def __init__(self,
step: float = None,
tol: float = 0.,
max_iter: int = 100,
verbose: bool = True,
print_every: int = 10,
record_every: int = 1,
linesearch_step_increase: float = 2.,
linesearch_step_decrease: float = 0.5,
modified=False):
SolverFirstOrder.__init__(self,
step=step,
tol=tol,
max_iter=max_iter,
verbose=verbose,
print_every=print_every,
record_every=record_every)
self.linesearch_step_increase = linesearch_step_increase
self.linesearch_step_decrease = linesearch_step_decrease
self.modified = modified
self.model_ssc = None
self.prox_ssc = self._ProxStandardSC()
self._th_gain = 0
def __init__(self,
step: float = None,
epoch_size: int = None,
rand_type="unif",
tol=0.,
max_iter=100,
verbose=True,
print_every=10,
record_every=1,
seed=-1):
self._step = None
# We must first construct SolverSto (otherwise self.step won't
# work in SolverFirstOrder)
SolverSto.__init__(self,
epoch_size=epoch_size,
rand_type=rand_type,
seed=seed)
SolverFirstOrder.__init__(self,
step=step,
tol=tol,
max_iter=max_iter,
verbose=verbose,
print_every=print_every,
record_every=record_every)
def set_prox(self, prox: Prox):
"""Set proximal operator in the solver.
Parameters
----------
prox : `Prox`
The proximal operator of the penalization function
Returns
-------
output : `Solver`
The solver with given prox
Notes
-----
In some solvers, ``set_model`` must be called before
``set_prox``, otherwise and error might be raised.
"""
if type(prox) is ProxZero:
SolverFirstOrder.set_prox(self, prox)
self._set("_prox_grad", lambda x: x)
elif type(prox) is ProxL2Sq:
SolverFirstOrder.set_prox(self, prox)
self._set("_prox_grad", lambda x: prox.strength * x)
else:
raise ValueError("BFGS only accepts ProxZero and ProxL2sq "
"for now")
return self
def set_prox(self, prox: list):
"""
Parameters
----------
prox : list[Prox]
List of all proximal operators of the model
"""
prox = CompositeProx(prox)
SolverFirstOrder.set_prox(self, prox)
return self
def __init__(self, step: float = None, tol: float = 0.,
max_iter: int = 100, linesearch: bool = True,
linesearch_step_increase: float = 2.,
linesearch_step_decrease: float = 0.5,
verbose: bool = True, print_every: int = 10,
record_every: int = 1):
SolverFirstOrder.__init__(self, step=step, tol=tol, max_iter=max_iter,
verbose=verbose, print_every=print_every,
record_every=record_every)
self.linesearch = linesearch
self.linesearch_step_increase = linesearch_step_increase
self.linesearch_step_decrease = linesearch_step_decrease
def _handle_history(self, n_iter: int, force: bool = False, **kwargs):
"""Updates the history of the solver.
"""
if n_iter == 0:
self._set('_initial_n_hessiannorm_calls',
self.model.n_calls_hessian_norm)
hessiannorm_calls = self.model.n_calls_hessian_norm - \
self._initial_n_hessiannorm_calls
SolverFirstOrder._handle_history(self, n_iter, force=force,
n_calls_hessiannorm=hessiannorm_calls,
**kwargs)
def __init__(self,
tol: float = 1e-10,
max_iter: int = 10,
verbose: bool = True,
print_every: int = 1,
record_every: int = 1):
SolverFirstOrder.__init__(self,
step=None,
tol=tol,
max_iter=max_iter,
verbose=verbose,
print_every=print_every,
record_every=record_every)
self._prox_grad = None
def _initialize_values(self, x0, step):
step, obj, x, prev_x, prev_grad_x_ssc, grad_x_ssc = \
SolverFirstOrder._initialize_values(self, x0, step,
n_empty_vectors=3)
grad_x_ssc = self.model_ssc.grad(x, out=grad_x_ssc)
return step, obj, x, prev_x, prev_grad_x_ssc, grad_x_ssc
def __init__(self,
step: float = None,
tol: float = 0.,
max_iter: int = 1000,
surrelax=1.,
verbose: bool = True,
print_every: int = 10,
record_every: int = 1):
SolverFirstOrder.__init__(self,
step=step,
tol=tol,
max_iter=max_iter,
verbose=verbose,
print_every=print_every,
record_every=record_every)
self.surrelax = surrelax
def initialize_values(self, x0, step):
step, obj, x, x_old = \
SolverFirstOrder._initialize_values(self, x0, step,
n_empty_vectors=1)
z_list = [np.zeros_like(x) for _ in range(self.prox.n_proxs)]
z_old_list = [np.zeros_like(x) for _ in range(self.prox.n_proxs)]
return x, x_old, z_list, z_old_list, obj, step
def set_prox(self, prox: Prox):
"""Set prox in the solver
Parameters
----------
prox : `Prox`
Sets the prox in the solver
Returns
-------
output : `Solver`
The `Solver` with given prox
"""
SolverFirstOrder.set_prox(self, prox)
SolverSto.set_prox(self, prox)
return self
def set_model(self, model: Model):
"""Set model in the solver
Parameters
----------
model : `Model`
Sets the model in the solver. The model gives the first
order information about the model (loss, gradient, among
other things)
Returns
-------
output : `Solver`
The `Solver` with given model
"""
SolverFirstOrder.set_model(self, model)
SolverSto.set_model(self, model)
return self
def _initialize_values(self, x0=None, step=None):
if step is None:
if self.step is None:
if self.linesearch:
# If we use linesearch, then we can choose a large
# initial step
step = 1e9
else:
raise ValueError("No step specified.")
step, obj, x, prev_x, x_new = \
SolverFirstOrder._initialize_values(self, x0, step,
n_empty_vectors=2)
return x, prev_x, x_new, step, obj