def _update(self):
with self.device:
if self.accelerate or self.proxg is not None:
backend.copyto(self.x_old, self.x)
if self.accelerate:
backend.copyto(self.x, self.z)
gradf_x = self.gradf(self.x)
util.axpy(self.x, -self.alpha, gradf_x)
if self.proxg is not None:
backend.copyto(self.x, self.proxg(self.alpha, self.x))
if self.accelerate:
t_old = self.t
self.t = (1 + (1 + 4 * t_old**2)**0.5) / 2
backend.copyto(
self.z,
self.x + (t_old - 1) / self.t * (self.x - self.x_old))
xp = self.device.xp
if self.accelerate or self.proxg is not None:
self.resid = util.asscalar(
xp.linalg.norm((self.x - self.x_old) / self.alpha**0.5))
else:
self.resid = util.asscalar(xp.linalg.norm(gradf_x))
def gradf(x):
with self.x_device:
gradf_x = self.A.N(x) - AHy
if self.lamda != 0:
if self.z is None:
util.axpy(gradf_x, self.lamda, x)
else:
util.axpy(gradf_x, self.lamda, x - self.z)
return gradf_x
def _update(self):
self.minL(self.mu)
if self.g is not None:
device = backend.get_device(self.u)
xp = device.xp
with device:
util.axpy(self.u, self.mu, self.g(self.x))
backend.copyto(self.u, xp.clip(self.u, 0, np.infty))
if self.h is not None:
util.axpy(self.v, self.mu, self.h(self.x))
def _update(self):
backend.copyto(self.u_old, self.u)
backend.copyto(self.x_old, self.x)
# Update dual.
delta_u = self.A(self.x_ext)
util.axpy(self.u, self.sigma, delta_u)
backend.copyto(self.u, self.proxfc(self.sigma, self.u))
# Update primal.
with self.x_device:
delta_x = self.AH(self.u)
if self.gradh is not None:
delta_x += self.gradh(self.x)
util.axpy(self.x, -self.tau, delta_x)
backend.copyto(self.x, self.proxg(self.tau, self.x))
# Update step-size if neccessary.
if self.gamma_primal > 0 and self.gamma_dual == 0:
with self.x_device:
xp = self.x_device.xp
theta = 1 / (
1 + 2 * self.gamma_primal * xp.amin(xp.abs(self.tau)))**0.5
self.tau *= theta
with self.u_device:
self.sigma /= theta
elif self.gamma_primal == 0 and self.gamma_dual > 0:
with self.u_device:
xp = self.u_device.xp
theta = 1 / (
1 + 2 * self.gamma_dual * xp.amin(xp.abs(self.sigma)))**0.5
self.sigma *= theta
with self.x_device:
self.tau /= theta
else:
theta = self.theta
# Extrapolate primal.
with self.x_device:
xp = self.x_device.xp
x_diff = self.x - self.x_old
backend.copyto(self.x_ext, self.x + theta * x_diff)
x_diff_norm = xp.linalg.norm(x_diff / self.tau**0.5).item()
with self.u_device:
xp = self.u_device.xp
u_diff = self.u - self.u_old
u_diff_norm = xp.linalg.norm(u_diff / self.sigma**0.5).item()
self.resid = x_diff_norm**2 + u_diff_norm**2
def _get_ConjugateGradient(self):
I = linop.Identity(self.x.shape)
AHA = self.A.H * self.A
AHy = self.A.H(self.y)
if self.lamda != 0:
AHA += self.lamda * I
if self.z is not None:
util.axpy(AHy, self.lamda, self.z)
self.alg = ConjugateGradient(
AHA, AHy, self.x, P=self.P, max_iter=self.max_iter)
def gradf(x):
with self.y_device:
r = self.A(x)
r -= self.y
with self.x_device:
gradf_x = self.A.H(r)
if self.lamda != 0:
if self.z is None:
util.axpy(gradf_x, self.lamda, x)
else:
util.axpy(gradf_x, self.lamda, x - self.z)
return gradf_x
def _get_ConjugateGradient(self):
I = linop.Identity(self.x.shape)
AHA = self.A.H * self.A
AHy = self.A.H(self.y)
if self.lamda != 0:
if self.R is None:
AHA += self.lamda * I
else:
AHA += self.lamda * self.R.H * self.R
if self.mu != 0:
AHA += self.mu * I
util.axpy(AHy, self.mu, self.z)
self.alg = ConjugateGradient(
AHA, AHy, self.x, P=self.P, max_iter=self.max_iter)
def _update(self):
xp = self.device.xp
with self.device:
x_old = self.x.copy()
if self.accelerate:
backend.copyto(self.x, self.z)
# Perform update
util.axpy(self.x, -self.alpha, self.gradf(self.x))
if self.proxg is not None:
backend.copyto(self.x, self.proxg(self.alpha, self.x))
if self.accelerate:
t_old = self.t
self.t = (1 + (1 + 4 * t_old**2)**0.5) / 2
backend.copyto(
self.z, self.x + ((t_old - 1) / self.t) * (self.x - x_old))
self.resid = xp.linalg.norm(self.x - x_old).item() / self.alpha
def _update(self):
x_old = self.x.copy()
