| |
| """CPU fp32 reference of StreamPETR (Autoware camera_streampetr): the ground truth every PCC / output-agreement gate |
| compares against (PLAN.md 0.3 item 1; research/streampetr/SPEC.md section 6). |
| |
| Layout (importable without ttnn; torch / onnx are imported inside the functions that need them): |
| |
| - ``config.py`` the HF model parameters + the node's post-processing / behaviour parameters this port reproduces, |
| each with its source line (autoware_universe main @ 9ceaccf, HF camera_streampetr @ v1.0). |
| - ``weights.py`` ``StreamPETRWeights``: the three ONNX files parsed as data with the vendored ``ttaw.weights`` |
| (initializers addressed by their consuming node; BN already folded by the exporter, so no fold |
| in the port), plus a graph check that the files are the deployed networks. |
| - ``model.py`` ``Backbone`` (VoVNet-99-eSE + CPFPN), ``PositionEmbedding``, ``Head`` (temporal memory, 6-layer |
| decoder, shared heads, the memory update): the three graphs in torch fp32 with per-module taps. |
| - ``port_form.py`` the TT port's exact rewrites (relative-motion memory, calibration constants, hoisted K / V, |
| the top-256 as a stable sort, merged branch inputs) with torch emulations proven against the |
| deployed form. |
| - ``pipeline.py`` ``StreamPETRReference``: the node end to end (stateless ``run_frame``, stateful API-equivalent |
| ``__call__``) on top of ``host`` (pre-processing, calibration, stream state, post-processing). |
| - ``ort.py`` the ONNX Runtime oracle of the three graphs (research and tests only; onnxruntime is NOT a |
| runtime dependency of the image). |
| |
| The ``to_dict()`` of this reference's output on the shipped sample is stored as |
| ``samples/<sample stem>.reference.json``: ``server/smoke_test.py`` compares the served output with it. |
| """ |
|
|
