[TensorFlow 2] A Simple Baseline for Bayesian Uncertainty in Deep Learning: SWA-Gaussian (SWAG)

TensorFlow implementation of "A Simple Baseline for Bayesian Uncertainty in Deep Learning"

Concept

Algorithm to utilize the SWAG [1].

Equation for the weight sampling from SWAG [1].

The red color and the blue color represent the initial state and current state respectively.

Method	Accuracy	Precision	Recall	F1-Score
Final Epoch	0.99230	0.99231	0.99222	0.99226
Best Loss	0.99350	0.99350	0.99338	0.99344
SWAG (S = 30)	0.99310	0.99305	0.99299	0.99302
SWAG (Last Momentum)	0.99340	0.99340	0.99330	0.99335

Method	Accuracy	Precision	Recall	F1-Score
Final Epoch	0.73130	0.73349	0.73130	0.73147
Best Loss	0.73240	0.73205	0.73240	0.73099
SWAG (S = 30)	0.74100	0.74622	0.74100	0.74260
SWAG (Last Momentum)	0.73490	0.73888	0.73490	0.73561

[1] Wesley Maddox et al. (2019). A Simple Baseline for Bayesian Uncertainty in Deep Learning. arXiv preprint arXiv:1902.02476.