File size: 2,026 Bytes
aec3f16 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 | # SPDX-License-Identifier: Apache-2.0
"""CPU reference of SceneSeg (Autoware VisionPilot): the ground truth every PCC / output-agreement gate of the TT port
compares against. Importable without ttnn; torch / onnx / onnxruntime are imported inside the functions that need
them.
- :mod:`.config` what the port reproduces: artifact pins (ONNX / .pth sha256, Drive ids, the re-hosted weights
repo), network constants (COMPILE), the channel-order variants (LOAD), the node's pre- / post-
processing constants, the tap names with their ONNX tensors; every value with its source.
- :mod:`.weights` the deployed ONNX's initializers addressed by consuming node (vendored ``ttaw.weights.OnnxWeights``;
no BN node is left in the export, so nothing is folded), checked against :data:`.weights.ARCH`
(108 layers written from the architecture); the same table from the upstream ``.pth`` /
re-hosted safetensors with the BatchNorms folded in fp64 (cross-check). The TT port loads its
weights with the same function.
- :mod:`.model` ``SceneSegReference``: the deployed graph in fp32 torch, module by module (EfficientNet-B0
backbone, SceneContext, SceneNeck, SceneSegHead), with per-module taps
(``ttaw.golden.TapRegistry``) and teacher-forced partial runs for the TT module gates.
- :mod:`.ort` the onnxruntime oracle of the deployed ONNX with taps as extra outputs (research and tests only).
- :mod:`.pipeline` ``ReferencePipeline``: image -> host pre-processing -> network -> host post-processing ->
``SceneSegMask``, the same output as ``SceneSeg.__call__``.
- :mod:`.assets` where tests find the weights without downloading.
Pre- and post-processing live in ``tt_sceneseg.host`` (shared with the Python API). The ``to_dict()`` of the
pipeline's output on the shipped sample is stored as ``samples/<stem>.<variant>.reference.json``.
"""
|