Uncertainty quantification for mobility analysis

About

This repository contains library code for training a surjective normalizing flow for out-of-distribution detection using epistemic uncertainty estimates.

Installation

To install the latest GitHub , just call the following on the command line:

pip install git+https://github.com/irmlma/uncertainty-quantification-for-mobility-analysis@<TAG>

This installs the library as well as executables in your current (virtual) environment.

Usage

Having installed as described above you can train and make predictions using the provided exectuables.

Train the model using the provided config file via:

uqma-train
    --config=configs/config.py \
    --infile=<FILE> \
    --outfile=<PARAMS_FILE>

where is a COMMA-separated file of numerical values which correspond to the features obtained from transforming inputs through a neural network that has residual connections and was trained using spectral-normalization and <PARAMS_FILE> is some file to which parameters are written (see Dirmeier et al. (2023)).

To make predictions for epistemic uncertainty estimates, call:

uqma-predict
    --params=<PARAMS_FILE> \
    --infile=<FILE> \
    --outfile=<PARAMS_FILE>

where <PARAMS_FILE> is the same file as before, is a features file and is a file where results are written to.

Citation

If you find our work relevant to your research, please consider citing

@article{dirmeier2023uncertain,
  title={Uncertainty quantification and out-of-distribution detection using surjective normalizing flows},
  author={Simon Dirmeier and Ye Hong and Yanan Xin and Fernando Perez-Cruz},
  year={2023},
  journal={arXiv preprint arXiv:2311.00377}
}

Author

Simon Dirmeier sfyrbnd @ pm me