attribench.masking.image.ImageMasker

class attribench.masking.image.ImageMasker(masking_level)[source]

Bases: Masker

Abstract base class for all image maskers.

Image maskers are specific maskers for image data. They can be used to mask features on the feature level or the pixel level. If the masker is used on the feature level, the attributions must have the same shape as the samples. If the masker is used on the pixel level, the attributions must have the same shape as the samples, except the channel dimension must be 1.

If the masker is used on the feature level, masking a feature means masking a specific input feature for all samples (i.e. one color value for one pixel). If the masker is used on the pixel level, masking a feature means masking a specific pixel for all samples (i.e. all color values for one pixel).

If segmented samples are provided, the masker will use the segments to mask features. This means that masking a feature will mask all pixels belonging to the same segment. The attribution value of a segment is defined as the average attribution value of all features in that segment.

Parameters:

masking_levelstr: Either "feature" or "pixel". If "feature", the masker will mask features. If "pixel", the masker will mask pixels.

Raises:

ValueError: If masking_level is not "feature" or "pixel".

Methods

`get_num_features`	Return the number of features in the samples.
`mask_bot`	Mask the `k` least important features, according to the attributions.
`mask_rand`	Mask `k` random features.
`mask_top`	Mask the `k` most important features, according to the attributions.
`set_batch`	Set the batch of samples and attributions to use for masking.

get_num_features()[source]

Return the number of features in the samples.

Returns:

int: Number of features in the samples.

mask_bot(k)[source]

Mask the k least important features, according to the attributions.

Parameters:

kint: Number of features to mask.

Returns:

torch.Tensor: Samples with the bottom k features masked.

mask_rand(k, return_indices=False)[source]

Mask k random features.

Parameters:

kint: Number of features to mask.
return_indicesbool, optional: Whether to return the indices of the masked features, by default False

Returns:

torch.Tensor: Samples with k random features masked.

Return type:

Union[Tensor, Tuple[Tensor, Tensor]]

mask_top(k)[source]

Mask the k most important features, according to the attributions.

Parameters:

kint: Number of features to mask.

Returns:

torch.Tensor: Samples with the top k features masked.

set_batch(samples, attributions=None, segmented_samples=None)[source]

Set the batch of samples and attributions to use for masking. Optionally also set the segmented samples.

Parameters:

samplestorch.Tensor: Samples to use for masking.
attributionstorch.Tensor, optional: Attributions to use for masking, by default None If None, the mask_top() and mask_bot() methods will not be available.
segmented_samplestorch.Tensor, optional: Segmented samples to use for masking, by default None