Fine-Tuning Dog Silhouette Modeling: Taming Overfitting with L2 Regularization

Against overfitting, I added L2 regularization for the convolution kernel. It should suppress excessive growth of weight values and, consequently, strong reactions to any elements of the image, making it gradual.

1e-4 for 100 epochs

For a network with 25k variables and a regularization coefficient of 1e-4, a mean square equal to 0.4 will be considered an error of 1. And weights above 1 will be:

1e-5 for 50 epochs with softened decay 1e-3 for 100 epochs with intensified decay

Erroneous elements persist with the tested regularization parameters.

Model loss:

Model loss:

Author — Egor Zyryanov

Huggingface, Morevorot, Deviousrage