l1_l2_kappa_influence.md

l1, l2 & (CEM)kappa Influence on the Explanation Generation

For the provided optimizers, it might is essential to chose the right combination of hyperparameters - namely l1, l2 and kappa. As you will see in this tutorial, the configuration of the latters heavily influences the result that is generated.

Experiment Setup

loss_class = maxi.loss.TF_CEMLoss
optimizer_class = maxi.optimizer.AdaExpGradOptimizer
gradient_class = maxi.gradient.TF_Gradient

loss_kwargs = {"mode": "PP", "c": 1, "gamma": 3, "K": 2}
optimizer_kwargs = {
    "l1": 0.05,
    "l2": 0.005,
    "channels_first": False,
}
gradient_kwargs = {"mu": None}

Original Image

kappa Influence

Kappa = 2

Kappa = 10

Kappa = 20

l1 Influence

l1 = 0.000005

l1 = 0.005

l1 = 5.0

l2 Influence

l2 = 0.0000005

l2 = 0.005

l2 = 5.0