API Documentation



sparse.solvers.dual_primal(X, y[, max_iter, ...])

Chambolle-Pock algorithm applied to the dual: interpolation on the primal update. Parameters ---------- X : np.array, shape (n_samples, n_features) Design matrix. y : np.array, shape (n_samples,) Observation vector. max_iter : int, optional (default=1000) Maximum number of Chambolle-Pock iterations. f_store : int, optional (default=10) Primal iterates w are stored every f_store iterations. prox : callable, optional (default=shrink) Proximal operator of the minimized regularizer. By default, a shrink is used, corresponding to the convex L1 regularizer. It is given (w, tau) as input (the primal iterate and the primal stepsize). ret_all : bool, optional (default=True) If True, return all stored primal iterates. callback : callable or None, optional (default=None) Callable called on primal iterate w every f_store iterations. memory : int, optional (default=10) If callback is not None and its value did not decrease for the last memory stored iterates, the algorithm is early stopped. step_ratio : float, optional (default=1) Ratio betwen primal and dual stepsizes (tau/sigma). If step_ratio=1, both stepsizes are equal. rho : float, optional (default=0.99) The product of the step sizes is equal to rho / norm(X, ord=2)**2 verbose : bool, optional (default=False) Verbosity of the algorithm. Returns ------- w : np.array, shape (n_features,) Last or best primal iterate. theta : np.array, shape (n_samples,) Last or best dual iterate. crits : np.array, shape (max_iter // f_store,) Value of callback along iterations. all_w : np.array, shape (max_iter // f_store, n_features) Primal iterates every f_store iterations. Returned only if ret_all is True.