btb.tuning package¶
Subpackages¶
Submodules¶
Module contents¶
Top level of the tuning module.
-
class
btb.tuning.
BooleanHyperParam
(default=False)[source]¶ Bases:
btb.tuning.hyperparams.base.BaseHyperParam
BooleanHyperParam class.
The BooleanHyperParam class is responsible for the transformation of boolean values in to normalized search space of \([0, 1]\), providing the ability to sample values of those and to inverse transform from the search space into the hyperparameter space.
- Hyperparameter space:
{True, False}
- Parameters
default (bool) – Default boolean value for the hyperparameter. Defaults to
False
.
-
cardinality
= 2¶
-
dimensions
= 1¶
-
sample
(n_samples)[source]¶ Generate sample values in the hyperparameter search space \({0, 1}\).
- Parameters
n_samples (int) – Number of values to sample.
- Returns
2D array with shape of
(n_samples, 1)
with normalized values inside the search space \({0, 1}\).- Return type
numpy.ndarray
Example
The example below shows simple usage case where a BooleanHyperParam is being created and it’s
sample
method is being called with a number of samples to be obtained.>>> instance = BooleanHyperParam() >>> instance.sample(2) array([[1], [1]])
-
class
btb.tuning.
CategoricalHyperParam
(choices, default=<object object>)[source]¶ Bases:
btb.tuning.hyperparams.base.BaseHyperParam
CategoricalHyperParam Class.
The CategoricalHyperParam class is responsible for the transform of categorical values in to normalized search space and provides the inverse transform from search space to hyperparameter space. Also provides a method that generates samples of those.
- Hyperparameter space:
\(h_1, h_2,... h_K\) where K is the number of categories.
- Search Space:
\(\{ 0, 1 \}^K\) where K is the number of categories.
- Parameters
choices (list) – List of values that the hyperparameter can be.
default (str or None) – Default value for the hyperparameter to take. Defaults to the first item in
choices
-
NO_DEFAULT
= <object object>¶
-
sample
(n_samples)[source]¶ Generate sample values in the hyperparameter search space of
[0, 1]^K
.- Parameters
n_samples (int) – Number of values to sample.
- Returns
2D array with shape of (n_samples, self.dimensions) with normalized values inside the search space \([0, 1]^K\).
- Return type
numpy.ndarray
Example
The example below shows simple usage case where a CategoricalHyperParam is being created with three possible values, (Cat, Dog, Tiger), and it’s method
sample
is being called with a number of samples to be obtained. Anumpy.ndarray
with values from the search space is being returned.>>> instance = CategoricalHyperParam(choices=['Cat', 'Dog', 'Tiger']) >>> instance.sample(2) array([[1, 0, 0], [0, 1, 0]])
-
class
btb.tuning.
GCPEiTuner
(tunable, maximize=True, num_candidates=1000, min_trials=5, length_scale=0.1)[source]¶ Bases:
btb.tuning.metamodels.gaussian_process.GaussianCopulaProcessMetaModel
,btb.tuning.tuners.gaussian_process.GPEiTuner
Gaussian Copula Process Expected Improvement Tuner.
This class uses a
GaussianProcessRegressor
model from thesklearn.gaussian_process
package, using anExpectedImprovement
function to return the better configurations predicted from the meta model that converts the input data using aUnivariate
copula.
-
class
btb.tuning.
GCPTuner
(tunable, maximize=True, num_candidates=1000, min_trials=5, length_scale=0.1)[source]¶ Bases:
btb.tuning.metamodels.gaussian_process.GaussianCopulaProcessMetaModel
,btb.tuning.tuners.gaussian_process.GPTuner
Gaussian Copula Process Tuner.
This class uses a
GaussianProcessRegressor
model from thesklearn.gaussian_process
package, using anumpy.argmax
function to return the better configurations predicted from the meta model that converts the input data using aUnivariate
copula.
-
class
btb.tuning.
GPEiTuner
(tunable, maximize=True, num_candidates=1000, min_trials=5, length_scale=0.1)[source]¶ Bases:
btb.tuning.metamodels.gaussian_process.GaussianProcessMetaModel
,btb.tuning.acquisition.expected_improvement.ExpectedImprovementAcquisition
,btb.tuning.tuners.base.BaseMetaModelTuner
Gaussian Process Expected Improvement Tuner.
This class uses a
GaussianProcessRegressor
model from thesklearn.gaussian_process
package, using anExpectedImprovement
function to return the better configurations predicted from the model.
-
class
btb.tuning.
