maskers

Classes:

Name	Description
BatchwiseChannelwisePatchwisePercentMasker	Constructs a mask based on a percentage of the highest values in the input tensor per batch element, per color channel, per patch.
BatchwisePercentMasker	Constructs a mask based on a percentage of the highest values in the input tensor per batch element.
Masker	Defines an interface for masking input values.
PercentMasker	Constructs a mask based on a percentage of the highest values in the input tensor.
ThresholdMasker	Defines an interface for constructing a mask from a threshold computed over an input tensor.

BatchwiseChannelwisePatchwisePercentMasker

Bases: PercentMasker

Constructs a mask based on a percentage of the highest values in the input tensor per batch element, per color channel, per patch.

This class is designed to be used with 2D convolutional layers that output tensors of shape (batch_size, color_channels * patch_area, patches_high, patches_wide).

Parameters:

Name	Type	Description	Default
percent_to_mask	float	A scalar value between 0 and 1 inclusive, indicating the percentage of the largest input values to mask.	required
color_channels	int	The number of color channels in the input tensor.	required
patch_size	int | tuple[int, int] | Sequence[int]	The dimension of the non-overlapping rectangular patches the image is tiled with.	required

Methods:

Name	Description
forward	Construct a mask over the input tensor.
reset_parameters	Reinitialize parameters and buffers.

forward

forward(
    input: Tensor, selector_mask: Tensor | None = None
) -> torch.Tensor

Construct a mask over the input tensor.

Parameters:

Name	Type	Description	Default
input	Tensor	The tensor over which to compute the thresholds.	required
selector_mask	Tensor | None	Unsupported. In the parent class, a mask tensor to use to select a subset of the elements of input when computing the threshold. torch.Tensor.masked_select results in a flattened tensor, and because BatchwiseChannelwisePatchwisePercentMasker operates per patch, per color channel, it cannot accept a flattened input.	None

Returns:

Type	Description
torch.Tensor	A mask tensor of the same shape as input.

Raises:

Type	Description
ValueError	If selector_mask is not None.

reset_parameters

reset_parameters() -> None

Reinitialize parameters and buffers.

This method is useful for initializing tensors created on the meta device.

Raises:

Type	Description
ValueError	If percent_to_mask is not a scalar.
ValueError	If percent_to_mask is not between 0 and 1 inclusive.

BatchwisePercentMasker

Bases: PercentMasker

Constructs a mask based on a percentage of the highest values in the input tensor per batch element.

Methods:

Name	Description
forward	Construct a mask over the input tensor.
reset_parameters	Reinitialize parameters and buffers.

forward

forward(
    input: Tensor, selector_mask: Tensor | None = None
) -> torch.Tensor

Construct a mask over the input tensor.

Parameters:

Name	Type	Description	Default
input	Tensor	The tensor over which to compute the threshold(s).	required
selector_mask	Tensor | None	A mask tensor to use to select a subset of the elements of input when computing the threshold. If None, all elements of input are used.	None

Returns:

Type	Description
torch.Tensor	A mask tensor of the same shape as input.

reset_parameters

reset_parameters() -> None

Reinitialize parameters and buffers.

This method is useful for initializing tensors created on the meta device.

Raises:

Type	Description
ValueError	If percent_to_mask is not a scalar.
ValueError	If percent_to_mask is not between 0 and 1 inclusive.

Masker

Bases: Module, ABC

Defines an interface for masking input values.

Maskers should return a boolean tensor of the same shape as the input tensor.

Added in version 0.30.0.

Methods:

Name	Description
reset_parameters	Reinitialize parameters and buffers.

reset_parameters abstractmethod

reset_parameters() -> None

Reinitialize parameters and buffers.

This method is useful for initializing tensors created on the meta device.

PercentMasker

Bases: ThresholdMasker

Constructs a mask based on a percentage of the highest values in the input tensor.

Parameters:

Name	Type	Description	Default
percent_to_mask	float	A scalar value between 0 and 1 inclusive, indicating the percentage of the largest input values to mask.	required

Methods:

Name	Description
forward	Construct a mask over the input tensor.
reset_parameters	Reinitialize parameters and buffers.

forward

forward(
    input: Tensor, selector_mask: Tensor | None = None
) -> torch.Tensor

Construct a mask over the input tensor.

Parameters:

Name	Type	Description	Default
input	Tensor	The tensor over which to compute the threshold(s).	required
selector_mask	Tensor | None	A mask tensor to use to select a subset of the elements of input when computing the threshold. If None, all elements of input are used.	None

Returns:

Type	Description
torch.Tensor	A mask tensor of the same shape as input.

reset_parameters

reset_parameters() -> None

Reinitialize parameters and buffers.

This method is useful for initializing tensors created on the meta device.

Raises:

Type	Description
ValueError	If percent_to_mask is not a scalar.
ValueError	If percent_to_mask is not between 0 and 1 inclusive.

ThresholdMasker

Bases: Masker

Defines an interface for constructing a mask from a threshold computed over an input tensor.

Methods:

Name	Description
forward	Construct a mask over the input tensor.
reset_parameters	Reinitialize parameters and buffers.

forward

forward(
    input: Tensor, selector_mask: Tensor | None = None
) -> torch.Tensor

Construct a mask over the input tensor.

Parameters:

Name	Type	Description	Default
input	Tensor	The tensor over which to compute the threshold(s).	required
selector_mask	Tensor | None	A mask tensor to use to select a subset of the elements of input when computing the threshold. If None, all elements of input are used.	None

Returns:

Type	Description
torch.Tensor	A mask tensor of the same shape as input.

reset_parameters abstractmethod

reset_parameters() -> None

Reinitialize parameters and buffers.

This method is useful for initializing tensors created on the meta device.