wingbeats.xai package

Submodules

wingbeats.xai.xai module

Library for Explainable AI

wingbeats.xai.xai.compute_gradients(model, X, target_ind=None)[source]

Compute the gradients of model output at index target_ind w.r.t. input(s) X. If target_ind not provided, the highest activation is considered.

Parameters

  • model (tf.Model) – Model to compute gradients on.

  • X (list or array) – Signal or batch of signals.

  • target_ind (int, optional) – Class index w.r.t. which to compute the gradients. Defaults to None (highest activation).

Returns

Gradients and also the prediction(s).

wingbeats.xai.xai.flip_pixels(model, image, target_ind, means, std_devs, percent_to_flip=0.25, percent_step=0.005, heatmap=None)[source]

Apply Pixel Flipping to an image of dimensions 1 x m x n x 3 of Tensor type. Percent_to_flip % of the pixels are chosen in percent_step % batches to be replaced with a randomly chosen value from the normal distribution with params. means and std_devs (one for every pixel). Parameter heatmap specifies the attention map type that is applied (saliency_map, guided_back_prop or None for random flipping). The model needs to be stripped of its softmax layer.

Parameters

  • model (tf.Model) – Model to make predictions.

  • image (array or Tensor) – Input image to flip pixels on.

  • target_ind (int) – Class index w.r.t. which to compute the gradients.

  • means (array) – Matrix of mean values.

  • std_devs (array) – Matrix of standard deviations.

  • percent_to_flip (float) – Percent of pixels to flip in total. Defaults to 0.25.

  • percent_step (float) – Percentual increment of flipped pixels. Defaults to 0.005.

  • heatmap (array.) – Attention map according to which to flip the pixels. Defaults to None (random flipping).

Returns

Prediction scores w.r.t. every species and every flipping step, indexes of the flipped pixels (as 1D array) and output image after flipping the pixels

wingbeats.xai.xai.guided_back_prop(analyzer, model, X, target_ind=None)[source]

Compute the GBP map(s) of model output(s) w.r.t. input(s) X.

Parameters

  • analyzer (Class created by factory method from innvestigate.) – Gradient based analyzer, i.e. from innvestigate.innvestigate.analyzer import gradient_based as G; analyzer = G.GuidedBackprop(model).

  • model (tf.Model) – Model to compute gradients on.

  • X (list or array) – Signal or batch of signals.

  • target_ind (int, optional) – Class index w.r.t. which to compute the gradients. Defaults to None (highest activation).

Returns

Heatmap(s) with the same dimensions as the input sample and the corresponding predictions.

wingbeats.xai.xai.integral_approximation(gradients)[source]

Approximate integral with the trapezoid rule.

Parameters

gradients (array) – Discrete gradient values to integrate.

Returns

Integral approximation

wingbeats.xai.xai.integrated_gradients(model, x, baseline, steps=50, target_ind=None)[source]

Compute attention map through the Integrated Gradients method.

Parameters

  • model (tf.Model) – Model to compute gradients on.

  • x (list or array) – Signal or batch of signals.

  • baseline (array) – Starting image i.e. zero image for interpolation.

  • steps (int) – Number of discrete steps (= number of interpolated images). Defaults to 50.

  • target_ind (int, optional) – Class index w.r.t. which to compute the gradients. Defaults to None (highest activation).

wingbeats.xai.xai.interpolate_images(baseline, image, alphas)[source]

Compute a scale of images of increasing intensity between baseline and original image.

Parameters

  • baseline (array) – Starting image i.e. zero image.

  • image (array) – Original (target) image.

  • alphas – Incremental intensity steps.

Type

list

Returns

Images of increasing intensities.

Return type

array

wingbeats.xai.xai.quantile_maps(heatmaps, quantiles, nbins=50)[source]

Compute quantile maps out of the pixel distributions from a list of heatmaps.

Parameters

  • heatmaps (list) – Distribution of heatmaps to compute statistics on.

  • quantiles (list) – Quantiles to be computed i.e. 0.25, 0.5, 0.75.

  • nbins (int) – Number of bins in histograms. Defaults to 50.

Returns

Quantile heatmaps.

wingbeats.xai.xai.saliency_map(model, X, target_ind=None)[source]

Compute the saliency map(s) of model output(s) w.r.t. input(s) X.

Parameters

  • model (tf.Model) – Model to compute gradients on.

  • X (list or array) – Signal or batch of signals.

  • target_ind (int, optional) – Class index w.r.t. which to compute the gradients. Defaults to None (highest activation).

Returns

Heatmap(s) with the same dimensions as the input sample and the corresponding predictions.

Module contents