GPTuner
(tunable, maximize=True, num_candidates=1000, min_trials=5, length_scale=0.1)[source]¶ Bases:
btb.tuning.metamodels.gaussian_process.GaussianProcessMetaModel
,btb.tuning.acquisition.predicted_score.PredictedScoreAcquisition
,btb.tuning.tuners.base.BaseMetaModelTuner
Gaussian Process Tuner.
This class uses a
GaussianProcessRegressor
model from thesklearn.gaussian_process
package, using anumpy.argmax
function to return the better configurations predicted from the model.
-
class
btb.tuning.
FloatHyperParam
(min=None, max=None, default=None, include_min=True, include_max=True)[source]¶ Bases:
btb.tuning.hyperparams.numerical.NumericalHyperParam
FloatHyperParam class.
The FloatHyperParam class represents a single hyperparameter within a range of
float
numbers, wheremin
andmax
can take as value any float number within that range, havingmin
to be smaller thanmax
.- Hyperparameter space:
\(h_1, h_2,... h_n\) where \(h_i = i * (max - min) + min\)
- Search space:
\(s_1, s_2,... s_n\) where \(s_i = (i - min) / (max - min)\)
- Parameters
min (float) – Float number to represent the minimum value that this hyperparameter can take, by default is
None
which will take the system’s minimum float value possible.max (float) – Float number to represent the maximum value that this hyperparameter can take, by default is
None
which will take the system’s maximum float value possible.default (float) – Float number that represents the default value for the hyperparameter. Defaults to
self.min
include_min (bool) – Either or not to include the minimum value in the search space.
include_max (bool) – Either or not to include the maximum value in the search space.
-
cardinality
= inf¶
-
sample
(n_samples)[source]¶ Generate sample values in the hyperparameter search space \({0, 1}\).
- Parameters
n_samples (int) – Number of values to sample.
- Returns
2D array with shape of (n_samples, 1) with normalized values inside the search space \({0, 1}\).
- Return type
numpy.ndarray
Example
The example below shows simple usage case where a FloatHyperParam is being created with a range that goes from
0.1
to0.9
and it’ssample
method is being called with a number of samples to be obtained. Anumpy.ndarray
with values from the search space is being returned.>>> instance = FloatHyperParam(min=0.1, max=0.9) >>> instance.sample(2) array([[0.52058728], [0.00582452]])
-
class
btb.tuning.
IntHyperParam
(min=None, max=None, default=None, include_min=True, include_max=True, step=1)[source]¶ Bases:
btb.tuning.hyperparams.numerical.NumericalHyperParam
IntHyperParam class.
The IntHyperParam class represents a single hyperparameter within an range of
int
numbers, wheremin
andmax
can take as value anyint
number that compose this range havingmin
to be smaller thanmax
.- Hyperparameter space:
\(h_1, h_2,... h_n\) where \(h_i = min + (i - 1) * step\)
- Search space:
\(s_1, s_2,... s_n\) where \(s_i = \frac{interval}{2} + (i - 1) * interval\)
- Parameters
min (int) – Integer number to represent the minimum value that this hyperparameter can take, by default is
None
which will take the system’s minimum int value possible.max (int) – Integer number to represent the maximum value that this hyperparameter can take, by default is
None
which will take the system’s maximum int value possible.default (int) – Integer number that represents the default value for the hyperparameter. Defaults to
self.min
.step (int) – Increase amount to take for each sample. Defaults to 1.
include_min (bool) – Either or not to include the minimum value in the search space.
include_max (bool) – Either or not to include the maximum value in the search space.
-
dimensions
= 1¶
-
sample
(n_samples)[source]¶ Generate sample values in the hyperparameter search space of [0, 1).
- Parameters
n_samples (int) – Number of values to sample.
- Returns
2D array with shape of (n_samples, 1) with normalized values inside the search space \({0, 1}\).
- Return type
numpy.ndarray
Example
The example below shows simple usage case where a IntHyperParam is being created with a range that goes from
1
to4
and it’ssample
method is being called with a number of samples to be obtained. Anumpy.ndarray
with values from the search space is being returned.>>> instance = IntHyperParam(min=1, max=4) >>> instance.sample(2) array([[0.625], [0.375]])
-
class
btb.tuning.
Tunable
(hyperparams)[source]¶ Bases:
object
Tunable class.
The Tunable class represents a collection of
HyperParams
that need to be tuned as a whole, at once.-
hyperparams
¶ Dict of HyperParams.
-
cardinality
¶ Int or
np.inf
amount that indicates the number of combinations possible for this tunable.
- Parameters
hyperparams (dict) – Dictionary object that contains the name and the hyperparameter asociated to it.
-
cardinality
= 1
-
dimensions
= 0¶
-
classmethod
from_dict
(dict_hyperparams)[source]¶ Create an instance from a dictionary containing information over hyperparameters.
