hep-viz 0.1.5

A local visualizer for HEP data

hep-viz

hep-viz is a lightweight, local 3D event visualizer designed specifically for the new ColliderML datasets. It provides an interactive interface to explore High Energy Physics (HEP) events, visualize particle tracks, calorimeter hits, and inspect Graph Neural Network (GNN) outputs such as edge scores and truth labels.

It is built for researchers and developers working with ColliderML data, supporting both standard event visualization and graph-based GNN inspection.

Installation

pip

You can install hep-viz directly from PyPI:

pip install hep-viz

From Source

git clone https://github.com/your-username/hep-viz.git
cd hep-viz
pip install .

Conda

conda install finnbarwilson::hep-viz

How to Use

hep-viz can be used either through its Command Line Interface (CLI) for quick file inspection or via its Python API for integration into notebooks and scripts.

1. Command Line Interface (CLI)

The CLI is perfect for viewing data stored in local directories (e.g., Parquet files).

hep-viz view /path/to/data/directory

Options:

• --port <int>: Specify the port to run the server on (default: 8000).
• --browser / --no-browser: control whether the browser opens automatically.

SSH Port Forwarding

If you are running hep-viz on a remote server via SSH, the browser cannot open automatically. You will need to forward the port to your local machine.

Standard SSH Forwarding: Run this in a new terminal on your local machine:

ssh -L 8000:localhost:8000 user@remote-host

Nested SSH Forwarding (SSH inside SSH): If you are connected to a jump host (login node) and then to a compute node:

1. On your local machine: Forward port 8000 to the login node.

   ssh -L 8000:localhost:8000 user@login-node
   

2. On the login node: Forward port 8000 to the compute node where hep-viz is running.

   ssh -L 8000:localhost:8000 user@compute-node
   

3. Open http://localhost:8000 in your local browser.

2. Python API

You can use hep-viz directly within your Python scripts.

Standard Event Visualization (view)

Pass a dictionary where each key is a list of dictionaries (records) or a Hugging Face Dataset.

import hep_viz

# Data is a dictionary of lists (records) or Datasets
event_data = {
    "particles": [
        {"event_id": 0, "particle_id": 1, "pdg_id": 11, ...}, # Event 0
        {"event_id": 1, "particle_id": 1, "pdg_id": 13, ...}  # Event 1
    ],
    "tracker_hits": [
        {"event_id": 0, "x": 10.1, "y": 20.2, "z": 30.3, "particle_id": 1},
        ...
    ],
    "calo_hits": [ ... ],
    
    # Standard Reconstruction Tracks
    "tracks": [
        {
            "event_id": 0,
            "track_id": 1,
            "majority_particle_id": 1,  # CRITICAL: Links track to Truth Particle
            "d0": 0.1, 
            "z0": 1.2
        },
        ...
    ]
}

# Launch the visualizer
hep_viz.view(event_data, port=8000)

Graph/GNN Visualization (view_graph)

Use this for visualizing GNN outputs (edges and scores).

import hep_viz

hep_viz.view_graph(
    hits=hits_df,       # DataFrame/Dict with x, y, z
    edges=edges_array,  # 2xN array of indices
    scores=scores_array, # Score per edge
    truth=truth_array,   # (Optional) Truth label per edge
    port=8001
)

Input Data Requirements

hep-viz expects key names matching the ColliderML schema.

Standard Inputs

• particles: List of particle dictionaries. Must contain event_id, particle_id, pdg_id.
• tracker_hits: List of hit dictionaries. Must contain event_id, x, y, z, particle_id.
  • Note: hep-viz builds the 3D track visualization by grouping these hits by particle_id.
• calo_hits: List of calorimeter hit dictionaries. Must contain event_id, energy, x, y, z.

Track Inputs (Reconstruction)

To visualize reconstruction performance, you provide a tracks list. hep-viz uses this to annotate the truth tracks (e.g., highlighting which particles were successfully reconstructed).

• event_id: ID of the event this track belongs to.
• majority_particle_id: (Required) The particle_id of the truth particle this track matches. This is used to link your reconstructed track to the visual truth track.
• Optional metadata: d0, z0, phi, theta, qop.

Comparing Reconstruction Models

hep-viz allows you to compare multiple reconstruction algorithms by seeing which truth particles they successfully recovered.

Rule: Any key in your data dictionary that starts with tracks will be treated as a reconstruction collection.

Example:

data = {
    'particles': ...,
    'tracker_hits': ...,
    
    'tracks': standard_tracks_list,       # e.g. Baseline
    'tracks_gnn': my_gnn_tracks_list,     # Your Custom Model
    'tracks_ckf': ckf_tracks_list         # Another Model
}

hep_viz.view(data)

In the UI, a dropdown menu 'Reco Algorithm' will appear. Switching between options will update the visualization to show which truth particles are "Found" (colored) vs "Missed" (greyed out) by that specific algorithm.

File Naming (CLI)

For the CLI to automatically detect files, they should contain the category name (e.g., my_particles.parquet). You can split data across multiple files using the pattern events<start>-<end> (e.g., particles.events0-999.parquet).

---

License

This project is licensed under the MIT License.

Copyright (c) 2025 Finnbar Wilson

Built for the ColliderML community.