keras-explain 0.0.1

Explanation toolbox for Keras models.

Keras Explain

This package includes the majority of explanation tools for explaining Keras models predictions. Currently, only models with images on input are supported. It supports following approaches:

Gradient methods:

• GradCam [Selvaraju]
• Guided GradCam [Selvaraju]
• Guided back-propagation [Springenberg]
• Integrated gradients [Sundararajan]
• Saliency [Simonyan]
• Layer-wise relevance propagation [BETA] [Bach]

Model-independent methods:

• Prediction difference [Zintgraf]
• Basic graying out [Zeiler]
• LIME [Ribeiro]

All approaches are easy to apply to your model in two lines of code. If you have any suggestion for new approaches to be included in the package please do not hesitate to suggest. Also all improvements suggestions, bug reports and bug fixes are welcome.

Right now we are in the process of implementing the following approaches:

• Meaningful perturbation by Fong et al.
• Layer-wise relevance propagation - we are adding layers that are not supported yet.

Usage

Gradient methods

GradCam

from keras_explain.grad_cam import GradCam

explainer = GradCam(model, layer=None)
exp = explainer.explain(image, target_class)

Parameters:

• model - Keras model which is explained
• image - input which prediction is explained
• target_class - approach explains prediction for a target class
• layer - (optional) The index (index in model.layers) of the layer which prediction is explained. If not specified the last layer prediction is explained automatically.

Output:

• exp - explanation. GradCam mark only features which contribute to the classification in a target class.

Guided GradCam

from keras_explain.grad_cam import GuidedGradCam

explainer = GuidedGradCam(model, lyer=None)
exp = explainer.explain(image, target_class)

Parameters:

• model - Keras model which is explained
• image - input which prediction is explained
• target_class - approach explains prediction for a target class
• layer - (optional) The index (index in model.layers) of the layer which prediction is explained. If not specified the last layer prediction is explained automatically.

Output:

• exp - explanation. GuidedGradCam mark only features which contribute to the classification in a target class.

Guided back-propagation

from keras_explain.guided_bp import GuidedBP

explainer = GuidedBP(model)
exp = explainer.explain(image, target_class)

Parameters:

• model - Keras model which is explained
• image - input which prediction is explained
• target_class - approach explains prediction for a target class

Output:

• exp - explanation. Guided back-propagation mark only features which contribute to the classification in a target class.

Integrated gradients

from keras_explain.integrated_gradients import IntegratedGradients

explainer = IntegratedGradients(model)
exp = explainer.explain(image, target_class)

Parameters:

• model - Keras model which is explained
• image - input which prediction is explained
• target_class - approach explains prediction for a target class

Output:

• exp - explanation. Integrated gradients mark only features which contribute to the classification in a target class.

Saliency

from keras_explain.saliency import Saliency

explainer = Saliency(model, layer=None)
exp = explainer.explain(image, target_class)

Parameters:

• model - Keras model which is explained
• image - input which prediction is explained
• target_class - approach explains prediction for a target class
• layer - (optional) The index (index in model.layers) of the layer which prediction is explained. If not specified the last layer prediction is explained automatically.

Output:

• exp - explanation. Saliency mark only features which contribute to the classification in a target class.

Layer-wise relevance propagation [BETA]

This approach does not support all layers yet. We are currently implementing missing layers. If you wish you can implement any layer support yourself and submit it as a pull request. Since implementation is very custom any suggestion for improvement is welcome.

from keras_explain.lrp import LRP

explainer = LRP(model)
exp = explainer.explain(image, target_class)

Parameters:

• model - Keras model which is explained
• image - input which prediction is explained
• target_class - approach explains prediction for a target class

Output:

• exp - explanation. LRP mark only features which contribute to the classification in a target class.

###Model independent approaches

Prediction difference

from keras_explain.prediction_diff import PredictionDiff

explainer = PredictionDiff(model)
exp_pos, exp_neg = explainer.explain(image, target_class)

Parameters:

• model - Keras model which is explained
• image - input which prediction is explained
• target_class - approach explains prediction for a target class

Output:

• exp_pos - explanation with marked features which contribute to the classification in a target class.
• exp_neg - explanation with marked features which contribute against the classification in a target class.

Basic graying out

from keras_explain.graying_out import GrayingOut

explainer = GrayingOut(model)
exp_pos, exp_neg = explainer.explain(image, target_class)

Parameters:

• model - Keras model which is explained
• image - input which prediction is explained
• target_class - approach explains prediction for a target class

Output:

• exp_pos - explanation with marked features which contribute to the classification in a target class.
• exp_neg - explanation with marked features which contribute against the classification in a target class.

LIME

from keras_explain.lime_ribeiro import Lime

explainer = Lime(model)
exp_pos, exp_neg = explainer.explain(image, target_class)

Parameters:

• model - Keras model which is explained
• image - input which prediction is explained
• target_class - approach explains prediction for a target class

Output:

• exp_pos - explanation with marked features which contribute to the classification in a target class.
• exp_neg - explanation with marked features which contribute against the classification in a target class.