def test_space_api():
space = Space([(0.0, 1.0), (-5, 5),
("a", "b", "c"), (1.0, 5.0, "log-uniform"), ("e", "f")])
cat_space = Space([(1, "r"), (1.0, "r")])
assert isinstance(cat_space.dimensions[0], Categorical)
assert isinstance(cat_space.dimensions[1], Categorical)
assert_equal(len(space.dimensions), 5)
assert_true(isinstance(space.dimensions[0], Real))
assert_true(isinstance(space.dimensions[1], Integer))
assert_true(isinstance(space.dimensions[2], Categorical))
assert_true(isinstance(space.dimensions[3], Real))
assert_true(isinstance(space.dimensions[4], Categorical))
samples = space.rvs(n_samples=10, random_state=0)
assert_equal(len(samples), 10)
assert_equal(len(samples[0]), 5)
assert_true(isinstance(samples, list))
for n in range(4):
assert_true(isinstance(samples[n], list))
assert_true(isinstance(samples[0][0], numbers.Real))
assert_true(isinstance(samples[0][1], numbers.Integral))
assert_true(isinstance(samples[0][2], str))
assert_true(isinstance(samples[0][3], numbers.Real))
assert_true(isinstance(samples[0][4], str))
samples_transformed = space.transform(samples)
assert_equal(samples_transformed.shape[0], len(samples))
assert_equal(samples_transformed.shape[1], 1 + 1 + 3 + 1 + 1)
# our space contains mixed types, this means we can't use
# `array_allclose` or similar to check points are close after a round-trip
# of transformations
for orig, round_trip in zip(samples,
space.inverse_transform(samples_transformed)):
assert space.distance(orig, round_trip) < 1.e-8
samples = space.inverse_transform(samples_transformed)
assert_true(isinstance(samples[0][0], numbers.Real))
assert_true(isinstance(samples[0][1], numbers.Integral))
assert_true(isinstance(samples[0][2], str))
assert_true(isinstance(samples[0][3], numbers.Real))
assert_true(isinstance(samples[0][4], str))
for b1, b2 in zip(space.bounds,
[(0.0, 1.0), (-5, 5),
np.asarray(["a", "b", "c"]), (1.0, 5.0),
np.asarray(["e", "f"])]):
assert_array_equal(b1, b2)
for b1, b2 in zip(space.transformed_bounds,
[(0.0, 1.0), (-5, 5), (0.0, 1.0), (0.0, 1.0), (0.0, 1.0),
(np.log10(1.0), np.log10(5.0)), (0.0, 1.0)]):
assert_array_equal(b1, b2)
def test_normalize():
# can you pass a Space instance to the Space constructor?
space = Space([(0.0, 1.0), (-5, 5),
("a", "b", "c"), (1.0, 5.0, "log-uniform"), ("e", "f")])
space.set_transformer("normalize")
X = [[0., -5, 'a', 1., 'e']]
Xt = np.zeros((1, 5))
assert_array_equal(space.transform(X), Xt)
assert_array_equal(space.inverse_transform(Xt), X)
assert_array_equal(space.inverse_transform(space.transform(X)), X)
def test_normalize_types():
# can you pass a Space instance to the Space constructor?
space = Space([(0.0, 1.0), Integer(-5, 5, dtype=int), (True, False)])
space.set_transformer("normalize")
X = [[0., -5, False]]
Xt = np.zeros((1, 3))
assert_array_equal(space.transform(X), Xt)
assert_array_equal(space.inverse_transform(Xt), X)
assert_array_equal(space.inverse_transform(space.transform(X)), X)
assert isinstance(space.inverse_transform(Xt)[0][0], float)
assert isinstance(space.inverse_transform(Xt)[0][1], int)
assert isinstance(space.inverse_transform(Xt)[0][2], (np.bool_, bool))
def test_normalize_bounds():
bounds = [(-999, 189000), Categorical((True, False))]
space = Space(normalize_dimensions(bounds))
for a in np.linspace(1e-9, 0.4999, 1000):
x = space.inverse_transform([[a, a]])
check_limits(x[0][0], -999, 189000)
y = space.transform(x)
check_limits(y, 0., 1.)
for a in np.linspace(0.50001, 1e-9 + 1., 1000):
x = space.inverse_transform([[a, a]])
check_limits(x[0][0], -999, 189000)
y = space.transform(x)
check_limits(y, 0., 1.)
