explanation_model_example.md

Example: Explanation Model Implementation

In this tutorial, we will present an example of how one would implement a custom loss function module into the MAX library. The BaseExplanationModel plays an integral part in the optimization process as it poses as target function of the explanation (optimization problem).

Note that it is explicitly recommended to inheret from the provided base class BaseExplanationModel in order to comply with our API!

BaseExplanationModel Documentation

Similar to the other components, the base class displays the obligatory interface that our new explanation model has to follow. Beside the get_loss method that every loss function must implement, the following attributes have to be parsed and set: x0_generator, lower, upper. These attributes may differ between explanation model formulations and have to be chosen carefully.

An example implementation can be found here: CEMLoss Documentation

    # chose desired component classes for the loss, optimizer and gradient
    # set our 'Contrastive Explanation Method' as optimization algorithm
    loss_class = maxi.CEMLoss
    optimizer_class = maxi.AoExpGradOptimizer
    gradient_class = maxi.URVGradientEstimator

    # specify the configuration for the components
    # set the arguments for our loss method, we chose to generate the pertinent positive
    loss_kwargs = {"mode": "PP", "c": 100, "gamma": 300, "K": 30, "AE": AE}
    optimizer_kwargs = {"l1": 0.1, "l2": 30, "eta": 1.0}
    gradient_kwargs = {"mu": None, "num_iter": 150}

    # instantiate the "ExplanationGenerator" with our settings
    cem = maxi.ExplanationGenerator(
        loss=loss_class,
        optimizer=optimizer_class,
        gradient=gradient_class,
        loss_kwargs=loss_kwargs,
        optimizer_kwargs=optimizer_kwargs,
        gradient_kwargs=gradient_kwargs,
        num_iter=1500,
        save_freq=500,
    )

Back to main page