pimm#
pimm (particle imaging models) is a library for training, fine-tuning and running foundation models on point clouds from particle-imaging detectors. It contains the Panda and PoLAr-MAE models, readers for liquid argon TPC and water Cherenkov simulations, and one launcher for a workstation or a Slurm cluster.
curl -sSL https://raw.githubusercontent.com/DeepLearnPhysics/particle-imaging-models/main/install.sh | bash
Where to start#
- Run a released model
Load Panda or PoLAr-MAE from Hugging Face and label the points of an event. No training needed.
- Train on PILArNet-M
Install, train a tiny model on 100 events, then move to the full recipes.
- Bring your detector’s data
Write your events as
.npyfiles or add a reader, then train any pimm model on them.
What’s in pimm#
Area |
Contents |
|---|---|
Models |
Panda (self-distillation) and PoLAr-MAE (masked point modeling) pretraining; semantic segmentation, the Panda detector and PointGroup task models; Point Transformer v1–v3, SparseUNet, LitePT and Volt backbones |
Data |
PILArNet-M (HDF5 and Parquet), JAXTPC, LUCiD water Cherenkov, and a generic |
Training |
|
Models out |
Portable exports that |
This site documents pimm 0.5.2 on the main branch. pimm is research software; its APIs and recipes change between versions.