def test_space_api():
space = Space([(0.0, 1.0), (-5, 5),
("a", "b", "c"), (1.0, 5.0, "log-uniform"), ("e", "f")])
cat_space = Space([(1, "r"), (1.0, "r")])
assert isinstance(cat_space.dimensions[0], Categorical)
assert isinstance(cat_space.dimensions[1], Categorical)
assert_equal(len(space.dimensions), 5)
assert_true(isinstance(space.dimensions[0], Real))
assert_true(isinstance(space.dimensions[1], Integer))
assert_true(isinstance(space.dimensions[2], Categorical))
assert_true(isinstance(space.dimensions[3], Real))
assert_true(isinstance(space.dimensions[4], Categorical))
samples = space.rvs(n_samples=10, random_state=0)
assert_equal(len(samples), 10)
assert_equal(len(samples[0]), 5)
assert_true(isinstance(samples, list))
for n in range(4):
assert_true(isinstance(samples[n], list))
assert_true(isinstance(samples[0][0], numbers.Real))
assert_true(isinstance(samples[0][1], numbers.Integral))
assert_true(isinstance(samples[0][2], str))
assert_true(isinstance(samples[0][3], numbers.Real))
assert_true(isinstance(samples[0][4], str))
samples_transformed = space.transform(samples)
assert_equal(samples_transformed.shape[0], len(samples))
assert_equal(samples_transformed.shape[1], 1 + 1 + 3 + 1 + 1)
assert_array_equal(samples, space.inverse_transform(samples_transformed))
samples = space.inverse_transform(samples_transformed)
assert_true(isinstance(samples[0][0], numbers.Real))
assert_true(isinstance(samples[0][1], numbers.Integral))
assert_true(isinstance(samples[0][2], str))
assert_true(isinstance(samples[0][3], numbers.Real))
assert_true(isinstance(samples[0][4], str))
for b1, b2 in zip(space.bounds,
[(0.0, 1.0), (-5, 5),
np.asarray(["a", "b", "c"]), (1.0, 5.0),
np.asarray(["e", "f"])]):
assert_array_equal(b1, b2)
for b1, b2 in zip(space.transformed_bounds,
[(0.0, 1.0), (-5, 5), (0.0, 1.0), (0.0, 1.0), (0.0, 1.0),
(np.log10(1.0), np.log10(5.0)), (0.0, 1.0)]):
assert_array_equal(b1, b2)
def test_space_api():
space = Space([(0.0, 1.0), (-5, 5),
("a", "b", "c"), (1.0, 5.0, "log-uniform")])
assert_equal(len(space.dimensions), 4)
assert_true(isinstance(space.dimensions[0], Real))
assert_true(isinstance(space.dimensions[1], Integer))
assert_true(isinstance(space.dimensions[2], Categorical))
assert_true(isinstance(space.dimensions[3], Real))
samples = space.rvs(n_samples=10, random_state=0)
assert_equal(len(samples), 10)
assert_equal(len(samples[0]), 4)
samples_transformed = space.transform(samples)
assert_equal(samples_transformed.shape[0], len(samples))
assert_equal(samples_transformed.shape[1], 1 + 1 + 3 + 1)
assert_array_equal(samples, space.inverse_transform(samples_transformed))
for b1, b2 in zip(space.bounds,
[(0.0, 1.0), (-5, 5),
np.asarray(["a", "b", "c"]), (1.0, 5.0)]):
assert_array_equal(b1, b2)
for b1, b2 in zip(space.transformed_bounds,
[(0.0, 1.0), (-5, 5), (0.0, 1.0), (0.0, 1.0), (0.0, 1.0),
(np.log10(1.0), np.log10(5.0))]):
assert_array_equal(b1, b2)
def generate(self, dimensions, n_samples, random_state=None):
"""Creates latin hypercube samples with maxpro criterion.
Parameters
----------
dimensions : list, shape (n_dims,)
List of search space dimensions.
Each search dimension can be defined either as
- a `(lower_bound, upper_bound)` tuple (for `Real` or `Integer`
dimensions),
- a `(lower_bound, upper_bound, "prior")` tuple (for `Real`
dimensions),
- as a list of categories (for `Categorical` dimensions), or
- an instance of a `Dimension` object (`Real`, `Integer` or
`Categorical`).
n_samples : int
The order of the LHS sequence. Defines the number of samples.
random_state : int, RandomState instance, or None (default)
Set random state to something other than None for reproducible
results.
