Skip to content

Commit

Permalink
Refactor recast algorithms to use ParametrizedFamily (facebookresearc…
Browse files Browse the repository at this point in the history 
…h#124)
  • Loading branch information
@jrapin
jrapin authored Feb 26, 2019
1 parent 93520ae commit a729909
Show file tree
Hide file tree
Showing 2 changed files with 58 additions and 83 deletions.
1 change: 1 addition & 0 deletions nevergrad/optimization/base.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -354,6 +354,7 @@ def __call__(self, dimension: int, budget: Optional[int] = None, num_workers: in
assert self._optimizer_class is not None
run = self._optimizer_class(dimension=dimension, budget=budget, num_workers=num_workers) # pylint: disable=not-callable
assert hasattr(run, "_parameters")
assert isinstance(run._parameters, self.__class__) # type: ignore
run._parameters = self # type: ignore
run.name = repr(self)
return run
Expand Down
140 changes: 57 additions & 83 deletions nevergrad/optimization/recastlib.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -9,25 +9,23 @@
from scipy import optimize as scipyoptimize
from scipy import stats
from . import base
from .base import registry
from . import recaster
from . import sequences


class ScipyMinimizeBase(recaster.SequentialRecastOptimizer):
class _ScipyMinimizeBase(recaster.SequentialRecastOptimizer):

def __init__(self, dimension: int, budget: Optional[int] = None, num_workers: int = 1, method: Optional[str] = None) -> None:
def __init__(self, dimension: int, budget: Optional[int] = None, num_workers: int = 1) -> None:
super().__init__(dimension, budget=budget, num_workers=num_workers)
self.method = method
self._parameters = ScipyOptimizer()
self.multirun = 1 # work in progress
assert self.method is not None, "A method must be specified"
self.initial_guess: Optional[base.ArrayLike] = None
self.random_restart = False

def get_optimization_function(self) -> Callable[[Callable[[base.ArrayLike], float]], base.ArrayLike]:
# create a different sub-instance, so that the current instance is not referenced by the thread
# (consequence: do not create a thread at initialization, or we get a thread explosion)
subinstance = self.__class__(dimension=self.dimension, budget=self.budget, num_workers=self.num_workers)
subinstance._parameters = self._parameters
return subinstance._optimization_function

def _optimization_function(self, objective_function: Callable[[base.ArrayLike], float]) -> base.ArrayLike:
Expand All @@ -40,72 +38,57 @@ def _optimization_function(self, objective_function: Callable[[base.ArrayLike],
remaining = budget - self._num_ask
while remaining > 0: # try to restart if budget is not elapsed
options: Dict[str, int] = {} if self.budget is None else {"maxiter": remaining}
res = scipyoptimize.minimize(objective_function, best_x if not self.random_restart else
np.random.normal(0., 1., self.dimension), method=self.method, options=options, tol=0)
res = scipyoptimize.minimize(objective_function, best_x if not self._parameters.random_restart else
np.random.normal(0., 1., self.dimension), method=self._parameters.method, options=options, tol=0)
if res.fun < best_res:
best_res = res.fun
best_x = res.x
remaining = budget - self._num_ask
return best_x


@registry.register
class NelderMead(ScipyMinimizeBase):

def __init__(self, dimension: int, budget: Optional[int] = None, num_workers: int = 1) -> None:
super().__init__(dimension, budget=budget, num_workers=num_workers, method="Nelder-Mead")


@registry.register
class Powell(ScipyMinimizeBase):
def __init__(self, dimension: int, budget: Optional[int] = None, num_workers: int = 1) -> None:
super().__init__(dimension, budget, num_workers=num_workers, method="Powell")

class ScipyOptimizer(base.ParametrizedFamily):
"""Scripy optimizers in a ask and tell format
@registry.register
class RPowell(ScipyMinimizeBase):
def __init__(self, dimension: int, budget: Optional[int] = None, num_workers: int = 1) -> None:
super().__init__(dimension, budget, num_workers=num_workers, method="Powell")
self.random_restart = True


@registry.register
class Cobyla(ScipyMinimizeBase):
def __init__(self, dimension: int, budget: Optional[int] = None, num_workers: int = 1) -> None:
super().__init__(dimension, budget, num_workers=num_workers, method="COBYLA")


@registry.register
class RCobyla(ScipyMinimizeBase):
def __init__(self, dimension: int, budget: Optional[int] = None, num_workers: int = 1) -> None:
super().__init__(dimension, budget, num_workers=num_workers, method="COBYLA")
self.random_restart = True
Parameters
----------
method: str
Name of the method to use, among Nelder-Mead, COBYLA, SLSQP and Powell
random_restart: bool
whether to restart at a random point if the optimizer converged but the budget is not entirely
spent yet (otherwise, restarts from best point)
"""

recast = True
_optimizer_class = _ScipyMinimizeBase

@registry.register
class SQP(ScipyMinimizeBase):
def __init__(self, dimension: int, budget: Optional[int] = None, num_workers: int = 1) -> None:
super().__init__(dimension, budget, num_workers=num_workers, method="SLSQP")
def __init__(self, *, method: str = "Nelder-Mead", random_restart: bool = False):
assert method in ["Nelder-Mead", "COBYLA", "SLSQP", "Powell"], f"Unknown method '{method}'"
self.method = method
self.random_restart = random_restart
super().__init__()


