boundlab.diff.net.diff_net#

boundlab.diff.net.diff_net(net1, net2)[source]#

Merge two structurally-identical ONNX models by pairing initializer inputs.

For each aligned node pair in net1/net2, any input position where both sides consume initializers is replaced by a shared boundlab::diff_pair value in the merged graph. This generalizes beyond affine operators (e.g. Gemm / MatMul) to any operator that reads parameters as initializers.

At concrete runtime, diff_pair is a no-op that returns branch-1 values (net1 semantics). Under differential interpretation, the paired values are lifted into DiffExpr2.

Parameters:

  • net1 (ir.Model | str | Path) – ONNX models with the same graph structure. Each may be an onnx_ir.Model, a str, or a pathlib.Path.

  • net2 (ir.Model | str | Path) – ONNX models with the same graph structure. Each may be an onnx_ir.Model, a str, or a pathlib.Path.

Return type:

A merged onnx_ir.Model.

Examples

>>> from boundlab.diff.net import diff_net
>>> from pathlib import Path
>>> nets = Path("compare/veridiff/examples/nets")
>>> merged = diff_net(nets / "single_layer_ref.onnx", nets / "single_layer_alt.onnx")
>>> any(n.domain == "boundlab" and n.op_type == "diff_pair" for n in merged.graph)
True