boundlab.diff.expr.DiffExpr3#

class boundlab.diff.expr.DiffExpr3[source]#

Bases: object

A triple (x, y, diff) for differential zonotope verification.

x and y track each network’s activations independently. diff over-approximates f₁(x) − f₂(y).

For affine operations f(z) = W z + b:

x and y receive both weight and bias.
diff receives only the weight (bias cancels: (Wx+b)−(Wy+b) = W(x−y)).

For pure-linear operations (no bias), all three components are updated identically.

Methods

`__init__`
`diag`
`expand`
`flatten`
`flip`
`gather`
`narrow`
`permute`
`repeat`
`reshape`
`roll`
`scatter`
`squeeze`
`tile`
`transpose`
`unflatten`
`unsqueeze`

x: Expr#

y: Expr#

diff: Expr#

property shape: torch.Size#

__add__(other)[source]#

__mul__(other)[source]#

scatter(indices, output_shape)[source]#

gather(indices)[source]#

reshape(*shape)[source]#

permute(*dims)[source]#

transpose(dim0, dim1)[source]#

flatten(start_dim=0, end_dim=-1)[source]#

unflatten(dim, sizes)[source]#

squeeze(dim=None)[source]#

unsqueeze(dim)[source]#

narrow(dim, start, length)[source]#

expand(*sizes)[source]#

repeat(*sizes)[source]#

tile(*sizes)[source]#

flip(dims)[source]#

roll(shifts, dims)[source]#

diag(diagonal=0)[source]#

__init__(x, y, diff)#