calibration
#
Module Contents#
Functions#
|
Calculate the percentage of draws from the predicted distribution that |
|
Plot the calibration metric for a grid of p_X percentages, |
- calibration.get_p_Y_val_approx_mahalanobis(post_samples, y_mean, y_truth, cov)[source]#
Calculate the percentage of draws from the predicted distribution that encompasses the truth, for all of the examples in the validation set.
Parameters#
- post_samplesnp.array of shape [n_samples, n_sightlines, Y_dim]
BNN posterior samples
- y_meannp.array of shape [n_lenses, Y_dim]
Central prediction to use in the distance calculation
- y_truth: np.array of shape [n_lenses, Y_dim]
True values
- covfloat
Scale factor to use in the distance calculation
Notes#
Adapted from swagnercarena/ovejero
- calibration.plot_calibration(post_samples, y_mean, y_truth, cov, color_map=['#377eb8', '#4daf4a'], n_perc_points=20, figure=None, ls='--', legend=None, show_plot=True, block=True, title=None, dpi=200)[source]#
Plot the calibration metric for a grid of p_X percentages, with error bars obtained through jackknife sampling
Parameters#
See the docstring for get_p_Y_val_approx_mahalanobis. n_perc_points : int
Grid size of p_X (probability volume) to compare p_Y against
Notes#
Adapted from swagnercarena/ovejero