ibp

bound_ibp(model, ptb, data, target, n_classes=10, bound_upper=False, reuse_input=False)

Compute the lower and upper bounds of the model's output using the IBP method.

Parameters:

Name	Type	Description	Default
model	BoundedModule	The neural network model for which bounds are to be computed.	required
ptb	PerturbationLpNorm	The perturbation object defining the perturbation set.	required
data	Tensor	The input data tensor.	required
target	Tensor	The target labels tensor. Default is None.	required
n_classes	int	The number of classes for classification. Default is 10.	10
bound_upper	bool	Whether to compute the upper bound. Default is False.	False
reuse_input	bool	Whether to reuse the input data from previous bounding operation. Default is False.	False

Returns:

Type	Description
Tuple[Tensor, Tensor]	The lower and upper bounds of the model's output.

Source code in CTRAIN/bound/ibp.py

def bound_ibp(model, ptb, data, target, n_classes=10, bound_upper=False, reuse_input=False):
    """
    Compute the lower and upper bounds of the model's output using the IBP method.

    Args:
        model (auto_LiRPA.BoundedModule): The neural network model for which bounds are to be computed.
        ptb (auto_LiRPA.PerturbationLpNorm): The perturbation object defining the perturbation set.
        data (Tensor): The input data tensor.
        target (Tensor, optional): The target labels tensor. Default is None.
        n_classes (int, optional): The number of classes for classification. Default is 10.
        bound_upper (bool, optional): Whether to compute the upper bound. Default is False.
        reuse_input (bool, optional): Whether to reuse the input data from previous bounding operation. Default is False.

    Returns:
        (Tuple[Tensor, Tensor]): The lower and upper bounds of the model's output.
    """
    data = BoundedTensor(data, ptb=ptb)
    if target is not None:
        c = construct_c(data, target, n_classes)
    else:
        c = None
    if reuse_input:
        bound_input = None
    else:
        bound_input = (data,)
    lb, ub = model.compute_bounds(x=bound_input, IBP=True, method="IBP", C=c, bound_upper=bound_upper)
    return lb, ub