peal.visualization.visualize_counterfactual_gradients¶
Debug grids for gradient-based counterfactual search.
While a counterfactual explainer optimises an image (or a generator latent)
against the predictor, visualize_step() writes one labelled image grid
per step showing the input, the encoded/decoded images, the pixel and latent
gradients as heatmaps, the masks in play and the best counterfactual found so
far. create_label_image() renders the row captions as image tensors so
the whole grid can be saved with torchvision.utils.save_image.
Functions
Render a text caption as an image tensor of a given size. |