mighty.trainer.MaskTrainer

class mighty.trainer.MaskTrainer(image_shape, accuracy_measure=AccuracyArgmax(), learning_rate=0.1, show_progress=False)

Interpretable Explanations of Black Boxes by Meaningful Perturbation [1].

Train an occlusion mask that shows where a neural network “looks at” in the input space.

Parameters:

accuracy_measureAccuracy

Accuracy estimator.

image_shapetuple

The shape of an input image.

learning_ratefloat, optional

Optimizer learning rate. Default: 0.1

show_progressbool, optional

Show the training progress bar or not. Default: False

References

[1]

Fong, R. C., & Vedaldi, A. (2017). Interpretable explanations of black boxes by meaningful perturbation. In Proceedings of the IEEE International Conference on Computer Vision (pp. 3429-3437).

Methods

__init__(image_shape[, accuracy_measure, ...])
cpu()
cuda()
get_probability(outputs, label)	Returns the probability of the label class of the outputs.
train_mask(model, image, label_true)	Train a grayscale occlusion mask.

Attributes

l1_coeff
mask_size
max_iterations
tv_beta
tv_coeff

get_probability(outputs, label)

Returns the probability of the label class of the outputs.

Parameters:

outputstorch.Tensor

The output of a model.

labelint

The true class label.

Returns:

float

The probability of outputs[label].

train_mask(model, image, label_true)

Train a grayscale occlusion mask.

Parameters:

modelnn.Module

A neural network model.

image(C, H, W) torch.Tensor

An input image.

label_trueint

The true class label of the image.

Returns:

mask_upsampledtorch.Tensor

The occlusion mask.

image_perturbedtorch.Tensor

The input image with the mask applied.

loss_tracelist of float

A list of training losses.