boundlab.diff.zono3.interpret

boundlab.diff.zono3.interpret = <boundlab.interp.Interpreter object>

Differential-verification interpreter.

Feed it a DiffExpr3 (x, y, d) where x and y are the two networks’ zonotope expressions and d over-approximates their difference, or a plain Expr for standard zonotope interpretation.

Examples

Differential mode (DiffExpr3 input):

>>> import torch
>>> from torch import nn
>>> import boundlab.expr as expr
>>> from boundlab.diff.expr import DiffExpr3
>>> from boundlab.diff.zono3 import interpret
>>> model = nn.Linear(4, 3)
>>> op = interpret(model)
>>> x = expr.ConstVal(torch.randn(4)) + expr.LpEpsilon([4])
>>> y = expr.ConstVal(torch.randn(4)) + expr.LpEpsilon([4])
>>> out = op(DiffExpr3(x, y, x - y))
>>> out.diff.ub().shape
torch.Size([3])

Fallback mode (plain Expr) matches standard zonotope interpretation:

>>> z = expr.ConstVal(torch.randn(4)) + expr.LpEpsilon([4])
>>> z_out = op(z)
>>> z_out.ub().shape
torch.Size([3])