def create_aop_denoiser(Omega, noisy, nu, alpha):
Omega_sym = T.matrix('Omega')
noisy_shared = theano.shared(noisy.astype(np.float32))
denoised = theano.shared(np.copy(noisy.astype(np.float32)))
denoised_sym = T.matrix("denoised")
Omega_normal = normalizing_AOP(Omega_sym, noisy.shape[0])
cost = (((noisy_shared - denoised_sym) ** 2).sum() / denoised_sym.shape[1]) + \
alpha * normalized_l2norm_cost(logsquared_cost(T.dot(Omega_normal, denoised_sym), nu, axis=0), denoised_sym.shape[1])
grad = theano.grad(cost, denoised_sym)
cg_denoising = CG(denoised,
cost,
grad,
denoised_sym,
k=noisy.shape[0],
n=noisy.shape[1],
t_init=1,
rho=0.9,
max_iter_line_search=125,
other_givens={Omega_sym: Omega})
return cg_denoising, denoised
def __init__(
self,
n, # input dimension
k, # subspace rank
p, # p for pseudo-norm
mu, # Smoothing
t, # Step size
phi, # This is the foreground weighting factor.
y_iters=5,
U_init=None,
step_size_func=lambda iter: 1.0 / iter,
median_filter_radius=3):
self.n = n
self.k = k
self.p = p
self.mu = mu
self.y_iters = y_iters
self.phi = theano.shared(np.float32(phi))
# Step size
self.t = theano.shared(np.float32(t))
self.image_shape = None
self.step_size_func = step_size_func
self.median_filter_radius = median_filter_radius
if U_init == None:
# Initialize with random subspace
U = npr.randn(n, k).astype(np.float32)
U, _ = npl.qr(U)
else:
U = U_init
# subspace basis
self.U_sym = T.fmatrix("U")
self.U_shared = theano.shared(U.astype(np.float32))
# foreground weights
self.W_shared = theano.shared(np.ones((n, 1), dtype=np.float32))
# subspace coordinates of current data point
self.y_sym = T.fmatrix("y")
self.y_shared = theano.shared(npr.randn(k, 1).astype(np.float32))
# current image
self.image_sym = T.fmatrix("image")
self.image_shared = theano.shared(np.zeros((n, 1), dtype=np.float32))
self.init_y_func = T.dot(self.U_shared.T, self.image_shared)
self.init_y = theano.function(
[], [], updates={self.y_shared: self.init_y_func})
# pROST cost function
self.cost = (self.W_shared *
#reconstruction error || I - Uy||_(p,mu)
((((self.image_sym - T.dot(self.U_sym, self.y_sym))) ** 2 + self.mu) ** (self.p / 2))) \
.sum()
self.reconstruction_func = T.dot(self.U_sym, self.y_sym)
self.error_image_func = T.abs_(self.image_sym -
T.dot(self.U_sym, self.y_sym))
self.segmentation_func = self.error_image_func > self.t
self.grad_y = theano.grad(self.cost, self.y_sym)
self.grad_U = theano.grad(self.cost, self.U_sym)
self.optimize_y = CG(
self.y_shared,
self.cost,
self.grad_y,
self.y_sym,
self.k,
1,
t_init=0.1,
rho=0.5,
max_iter_line_search=200, # this is a bit excessive
other_givens={
self.U_sym: self.U_shared,
self.image_sym: self.image_shared
})
self.optimize_U = GrassmanGD(self.U_shared,
self.y_shared,
self.cost,
self.grad_U,
self.U_sym,
rho=0.6,
max_iter_line_search=5,
other_givens={
self.image_sym: self.image_shared,
self.y_sym: self.y_shared
},
step_size_func=self.step_size_func)
self.recerrsegweight = theano.function(
[], [
self.reconstruction_func, self.error_image_func,
self.segmentation_func
],
givens={
self.U_sym: self.U_shared,
self.y_sym: self.y_shared,
self.image_sym: self.image_shared
})
self.error_image = None
self.reconstruction = None
self.segmentation = None
