Models#

pimm has two pretrained model families for LArTPC point clouds. Each release on Hugging Face is a portable export that pimm.from_pretrained loads in one call.

Model

Hugging Face repository

Returns

Classes

Panda Base

DeepLearnPhysics/Panda-Base

per-point features (encoder only)

—

Panda Semantic

DeepLearnPhysics/Panda-Semantic

seg_logits, one row per point

shower, track, Michel, delta, low-energy deposit

Panda Particle

DeepLearnPhysics/Panda-Particle

particle masks and classes

photon, electron, muon, pion, proton

Panda Interaction

DeepLearnPhysics/Panda-Interaction

interaction masks

—

PoLAr-MAE Pretrain

DeepLearnPhysics/PoLAr-MAE-Pretrain

masked-reconstruction losses

—

PoLAr-MAE Semantic

DeepLearnPhysics/PoLAr-MAE-Semantic

seg_logits, one row per point

shower, track, Michel, delta

import pimm

model = pimm.from_pretrained("DeepLearnPhysics/Panda-Semantic", device="cuda")

An export rebuilds the architecture and loads the weights. It doesn’t include the input pipeline; Panda and PoLAr-MAE list the transforms each model’s recipes use.