@registry.register
class RSQP(ScipyMinimizeBase):
def __init__(self, dimension: int, budget: Optional[int] = None, num_workers: int = 1) -> None:
super().__init__(dimension, budget, num_workers=num_workers, method="SLSQP")
self.random_restart = True
NelderMead = ScipyOptimizer(method="Nelder-Mead").with_name("NelderMead", register=True)
Powell = ScipyOptimizer(method="Powell").with_name("Powell", register=True)
RPowell = ScipyOptimizer(method="Powell", random_restart=True).with_name("RPowell", register=True)
Cobyla = ScipyOptimizer(method="COBYLA").with_name("Cobyla", register=True)
RCobyla = ScipyOptimizer(method="COBYLA", random_restart=True).with_name("RCobyla", register=True)
SQP = ScipyOptimizer(method="SLSQP").with_name("SQP", register=True)
RSQP = ScipyOptimizer(method="SLSQP", random_restart=True).with_name("RSQP", register=True)


@registry.register
class BO(recaster.SequentialRecastOptimizer):
class _BO(recaster.SequentialRecastOptimizer):

def __init__(self, dimension: int, budget: Optional[int] = None, num_workers: int = 1) -> None:
super().__init__(dimension, budget=budget, num_workers=num_workers)
self.qr = "none"
self._parameters = ParametrizedBO()

def get_optimization_function(self) -> Callable[[Callable[[base.ArrayLike], float]], base.ArrayLike]:
# create a different sub-instance, so that the current instance is not referenced by the thread
# (consequence: do not create a thread at initialization, or we get a thread explosion)
subinstance = self.__class__(dimension=self.dimension, budget=self.budget, num_workers=self.num_workers)
subinstance._parameters = self._parameters
return subinstance._optimization_function

def _optimization_function(self, objective_function: Callable[[base.ArrayLike], float]) -> base.ArrayLike:
Expand All @@ -119,23 +102,23 @@ def my_obj(**kwargs: Any) -> float:
for i in range(self.dimension):
bounds[str(i)] = (0., 1.)
bo = BayesianOptimization(my_obj, bounds)
if self.qr != "none":
if self._parameters.qr != "none":
points_dict: Dict[str, List[base.ArrayLike]] = {}
for i in range(self.dimension):
points_dict[str(i)] = []
budget = int(np.sqrt(self.budget))
sampler: Optional[sequences.Sampler] = None
if self.qr == "qr":
if self._parameters.qr == "qr":
sampler = sequences.HammersleySampler(self.dimension, budget=budget, scrambling=True)
elif self.qr == "mqr":
elif self._parameters.qr == "mqr":
sampler = sequences.HammersleySampler(self.dimension, budget=budget - 1, scrambling=True)
elif self.qr == "lhs":
elif self._parameters.qr == "lhs":
sampler = sequences.LHSSampler(self.dimension, budget=budget)
elif self.qr == "r":
elif self._parameters.qr == "r":
sampler = sequences.RandomSampler(self.dimension, budget=budget)
assert sampler is not None
for i in range(budget):
if self.qr == "mqr" and not i:
if self._parameters.qr == "mqr" and not i:
s = [0.5] * self.dimension
else:
s = list(sampler())
Expand All @@ -144,42 +127,33 @@ def my_obj(**kwargs: Any) -> float:
points_dict[str(j)].append(s[j])
bo.explore(points_dict)
assert self.budget is not None
budget = self.budget - (budget if self.qr != "none" else 0)
ip = 1 if self.qr == "none" else 0
budget = self.budget - (budget if self._parameters.qr != "none" else 0)
ip = 1 if self._parameters.qr == "none" else 0
bo.maximize(n_iter=budget - ip, init_points=ip)
# print [bo.res['max']['max_params'][str(i)] for i in xrange(self.dimension)]
v = [stats.norm.ppf(bo.res['max']['max_params'][str(i)]) for i in range(self.dimension)]
v = [min(max(v_, -100), 100) for v_ in v]
return v


@registry.register
class RBO(BO):

def __init__(self, dimension: int, budget: Optional[int] = None, num_workers: int = 1) -> None:
super().__init__(dimension, budget=budget, num_workers=num_workers)
self.qr = "r"
class ParametrizedBO(base.ParametrizedFamily):
"""Bayesian optimization
qr: str
TODO
"""

@registry.register
class QRBO(BO):
recast = True
_optimizer_class = _BO

def __init__(self, dimension: int, budget: Optional[int] = None, num_workers: int = 1) -> None:
super().__init__(dimension, budget=budget, num_workers=num_workers)
self.qr = "qr"
def __init__(self, *, qr: str = "none"):
assert qr in ["r", "qr", "mqr", "lhs", "none"]
self.qr = qr
super().__init__()


@registry.register
class MidQRBO(BO):

def __init__(self, dimension: int, budget: Optional[int] = None, num_workers: int = 1) -> None:
super().__init__(dimension, budget=budget, num_workers=num_workers)
self.qr = "mqr"


@registry.register
class LBO(BO):

def __init__(self, dimension: int, budget: Optional[int] = None, num_workers: int = 1) -> None:
super(LBO, self).__init__(dimension, budget=budget, num_workers=num_workers)
self.qr = "lhs"
BO = ParametrizedBO().with_name("BO", register=True)
RBO = ParametrizedBO(qr="r").with_name("RBO", register=True)
QRBO = ParametrizedBO(qr="qr").with_name("QRBO", register=True)
MidQRBO = ParametrizedBO(qr="mqr").with_name("MidQRBO", register=True)
LBO = ParametrizedBO(qr="lhs").with_name("LBO", register=True)

0 comments on commit a729909

Please sign in to comment.