passage-reg-repvit-m1 / docs /RELEASE_DETAILS.md
geoffrey-g-delhomme's picture
Clarify candidate usage, methods and 256x256 scope
8155c1f
|
Raw History Blame Contribute Delete
6.84 kB
metadata
license: mit
pipeline_tag: image-to-image
tags:
  - passage
  - path-planning
  - onnx
  - private-release
datasets:
  - thalesgroup/passage

PASSAGE Reg: release details

1. Run the candidate

This private release contains the three measured 256 × 256 ONNX exports, their embedded trained parameters, frozen preprocessing and decoder, and a CPU inference command for new input scenes. Seed 2026 is the original reg-bellman-descent candidate; 1234 and 3407 are separately trained replicas. PASSAGE code and trained weights are licensed under MIT. Third-party software and terrain data retain the terms in NOTICE.md.

From the downloaded repository, install the checked dependencies following INSTALL.md, then run:

mkdir predictions
cpu-env/bin/python -I -B predict.py --seed 2026 \
  --input validation-input-only.npz --output predictions/example.json

The CPU functional check records one exposed validation scene per supplied seed. The command uses the NumPy reference interpreter included in onnx, with GPU visibility disabled and one CPU thread. It imports no native SDK or ONNX Runtime. It loads the selected model once and applies the frozen preprocessing and decoder to each scene. The JSON output contains routes, statuses, fallback flags, work counters and artifact hashes. It reports no latency or benchmark score; this CPU path is a functional reference and its numerical outputs may differ from native engines.

When using snapshot_download without local_dir, run the script through its path inside the returned snapshot directory. The adapter keeps that directory as its default root even when the script is a cache symlink; --root can also select it explicitly.

2. Inputs and outputs

Supply an NPZ archive readable with numpy.load(..., allow_pickle=False), containing exactly these arrays:

Array Type and shape Meaning
sample_ids Unicode strings [N] Nonempty unique scene identifiers
terrain Finite float64 [N,256,256] Raw elevation in metres
obstacles Boolean [N,256,256] Blocked destinations
starts Integer [N,2] Start row, column
goals Integer [N,2] Goal row, column

Endpoints must be free, distinct and within the grid. The decoder receives only the predicted field, obstacles and endpoints. It does not receive terrain or an oracle route. Statuses include a successful route, no path and an exhausted work budget, as produced by the frozen decoder. The tool refuses an existing output filename.

The included input-only example comes from one exposed validation scene at dataset revision c09cc5a21a55a93e7c09d47368d5f5bf75cb8fae. It contains no reference path or label. Its provenance and hash are in validation-input.json. The terrain example is subject to JAXA and applicable third-party data terms, separately from the MIT-licensed code and weights.

3. Measured exports and frozen implementation

Seed Role ONNX SHA-256
1234 separately trained replica cd0a9a5f98cc927f2b00e6b5357e63e4fcc01c2e9168a94220161a484dd6b335
2026 original frozen candidate 6323f1062c48f78db90e6d4a4aa9fc534e9fb3e0b579f731b60f698a5c563982
3407 separately trained replica b25ec41481c0df8a986d516271d6712e64e4ffab63571561447bec0f9f79e53a

passage.json describes the supplied release. reg-candidates.json lists every bundle artifact and its measured identities. sources.json identifies the frozen source files; runtime-configuration.json gives input calibration and decoder arguments. The command checks the selected bundle's hashes and the frozen decoder before execution. SHA256SUMS covers the distributed files.

The original checkpoint hashes document training provenance. Checkpoint pickle files, optimiser state, native TensorRT engines, SDKs, drivers, wheels and runtime images are not distributed. Seg retains the original two-class source graphs and the separately identified GRN rewrites for lineage and native build recipes; the inference graph still emits only the raw path-class logit. Reg emits its scalar cost-to-go field.

4. Native execution and scientific scope

The execution record and ledger preserve measured native model, engine and common-image identities. Their historical receipt hashes refer to S sessions on 1,024 scenes; the separate 40,000-scene deployment binding does not itself establish campaign completion or a score. The CPU command is a new portable entry point using the same graph, preprocessing and decoder; it is not the historical native timing harness.

A native inference API is now provided through native/infer_existing.py. It reuses the existing loaders/runtime and exact single-request function without requiring checkpoint files or loading pickle. It needs a compatible complete evaluator/inference package environment and the separately acquired vendor runtime; TensorRT also requires the exact selected compiled manifest and engine. The historical qualification preflight remains a different workflow. Native build argument templates preserve measured options; byte-identical engine reconstruction, fused arithmetic precision and general edge compatibility of this distributed package are not established. Installing vendor components follows their own terms. A native functional smoke now records one exposed validation scene per export on each of the three existing targets, with successful execution and cleanup. This is not a general numerical-parity, precision, timing or benchmark admission.

5. Licence and source notices

The MIT licence applies to PASSAGE's original contributions and PASSAGE-trained parameters, including these ONNX exports. NOTICE.md preserves ConvNeXt V2, RepViT and JAXA/AW3D30 attribution and explains the separate source-data terms. INFERENCE_NOTICE.md records inference-package provenance and dependency scope. No SDK, third-party code or source terrain is relicensed as MIT by the model metadata.

The exact training-source origins of UPerNet and DropPath remain undocumented; this release does not include that architecture source or assign a licence to unidentified third-party code. No upstream ImageNet pretrained weights are included. The release does not assert that a JAXA commercial-use notification has been sent or determine whether a particular learned-weight use falls within those data terms.