Classifying Collider Events with Point Clouds

Implementation of architectures in Comparing Point Cloud Strategies for Collider Event Classification [2212.10659]

---

If you have any questions at all about the code or using these kinds of architectures in your own analysis I'm always open to chatting, just let me know at delon@stanford.edu

Basics

To get started create a Anaconda environment with the provided environment.yml file

conda env create -f environment.yml
conda activate pcec
python -m ipykernel install --user --name pcec --display-name "pcec"

Models are implemented with Keras/Tensorflow. An simple example of using one our pairwise architecture recommend in the paper for event classification is found in SimpleArchitectureDemo.ipynb. Sample data files are stored in the data/ folder.

Architectures

The architectures are implemented in Architectures.py as tf.keras.Model with the following names

• Particlewise Architecture: DeepSet.
  • width: the number of nodes in the hidden layers of the neural networks parameterizing F and Φ
  • depth: the number of layers in each of the neural networks parameterizing F and Φ
  • latent_dim: the dimension of the latent space that particles are mapped to
• Pairwise/Tripletwise Architecture: Pairwise/Tripletwise.
  • width: the number of nodes in the hidden layers of the neural networks parameterizing F
  • depth: the number of layers in each of the neural networks parameterizing F
  • ec_widths: a list containing the number of nodes in each layer of Φ
  • num_particles: the number of particles you pad your events to (see N_PAD below)
• (Iterated) Nonlinear Pairwise Architecture: IteratedPiPairwise:
  • depth: the number of layers in each of the neural networks parameterizing F⁽ⁱ⁾
  • width: the number of nodes in the hidden layers of the neural networks parameterizing F⁽ⁱ⁾
  • ec_widths: a list of lists where each list contains the number of nodes in each layer of Φ⁽ⁱ⁾
  • num_particles: the number of jets you pad your events to (see N_PAD below)
• Nested Concatenation Architecture: NestedConcat:
  • L: number of nested Deep Sets to use. When L=1 then NestedConcat is equivalent to DeepSet
  • width: the number of layers in each of the neural networks parameterizing F⁽ⁱ⁾ and Φ⁽ⁱ⁾
  • depth: the number of layers in each of the neural networks parameterizing F⁽ⁱ⁾ and Φ⁽ⁱ⁾ (note: Φ⁽ⁱ⁾ when i≠L have depth-1 layers)
  • latent_dim: the dimension of the latent space that particles are mapped to

You can get any of these models just by calling them:

from Architectures import *
classifier = DeepSet(depth=3, width=128, latent_dim=64)

These have all the methods defined in the tensorflow documentation for tf.keras.Model, the most important of which are fit for training the model and predict for classifying events

Parameters used in paper are stored in dictionaries at the very bottom of the Architectures.py file. For example if you want to create the Pairwise architecture with the same parameter as used in the experiments in the paper you could run

from Architectures import *
classifier = classifiers['pairwise'](**model_params_dict['pairwise'])

Data

Input data should have shape (N_events, N_PAD, N_features).

• N_events is the number of collider events in your dataset
• N_PAD is the number of particles you zero-pad each event. The zero padded particles are ignored by the architecture through masking layers but required for computationally efficient implementation. N_PAD should be larger than or equal to the maximum number of particles you have in one event in your dataset. If you have a dataset with shape (N_events, -1, N_features) where -1 implies that each event has a variable number of particles you can call tensorflow.keras.preprocessing.sequence.pad_sequences(events, dtype='float32', padding='post', maxlen=N_PAD) to generate a padded version of your dataset suitable to be passed to the architectures.
• N_features is the number of features you describe each particle in your event with. We recommend mimicing the features chosen in the paper described in equation (34)

Generating Figures in Paper

The (very messy) code used to generate results in paper can be found in the directory paper_code/. The scripts to train the models is in paper_code/gen_models.py and the code to generate the figures in the paper are scattered throughout the other files. Namely:

• Fig 3 and Table 1 from gen_plots.py
• Table 3 from latent_scan.py and latent_scan_aux.ipynb
• Fig 4 and Fig 5 from l2l8.ipynb
• Fig 6 from spearman_rank_analysis.ipynb
• Fig 7 from ditau_study.ipynb

All this code and the paper_code folder itself is truly quite messy so if you have any questions about anything in this folder please don't hesitate to reach out at delon@stanford.edu.