search

NoahVl / Explaining-In-Style-Reproducibility-Study

Re-implementation of the StylEx paper, training a GAN to explain a classifier in StyleSpace, paper by Lang et al. (2021).
Other
36 stars 8 forks source link

KL Div 0 #5

Closed NoahVl closed 2 years ago

NoahVl commented 2 years ago

The KL Divergence loss seems to be 0 for the MNIST one vs all classifier for all timesteps. It used to be a bit higher than 0 when we used 10 classes instead of 2. It might just be caused by the high accuracy of the classification model, however it seems strange that it is 0 for all timesteps.

NoahVl commented 2 years ago

Fixed in the aptly named: https://github.com/NoahVl/Explaining-In-Style-Reproducibility-Study/commit/e169aaa927ccfdfbf0379bff885bc8dd8073816f log_target was not yet used.