This repo is for keras version application of the paper,
"A Simple Baseline for Bayesian Uncertainty in Deep Learning" (Maddox et al., 2019).
https://ajlab402.blogspot.com/2021/08/pr402-6-loss-surface-simplexes-for-mode.html

Just to get some basic concepts of the SWA-Gaussian, I brought a simple model from keras tutorial, and changed it a little bit.
Link : https://www.tensorflow.org/tutorials/images/cnn?hl=ko

CIFAR100 dataset is used to test the performance of SWAG. Also you can compare with SWA.

SWAG Keras_application.py

It contains all of the code.
This is just an application of the alogrithm below:

---

---

Training history is as below:
The y-axis is validation accuracy.