This repository contains the reproduction of our performance benchmarks (table 1) in the paper: Uncertainty Wizard; Fast and User-Friendly Neural Network Uncertainty Quantification (preprint available upon request).
For training, the scripts run_in_ubu_2xgpu.py and run_on_win_2080_ti.py were used.
So created artefacts were manually moved into an according subfolder in results/.
The final results (as shown an the paper and in this readme below) can be extracted using table_plotter.py
On the ubuntu machine, the training scripts were executed in a docker container, created using the Dockerfile
in this repository.
On the windows machine, a python 3.6 (64 bit) environment was used.
For questions regarding uncertainty-wizard please refer to the
library repository.
CPU = Threadripper 1920X, GPU 0 = GTX 1060, GPU 1 = GTX 1070Ti
Context: mainprocess
Total training time: 5h 10min
Avg load CPU: 8.9%
Avg load GPU 0: 0.0%
Avg load GPU 1: 45.42%
Context: multiprocess
Total training time: 2h 57min
Avg load CPU: 14.28%
Avg load GPU 0: 0.09%
Avg load GPU 1: 92.24%
Context: multigpu
Total training time: 1h 41min
Avg load CPU: 26.77%
Avg load GPU 0: 98.19%
Avg load GPU 1: 92.41%
CPU = i7 9700, GPU 0 = RTX 2080Ti
Context: mainprocess
Total training time: 3h 7min
Avg load CPU: 14.74%
Avg load GPU 0: 46.93%
Context: multiprocess
Total training time: 2h 32min
Avg load CPU: 50.31%
Avg load GPU 0: 83.9%