[tmva][sofie] Extend PyTorch parser with 10 new operators

[AdityaDRathore]

Description

The PyTorch parser currently supports only 6 ONNX operators (Gemm, Conv, Relu, Selu, Sigmoid, Transpose), which limits its ability to parse common model architectures that use activations like Tanh or Softmax, normalization layers, or element-wise arithmetic. This PR extends the parser to support 10 additional operators, enabling a significantly wider range of PyTorch models to be parsed directly from .pt files.

Relates to #21527.

Motivation

Models using MobileNet-style architectures, Transformer activations, or residual connections fail to parse because the required operators are not registered in mapPyTorchNode. Extending operator coverage is a prerequisite for the broader goal of improving the PyTorch and Keras parsers in SOFIE

Changes or fixes

1. 10 new operators added to mapPyTorchNode in RModelParser_PyTorch.cxx:

   • Stateless activations: Tanh
   • Stateful activations: Softmax (axis), LeakyRelu (alpha)
   • Binary arithmetic: Add, Sub, Mul
   • Matrix operations: MatMul (mapped to ROperator_Gemm without bias)
   • Structural: Flatten (axis), Reshape (allowzero)
   • Normalization: BatchNormalization (epsilon, momentum, 5 inputs)

2. Tensor name collision workaround: All Python-extracted tensor names in Parse() now normalize dots to underscores (. → _) via .replace('.','_') before reaching UTILITY::Clean_name(). This prevents TorchScript's dotted intermediate names (e.g., input.1) from colliding with sequential intermediates (e.g., input1) after Clean_name() erases the dot. The root cause in Clean_name() is tracked in [tmva][sofie] Clean_name() erases dots causing tensor name collisions in PyTorch parser #21527.

3. 2 new end-to-end test models added to TestRModelParserPyTorch.C and generatePyTorchModels.py:

   • ACTIVATION_MODEL: Linear → Tanh → Linear → LeakyReLU → Linear → Softmax — validates Tanh, LeakyRelu, Softmax
   • BATCHNORM_MODEL: BatchNorm1d → Flatten → Linear → Mul → Add — validates BatchNormalization, Flatten, Mul, Add

4. Updated AddNeededStdLib / AddBlasRoutines in Parse() for all new operators requiring cmath or BLAS routines.

Operator coverage summary

Operator	MakePyTorch* Function	ROperator Class	Attributes	Tested
onnx::Tanh	MakePyTorchTanh	ROperator_Tanh<float>	—	✅
onnx::Softmax	MakePyTorchSoftmax	ROperator_Softmax	axis (default: -1)	✅
onnx::LeakyRelu	MakePyTorchLeakyRelu	ROperator_LeakyRelu<float>	alpha (default: 0.01)	✅
onnx::Add	MakePyTorchAdd	ROperator_BasicBinary<float,Add>	—	✅
onnx::Mul	MakePyTorchMul	ROperator_BasicBinary<float,Mul>	—	✅
onnx::Flatten	MakePyTorchFlatten	ROperator_Reshape	axis (default: 1)	✅
onnx::BatchNormalization	MakePyTorchBatchNormalization	ROperator_BatchNormalization<float>	epsilon, momentum	✅
onnx::Sub	MakePyTorchSub	ROperator_BasicBinary<float,Sub>	—	—
onnx::MatMul	MakePyTorchMatMul	ROperator_Gemm<float>	—	—
onnx::Reshape	MakePyTorchReshape	ROperator_Reshape	allowzero	—

Note on untested operators: Sub and MatMul share the same ROperator classes as Add and Gemm respectively, and are exercised through the same CPython extraction pipeline. Reshape is implemented but testing is deferred due to unpredictable torch.jit.script ONNX export behavior for structural operations. All three can be validated with dedicated ONNX model tests in a follow-up.

Files Changed

File	Change
tmva/sofie_parsers/src/RModelParser_PyTorch.cxx	10 new MakePyTorch* functions, extended mapPyTorchNode, updated Parse() routines and name normalization
tmva/sofie/test/TestRModelParserPyTorch.C	2 new test cases: ACTIVATION_MODEL, BATCHNORM_MODEL
tmva/sofie/test/generatePyTorchModels.py	2 new model generators: generateActivationModel(), generateBatchNormModel()

Checklist

cc: @lmoneta @sanjibansg @guitargeek @devajithvs