Returns
-------
np.array, shape=(n_dim, n_samples)
LHS set
"""
rng = check_random_state(random_state)
space = Space(dimensions)
transformer = space.get_transformer()
n_dim = space.n_dims
space.set_transformer("normalize")
h = self._lhs_normalized(n_dim, n_samples, rng)
self.num_pts = n_samples
self.dim = n_dim
if self.use_gradient:
print('Using gradient descent')
bounds = [(0, 1)] * len(dimensions) * self.num_pts
h_opt = minimize(self.maxpro_criter,
h,
jac=self.maxpro_grad,
bounds=bounds)
h_opt = h_opt['x'].reshape(n_samples, n_dim)
else:
print('Using naive method')
best = 1e+6
for i in range(self.iterations):
h = self._lhs_normalized(n_dim, n_samples, i * rng)
criter = self.maxpro_criter(h)
if best > criter:
best = criter
h_opt = h.copy()
h_opt = space.inverse_transform(h_opt)
space.set_transformer(transformer)
return h_opt
class Optimizer(object):
"""Run bayesian optimisation loop.
An `Optimizer` represents the steps of a bayesian optimisation loop. To
use it you need to provide your own loop mechanism. The various
optimisers provided by `skopt` use this class under the hood.
Use this class directly if you want to control the iterations of your
bayesian optimisation loop.
Parameters
----------
* `dimensions` [list, shape=(n_dims,)]:
List of search space dimensions.
Each search dimension can be defined either as
- a `(lower_bound, upper_bound)` tuple (for `Real` or `Integer`
dimensions),
- a `(lower_bound, upper_bound, "prior")` tuple (for `Real`
dimensions),
- as a list of categories (for `Categorical` dimensions), or
- an instance of a `Dimension` object (`Real`, `Integer` or
`Categorical`).
* `base_estimator` ["GP", "RF", "ET", "GBRT" or sklearn regressor, default="GP"]:
Should inherit from `sklearn.base.RegressorMixin`.
In addition the `predict` method, should have an optional `return_std`
argument, which returns `std(Y | x)`` along with `E[Y | x]`.
If base_estimator is one of ["GP", "RF", "ET", "GBRT"], a default
surrogate model of the corresponding type is used corresponding to what
is used in the minimize functions.
* `n_random_starts` [int, default=10]:
DEPRECATED, use `n_initial_points` instead.
* `n_initial_points` [int, default=10]:
Number of evaluations of `func` with initialization points
before approximating it with `base_estimator`. Points provided as
`x0` count as initialization points. If len(x0) < n_initial_points
additional points are sampled at random.
* `acq_func` [string, default=`"gp_hedge"`]:
Function to minimize over the posterior distribution. Can be either
- `"LCB"` for lower confidence bound.
- `"EI"` for negative expected improvement.
- `"PI"` for negative probability of improvement.
- `"gp_hedge"` Probabilistically choose one of the above three
acquisition functions at every iteration.
- The gains `g_i` are initialized to zero.
- At every iteration,
- Each acquisition function is optimised independently to
propose an candidate point `X_i`.
- Out of all these candidate points, the next point `X_best` is
chosen by $softmax(\eta g_i)$
- After fitting the surrogate model with `(X_best, y_best)`,
the gains are updated such that $g_i -= \mu(X_i)$
- `"EIps" for negated expected improvement per second to take into
account the function compute time. Then, the objective function is
assumed to return two values, the first being the objective value and
the second being the time taken in seconds.
- `"PIps"` for negated probability of improvement per second. The
return type of the objective function is assumed to be similar to
that of `"EIps
* `acq_optimizer` [string, `"sampling"` or `"lbfgs"`, default=`"auto"`]:
Method to minimize the acquistion function. The fit model
is updated with the optimal value obtained by optimizing `acq_func`
with `acq_optimizer`.
- If set to `"auto"`, then `acq_optimizer` is configured on the
basis of the base_estimator and the space searched over.
If the space is Categorical or if the estimator provided based on
tree-models then this is set to be "sampling"`.
- If set to `"sampling"`, then `acq_func` is optimized by computing
`acq_func` at `n_points` randomly sampled points.
- If set to `"lbfgs"`, then `acq_func` is optimized by
- Sampling `n_restarts_optimizer` points randomly.
