index.rst

keravis API Documentation

Convolutional layer activations

conv_layer_activations(model, layer, test_img, nested_model=None, title=None)

Visualize activations of layer corresponding to test_img in a grid

Args:

model : keras.Model

Model.

layer : str

Layer whose activations to visualize.

test_img : ndarray

Image for which to look at activations of.

nested_model : str, default=None

Name of nested model, if any.

title : str, default=None

Title of the figure.

2-dimensional feature space representations

feature_space(model, dataset=None, X=None, y=None, kind='tsne', title=None)

Visualize feature space of model on a set of images X in 2-dimensional space using tSNE or PCA

Args:

model : keras.Model

Model.

dataset : keras.preprocessing.image.DataIterator, default=None

Batched dataset. If given, X and y are ignored.

X : ndarray, default=None

Set of images.

y : ndarray, default=None

Set of labels.

kind : str, default='tsne'

Type of plot. One of 'tsne' or 'pca'.

title : str, default=None

Title of the figure.

Saliency maps

saliency_backprop(model, test_img, class_idx=0, title=None)

Visualize the saliency map of test_img using vanilla backprop

Args:

model : keras.Model

Model.

test_img : ndarray

Image for which to find saliency map of.

class_idx : int, default=0

Class index of image.

title : str, default=None

Title of the figure.

saliency_guided_backprop(model, test_img, class_idx=0, title=None)

Visualize the saliency map of test_img using guided backprop

Args:

model : keras.Model

Model.

test_img : ndarray

Image for which to find saliency map of.

class_idx : int, default=0

Class index of image.

title : str, default=None

Title of the figure.

saliency_occlusion(model, test_img, class_idx=0, title=None)

Visualize the saliency map of test_img using occlusion

Args:

model : keras.Model

Model.

test_img : ndarray

Image for which to find saliency map of.

class_idx : int, default=0

Class index of image.

title : str, default=None

Title of the figure.

Generated image that maximally activates classifier output

maximal_class_score_input(model, class_idx, dim, title=None)

Visualize a generated image corresponding to a maximal class score of class_idx

Args:

model : keras.Model

Model.

class_idx : int

Class index for which to find maximally activating image.

dim : tuple

(width,height,channels) of generated image.

title : str, default=None

Title of the figure.

Patches in a set of images that maximally activate an intermediate neuron

maximally_activating_patches(model, layer, dataset=None, X=None, nested_model = None, channel=None, title=None)

Visualizes maximally activating patches in X of a random intermediate neuron in layer, channel

Args:

model : keras.Model

Model.

layer : str

Layer whose activations to visualize.

dataset : keras.preprocessing.image.DataIterator, default=None

Batched dataset. If given, X and y are ignored.

X : ndarray, default=None

Set of images.

nested_model : str, default=None

Name of nested model, if any.

channel : int, default=None

Channel index. If not given, channel is randomly sampled.

title : str, default=None

Title of the figure.