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 |
|
per-point features (encoder only) |
— |
Panda Semantic |
|
|
shower, track, Michel, delta, low-energy deposit |
Panda Particle |
|
particle masks and classes |
photon, electron, muon, pion, proton |
Panda Interaction |
|
interaction masks |
— |
PoLAr-MAE Pretrain |
|
masked-reconstruction losses |
— |
PoLAr-MAE Semantic |
|
|
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.