- `"lbfgs"` is run for 20 iterations with these points as initial
points to find local minima.
- The optimal of these local minima is used to update the prior.
* `random_state` [int, RandomState instance, or None (default)]:
Set random state to something other than None for reproducible
results.
* `acq_func_kwargs` [dict]:
Additional arguments to be passed to the acquistion function.
* `acq_optimizer_kwargs` [dict]:
Additional arguments to be passed to the acquistion optimizer.
Attributes
----------
* `Xi` [list]:
Points at which objective has been evaluated.
* `yi` [scalar]:
Values of objective at corresponding points in `Xi`.
* `models` [list]:
Regression models used to fit observations and compute acquisition
function.
* `space`
An instance of `skopt.space.Space`. Stores parameter search space used
to sample points, bounds, and type of parameters.
"""
def __init__(self,
dimensions,
base_estimator="gp",
n_random_starts=None,
n_initial_points=10,
acq_func="gp_hedge",
acq_optimizer="auto",
random_state=None,
acq_func_kwargs=None,
acq_optimizer_kwargs=None):
self.rng = check_random_state(random_state)
# Configure acquisition function
self.acq_func = acq_func
# monitor runtime
self.regressiontime = None
self.actime = None
self.tt_regressiontime = []
self.tt_actime = []
self.randompoints = []
self.values_ = None
# Store and creat acquisition function set
self.acq_func_kwargs = acq_func_kwargs
allowed_acq_funcs = ["gp_hedge", "EI", "LCB", "PI", "EIps", "PIps"]
if self.acq_func not in allowed_acq_funcs:
raise ValueError("expected acq_func to be in %s, got %s" %
(",".join(allowed_acq_funcs), self.acq_func))
# treat hedging method separately
if self.acq_func == "gp_hedge":
self.cand_acq_funcs_ = ["EI", "LCB", "PI"]
self.gains_ = np.zeros(3)
else:
self.cand_acq_funcs_ = [self.acq_func]
if acq_func_kwargs is None:
acq_func_kwargs = dict()
self.eta = acq_func_kwargs.get("eta", 1.0)
# Configure counters of points
# Check `n_random_starts` deprecation first
if n_random_starts is not None:
warnings.warn(("n_random_starts will be removed in favour of "
"n_initial_points."), DeprecationWarning)
n_initial_points = n_random_starts
if n_initial_points < 0:
raise ValueError("Expected `n_initial_points` >= 0, got %d" %
n_initial_points)
self._n_initial_points = n_initial_points
self.n_initial_points_ = n_initial_points
# Configure estimator
# build base_estimator if doesn't exist
if isinstance(base_estimator, str):
base_estimator = cook_estimator(base_estimator,
space=dimensions,
random_state=self.rng.randint(
0,
np.iinfo(np.int32).max))
# check if regressor
if not is_regressor(base_estimator) and base_estimator is not None:
raise ValueError("%s has to be a regressor." % base_estimator)
# treat per second acqusition function specially
is_multi_regressor = isinstance(base_estimator, MultiOutputRegressor)
if "ps" in self.acq_func and not is_multi_regressor:
self.base_estimator_ = MultiOutputRegressor(base_estimator)
else:
self.base_estimator_ = base_estimator
# Configure optimizer
# decide optimizer based on gradient information
if acq_optimizer == "auto":
if has_gradients(self.base_estimator_):
acq_optimizer = "lbfgs"
else:
acq_optimizer = "sampling"
if acq_optimizer not in ["lbfgs", "sampling"]:
raise ValueError("Expected acq_optimizer to be 'lbfgs' or "
"'sampling', got {0}".format(acq_optimizer))
if (not has_gradients(self.base_estimator_)
and acq_optimizer != "sampling"):
raise ValueError("The regressor {0} should run with "
"acq_optimizer"
"='sampling'.".format(type(base_estimator)))
self.acq_optimizer = acq_optimizer
# record other arguments
if acq_optimizer_kwargs is None:
acq_optimizer_kwargs = dict()
self.n_points = acq_optimizer_kwargs.get("n_points", 10000)
self.n_restarts_optimizer = acq_optimizer_kwargs.get(
"n_restarts_optimizer", 5)
n_jobs = acq_optimizer_kwargs.get("n_jobs", 1)
self.n_jobs = n_jobs
self.acq_optimizer_kwargs = acq_optimizer_kwargs
# Configure search space
# normalize space if GP regressor
if isinstance(self.base_estimator_, GaussianProcessRegressor):
dimensions = normalize_dimensions(dimensions)
self.space = Space(dimensions)
# record categorical and non-categorical indices
self._cat_inds = []
self._non_cat_inds = []
for ind, dim in enumerate(self.space.dimensions):
if isinstance(dim, Categorical):
self._cat_inds.append(ind)
else:
self._non_cat_inds.append(ind)
# Initialize storage for optimization
self.models = []
self.Xi = []
self.yi = []
