Welcome to py_ciu_image documentation!

This package implements the Contextual Importance and Utility (CIU) method.

Classes:

CIUimage.CIUimage: The CIUimage class implements the Contextual Importance and Utility method for explaining image classification.

Example:

# Example code using the module
import CIUimage as CIU
out_names = ["NonBleeding", "Bleeding"] # Can also be "None".
ciu_object = CIU.CIUimage(model, out_names)
ciu_sp_result = ciu_object.explain(img_to_xplain)

class CIUimage.CIUimage.CIUimage(model, out_names=None, predict_function=None, background_color=(190, 190, 190), strategy='straight', neutralCU=0.5, segments=None, nbr_segments=50, compactness=10, debug=False)

This class implements the Contextual Importance and Utility (CIU) explainable AI method for explaining image classifications.

Parameters:

model – ML model to use.
out_names (list) – List of output class names to be used.
predict.function – Function that takes a list of images and return a numpy.ndarray with output probabilities. If this is None, then it is set to model.predict_on_batch by default.
background_color – Background color to use for “transparent”, in RGB. In the future this will be modified for supporting more than one different colors, patterns or other perturbation methods.
strategy (str) – Defines CIU strategy. Either “straight” or “inverse”.
neutralCU – CU value that is considered “neutral” and that provides a limit between negative and positive influence in the “Contextual influence” calculation CIx(CU - neutralCU).
segments – np.array of same dimensions as image, with segment index for every pixel. The default is None, which signifies that the default SLIC method will be used for creating superpixels.
nbr_segments (int) – The amount of target segments to be used by the SLIC algorithm.
compactness (int) – The compactness of the segments accounting for proximity or RGB values. The default is 10 and logarithmic.
debug (bool) – Display variable values and messages for debugging purposes.

explain(image, strategy=None)

Calculate CIU values for the given image.

Parameters:

image – Image object to explain.
strategy (str) – Defines CIU strategy. Either “straight” or “inverse”. The default is “None”, which causes self.strategy (from constructor) to be used instead.

image_influential_segments_only(ind_output=0, Cinfl_limit=None, type='why', CI_limit=0.5, CU_limit=0.51)

Method that returns a version of the explained image that shows only superpixels above (for type=”why”) or below (for type=”whynot”) threshold values. Threshold values can be either on Contextual influence (Cinfl_limit) OR a combination of CI values equal or over CI_limit and CU values equal or over/under (depending on type) CU_limit.

REMARK: Check proper operation with CI and CU; meanwhile, it’s recommended to use Cinfl_limit instead.

Parameters:

ind_output (int) – Index of output class to use.
Cinfl_limit (float) – Contextual influence limit to use. If None, then use CI&CU limits instead.
type (str) – Can take values “why” or “whynot”.
CI_limit (float) – Inclusion limit (>=) for Contextual Importance (CI).
CU_limit (float) – Inclusion limit (>=) for Contextual Utility (CU).

Welcome to py_ciu_image documentation!

Indices and tables