Featurize Categorical Hyperparameters

[leix28]

• BTB version: 0.2.5-dev
• Python version: 3.6
• Operating System: Mac OS 10.14

Description

Is the categorical hyperparameter handled correctly in BTB?

Right now, categorical hyperparamters are represented as a single float number. For example, if I try to tune

• a string hyperparameter optimizer in ['adam', 'sgd', 'rmsprop']
• a float hyperparameter x in [0, 1]

and I get a few hyperparamter and score pairs

• `(optimizer='adam', x=0.5), score=0.8
• `(optimizer='adam', x=0.6), score=0.9
• `(optimizer='sgd', x=0.5), score=0.3

BTB will convert these pairs to

• (0.85, 0.5), 0.8 # 0.85 is the average score of 'adam'
• (0.85, 0.6), 0.9
• (0.3, 0.5), 0.3 # 0.3 is the average score of 'sgd'

and use GaussianProcessRegressor to fit these 3 pairs.

I'm not sure if this is the correct way to featurize categorical hyperparameters.

How to change the categorical hyperparamter fit_transform and inverse_transform?

A more intuitive way to featurize categorical hyperparamters is one-hot representation. But changing the fit_transform is CatHyperParameter is hard.

[pvk-developer]

This issue is deprecated after the 0.3.0 realease. We are now using OneHotEncoder for CategoricalHyperParameters , reffer to #129 #131