# Initialize cache for `ask` method responses
# This ensures that multiple calls to `ask` with n_points set
# return same sets of points. Reset to {} at every call to `tell`.
self.cache_ = {}
def copy(self, random_state=None):
"""Create a shallow copy of an instance of the optimizer.
Parameters
----------
* `random_state` [int, RandomState instance, or None (default)]:
Set the random state of the copy.
"""
optimizer = Optimizer(
dimensions=self.space.dimensions,
base_estimator=self.base_estimator_,
n_initial_points=self.n_initial_points_,
acq_func=self.acq_func,
acq_optimizer=self.acq_optimizer,
acq_func_kwargs=self.acq_func_kwargs,
acq_optimizer_kwargs=self.acq_optimizer_kwargs,
random_state=random_state,
)
if hasattr(self, "gains_"):
optimizer.gains_ = np.copy(self.gains_)
if self.Xi:
optimizer._tell(self.Xi, self.yi)
return optimizer
def ask(self, n_points=None, nrandom=2, strategy="cl_min"):
"""Query point or multiple points at which objective should be evaluated.
* `n_points` [int or None, default=None]:
Number of points returned by the ask method.
If the value is None, a single point to evaluate is returned.
Otherwise a list of points to evaluate is returned of size
n_points. This is useful if you can evaluate your objective in
parallel, and thus obtain more objective function evaluations per
unit of time.
* `strategy` [string, default=`"cl_min"`]:
Method to use to sample multiple points (see also `n_points`
description). This parameter is ignored if n_points = None.
Supported options are `"cl_min"`, `"cl_mean"` or `"cl_max"`.
- If set to `"cl_min"`, then constant liar strtategy is used
with lie objective value being minimum of observed objective
values. `"cl_mean"` and `"cl_max"` means mean and max of values
respectively. For details on this strategy see:
https://hal.archives-ouvertes.fr/hal-00732512/document
With this strategy a copy of optimizer is created, which is
then asked for a point, and the point is told to the copy of
optimizer with some fake objective (lie), the next point is
asked from copy, it is also told to the copy with fake
objective and so on. The type of lie defines different
flavours of `cl_x` strategies.
"""
if n_points is None:
return self._ask()
# if nrandom is None:
# return self._ask()
# else:
# print("You are using serail BO now, the nrandom must be None! Reset nrandom")
# nrandom = None
# return self._ask()
supported_strategies = ["cl_min", "cl_mean", "cl_max"]
if not (isinstance(n_points, int) and n_points > 0):
raise ValueError("n_points should be int > 1, got " +
str(n_points))
if strategy not in supported_strategies:
raise ValueError("Expected parallel_strategy to be one of " +
str(supported_strategies) + ", " +
"got %s" % strategy)
# Caching the result with n_points not None. If some new parameters
# are provided to the ask, the cache_ is not used.
if (n_points, strategy) in self.cache_:
return self.cache_[(n_points, strategy)]
# Copy of the optimizer is made in order to manage the
# deletion of points with "lie" objective (the copy of
# oiptimizer is simply discarded)
opt = self.copy(random_state=self.rng.randint(0,
np.iinfo(np.int32).max))
X = []
if n_points > nrandom:
for i in range(n_points - nrandom):
x = opt.ask()
X.append(x)
# print(f'1step Selected {len(X)} points')
ti_available = "ps" in self.acq_func and len(opt.yi) > 0
ti = [t for (_, t) in opt.yi] if ti_available else None
if strategy == "cl_min":
y_lie = np.min(opt.yi) if opt.yi else 0.0 # CL-min lie
t_lie = np.min(ti) if ti is not None else log(
sys.float_info.max)
elif strategy == "cl_mean":
y_lie = np.mean(opt.yi) if opt.yi else 0.0 # CL-mean lie
t_lie = np.mean(ti) if ti is not None else log(
sys.float_info.max)
else:
y_lie = np.max(opt.yi) if opt.yi else 0.0 # CL-max lie
t_lie = np.max(ti) if ti is not None else log(
sys.float_info.max)
