README-CONFIGURATION.md

⚙️ Configure ⚙️

The tool can be configured using the files:

config/config_dirs.py

This file defines the location of the inputs and the outputs of the tool.

config/config_data.py

Ths file defines configurations for some constants used by the tool.

EXPECTED_LABEL: The label to use to filter the data for the experiment.

NUM_INPUTS: Number of inputs from data set to be consider for comparison.

INPUT_MAXLEN: The size of the texts with paddings in the dataset.

VOCAB_SIZE: The size of dictionary used for text tokenization.

USE_RGB: If use RGB images or grey scaled.

config/config_general.py

Ths file defines configurations common to both the heatmaps and the featuremaps.

CLUSTERS_SORT_METRIC: The preference for the clusters when sampling them to show some of their images. If no sorting is provided, the tool draw a random sample.

CLUSTER_SIMILARITY_METRIC: The similarity metric to use when comparing different clusters.

config/config_heatmaps.py

This fle defines the configuration for the heatmaps.

APPROACH: The processing mode to use when generating the heatmaps [Original, LocalLatentMode, GlobalLatentMode].

EXPLAINERS: The list of explainers to use when generating the contributions.

DIMENSIONALITY_REDUCTION_TECHNIQUES: The dimensionality reduction techniques to use to project the contributions in the two-dimensional latent space. The tool will experiment with the different techniques and choose the best configuration according to the silhouette score of the corresponding clusters.

CLUSTERING_TECHNIQUE: The clustering technique to use when grouping the contributions.

ITERATIONS: The number of iterations to use when running the experiment.

config/config_featuremaps.py

This fle defines the configuration for the featuremaps.

CASE_STUDY: MNIST or IMDB.

NUM_CELLS: The size of the featuremaps.

BITMAP_THRESHOLD: The threshold for luminosity metric computation.

ORIENTATION_THRESHOLD: The threshold for orientation metric computation.

FEATUREMAPS_CLUSTERS_MODE: The clustering technique to use on the featuremaps [ORIGINAL, REDUCED, CLUSTERED].

MAP_DIMENSIONS: The list of dimensions to be considered for generating feature maps.

config/config_outputs.py

Ths file defines configurations for visualisations.

IMG_SIZE: The size of image of an input.

NUM_SAMPLES: The number of samples for human study.

ORIENTATION_THRESHOLD: the threshold for orientation metric computation.