ML Framework for developing models for HHH-->4b 2g
An multi head attention network is trained to distinguish jets coming from Higgs Bosons and from generic QCD jets produced at p-p collisions.
A neptune based monitoring system is integrated to the framework for online monitoring of tarining step , logging evaluation mentics and plots and comparison studies
We use conda
for managing the pacakges used for the development.
For instantiation of the environment one can use the yaml
config provided
conda env create -f env.yaml
We use root files as entry point for data via. Uproot. The workflow is based on PyTorch framework.
Setting up a remote jupyter notebook session [ set it up of a machine with GPU support for quicker dev cycle ]
<local> $ <login to server >
<server> $ cd <workdir>
<server> $ jupyter notebook --no-browser --port=<PORT> #change port number if the posrt is busy
Connecting to remote jupyter notebook session
<local> $ ssh -L 8080:localhost:<PORT> <REMOTE_USER>@<REMOTE_HOST>
open http://localhost:8080/ to go to the Jupyter Notebook web interface. If password is asked , give the token shown at the end of link shown while launching the notebook in server
Launch tensorbord for monitoring the training
tensorboard --logdir <LOGDIR> --port=<PORT> >& /dev/null &
#tensorboard --logdir workarea/ml/attention4HHH/checkpoints/trippleHiggsVsQCD/lightning_logs/ --port=8008 >& /dev/null &
Development was based on these course notebooks.