Closed jeremyfix closed 4 years ago
No no , actually , this is not an issue ; The FashionMnist dataset is already normalized in pytorch which is unfortunate from a pedagogical point of view.
So, since FashionMnist is normalized, we have the expected consistent behaviors for the training. Were the dataset not normalized we could have illustrated non consistent behaviors of bad conditioned optimization problem.
I do not know what I did but the illustrations on the interest of using normalization are not the right ones;
In the subject, the figures with "raining accuracy for logistic regression without normalization of the input" are actually the ones with normalization !!!
also there is some missing texts when introducing the normalization ; There is no sentence pointing the difference;