# Lie to the optimizer.
if "ps" in self.acq_func:
# Use `_tell()` instead of `tell()` to prevent repeated
# log transformations of the computation times.
opt._tell(x, (y_lie, t_lie))
else:
opt._tell(x, y_lie)
#print(f'Selected {len(X)} evaluated points')
if nrandom == 0:
pass
elif nrandom > 0:
for j in range(nrandom):
x_random = opt._ask_random()
X.append(x_random)
self.randompoints.append(x_random)
else:
ValueError("nrandom muss be nonnegative integer, but got ",
nrandom)
#print(f'Selected {len(X)} points')
else:
for i in range(n_points):
x = opt.ask()
X.append(x)
ti_available = "ps" in self.acq_func and len(opt.yi) > 0
ti = [t for (_, t) in opt.yi] if ti_available else None
if strategy == "cl_min":
y_lie = np.min(opt.yi) if opt.yi else 0.0 # CL-min lie
t_lie = np.min(ti) if ti is not None else log(
sys.float_info.max)
elif strategy == "cl_mean":
y_lie = np.mean(opt.yi) if opt.yi else 0.0 # CL-mean lie
t_lie = np.mean(ti) if ti is not None else log(
sys.float_info.max)
else:
y_lie = np.max(opt.yi) if opt.yi else 0.0 # CL-max lie
t_lie = np.max(ti) if ti is not None else log(
sys.float_info.max)
# Lie to the optimizer.
if "ps" in self.acq_func:
# Use `_tell()` instead of `tell()` to prevent repeated
# log transformations of the computation times.
opt._tell(x, (y_lie, t_lie))
else:
opt._tell(x, y_lie)
#print(f'Selected {len(X)} points')
self.cache_ = {(n_points, strategy): X} # cache_ the result
return X
def _ask(self):
"""Suggest next point at which to evaluate the objective.
Return a random point while not at least `n_initial_points`
observations have been `tell`ed, after that `base_estimator` is used
to determine the next point.
"""
if self._n_initial_points > 0 or self.base_estimator_ is None:
# this will not make a copy of `self.rng` and hence keep advancing
# our random state.
return self.space.rvs(random_state=self.rng)[0]
else:
if not self.models:
raise RuntimeError("Random evaluations exhausted and no "
"model has been fit.")
next_x = self._next_x
min_delta_x = min(
[self.space.distance(next_x, xi) for xi in self.Xi])
if abs(min_delta_x) <= 1e-8:
warnings.warn("The objective has been evaluated "
"at this point before.")
# return point computed from last call to tell()
return next_x
def _ask_random(self):
"""Suggest next point at which to evaluate the objective.
Return a random point while not at least `n_initial_points`
observations have been `tell`ed, after that `base_estimator` is used
to determine the next point.
"""
X = self.space.transform(
self.space.rvs(n_samples=self.n_points, random_state=self.rng))
next_xs_ = []
for cand_acq_func in self.cand_acq_funcs_:
# Find the minimum of the acquisition function by randomly
# sampling points from the space
next_x = X[np.argmin(self.values_)]
transformed_bounds = np.array(self.space.transformed_bounds)
if not self.space.is_categorical:
next_x = np.clip(next_x, transformed_bounds[:, 0],
transformed_bounds[:, 1])
next_xs_.append(next_x)
next_x = next_xs_[0]
# note the need for [0] at the end
_next_x = self.space.inverse_transform(next_x.reshape((1, -1)))[0]
# if not self.models:
# raise RuntimeError("Random evaluations exhausted and no "
# "model has been fit.")
next_x = _next_x
min_delta_x = min([self.space.distance(next_x, xi) for xi in self.Xi])
if abs(min_delta_x) <= 1e-8:
warnings.warn("The objective has been evaluated "
"at this point before.")
# return point computed from last call to tell()
return next_x
def tell(self, x, y, fit=True):
"""Record an observation (or several) of the objective function.
Provide values of the objective function at points suggested by `ask()`
or other points. By default a new model will be fit to all
observations. The new model is used to suggest the next point at
which to evaluate the objective. This point can be retrieved by calling
`ask()`.
