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:

docker build https://github.com/irmlma/uncertainty-quantification-snf.git#<TAG> -t uqma

where is, e.g., v0.1.1.

Example usage

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

You can then run the container using

docker run uqma --help

Train the model using the provided config file via:

docker run -v <<some path>>:/mnt \
  uqma \
  --mode=train \
  --config=/mnt/<<config.py>> \
  --infile=/mnt/<<train_dataset.csv>> \
  --outfile=/mnt/<<outfile.pkl>>

where

• <<some path> is a local path you want to mount to /mnt/ to make it accessible to Docker,
• <<config.py>> is a config file that is following the template in configs/config.py,
• <<train_dataset.csv>> is a comma-separated file of numerical values which correspond to the features obtained from transforming inputs through a neural network,
• <<outfile.pkl>> is the outfile to which parameter and meta data is saved.

To make predictions for epistemic uncertainty estimates, call:

docker run -v <<some path>>:/mnt \
  uqma \
  --mode=predict \
  --config=/mnt/<<config.py>> \
  --infile=/mnt/<<test_dataset.csv>> \
  --outfile=/mnt/<<outfile.pkl>> \
  --checkpopint=/mnt/<<checkpoint>>

where

• <<some path> is the same as above,
• <<config.py>> is the same as above,
• <<test_dataset.csv>> is a data set for which you want to evaluate if it is OoD,
• <<outfile.pkl>> is the name of the outfile,
• <<checkpoint>> is the parameter file obtained through the training (i.e., in this case <<outfolder>>/params.pkl).

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