Skip to content

CODE_RELEASE_STATUS.md

Latest commit

 

History

History
120 lines (101 loc) · 7.04 KB

CODE_RELEASE_STATUS.md

File metadata and controls

120 lines (101 loc) · 7.04 KB
Raw

Release Status

Updated: 2026-10-06. This page describes the public source package, its tested scope and its remaining resource restrictions. Historical execution logs and machine-specific package inventories are not part of the release.

Source And Resources

Component Status
Core package CPU-only protocol validation, count-based F6 metrics, result merge and summaries
Included forecasters II-World, GenieDrive, OccWorld, COME, DOME and SparseWorld-TC
Included upstream integrations ALOcc, STCOcc, FusionOcc, SDGOcc, CVT-Occ, EFFOcc and FlashOcc
Additional source OccFusion source and existing helpers; not part of the seven-source runtime gate
Dataset assets All 75 archives published and remote size/checksum verified; all three datasets pin the verified index revision in the resource registry
Checkpoints Exact file guidance and available runtime/official digests recorded; GenieDrive identity matched; remaining unverified references are explicit
Citation NeurIPS 2026; arXiv:2512.15621

The root MIT license covers original OccStress code only. Retained third-party components have their own terms; see licenses. Data and checkpoint availability are separate from source availability.

Validation Scope

CPU regressions cover protocol construction, dataset paths, metric accumulation, result validation, launch/config guards, packaging and source availability. The download tool pins a dataset commit, rejects incomplete selections, resumes through the Hub client and verifies archive sizes/checksums. Its offline checks do not require a model framework or a GPU. The full-source profile runs checks for all restored implementations. The optional filtered profile still skips tests requiring its omitted source.

Small A100 runtime gates used real checkpoints and data: clean plus one corruption on each dataset, two anchors per setting and six future horizons. They validate installation and code migration, not full-paper reproduction.

Forecaster Completed small-sample checks Limitations
II-World Six metric comparisons matched; Waymo/CARLA confusion matrices matched nuScenes native path does not return confusion matrices
GenieDrive Six metric comparisons matched; Waymo/CARLA confusion matrices matched No matrix-parity claim for nuScenes
DOME Six deterministic comparisons matched in metrics and matrices Default autotuning showed about 0.03 pp variation
COME Six deterministic comparisons matched in metrics and matrices Default autotuning showed about 0.094 pp variation
OccWorld Six deterministic comparisons matched in metrics and matrices Default mode showed about 0.007-0.023 pp variation
SparseWorld-TC Six GPU smoke checks; Waymo/CARLA preprocessing matched Not a GPU-prediction parity claim

COME/OccWorld comparisons used old and release code in the same new environment, not old-versus-new PyTorch environments. Deterministic options are opt-in and do not change historical defaults. The training entrypoints inherited from model repositories are not covered by these evaluation gates.

Upstream gates exported five states per clean/corrupted setting and passed them through one GenieDrive H4/F6 anchor, checking 200x200x16 grids, 18-class labels and finite final metrics. Tested sources/settings were:

  • ALOcc and STCOcc on nuScenes: camera Brightness-hard.
  • FusionOcc and SDGOcc on nuScenes: camera Brightness-hard and LiDAR beam_missing-heavy, respectively.
  • CVT-Occ on Waymo: brightness-hard.
  • EFFOcc on Waymo: camera Brightness-hard and LiDAR beam_missing-heavy.
  • FlashOcc/EFFOcc on CARLA: camera Brightness-hard and LiDAR beam_missing-heavy; the optional fusion-model camera stress was checked separately.

STCOcc used the existing experiment checkpoint iter_63288.pth, not an asserted official model-zoo download. FusionOcc/SDGOcc used isolated clones of a working legacy environment with rebuilt extensions, not fresh dependency solves. Other listed gates used new environment prefixes. A100 checks do not establish H100 or RTX 5090 compatibility; see environment recipes.

Dataset And Metric Boundaries

Distribution scope records per-source conditions. The external preparation guide includes a direct 360-frame CARLA GT converter and safe mount/reference checks.

  • Set OCCSTRESS_DATA_ROOT for the shared benchmark layout. Raw sensors, official GT and auxiliary resources use OCCSTRESS_EXTERNAL_ROOT and are external dependencies, not bundled source assets.
  • Manual generators, the shared upstream builder, ALOcc/STCOcc/CARLA wrappers and both Waymo EFFOcc builders use the namespaced release layout and relative protocol/event references. Upstream builders consume existing canonical exports, not the raw sensor data or global release catalog. See construction boundaries.
  • The published release includes 975 protocols, 326 asset conditions and 1,654,143 asset files. Its 75 archives total 40.95 GB compressed, including metadata. All remote archive sizes/digests matched; anonymous CARLA/STCOcc downloads passed archive checks. Loader checks sampled completed assets; no exhaustive payload rehash is implied.
  • Unified results use occstress-present-gt-v1. Historical paper metrics may handle absent classes and zero IoU differently; do not relabel them as the unified convention. Full-paper scores are reported in the paper.
  • Main temporal regimes combine position and duration. Position sweeps control a single affected state; an unobserved t=-2.0 state is not robustness gain.
  • SparseWorld-TC is camera-direct and has no manual/point-source interface. EFFOcc camera stress retains clean LiDAR and is not a camera-only model.

Before Publishing

  1. Retain all applicable third-party notices and operator-specific terms.
  2. Check the live dataset package index for the selected track. Keep unresolved weight identities and optional experiment-weight mirrors explicit. Robo3D fog lookup tables remain external assets.
  3. Build a clean public candidate, run the checks below, review its diff and freeze the release commit. GitHub CI and tagging happen when publishing.
PYTHONDONTWRITEBYTECODE=1 python -m unittest discover -s tests -v
PYTHONDONTWRITEBYTECODE=1 python tools/check_release.py --public --syntax
python tools/check_documentation.py
python tools/check_formal_configs.py
python tools/update_release_metadata.py
python tools/release_status.py

These checks do not download data, run GPU jobs or certify legal compliance. tools/release_status.py --require-downloads intentionally fails while any resource is uploading, pending or identity-unverified. This includes optional upstream re-export weights and is stricter than a code-only release. Full-paper reruns and filling results/reference are not required for this source release. Runtime trees with extensions, caches or data mounts must not be published as source archives.