To add observations without fitting a new model set `fit` to False.
To add multiple observations in a batch pass a list-of-lists for `x`
and a list of scalars for `y`.
Parameters
----------
* `x` [list or list-of-lists]:
Point at which objective was evaluated.
* `y` [scalar or list]:
Value of objective at `x`.
* `fit` [bool, default=True]
Fit a model to observed evaluations of the objective. A model will
only be fitted after `n_initial_points` points have been told to
the optimizer irrespective of the value of `fit`.
"""
check_x_in_space(x, self.space)
self._check_y_is_valid(x, y)
# take the logarithm of the computation times
if "ps" in self.acq_func:
if is_2Dlistlike(x):
y = [[val, log(t)] for (val, t) in y]
elif is_listlike(x):
y = list(y)
y[1] = log(y[1])
return self._tell(x, y, fit=fit)
def _tell(self, x, y, fit=True):
"""Perform the actual work of incorporating one or more new points.
See `tell()` for the full description.
This method exists to give access to the internals of adding points
by side stepping all input validation and transformation."""
if "ps" in self.acq_func:
if is_2Dlistlike(x):
self.Xi.extend(x)
self.yi.extend(y)
self._n_initial_points -= len(y)
elif is_listlike(x):
self.Xi.append(x)
self.yi.append(y)
self._n_initial_points -= 1
# if y isn't a scalar it means we have been handed a batch of points
elif is_listlike(y) and is_2Dlistlike(x):
self.Xi.extend(x)
self.yi.extend(y)
self._n_initial_points -= len(y)
elif is_listlike(x):
self.Xi.append(x)
self.yi.append(y)
self._n_initial_points -= 1
else:
raise ValueError("Type of arguments `x` (%s) and `y` (%s) "
"not compatible." % (type(x), type(y)))
# optimizer learned something new - discard cache
self.cache_ = {}
# after being "told" n_initial_points we switch from sampling
# random points to using a surrogate model
#mprint("starting fit gpr......")
if (fit and self._n_initial_points <= 0
and self.base_estimator_ is not None):
#print('start finding next x.....')
transformed_bounds = np.array(self.space.transformed_bounds)
est = clone(self.base_estimator_)
start_clock = time.time()
with warnings.catch_warnings():
warnings.simplefilter("ignore")
est.fit(self.space.transform(self.Xi), self.yi)
self.regressiontime = time.time() - start_clock
self.tt_regressiontime.append(self.regressiontime)
if hasattr(self, "next_xs_") and self.acq_func == "gp_hedge":
self.gains_ -= est.predict(np.vstack(self.next_xs_))
self.models.append(est)
#mprint("finished fit gpr......")
# even with BFGS as optimizer we want to sample a large number
# of points and then pick the best ones as starting points
start_clock_ac = time.time()
X = self.space.transform(
self.space.rvs(n_samples=self.n_points, random_state=self.rng))
self.next_xs_ = []
for cand_acq_func in self.cand_acq_funcs_:
values = _gaussian_acquisition(
X=X,
model=est,
y_opt=np.min(self.yi),
acq_func=cand_acq_func,
acq_func_kwargs=self.acq_func_kwargs)
self.values = values
# Find the minimum of the acquisition function by randomly
# sampling points from the space
if self.acq_optimizer == "sampling":
next_x = X[np.argmin(values)]
# Use BFGS to find the mimimum of the acquisition function, the
# minimization starts from `n_restarts_optimizer` different
# points and the best minimum is used
elif self.acq_optimizer == "lbfgs":
x0 = X[np.argsort(values)[:self.n_restarts_optimizer]]
with warnings.catch_warnings():
warnings.simplefilter("ignore")
results = Parallel(n_jobs=self.n_jobs)(
delayed(fmin_l_bfgs_b)(
gaussian_acquisition_1D,
x,
args=(est, np.min(self.yi), cand_acq_func,
self.acq_func_kwargs),
bounds=self.space.transformed_bounds,
approx_grad=False,
maxiter=20) for x in x0)
cand_xs = np.array([r[0] for r in results])
cand_acqs = np.array([r[1] for r in results])
next_x = cand_xs[np.argmin(cand_acqs)]
self.actime = start_clock_ac - time.time()
self.tt_actime.append(self.actime)
