Internal sorting of parameter keys leads to unexpected behaviour when registering an array #530
Description
Describe the bug
Internally, the target space sorts the keys of the pbounds dict before creating the target space bounds. This is presumably because python dicts were unordered before python 3.7 and ensured that there was a canonical order to the parameters. However, after python 3.7 dict keys are now ordered, making this sorting unnecessary. Moreover, the fact that the internal representation is changed can lead to unexpected behaviour when registering a numpy array (see example below). This seems very likely to happen when using a DataFrame representation of data.
To Reproduce
Ex:
from sklearn import datasets
import pandas as pd
from bayes_opt import BayesianOptimization
df = pd.DataFrame(datasets.load_iris().data, columns=datasets.load_iris().feature_names)
df.columns = [col.replace(' (cm)', '') for col in df.columns]
df.columns = [col.replace(' ', '_') for col in df.columns]
pbounds = {
'sepal_length': (df['sepal_length'].min(), df['sepal_length'].max()),
'sepal_width': (df['sepal_width'].min(), df['sepal_width'].max()),
'petal_length': (df['petal_length'].min(), df['petal_length'].max()),
'petal_width': (df['petal_width'].min(), df['petal_width'].max())
}
optimizer = BayesianOptimization(
f=None,
pbounds=pbounds,
random_state=1,
)
optimizer.register(df.iloc[0].values, 0)
print(dict(optimizer.space.array_to_params(optimizer.space.params[0])))
print(dict(df.iloc[0]))Output:
{'petal_length': np.float64(5.1), 'petal_width': np.float64(3.5), 'sepal_length': np.float64(1.4), 'sepal_width': np.float64(0.2)}
{'sepal_length': np.float64(5.1), 'sepal_width': np.float64(3.5), 'petal_length': np.float64(1.4), 'petal_width': np.float64(0.2)}
Expected behavior
The optimizer should not change the order of the keys provided by the user. Registering a numpy array instead of a dict should assume the same order as given by the pbounds.