[tmva][sofie] Generate ONNX models and refs instead of binaries in repo

[guitargeek]

The SOFIE ONNX unit tests relied on 145 binary .onnx files checked into tmva/sofie/test/input_models, plus ~108 frozen reference headers (input_models/references/*.ref.hxx) whose expected outputs were computed once from the exact weights inside those binaries. This made it impossible to tell from the repository what the models contain, and impossible to regenerate them (some were exported with pytorch versions as old as 1.5).

Replace both with a single script, tmva/sofie/test/generate_input_models.py:

• Each model is built with the onnx helper API in a make_() function, so graph structure, attributes, shapes and initializers are readable and reviewable. Large weight tensors that used to be opaque random blobs are seeded-random via _random_tensor(), with pytorch-like 1/sqrt(fan_in) scaling.
• The inputs for the value-based tests are defined in TEST_INPUTS, and the expected outputs are computed with onnx's ReferenceEvaluator and written to references/.ref next to the generated models. Where the evaluator is wrong or unimplemented (MaxPool with asymmetric padding, Mean with multidirectional broadcasting, bidirectional or batchwise RNN/LSTM/GRU), numpy fallbacks implement the ONNX operator definitions directly.
• The script runs as the new SofieGenerateModels_ONNX ctest, which the other SOFIE ONNX tests require as a fixture. CMake gets the model list at configure time from generate_input_models.py --list, which works without the onnx module.

TestCustomModelsFromONNX.cxx and TestCladAutodiff.cxx read the inputs and expected outputs at runtime through a small reader in test_helpers.h instead of including frozen headers. SOFIE's results are thus checked against an independent implementation instead of a snapshot of its own past output.

While migrating, all computed references were validated to reproduce the frozen .ref.hxx values with the original weights before switching to seeded-random weights.

The ONNX model tests now require the onnx python package; they are disabled with a warning if it is missing or has the broken version 1.19.0. The TMVA_SOFIE_ONNX.C tutorial input is likewise generated at configure time instead of being copied from the test sources.

🤖 Done with the help of AI.

[github-actions]

Test Results

    23 files      23 suites   3d 15h 55m 55s ⏱️
 3 875 tests  3 875 ✅ 0 💤 0 ❌
78 807 runs  78 807 ✅ 0 💤 0 ❌

Results for commit b8d4f28.