Infinite-Task Learning

Codecov CircleCI

With applicatons to Infinite Quantile Regression, Infinite Level Set Estimation, and Infinite Joint Cost Sensitive Learning.


  • TexLive (full) 2016 or later
  • Git and Sconstruct
  • python and pip 3
  • all the python packages listed in the file requirements.txt


To copy the repository type in a terminal

git clone git@bitbucket.org:RomainBrault/itl.git

The project is composed of a folder 'experiments' which contains the numerical experiments of the paper and the results. The folder 'doc' contains different version of the paper and various drafts. To install the library run

pip install .



To generate the requirements.txt file automatically use pigar

pigar -P itl -p ./requirements.txt


To compile the paper with XeLaTeX and biber, go into the paper folder

cd doc/NIPS_2018

and run the Sconstruct file


The resulting pdf should be in the subfolder build, and the temporary files in the subfolder build.


The experiments of the paper can be reproduced using the library in the folder demos/NIPS_2018/. Simply run

pip install .

to install the python library and run one of the files in the subfolder experiments/itl/demos/. E.g.

python demos/NIPS_2018/icsl_vs.py --show --save_graph=./tflog

When compiling the paper for the first time, the figure are generated automatically by running python on the files present in the folder demos/NIPS_2018/. The results are stored in the folder doc/NIPS_2018/build/src/fig/. If the pdf or eps files are already presents, the figures are not generated, and latex will use the one provided in the aformentioned folder. For the sake of fast installation we provide a cache version of our build folder.

the computation graph can be visualized using

tensorboard --localdir ./tflog

and opening a browser (usually) at http://localhost:6006

Contact and authors

  • Florence d'Alche-Buc, Telecom ParisTech, LTCI
  • Romain Brault: mail@romainbrault.com (corresponding author), Centrale-Supelec, L2S
  • Alex Lambert: alex.lambert@telecom-paristech.fr (corresponding author), Telecom ParisTech, LTCI
  • Maxime Sangnier, Universite Pierre et Marie Curie
  • Zoltan Szabo, Ecole Polytechnique, CMAP