def __init__(self,
A,
y,
proxg,
eps,
x=None,
G=None,
max_iter=100,
tau=None,
sigma=None,
show_pbar=True):
self.y = y
self.x = x
self.y_device = backend.get_device(y)
if self.x is None:
with self.y_device:
self.x = self.y_device.xp.zeros(A.ishape, dtype=self.y.dtype)
self.x_device = backend.get_device(self.x)
if G is None:
self.max_eig_app = MaxEig(A.H * A,
dtype=self.x.dtype,
device=self.x_device,
show_pbar=show_pbar)
proxfc = prox.Conj(prox.L2Proj(A.oshape, eps, y=y))
else:
proxf1 = prox.L2Proj(A.oshape, eps, y=y)
proxf2 = proxg
proxfc = prox.Conj(prox.Stack([proxf1, proxf2]))
proxg = prox.NoOp(A.ishape)
A = linop.Vstack([A, G])
if tau is None or sigma is None:
max_eig = MaxEig(A.H * A,
dtype=self.x.dtype,
device=self.x_device,
show_pbar=show_pbar).run()
tau = 1
sigma = 1 / max_eig
with self.y_device:
self.u = self.y_device.xp.zeros(A.oshape, dtype=self.y.dtype)
alg = PrimalDualHybridGradient(proxfc,
proxg,
A,
A.H,
self.x,
self.u,
tau,
sigma,
max_iter=max_iter)
super().__init__(alg, show_pbar=show_pbar)
def test_L2Proj(self):
shape = [6]
epsilon = 1.0
P = prox.L2Proj(shape, epsilon)
x = util.randn(shape)
y = P(1.0, x)
npt.assert_allclose(y, x / np.linalg.norm(x.ravel()))