def __init__(self,
oracle,
x_0,
H_0=None,
tolerance=1e-4,
line_search_options=None,
stopping_criteria='grad_rel',
trace=True):
super(BFGS, self).__init__(oracle, x_0, stopping_criteria, trace)
self.tolerance = tolerance
self.line_search_tool = get_line_search_tool(line_search_options)
self.x_k = x_0.copy()
self.grad_norm_0 = np.linalg.norm(self.oracle.grad(x_0))
if H_0 is None:
self.H_k = np.eye(x_0.shape[0], dtype=x_0.dtype)
else:
self.H_k = H_0.copy()
if self.line_search_tool._method == 'Constant':
self.alpha_0 = self.line_search_tool.c
else:
self.alpha_0 = 1.
self.grad_k = self.oracle.grad(self.x_k)
def __init__(self, oracle, x_0, s_distr, tolerance=1e-10, B_0=None,
line_search_options=None, stopping_criteria='grad_rel',
store_search_directions=False, update_every_iteration=False,
trace=True):
super(RBFGS, self).__init__(oracle, x_0, stopping_criteria, trace)
if store_search_directions and \
not isinstance(s_distr, (ConstantDistribution, CustomDiscrete)):
raise ValueError('Storing search directions only available for ConstantDistribution and CustomDiscrete')
self.s_distr = s_distr
self.d = x_0.shape[0]
if B_0 is not None:
self.B_0 = B_0.copy()
else:
self.B_0 = np.eye(self.d)
self.B_k = self.B_0.copy()
self.x_0 = x_0.copy()
self.x_k = x_0.copy()
self.id_mat = np.eye(self.d)
self.line_search_tool = get_line_search_tool(line_search_options)
self.grad_norm_0 = np.linalg.norm(self.oracle.grad(x_0))
self.tolerance = tolerance
self.store_search_directions = store_search_directions
self.update_every_iteration = update_every_iteration
self.iter_count = 0
if store_search_directions:
self.prev_directions = np.random.normal(
0., 1., (self.d, self.s_distr.mat.shape[1]))
def __init__(self, oracle, x_0, tolerance=1e-5, line_search_options=None,
stopping_criteria='grad_rel', trace=True):
super(GradientDescent, self).__init__(oracle, x_0, stopping_criteria, trace)
self.x_k = self.x_0.copy()
self.grad_norm_0 = np.linalg.norm(self.oracle.grad(x_0))
self.tolerance = tolerance
self.line_search_tool = get_line_search_tool(line_search_options)
try:
self.alpha_k = self.line_search_tool.alpha_0
except AttributeError:
self.alpha_k = self.line_search_tool.c
def __init__(self,
oracle,
x_0,
tolerance=1e-5,
line_search_options=None,
stopping_criteria='grad_rel',
trace=True):
super(Nesterov, self).__init__(oracle, x_0, stopping_criteria, trace)
self.x_k = self.x_0.copy()
self.y_k = self.x_0.copy()
self.grad_norm_0 = np.linalg.norm(self.oracle.grad(x_0))
self.line_search_tool = get_line_search_tool(line_search_options)
self.tolerance = tolerance
self.lam = 0.5
def __init__(self,
oracle,
x_0,
tolerance=1e-4,
memory_size=10,
line_search_options=None,
stopping_criteria='grad_rel',
trace=True):
super(LBFGS, self).__init__(oracle, x_0, stopping_criteria, trace)
self.x_k = self.x_0.copy()
self.grad_norm_0 = np.linalg.norm(self.oracle.grad(x_0))
self.tolerance = tolerance
self.line_search_tool = get_line_search_tool(line_search_options)
self.memory_size = memory_size
self.lbfgs_queue = deque(maxlen=memory_size)
if self.line_search_tool._method == 'Constant':
self.alpha_0 = self.line_search_tool.c
else:
self.alpha_0 = 1.