PicoDet-M A-J 5-Epoch Adapter Experiments
This repository publishes the canonical model artifacts associated with the 43
completed experiments in AJ_EPOCH5_REPORT_20260906.md.
Contents
- 43 experiment variants × epochs 1–5 = 215 canonical full Paddle checkpoints.
- The corresponding 215 adapter-only checkpoints.
- Per-epoch state/completion metadata and SHA-256 identities.
- The frozen experiment manifest, readable report, result table, and analysis evidence needed to interpret each model.
- The common clean PicoDet-M W0 checkpoint and shared initialization payloads required by the published model family, when applicable.
Failed or superseded retries, duplicate evaluation copies, optimizer states, raw predictions, and runtime logs are excluded. E5 is the predeclared primary checkpoint; E1–E4 are supplied for learning-curve analysis rather than post-hoc checkpoint selection.
Layout
models/<experiment-id>/e01..e05/
model.pdparams
adapter_only.pdparams
state.json
completion.json
reports/
configs/
shared/
MANIFEST.json
The full checkpoint contains W0 plus the experiment-specific modules. The adapter-only checkpoint must be loaded with the exact W0 and runtime/configuration documented in this repository.
License and use
These models were trained and evaluated on nuImages and derived corruption images. The controlling nuImages Dataset Terms apply to the source and derived data and describe the dataset as non-commercial. This research release does not grant broader rights or imply endorsement by Motional. Users must review the current terms and obtain any commercial license that their use requires.
Evaluation boundary
Reported AP is COCO bbox AP@[0.50:0.95] on the frozen 650-source development split across 60 corruption cells; Clean is reported separately. Results use one seed, and the cells reuse source scenes. Small differences are not claims of statistical significance or external generalization.