Class method that creates an instance from a dictionary that describes the type of a hyperparameter, the range or values that this can have and the default value of the hyperparameter.
- Parameters
dict_hyperparams (dict) –
A python dictionary containing as key the given name for the hyperparameter and as value a dictionary containing the following keys:
- Type (str):
bool
forBoolHyperParam
,int
forIntHyperParam
,float
forFloatHyperParam
,str
forCategoricalHyperParam
.
- Range or Values (list):
Range / values that this hyperparameter can take, in case of
CategoricalHyperParam
those will be used as thechoices
, forNumericalHyperParams
themin
value will be used as the minimum value and themax
value will be used as themaximum
value.
- Default (str, bool, int, float or None):
The default value for the hyperparameter.
- Returns
A
Tunable
instance with the given hyperparameters.- Return type
-
hyperparams
= None
-
inverse_transform
(values)[source]¶ Invert one or more hyperparameter configurations.
Invert one or more hyperparameter configurations from the normalized search space \([0, 1]^K\) to the original hyperparameter space.
- Parameters
values (array-like) – 2D array of normalized values with shape
(n, dimensions)
wheredimensions
is the sum of dimensions from all theHyperParams
that compose thistunable
.- Returns
pandas.DataFrame
Example
The example below shows a simple usage of a Tunable class which will inverse transform a valid data from a 2D list and a
pandas.DataFrame
will be returned.>>> from btb.tuning.hyperparams.boolean import BooleanHyperParam >>> from btb.tuning.hyperparams.categorical import CategoricalHyperParam >>> from btb.tuning.hyperparams.numerical import IntHyperParam >>> chp = CategoricalHyperParam(['cat', 'dog']) >>> bhp = BooleanHyperParam() >>> ihp = IntHyperParam(1, 10) >>> hyperparams = { ... 'chp': chp, ... 'bhp': bhp, ... 'ihp': ihp ... } >>> tunable = Tunable(hyperparams) >>> values = [ ... [1, 0, 0, 0.95], ... [0, 1, 1, 0.05] ... ] >>> tunable.inverse_transform(values) chp bhp ihp 0 cat False 10 1 dog True 1
-
names
= None¶
-
sample
(n_samples)[source]¶ Sample values in the hyperparameters space for this tunable.
- Parameters
n_samlpes (int) – Number of values to sample.
- Returns
2D array with shape of
(n_samples, dimensions)
wheredimensions
is the sum of dimensions from all theHyperParams
that compose thistunable
.- Return type
numpy.ndarray
Example
The example below shows a simple usage of a Tunable class which will generate 2 samples by calling it’s sample method. This will return a
numpy.ndarray
.>>> from btb.tuning.hyperparams.boolean import BooleanHyperParam >>> from btb.tuning.hyperparams.categorical import CategoricalHyperParam >>> from btb.tuning.hyperparams.numerical import IntHyperParam >>> chp = CategoricalHyperParam(['cat', 'dog']) >>> bhp = BooleanHyperParam() >>> ihp = IntHyperParam(1, 10) >>> hyperparams = { ... 'chp': chp, ... 'bhp': bhp, ... 'ihp': ihp ... } >>> tunable = Tunable(hyperparams) >>> tunable.sample(2) array([[0. , 1. , 0. , 0.45], [1. , 0. , 1. , 0.95]])
-
transform
(values)[source]¶ Transform one or more hyperparameter configurations.
Transform one or more hyperparameter configurations from the original hyperparameter space to the normalized search space.
- Parameters
values (pandas.DataFrame, pandas.Series, dict, list(dict), 2D array-like) – Values of shape
(n, len(self.hyperparams))
.- Returns
2D array of shape
(len(values), dimensions)
wheredimensions
is the sum of dimensions from all theHyperParams
that compose thistunable
.- Return type
numpy.ndarray
Example
The example below shows a simple usage of a Tunable class which will transform a valid data from a 2D list and a
numpy.ndarray
is being returned.>>> from btb.tuning.hyperparams.boolean import BooleanHyperParam >>> from btb.tuning.hyperparams.categorical import CategoricalHyperParam >>> from btb.tuning.hyperparams.numerical import IntHyperParam >>> chp = CategoricalHyperParam(['cat', 'dog']) >>> bhp = BooleanHyperParam() >>> ihp = IntHyperParam(1, 10) >>> hyperparams = { ... 'chp': chp, ... 'bhp': bhp, ... 'ihp': ihp ... } >>> tunable = Tunable(hyperparams) >>> values = [ ... ['cat', False, 10], ... ['dog', True, 1], ... ] >>> tunable.transform(values) array([[1. , 0. , 0. , 0.95], [0. , 1. , 1. , 0.05]])
-