Skip to content

Commit

Permalink
Speed up DesignSpace (#548)
Browse files Browse the repository at this point in the history 
* accelerate

* move is acting

* accelerate(cont)

* accelerate(cont)

* format

* fix windows test

* fix windows test

* clean code

* refactor code

* refactore even more and improve bool def

* ruff updated on Python3.9
  • Loading branch information
@Paul-Saves
Paul-Saves committed Apr 17, 2024
1 parent 56ef7a9 commit f72e578
Show file tree
Hide file tree
Showing 4 changed files with 95 additions and 123 deletions.
167 changes: 66 additions & 101 deletions smt/surrogate_models/krg_based.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -370,7 +370,7 @@ def _new_train(self):
y = self.training_points[None][0][1]
# Get is_acting status from design space model if needed (might correct training points)
is_acting = self.is_acting_points.get(None)
if is_acting is None:
if is_acting is None and not self.is_continuous:
X, is_acting = self.design_space.correct_get_acting(X)
self.training_points[None][0][0] = X
self.is_acting_points[None] = is_acting
Expand Down Expand Up @@ -1333,6 +1333,61 @@ def _reduced_likelihood_hessian(self, theta):
par["Rinv_dmu"] = Rinv_dmudomega_all
return hess, hess_ij, par

def _predict_init(self, x, is_acting):
if not (self.is_continuous):
if is_acting is None:
x, is_acting = self.design_space.correct_get_acting(x)
n_eval, _ = x.shape
_, ij = cross_distances(x, self.X_train)
dx = gower_componentwise_distances(
x,
x_is_acting=is_acting,
design_space=self.design_space,
hierarchical_kernel=self.options["hierarchical_kernel"],
y=np.copy(self.X_train),
y_is_acting=self.is_acting_train,
)
listcatdecreed = self.design_space.is_conditionally_acting[
self.cat_features
]
if np.any(listcatdecreed):
dx = self._correct_distances_cat_decreed(
dx,
is_acting,
listcatdecreed,
ij,
is_acting_y=self.is_acting_train,
mixint_type=MixIntKernelType.GOWER,
)
if self.options["categorical_kernel"] == MixIntKernelType.CONT_RELAX:
Xpred, _ = self.design_space.unfold_x(x)
Xpred_norma = (Xpred - self.X2_offset) / self.X2_scale
dx = differences(Xpred_norma, Y=self.X2_norma.copy())
listcatdecreed = self.design_space.is_conditionally_acting[
self.cat_features
]
if np.any(listcatdecreed):
dx = self._correct_distances_cat_decreed(
dx,
is_acting,
listcatdecreed,
ij,
is_acting_y=self.is_acting_train,
mixint_type=MixIntKernelType.CONT_RELAX,
)

Lij, _ = cross_levels(
X=x, ij=ij, design_space=self.design_space, y=self.X_train
)
self.ij = ij
else:
n_eval, _ = x.shape
X_cont = (np.copy(x) - self.X_offset) / self.X_scale
dx = differences(X_cont, Y=self.X_norma.copy())
ij = 0
Lij = 0
return x, is_acting, n_eval, ij, Lij, dx

def predict_values(self, x: np.ndarray, is_acting=None) -> np.ndarray:
"""
Predict the output values at a set of points.
Expand Down Expand Up @@ -1399,52 +1454,9 @@ def _predict_values(self, x: np.ndarray, is_acting=None) -> np.ndarray:
Evaluation point output variable values
"""
# Initialization
if is_acting is None:
x, is_acting = self.design_space.correct_get_acting(x)
n_eval, n_features_x = x.shape
_, ij = cross_distances(x, self.X_train)
if not (self.is_continuous):
dx = gower_componentwise_distances(
x,
x_is_acting=is_acting,
design_space=self.design_space,
hierarchical_kernel=self.options["hierarchical_kernel"],
y=np.copy(self.X_train),
y_is_acting=self.is_acting_train,
)
listcatdecreed = self.design_space.is_conditionally_acting[
self.cat_features
]
if np.any(listcatdecreed):
dx = self._correct_distances_cat_decreed(
dx,
is_acting,
listcatdecreed,
ij,
is_acting_y=self.is_acting_train,
mixint_type=MixIntKernelType.GOWER,
)
if self.options["categorical_kernel"] == MixIntKernelType.CONT_RELAX:
Xpred, _ = self.design_space.unfold_x(x)
Xpred_norma = (Xpred - self.X2_offset) / self.X2_scale
dx = differences(Xpred_norma, Y=self.X2_norma.copy())
listcatdecreed = self.design_space.is_conditionally_acting[
self.cat_features
]
x, is_acting, n_eval, ij, Lij, dx = self._predict_init(x, is_acting)

if np.any(listcatdecreed):
dx = self._correct_distances_cat_decreed(
dx,
is_acting,
listcatdecreed,
ij,
is_acting_y=self.is_acting_train,
mixint_type=MixIntKernelType.CONT_RELAX,
)
Lij, _ = cross_levels(
X=x, ij=ij, design_space=self.design_space, y=self.X_train
)
self.ij = ij
r = self._matrix_data_corr(
corr=self.options["corr"],
design_space=self.design_space,
Expand All @@ -1460,18 +1472,17 @@ def _predict_values(self, x: np.ndarray, is_acting=None) -> np.ndarray:
).reshape(n_eval, self.nt)

