Models

Choose a published model by the output you need. Every repository below is a portable pimm export: from_pretrained() downloads the weights, rebuilds the architecture from config.json, loads strictly, switches to eval mode, and returns the model.

import pimm

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

Important

The portable loader reconstructs the architecture, not an executable data pipeline. An export made from a full run may retain a sanitized data section, but from_pretrained does not build or apply it. Reproduce the coordinate, energy, sampling, and collation steps from the release recipe before calling a model. See Prepare an input batch.

Choose a model

Goal	Model	What forward returns in eval mode	Published metric
Build a Panda task model	DeepLearnPhysics/Panda-Base	A Point representation from an encoder-only PT-v3m2; no task prediction	Not applicable
Label every point	DeepLearnPhysics/Panda-Semantic	seg_logits, shape \((N, 5)\)	See TODO below
Find particles and assign PID	DeepLearnPhysics/Panda-Particle	Query masks/classes; postprocess() produces point-level instances and PID	See TODO below
Group points by interaction	DeepLearnPhysics/Panda-Interaction	Query masks/classes; postprocess() produces point-level interaction instances	See TODO below
Continue PoLAr-MAE pretraining	DeepLearnPhysics/PoLAr-MAE-Pretrain	Self-supervised loss terms, not a downstream prediction	Not applicable
Run four-class PoLAr-MAE segmentation	DeepLearnPhysics/PoLAr-MAE-Semantic	seg_logits, shape \((N, 4)\)	\(\mathrm{mF1} \approx 0.82\), as reported by the model card

\(N\) is the total number of points in the packed batch. The Hub repository names above use their canonical capitalization; Hub lookup itself is case-insensitive.

The two maintained collections are the quickest way to see new releases:

• Panda collection

• PoLAr-MAE collection

TODO

Add signed-off, held-out Panda metrics and their evaluation protocols to the model cards, then replace the three TODO cells above. Each value needs its split, preprocessing contract, and metric definition.

Output and label lookup

Model	Registry type	Labels or objective	Next step
Panda Base	PT-v3m2	Masked-point pretrained encoder; no task head	Attach a head and fine-tune
Panda Semantic	DefaultSegmentorV2	shower, track, Michel, delta, LED	output["seg_logits"].argmax(-1)
Panda Particle	detector-v4	photon, electron, muon, pion, proton, LED	postprocess()
Panda Interaction	detector-v4	stuff, thing	postprocess()
PoLAr-MAE Pretrain	PoLAr-MAE	masked geometry reconstruction and energy infilling	Read loss, chamfer_loss, energy_loss
PoLAr-MAE Semantic	PoLArMAE-SemSeg	shower, track, Michel, delta	output["seg_logits"].argmax(-1)

The model cards are the authority for provenance and intended use. The recipe configs are the authority for preprocessing and training settings.

Open the matching recipe

Model	In-repository recipe
Panda Base	pretrain-sonata-v1m1-pilarnet-smallmask.py
Panda Semantic	semseg-pt-v3m2-pilarnet-ft-5cls-fft.py
Panda Particle	detector-v5-pt-v3m2-ft-pid-fft-detector.py
Panda Interaction	detector-v5-pt-v3m2-ft-vtx-fft-detector.py
PoLAr-MAE Pretrain	pretrain-polarmae-pilarnet.py
PoLAr-MAE Semantic	semseg-polarmae-pilarnet-fft-reproduce.py

Next

Goal	Guide
Run inference or fine-tune released weights	Load, run, and fine-tune a model
Publish portable weights	Export and publish