Input Modifiers

class tf_keras_vis.activation_maximization.input_modifiers.InputModifier

Bases: ABC

Abstract class for defining an input modifier.

abstract __call__(seed_input)

Implement modification to the input before processing gradient descent.

Parameters:

seed_input – An numpy.ndarray or a tf.Tensor that indicates a value to input to model.

Returns:

An numpy.ndarray or a tf.Tensor.

Raises:

NotImplementedError – This method must be overwritten.

Return type:

ndarray | Tensor

class tf_keras_vis.activation_maximization.input_modifiers.Jitter(jitter=8)

Bases: InputModifier

An input modifier that introduces random jitter. Jitter has been shown to produce crisper activation maximization images.

Parameters:

jitter – The amount of jitter to apply. Defaults to 8.

class tf_keras_vis.activation_maximization.input_modifiers.Rotate(axes=(1, 2), degree=3.0)

Bases: InputModifier

An input modifier that introduces random rotation.

Parameters:

• axes – The two axes that define the plane of rotation. Defaults to (1, 2).

• degree – The amount of rotation to apply. Defaults to 3.0.

Raises:

ValueError – When axes is not a tuple of two ints.

class tf_keras_vis.activation_maximization.input_modifiers.Rotate2D(degree=3.0)

Bases: Rotate

An input modifier for 2D that introduces random rotation.

Parameters:

degree – The amount of rotation to apply. Defaults to 3.0.

class tf_keras_vis.activation_maximization.input_modifiers.Scale(low=0.9, high=1.1)

Bases: InputModifier

An input modifier that introduces randam scaling.

Parameters:

• low (float, optional) – Lower boundary of the zoom factor. Defaults to 0.9.

• high (float, optional) – Higher boundary of the zoom factor. Defaults to 1.1.