| |
| """CPU reference of Point Transformer V3 (Autoware autoware_ptv3, weights ``AutowareFoundation/ptv3@v4.0``, node |
| semantics @5eb9286) -- 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 yamls of the weights, the node's |
| ``ptv3.param.yaml``, class LUTs) and the port's COMPILE constants (stage capacities, patch 512, BEV |
| 256 x 256, 500 proposals), each with its source; |
| - ``weights.py`` the three ONNX graphs read as data with the vendored ``ttaw.weights.OnnxWeights`` (by consuming |
| node), every wiring assumption asserted from the graph, BNs folded in float64 where the port folds |
| them (the pooling BN stays an affine after the max); |
| - ``model.py`` fp32 torch rebuild of the encoder, seg head and det head on the host tables (gather-form sparse |
| conv, clip-padded patch attention, gather + max pooling / BEV scatter, ONNX-semantics TopK with |
| teacher forcing) with per-module taps (``TAP_NAMES``); |
| - ``pipeline.py`` ``Reference``: ``host.preprocess`` -> model -> ``host.postprocess`` (golden layout of |
| research/ptv3/public_data); |
| - ``ort.py`` oracles on the unmodified graphs: ONNX Runtime for the det head (with intermediates), and |
| ``onnx.reference.ReferenceEvaluator`` + numpy ``autoware::`` ops for the encoder / seg head |
| (research / tests only). |
| |
| Pre- and post-processing live in ``tt_ptv3.host`` (shared with the device path, so TT and CPU see identical tables). |
| ``samples/pandaset_019_f40.reference.json`` (and ``.seg.``) is ``PTv3Output.to_dict()`` of this reference on the |
| shipped sample (``code/scripts/make_samples.py``): ``server/smoke_test.py`` compares the served output with it. |
| """ |
|
|
