Status and known limitations#
This page describes the current repository. It is not a promise of dates or future work. Pin a release or commit for research and validate the exact model, data, and execution path you use.
Current capability#
Area |
Status |
|---|---|
sparse, variable-length 3D point-cloud inputs |
primary supported modality |
local single-GPU and multi-GPU DDP training |
supported path |
Submitit-backed Slurm batch and interactive launch |
implemented; site profile must be validated locally |
standard split checkpoints and same-topology loader-state resume |
implemented |
topology-changing model/optimizer restore |
implemented with saved-epoch restart rather than exact loader-cursor continuation |
portable safetensors/Hub exports |
implemented; loading reconstructs the model, not a preprocessing pipeline or trainer state |
FSDP2 |
experimental; no repository-wide model parity matrix |
standalone evaluation |
available through |
2D wire-plane, optical waveform, and other detector modalities |
not currently supported by the common workflow |
Platform boundary#
The full locked training environment targets Linux x86-64, Python 3.10, PyTorch 2.10, CUDA 12.6, and an NVIDIA driver compatible with that runtime. macOS can use the launcher-only environment; the native sparse/CUDA stack is not installed there. Other architectures, Python/PyTorch/CUDA combinations, and non-NVIDIA accelerators are not covered by the committed lock/wheels.
Scientific boundaries#
A common
Pointrepresentation does not standardize detector coordinates, units, calibration, feature order, labels, or selection.A committed recipe is not automatically a supported benchmark. Several are active research variants without signed-off metrics or stability guarantees.
Published exports reconstruct architecture and weights, not the transform pipeline, class interpretation, dataset license, or evaluation protocol.
PILArNet-M v1 and v2 have a standard downloader. The HDF5 reader accepts v3, but pimm does not provide a standard public v3 download path.
The Parquet reader does not support
test_mode; use the HDF5 reader for the current voxelized/augmented test path.Two committed Panda/Sonata Parquet configs inherit a base file absent from this revision and cannot currently resolve; see the config catalog.
Exact resume depends on dataloader state plus the same world-size/worker topology and does not guarantee deterministic kernels outside pimm’s control.
User-facing gaps that should stay visible#
TODO
Replace the gaps below only with measured values or an agreed project policy. Until then, they remain visible here rather than being implied as features.
Gap |
What is needed before claiming completion |
|---|---|
canonical software citation |
approved authors/ORCIDs, |
model chooser metrics |
signed-off held-out values, exact split/protocol, uncertainty, config and checkpoint revision |
data reference |
authoritative sizes/counts/checksums, licenses/citations, and coordinate/unit figures per revision |
evaluation artifacts |
one versioned prediction/metric schema per task plus a first-class evaluation CLI |
FSDP2 support level |
model-by-model multi-rank train/save/resume/evaluation parity matrix |
support/security policy |
maintainers, expected response scope, supported versions, private reporting route |
compatibility policy |
documented stability guarantees and migration/deprecation process for registry names/config/checkpoint schemas |
How priorities are chosen#
No formal public roadmap or release schedule is committed in this repository. Propose work through a GitHub issue with:
the user/research outcome;
the current blocking behavior;
the data/model/API and compatibility contract;
a bounded test or measurable acceptance criterion;
required ownership, hardware, data access, and publication decisions.
For implementation expectations, see Contributing.