Home
Documentation for Umami#
Umami is a framework which can be used for training (most) machine-learning-based taggers used in ATLAS FTAG:
The intended use is as a software using configuration files to carry out preprocessing, training and evaluation.
It is not written with its use as a library in mind, as its core functionalities are provided by separate python modules which in turn can be installed with pip
.
These are:
Umami is also a tagger, the Umami tagger (UT). Its architecture includes jet features (DL1 inputs) plus a DIPS-like block. The high-level tagger (UT) and the DIPS block are trained in a single training (different to e.g. DL1d developments).
Umami is hosted on CERN GitLab:
Docker images are available on CERN GitLab container registry and on Docker Hub:
An API reference can be found here.
Tutorials for Umami#
A general tutorial is provided as part of the Umami documentation. It makes use of the JetClass dataset and is provided here.
A corresponding step-by-step tutorial for Umami inside the ATLAS collaboration for jet flavour tagging is provided here
Additionally, at the FTAG Workshop in 2022 in Amsterdam, we gave a tutorial how to work with Umami. You can find the slides together with a recording of the talk here. Please note that since that tutorial, the Umami software has been developed further and might slightly deviate from the version covered in the tutorial.