AnalysisObjectTransformer Model

This repository contains the implementation of the AnalysisObjectTransformer model, a deep learning architecture designed for event classification with data from the CERN LHC. The model operates reconstructed-object and event-level features. MultiHeadAttention is used to extract the correlation between reconstructed objects such as jets (hadrons) or leptons in the final state, while event-level features capture the event summary, such as total hadronic energy or missing transverse energy. Achieves state-of-the-art performance on final states which can be summarized as jets accompanied by missing transverse energy.

Model Overview

The AnalysisObjectTransformer model is structured to process object-level features, in the case of jets: energy, mass, area, btag score, in any order (permutation invariance) and event-level features (HT, MET) to classify signal from background processes to enhance the sensitivity to rare BSM signatures.

Components

See here for complete architecure.

• Embedding Layers: Transform input data into a higher-dimensional space for subsequent processing.
• Attention Blocks (AttBlock): Utilize multi-head attention to capture dependencies between different elements of the input data.
• Class Blocks (ClassBlock): Extend attention mechanisms to incorporate class tokens, enabling the model to focus on class-relevant features. Implementation based on "Going deeper with transformers": https://arxiv.org/abs/2103.17239
• MLP Head: A sequence of fully connected layers that maps the output of the transformer blocks to the final prediction targets.

Usage

Firstly, clone the repository:

git clone https://huggingface.co/maciek-g/AnalysisObjectTransformer

You can then import the model object and use within standard PyTorch and PyTorch lightning training workflows.

from particle_transformer import AnalysisObjectTransformer

model = AnalysisObjectTransformer(input_dim_obj=..., input_dim_event=..., embed_dims=..., linear_dims1=..., linear_dims2=..., mlp_hidden_1=..., mlp_hidden_2=..., num_heads=...)

Parameter definitions

input_dim_obj: The number of features associated with each event object (features per jet, lepton etc..)

input_dim_events: The number of features associated with the event (Number of jets, total hadronic energy, total missing transverse energy etc..)

embed_dims: Sequence embedding dims