def _convert_space_hop_skopt(space):
dimensions = []
for name, specs in space.items():
specs = str(specs).split('\n')
method = specs[3].split(' ')[-1]
bounds = specs[4:]
if len(bounds) == 1:
bounds = bounds[0].split('range')[-1]
bounds = bounds.replace('(', '').replace(')', '').replace('}', '')
low, high = [float(v) for v in bounds.split(',')]
else:
vals = [
float(
b.split('Literal')[-1].replace('}', '').replace('{', ''))
for b in bounds
]
low = min(vals)
high = max(vals)
if method == 'randint':
dimensions.append(skopt.space.Integer(low, high, name=name))
elif method == 'uniform':
dimensions.append(
skopt.space.Real(low, high, name=name, prior='uniform'))
elif method == 'loguniform':
low, high = np.exp(low), np.exp(high)
dimensions.append(
skopt.space.Real(low, high, name=name, prior='log-uniform'))
else:
raise NotImplementedError
skopt_space = skopt.Space(dimensions)
return skopt_space
def _convert_to_param_space(df, param_cols, param_types):
dimensions = []
for colname, col_type in zip(param_cols, param_types):
if col_type == str:
dimensions.append(
skopt.space.Categorical(categories=df[colname].unique(),
name=colname))
elif col_type == float:
low, high = df[colname].min(), df[colname].max()
dimensions.append(skopt.space.Real(low, high, name=colname))
else:
raise NotImplementedError
skopt_space = skopt.Space(dimensions)
return skopt_space
def dispatch(self, *args, **kwargs):
import skopt
minim = self._obj
self._ensure_object(minim)
if len(args) > 0:
raise ValueError(
f"{minim.__class__.__name__}.to.skopt only accepts keyword arguments"
)
# First create the Space variable
def skoptReal(v):
low, high = v.bounds
return skopt.space.Real(low,
high,
transform="identity",
name=v.name)
space = skopt.Space(list(map(skoptReal, self.variables)))
# Extract sampled data-points
if "x" in kwargs:
Xi = kwargs.pop("x")
yi = kwargs.pop("y")
else:
Xi = np.array(self.data.x)
yi = np.array(self.data.y)
if "models" not in kwargs:
import sklearn
# We can't use categorial (SVC) since these are regression models
# fast, but should not be as accurate?
#model = sklearn.svm.LinearSVR()
# much slower, but more versatile
# I don't know which one is better ;)
model = sklearn.svm.SVR(cache_size=500)
#model = sklearn.svm.NuSVR(kernel="poly", cache_size=500)
# we need to fit to create auxiliary data
warnings.warn(
f"Converting to skopt without a 'models' argument forces "
f"{minim.__class__.__name__} to train a model for the sampled data. "
f"This may be slow depending on the number of samples...")
model.fit(Xi, yi)
kwargs["models"] = [model]
result = skopt.utils.create_result(Xi, yi, space=space, **kwargs)
return result