# LW-DETR mathematical reference Run `python3 examples/lw_detr.py` from the repository root. No installation, network, GPU, checkpoint or third-party package is required. The default entrypoint is `lw_detr.py`; it prints JSON to stdout and creates no output file. The dependency receipt is `requirements-lwDetr.txt`. This authored deterministic reference calculates window-major permutations and their inverse, structural local/global reach and ordered attention-pair counts, projector output geometry, grouped-query budgets and IoU-aware BCE forward values. It does not run attention, C2f, proposal ranking, Hungarian matching, training, inference or TensorRT. All attention counts include diagonal pairs per head and exclude projections, FFNs and memory movement. Exact divisible grids are required; there is no implicit padding or claim that arbitrary shapes are upstream configs. `windows`, `connectivity`, `projector`, `queries` and `loss` are the five excerpt keys. The repository's sourceRegions parser requires `# region: NAME` and `# endregion: NAME` (including the colon). The owning sync workflow generates `src/data/lwDetrCode.mjs` with `lwDetrCode` and `lwDetrTokens`, and public downloads. Do not hand-edit those generated files. Run `node --test test/lwDetr*.test.mjs` for mathematical and Python parity checks. Tiny uses 100 queries; other variants use 300. Training has 13 groups sharing weights, inference one. Tiny/small/medium produce stride-16 features; large/xlarge produce stride-8 and stride-32 features. Depth aggregation itself retains the encoder grid. Loss arithmetic follows the paper's equations, without reproducing implementation-specific detached targets, normalization, reductions or gradients. Endpoints use extended-real BCE, with `0 log 0 = 0`. Sources: [paper v1, §3 and Tables 1–3](https://arxiv.org/html/2406.03459v1), [official implementation](https://github.com/Atten4Vis/LW-DETR), especially `models/backbone/vit.py` and `models/lwdetr.py`. No pretrained output or new latency/AP measurement is included. Published T4 FP16 TensorRT results must not be inferred from these CPU calculations.