thudzj/DE-GP

Code for "Deep Ensemble as a Gaussian Process Posterior"

Deep Ensemble as a Gaussian Process Posterior [arXiv]

None

How to train DE-GP

UCI regression:

    python uci.py

FashionMNIST classification:

For DE

    python fmnist.py --method free

For rDE

    python fmnist.py --method reg

For RMS

    python fmnist.py --method anc

For DE-GP

    python fmnist.py --method our

For MC dropout

    python fmnist.py --method free --n_ensemble 1 --dropout 0.3 --n_dropout_inf 10

For snapshot ensemble

    python fmnist.py --method free --n_ensemble 1 --n_snapshots 10 --epochs 4

For logits-space deep ensemble

    python fmnist.py --method free --logits_mean

Requirement

• Python 3.6+
• Pytorch>=1.4.0
• theano==1.0.3, pymc3==3.5 (for VI and HMC baselines)
• jax==0.2.12, jaxlib==0.1.65, neural-tangents==0.3.6 (for NN-GP baseline)
• genrl