ksddescent.ksdd_gradient

ksddescent.ksdd_gradient(x0, score, step, kernel='gaussian', max_iter=1000, bw=1, store=False, verbose=False, clamp=None, beta=0.2)

Kernel Stein Discrepancy descent with gradient descent

Perform Kernel Stein Discrepancy descent with gradient descent. Since it uses gradient descent, a step size must be specified.

Parameters

x0torch.tensor, size n_samples x n_features

initial positions

scorecallable

function that computes the score

stepfloat

step size

max_iterint

max numer of iters

bwfloat

bandwidth of the stein kernel

storesNone or list of ints

whether to stores the iterates at the indexes in the list

verbose: bool

wether to print the current loss

clamp:

if not None, should be a tuple (a, b). The points x are then constrained to stay in [a, b]

Returns

x: torch.tensor, size n_samples x n_features

The final positions

loss_listlist of floats

List of the loss values during iterations

References

A.Korba, P-C. Aubin-Frankowski, S.Majewski, P.Ablin. Kernel Stein Discrepancy Descent International Conference on Machine Learning, 2021.