional.dropout()

Parameters:

x (Tensor) – Tensor of any shape

Return type:

Tensor

Returns:

Tensor of same shape as x

Probabilistic Dropout 1d

Randomly zero out entire 1d feature maps.

The ProbabilisticDropout1d class is imported using the following command:

>>> from UQpy.scientific_machine_learning import ProbabilisticDropout1d
class ProbabilisticDropout1d(p=0.5, inplace=False, dropping=True, **kwargs)[source]

Randomly zero out entire channels with probability \(p\)

A channel is a 1D feature map

Parameters:

  • p (float) – Probability of a channel to be zeroed. Default: 0.5

  • inplace (bool) – If True, will do this operation in-place. Default: False

  • dropping (bool) – If True, will perform dropout, otherwise acts as identity function. Default: True

Shape:

  • Input: \((N, C, L)\) or \((C, L)\)

  • Output: \((N, C, L)\) or \((C, L)\) (same shape as input)

Example:

>>> dropout = sml.ProbabilisticDropout1d(p=0.6)
>>> input = torch.rand(10, 3, 200)
>>> output = dropout(input)
forward(x)[source]

Calls torch.nn.functional.dropout1d()

Parameters:

x (Tensor) – Tensor of shape \((N, C, L)\) or \((C, L)\)

Return type:

Tensor

Returns:

Tensor of shape \((N, C, L)\) or \((C, L)\) (same shape as x)

Probabilistic Dropout 2d

Randomly zero out entire 2d feature maps.

The ProbabilisticDropout2d class is imported using the following command:

>>> from UQpy.scientific_machine_learning import ProbabilisticDropout2d
class ProbabilisticDropout2d(p=0.5, inplace=False, dropping=True, **kwargs)[source]

Randomly zero out entire channels with probability \(p\)

A channel is a 2D feature map.

Parameters:

  • p (float) – Probability of a channel to be zeroed. Default: 0.5

  • inplace (bool) – If True, will do this operation in-place. Default: False

  • dropping (bool) – If True, will perform dropout, otherwise acts as identity function. Default: True

Shape:

  • Input: \((N, C, H, W)\)

  • Output: \((N, C, H, W)\)

Example:

>>> dropout = sml.ProbabilisticDropout2d(p=0.3)
>>> input = torch.rand(10, 5, 30, 40)
>>> output = dropout(input)
forward(x)[source]

Calls torch.nn.functional.dropout2d()

Parameters:

x (Tensor) – Tensor of shape \((N, C, H, W)\)

Return type:

Tensor

Returns:

Tensor of shape \((N, C, H, W)\)

Probabilistic Dropout 3d

Randomly zero out entire 3d feature maps.

The ProbabilisticDropout3d class is imported using the following command:

>>> from UQpy.scientific_machine_learning import ProbabilisticDropout3d
class ProbabilisticDropout3d(p=0.5, inplace=False, dropping=True, **kwargs)[source]

Randomly zero out entire channels with probability \(p\)

A channel is a 3D feature map.

Parameters:

  • p (float) – Probability of a channel to be zeroed. Default: 0.5

  • inplace (bool) – If True, will do this operation in-place. Default: False

  • dropping (bool) – If True, will perform dropout, otherwise acts as identity function. Default: True

Shape:

  • Input: \((N, C, D, H, W)\) or \((C, D, H, W)\)

  • Output: \((N, C, D, H, W)\) or \((C, D, H, W)\) (same shape as input)

Example:

>>> dropout = sml.ProbabilisticDropout3d(p=0.4)
>>> input = torch.rand(2, 4, 8, 16, 32)
>>> output = dropout(input)
forward(x)[source]

Calls torch.nn.functional.dropout3d()

Parameters:

x (Tensor) – Tensor of shape \((N, C, D, H, W)\) or \((C, D, H, W)\)

Return type:

Tensor

Returns:

Tensor of shape \((N, C, D, H, W)\) or \((C, D, H, W)\) (same shape as x)

Examples