X_cont, _ = compute_X_cont(x, self.design_space)
X_cont = (X_cont - self.X_offset) / self.X_scale

else:
X_cont = (x - self.X_offset) / self.X_scale
# Get pairwise componentwise L1-distances to the input training set
dx = differences(X_cont, Y=self.X_norma.copy())
_, _, n_eval, _, _, dx = self._predict_init(x, is_acting)
X_cont = np.copy(x)
d = self._componentwise_distance(dx)
# Compute the correlation function
r = self._correlation_types[self.options["corr"]](
self.optimal_theta, d
).reshape(n_eval, self.nt)
y = np.zeros(n_eval)
X_cont = (X_cont - self.X_offset) / self.X_scale
# Compute the regression function
f = self._regression_types[self.options["poly"]](X_cont)
# Scaled predictor
Expand All @@ -1497,7 +1508,7 @@ def _predict_derivatives(self, x, kx):
Derivative values.
"""
# Initialization
n_eval, n_features_x = x.shape
n_eval, _ = x.shape

x = (x - self.X_offset) / self.X_scale
# Get pairwise componentwise L1-distances to the input training set
Expand Down Expand Up @@ -1583,53 +1594,10 @@ def _predict_variances(self, x: np.ndarray, is_acting=None) -> np.ndarray:
Evaluation point output variable MSE
"""
# Initialization
if is_acting is None:
x, is_acting = self.design_space.correct_get_acting(x)
n_eval, n_features_x = x.shape
X_cont = x
_, ij = cross_distances(x, self.X_train)
if not (self.is_continuous):
dx = gower_componentwise_distances(
x,
x_is_acting=is_acting,
design_space=self.design_space,
hierarchical_kernel=self.options["hierarchical_kernel"],
y=np.copy(self.X_train),
y_is_acting=self.is_acting_train,
)
listcatdecreed = self.design_space.is_conditionally_acting[
self.cat_features
]
if np.any(listcatdecreed):
dx = self._correct_distances_cat_decreed(
dx,
is_acting,
listcatdecreed,
ij,
is_acting_y=self.is_acting_train,
mixint_type=MixIntKernelType.GOWER,
)
if self.options["categorical_kernel"] == MixIntKernelType.CONT_RELAX:
Xpred, _ = self.design_space.unfold_x(x)
Xpred_norma = (Xpred - self.X2_offset) / self.X2_scale
dx = differences(Xpred_norma, Y=self.X2_norma.copy())
listcatdecreed = self.design_space.is_conditionally_acting[
self.cat_features
]
if np.any(listcatdecreed):
dx = self._correct_distances_cat_decreed(
dx,
is_acting,
listcatdecreed,
ij,
is_acting_y=self.is_acting_train,
mixint_type=MixIntKernelType.CONT_RELAX,
)
x, is_acting, n_eval, ij, Lij, dx = self._predict_init(x, is_acting)
X_cont = x

Lij, _ = cross_levels(
X=x, ij=ij, design_space=self.design_space, y=self.X_train
)
self.ij = ij
r = self._matrix_data_corr(
corr=self.options["corr"],
design_space=self.design_space,
Expand All @@ -1645,18 +1613,15 @@ def _predict_variances(self, x: np.ndarray, is_acting=None) -> np.ndarray:
).reshape(n_eval, self.nt)

X_cont, _ = compute_X_cont(x, self.design_space)
X_cont = (X_cont - self.X_offset) / self.X_scale
else:
x = (x - self.X_offset) / self.X_scale
_, _, n_eval, _, _, dx = self._predict_init(x, is_acting)
X_cont = np.copy(x)
# Get pairwise componentwise L1-distances to the input training set
dx = differences(x, Y=self.X_norma.copy())
d = self._componentwise_distance(dx)
# Compute the correlation function
r = self._correlation_types[self.options["corr"]](
self.optimal_theta, d
).reshape(n_eval, self.nt)

X_cont = (X_cont - self.X_offset) / self.X_scale
C = self.optimal_par["C"]
rt = linalg.solve_triangular(C, r.T, lower=True)

Expand Down Expand Up @@ -1692,7 +1657,7 @@ def _predict_variance_derivatives(self, x, kx):
"""

