def _get_primal(self, dual_update):
if self.lower is not None:
dus = dual_update.shape
if len(dus) == 2 and not 1 in set(dus):
(U, s, V) = np.linalg.svd(dual_update)
sparse_s = get_st(s, lower=self.lower)
dual_update = get_multiplied_svd(U, s, V)
else:
dual_update = get_st(dual_update, lower=self.lower)
return dual_update / (self.H + self.delta)
def _get_primal(self, dual_update):
if self.lower is not None:
dus = dual_update.shape
if len(dus) == 2 and not 1 in set(dus):
(U, s, V) = np.linalg.svd(dual_update)
sparse_s = get_st(s, lower=self.lower)
dual_update = get_multiplied_svd(U, s, V)
else:
dual_update = get_st(dual_update, lower=self.lower)
H_inv = get_svd_power(self.H, -1)
return np.dot(H_inv, dual_update)
def _get_primal(self, dual_update):
if self.lower is not None:
dual_update = get_st(
dual_update, lower=self.lower)
return dual_update / self.H
def _get_primal(self, dual_update):
(self.U, dual_s, self.V) = np.linalg.svd(dual_update)
if self.lower is not None:
dual_s = get_st(dual_s, lower=self.lower)
self.s = get_lpn_grad(dual_s, self.q)
return get_multiplied_svd(self.U, self.s, self.V)
def get_update(self, parameters, gradient, eta):
unscaled = parameters - eta * gradient
if self.lower is not None:
unscaled = get_st(unscaled, lower=self.lower)
if self.window.is_full():
last_period = None
if self.weight > 0:
last = self.weight * self.window.dequeue()
second_last = self.window.get_items()[-1]
last_period = last + second_last
else:
last_period = self.window.dequeue()
unscaled += self.c * last_period
self.window.enqueue(np.copy(parameters))
self.num_rounds += 1
other_c = eta**(-1)
return (other_c + other_c * self.c)**(-1) * unscaled