brain-research / ncp

Reliable Uncertainty Estimates in Deep Neural Networks using Noise Contrastive Priors

Noise Contrastive Priors

This project contains the source code for Noise Contrastive Priors (NCP) by Danijar Hafner, Dustin Tran, Alex Irpan, Timothy Lillicrap, James Davidson.

Running the code

Install dependencies:

pip3 install numpy tensorflow tensorflow_probability matplotlib ruamel.yaml

Active learning on the toy data set:

python3 -m ncp.scripts.toy_active --seeds 20 --logdir /path/to/logdir

Active learning on the flights delay data set:

python3 -m ncp.scripts.flights_active \
  --seeds 10 \
  --dataset /path/to/dataset \
  --logdir /path/to/logdir

The data set path should be a file prefix for four Numpy files named <dataset>-train-inputs.npy, <dataset>-train-targets.npy, <dataset>-test-inputs.npy, <dataset>-test-targets.npy.