What Uncertainties Do We Need in Bayesian Deep Learning for Computer Vision?
Pytorch implementation of "What Uncertainties Do We Need in Bayesian Deep Learning for Computer Vision?", NIPS 2017
1. Usage
# Data Tree
config.data_dir/
└── config.data_name/
# Project Tree
WHAT
├── WHAT_src/
│ ├── data/ *.py
│ ├── loss/ *.py
│ ├── model/ *.py
│ └── *.py
└── WHAT_exp/
├── log/
├── model/
└── save/
1.1 Train
# L2 loss only
python train.py --uncertainty "normal" --drop_rate 0.
# Epistemic / Aleatoric
python train.py --uncertainty ["epistemic", "aleatoric"]
# Epistemic + Aleatoric
python train.py --uncertainty "combined"
1.2 Test
# L2 loss only
python train.py --is_train false --uncertainty "normal"
# Epistemic
python train.py --is_train false --uncertainty "epistemic" --n_samples 25 [or 5, 50]
# Aleatoric
python train.py --is_train false --uncertainty "aleatoric"
# Epistemic + Aleatoric
python train.py --is_train false --uncertainty "combined" --n_samples 25 [or 5, 50]
1.3 Requirements
-
Python3.7
-
Pytorch >= 1.0
-
Torchvision
-
distutils
2. Experiment
This is not official implementation.
2.1 Network & Datset
-
Autoencoder based on Bayesian Segnet
- Network depth 2 (paper 5)
- Drop_rate 0.2 (paper 0.5)
-
Fahsion MNIST / MNIST
- Input = Label (for autoencoder)
2.2 Results
2.2.1 PSNR
Combined > Aleatoric > Normal (w/o D.O) > Epistemic > Normal (w/ D.O)
2.2.2 Images
Input / Label
Combined
Aleatoric
Epistemic