boundlab.diff.onnx#
Merge two ONNX models into one paired graph for differential interpretation.
diff_net() walks two structurally identical graphs in lock-step and,
wherever both sides read a parameter from an initializer, replaces that input
by a shared boundlab::DiffPair value. At concrete-tensor runtime the
merged model still evaluates network 1; under
boundlab.diff.zono3.interpret the paired initializers become
DiffExpr2 values that carry both branches through
the rest of the graph.
Functions
Pair the initializers of two structurally identical ONNX models. |