# Update dual.
util.axpy(self.u, self.sigma, self.A(self.x_ext))
backend.copyto(self.u, self.proxfc(self.sigma, self.u))
# Update primal.
with self.x_device:
util.axpy(self.x, -self.tau, self.AH(self.u))
backend.copyto(self.x, self.proxg(self.tau, self.x))
# Update step-size if neccessary.
if self.gamma_primal > 0 and self.gamma_dual == 0:
with self.x_device:
xp = self.x_device.xp
theta = 1 / (1 + 2 * self.gamma_primal * self.tau_min)**0.5
self.tau *= theta
self.tau_min *= theta
with self.u_device:
self.sigma /= theta
elif self.gamma_primal == 0 and self.gamma_dual > 0:
with self.u_device:
xp = self.u_device.xp
theta = 1 / (1 + 2 * self.gamma_dual * self.sigma_min)**0.5
self.sigma *= theta
self.sigma_min *= theta
with self.x_device:
self.tau /= theta
else:
theta = self.theta
# Extrapolate primal.
with self.x_device:
xp = self.x_device.xp
x_diff = self.x - x_old
self.resid = xp.linalg.norm(x_diff / self.tau**0.5).item()
backend.copyto(self.x_ext, self.x + theta * x_diff)
def _update(self):
with self.device:
xp = self.device.xp
Ap = self.A(self.p)
pAp = xp.real(xp.vdot(self.p, Ap)).item()
if pAp <= 0:
self.not_positive_definite = True
return
self.alpha = self.rzold / pAp
util.axpy(self.x, self.alpha, self.p)
if self.iter < self.max_iter - 1:
util.axpy(self.r, -self.alpha, Ap)
if self.P is not None:
z = self.P(self.r)
else:
z = self.r
rznew = xp.real(xp.vdot(self.r, z))
beta = rznew / self.rzold
util.xpay(self.p, beta, z)
self.rzold = rznew
self.resid = self.rzold.item()**0.5
def _update(self):
with self.device:
xp = self.device.xp
Ap = self.A(self.p)
pAp = xp.real(xp.vdot(self.p, Ap))
if pAp == 0:
self.zero_gradient = True
return
self.alpha = self.rzold / pAp
util.axpy(self.x, self.alpha, self.p)
if self.iter < self.max_iter - 1:
util.axpy(self.r, -self.alpha, Ap)
if self.P is not None:
z = self.P(self.r)
else:
z = self.r
rznew = xp.real(xp.vdot(self.r, z))
beta = rznew / self.rzold
util.xpay(self.p, beta, z)
self.rzold = rznew
self.resid = util.asscalar(self.rzold)**0.5
def run(self):
with self.device:
xp = self.device.xp
img = self.x
step_size = self.init_step_size
fnorm = []
tnorm = []
cost = []
for iter in range(self.max_iter):
# calculate gradient of fidelity and regularization
f_new = self._update_fidelity(img)
util.axpy(f_new, self.lambda_t,
xp.squeeze(self._update_temporal_fd(img)))
f2_new = xp.vdot(f_new, f_new)
if iter == 0:
f2_old = f2_new
f_old = f_new
# conjugate gradient
beta = f2_new / (f2_old + xp.finfo(float).eps)
util.axpy(f_new, beta, f_old)
f2_old = f2_new
f_old = f_new
# update image
fnorm_t = self._calculate_fnorm(img)
tnorm_t = self._calculate_tnorm(img)
cost_t = fnorm_t + tnorm_t
step_size = self._line_search(img, f_new, cost_t, step_size)
util.axpy(img, step_size, f_old)
# TODO stop criteria
# if abs(np.vdot(update_old.flatten(), update_old.flatten())) * step_size < 1e-6:
# break
if step_size < 2e-3:
break
fnorm.append(fnorm_t)
tnorm.append(tnorm_t)
cost.append(cost_t)
print("Iter[%d/%d]\tStep:%.5f\tCost:%.3f" %
(iter + 1, self.max_iter, step_size, cost_t))
return img, fnorm, tnorm, cost
def _update(self):
self.min_lagrangian(self.x, self.u, self.mu)
util.axpy(self.u, -self.mu, self.constraints(self.x))
