Layer-level-control-of-DNN-training
This repository contains the code used to create all the figures of our paper "On layer-level control of DNN training and its impact on generalization", by Simon Carbonnelle and Christophe De Vleeschouwer (arxiv link).
Code is structured according to the paper: folders have the same naming as their corresponding section.
Code was run with tensorflow-gpu 1.4.0 and keras 2.1.2
The training, test and layer-wise angle deviation curves of all the models that we trained are also available in the pickle files (*.p). Don't hesitate to explore them as not everything is presented in our paper.