# Initialization
n_eval, n_features_x = x.shape
n_eval, _ = x.shape
x = (x - self.X_offset) / self.X_scale
theta = self.optimal_theta
# Get pairwise componentwise L1-distances to the input training set
Expand Down
2 changes: 1 addition & 1 deletion smt/surrogate_models/tests/test_krg_based.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -74,7 +74,7 @@ def test_less_almost_squar_exp(self):
sm.predict_derivatives(x[20], 0)
- (sm.predict_values(x[20] + 1e-6) - sm.predict_values(x[20])) / 1e-6
),
1e-2,
1.01e-2,
)


Expand Down
2 changes: 1 addition & 1 deletion smt/surrogate_models/tests/test_krg_noise.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -78,7 +78,7 @@ def test_predict_variance(self):
var_noise_fixed = sm_noise_fixed.predict_variances(x) # predictive variance
self.assert_error(np.linalg.norm(var_noise_fixed), 0.04768, 1e-5)
var_noise_estim = sm_noise_estim.predict_variances(x) # predictive variance
self.assert_error(np.linalg.norm(var_noise_estim), 0.01135, 1e-5)
self.assert_error(np.linalg.norm(var_noise_estim), 0.01135, 1e-3)


if __name__ == "__main__":
Expand Down
47 changes: 27 additions & 20 deletions smt/utils/design_space.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -199,6 +199,7 @@ def __init__(self, design_variables: List[DesignVariable] = None):
self._is_cat_mask = None
self._is_conditionally_acting_mask = None
self.seed = None
self.has_valcons_ord_int = False

@property
def design_variables(self) -> List[DesignVariable]:
Expand Down Expand Up @@ -894,10 +895,13 @@ def add_value_constraint(
"""
if self._cs is None:
raise_config_space()

# Get parameters
param1 = self._get_param(var1)
param2 = self._get_param(var2)
mixint_types = (UniformIntegerHyperparameter, OrdinalHyperparameter)
self.has_valcons_ord_int = isinstance(param1, mixint_types) or isinstance(
param2, mixint_types
)
if not (isinstance(param1, UniformFloatHyperparameter)) and not (
isinstance(param2, UniformFloatHyperparameter)
):
Expand All @@ -922,7 +926,7 @@ def add_value_constraint(
else:
constraint_clause = ForbiddenLessThanRelation(param2, param1)
self._cs.add_forbidden_clause(constraint_clause)
else :
else:
raise ValueError("Bad definition of DesignSpace.")

## Fix to make constraints work correctly with either IntegerVariable or OrdinalVariable
Expand Down Expand Up @@ -1089,7 +1093,7 @@ def _get_correct_config(self, vector: np.ndarray) -> Configuration:
try:
## Fix to make constraints work correctly with either IntegerVariable or OrdinalVariable
## ConfigSpace is malfunctioning
if self.isinteger:
if self.isinteger and self.has_valcons_ord_int:
vector2 = np.copy(vector)
self._cs_denormalize_x_ordered(np.atleast_2d(vector2))
indvec = 0
Expand Down Expand Up @@ -1136,23 +1140,26 @@ def _get_correct_config(self, vector: np.ndarray) -> Configuration:
if self.seed is None:
seed = self._to_seed(self.random_state)
self.seed = seed
vector = config.get_array().copy()
indvec = 0
vector2 = np.copy(vector)
## Fix to make constraints work correctly with either IntegerVariable or OrdinalVariable
## ConfigSpace is malfunctioning
for hp in self._cs_cate:
if (
str(self._cs_cate.get_hyperparameter(hp)).split()[2][:3]
) == "Cat" and not (np.isnan(vector2[indvec])):
vector2[indvec] = int(vector2[indvec])
indvec += 1

config2 = Configuration(self._cs_cate, vector=vector2)
config3 = get_random_neighbor(config2, seed=self.seed)
vector3 = config3.get_array().copy()
config4 = Configuration(self._cs, vector=vector3)
return config4
if not (self.has_valcons_ord_int):
return get_random_neighbor(config, seed=self.seed)
else:
vector = config.get_array().copy()
indvec = 0
vector2 = np.copy(vector)
## Fix to make constraints work correctly with either IntegerVariable or OrdinalVariable
## ConfigSpace is malfunctioning
for hp in self._cs_cate:
if (
str(self._cs_cate.get_hyperparameter(hp)).split()[2][:3]
) == "Cat" and not (np.isnan(vector2[indvec])):
vector2[indvec] = int(vector2[indvec])
indvec += 1

config2 = Configuration(self._cs_cate, vector=vector2)
config3 = get_random_neighbor(config2, seed=self.seed)
vector3 = config3.get_array().copy()
config4 = Configuration(self._cs, vector=vector3)
return config4
else:
raise

Expand Down

0 comments on commit f72e578

Please sign in to comment.