Login to a machine with a gpu:
ssh <username>@login.af.uchicago.edu
ssh <username>@lxplus-gpu.cern.ch
Clone the repo:
git clone https://github.com/LukeV37/ML_Tutorial_2025.git
Enter the directory cd ML_Tutorial_2025.
First, setup the virtual environment directly from the terminal using:
source setup.sh
Please be patient while pip installs packages and Pythia8 compiles...
If running on UChicago, request a jupyter lab instance from UChicago Analysis Facility.
If running on lxplus-gpu, start a jupyter server and use jupyter notebook list to open the notebook locally.
Open the jupyter notebook in the notebooks directory. Otherwise, run the code directly from the terminal using python in the scripts directory.
Tip
Ssh into the remote server and forward the port of your jupyter notebook server ssh -L XXXX:localhost:XXXX <username>@lxplus.cern.ch
Then start a jupyter server using jupyter notebook --no-browser --port=XXXX
Type jupyter notebook list and click the link to open the server locally in your browser.
No setup needed! Simply open the URL in your web browser and sign into your google account: https://colab.research.google.com/github/LukeV37/ML_Tutorial_2025/blob/main/notebooks/GoogleColab.ipynb
Important
To utilize GPU resources, click Runtime > Change runtime type > T4 GPU. Otherwise training will take longer on CPU.
This tutorial demonstrates Top Tagging by performing binary classification on ttbar vs diboson W+jets events. Particles are generated using pythia and the top 100 tracks - sorted by pT with features (pT, η, ϕ) - are input to neural networks. Three architectures are demonstrated: Multilayer Perceptron, DeepSets, and Transformer Encoder.
Note
With a large enough dataset, we can show that large transformer models are capable of learning topological structure of signal events.
Caution
Using low level particle information, the models perform best with more events. Be careful training on less than 100k events!