# lbfgs should handle this but just in case there are
# precision errors.
if not self.space.is_categorical:
next_x = np.clip(next_x, transformed_bounds[:, 0],
transformed_bounds[:, 1])
self.next_xs_.append(next_x)
if self.acq_func == "gp_hedge":
logits = np.array(self.gains_)
logits -= np.max(logits)
exp_logits = np.exp(self.eta * logits)
probs = exp_logits / np.sum(exp_logits)
next_x = self.next_xs_[np.argmax(self.rng.multinomial(
1, probs))]
else:
next_x = self.next_xs_[0]
# note the need for [0] at the end
self._next_x = self.space.inverse_transform(next_x.reshape(
(1, -1)))[0]
# Pack results
return create_result(self.Xi,
self.yi,
self.space,
self.rng,
models=self.models)
def _check_y_is_valid(self, x, y):
"""Check if the shape and types of x and y are consistent."""
if "ps" in self.acq_func:
if is_2Dlistlike(x):
if not (np.ndim(y) == 2 and np.shape(y)[1] == 2):
raise TypeError("expected y to be a list of (func_val, t)")
elif is_listlike(x):
if not (np.ndim(y) == 1 and len(y) == 2):
raise TypeError("expected y to be (func_val, t)")
# if y isn't a scalar it means we have been handed a batch of points
elif is_listlike(y) and is_2Dlistlike(x):
for y_value in y:
if not isinstance(y_value, Number):
raise ValueError("expected y to be a list of scalars")
elif is_listlike(x):
if not isinstance(y, Number):
print(type(y))
raise ValueError("`func` should return a scalar")
else:
raise ValueError("Type of arguments `x` (%s) and `y` (%s) "
"not compatible." % (type(x), type(y)))
def run(self, func, n_iter=1):
"""Execute ask() + tell() `n_iter` times"""
for _ in range(n_iter):
x = self.ask()
self.tell(x, func(x))
return create_result(self.Xi,
self.yi,
self.space,
self.rng,
models=self.models)
def get_regressiontime(self):
return (self.regressiontime)
def get_acntime(self):
return (self.actime)
def get_randompoint(self):
return (self.randompoints)
def SkoptCMAoptimizer(
func,
dimensions,
n_calls,
verbose=False,
callback=(),
x0=None,
sigma0=.5,
normalize=True,
):
'''
Optmizer based on CMA-ES algorithm.
This is essentially a wrapper fuction for the cma library function
to align the interface with skopt library.
Args:
func (callable): function to optimize
dimensions: list of tuples. search dimensions
n_calls: the number of samples.
verbose: if this func should be verbose
callback: the list of callback functions.
x0: inital values
if None, random point will be sampled
sigma0: initial standard deviation
normalize: whether optimization domain should be normalized
Returns:
`res` skopt.OptimizeResult object
The optimization result returned as a dict object.
Important attributes are:
- `x` [list]: location of the minimum.
- `fun` [float]: function value at the minimum.
- `x_iters` [list of lists]: location of function evaluation for each
iteration.
- `func_vals` [array]: function value for each iteration.
- `space` [Space]: the optimization space.
'''
specs = {
'args': copy.copy(inspect.currentframe().f_locals),
'function': inspect.currentframe().f_code.co_name,
}
if normalize:
dimensions = list(
map(lambda x: check_dimension(x, 'normalize'), dimensions))
space = Space(dimensions)
if x0 is None: x0 = space.transform(space.rvs())[0]
tempdir = tempfile.mkdtemp()
xi, yi = [], []
options = {
'bounds': np.array(space.transformed_bounds).transpose().tolist(),
'verb_filenameprefix': tempdir,
}
def delete_tempdir(self, *args, **kargs):
os.removedirs(tempdir)
return
model = cma.CMAEvolutionStrategy(x0, sigma0, options)
model.logger.__del__ = delete_tempdir
for i in range(n_calls):
if model.stop(): break
new_xi = model.ask()
new_xi_denorm = space.inverse_transform(np.array(new_xi))
new_yi = [func(x) for x in new_xi_denorm]
model.tell(new_xi, new_yi)
model.logger.add()
if verbose: model.disp()
xi += new_xi_denorm
yi += new_yi
results = create_result(xi, yi)
for f in callback:
f(results)
results = create_result(xi, yi, space)
model.logger.load()
results.cma_logger = model.logger
results.specs = specs
return results
