Per-intermediate shape inference correctness on every backend test
tagged shape, local_function or legacy
inference (collected via
onnx_light.onnx_lib.backend.test.case.collect_test_case,
same set exercised by
unittests/backend/test_backend_with_shape_inference.py).
Each case has its output and intermediate value_info shapes
stripped, then re-inferred with six runtimes:
onnx-light,
onnx_light.onnx_optim.shape_inference,
onnx.shape_inference,
onnx-shape-inference,
onnxruntime.transformers
(symbolic shape inference, exposed via onnxruntime.transformers.shape_infer_helper)
and
yobx.xshape.BasicShapeBuilder
(shape inference shipped with yet-another-onnx-builder).
The table reports the percentage of intermediates whose
elem_type and shape were correctly recovered, including
intermediates nested inside control-flow subgraphs
(If/Loop/Scan). Click a row
to see the per-intermediate detail for each runtime; the graph
input shapes fed to the runtimes are listed first for
context.