| |
| """CPU reference of FRNet (Autoware lidar_frnet) -- the ground truth every PCC / output-agreement gate compares |
| against (PLAN.md 0.3 item 1). It runs where ttnn is absent (the research venv) and never imports ttnn. |
| |
| - ``config.py`` the Autoware parameters this port reproduces (``ml_package_frnet_<variant>.param.yaml`` of the |
| weights, ``frnet.param.yaml`` node parameters) and the COMPILE constants, each with its source line; |
| - ``weights.py`` the deployed ONNX read as data with the vendored ``ttaw.weights.OnnxWeights`` (initializers by |
| consuming node; structure and every wiring assumption checked on the graph; BN folded in float64 |
| into the point layers, ``pre_norm`` kept as the host's scale / shift) -- the loader the TT port |
| uses too; |
| - ``model.py`` fp32 torch rebuild of the deployed graph with per-module taps (``ttaw.golden.TapRegistry``; names |
| in ``weights.MODULE_TAPS`` / ``OPT_IN_TAPS``), ported from research/frnet/scripts/frnet_torch_ref.py; |
| - ``rewrites.py`` the TT device dataflow (sorted + padded points, gather-form scatters with a zero sentinel row, the |
| exact log-step segmented max, K-split convs and point layers, interpolation-matrix up-sampling) |
| proven equal to ``model.py`` on the CPU, and a first-order precision emulation; |
| - ``ort.py`` ONNX Runtime golden of the deployed ONNX with matching intermediates (research / tests only: |
| onnxruntime is not in the image); |
| - ``pipeline.py`` the reference end to end: ``host/`` pre-processing -> torch / ORT / plan network -> ``host/`` |
| post-processing -> ``FRNetSegmentation``. |
| |
| Pre- and post-processing are not duplicated here: they live in ``tt_frnet.host`` (shared with the device path, so |
| TT and CPU see identical network inputs) and are tested against the research scripts. |
| |
| The reduced ``to_dict()`` of the ORT reference on a sample is stored per variant next to it as |
| ``samples/synthetic_<variant>.<variant>.reference.json``: ``server/smoke_test.py`` compares the served output with it. |
| """ |
|
|
