Josh Levy-Kramer
2016-04-14 15:24:08 UTC
Hi,
I've created a GridSearchCV like class, but instead of blindly calculating
the score at all parameter combinations, OptunitySearchCV uses cleaver
optimisation algorithms to find the best parameters. As you could guess it
leverages a library called Optunity:
http://optunity.readthedocs.org/en/latest/. The optimisation algorithms
used are specifically designed for time-expensive non-smooth functions.
I've crated a pull request which demonstrates the proof-of-concept of the
idea: https://github.com/scikit-learn/scikit-learn/pull/6662.
I think this should be included in the Scikit learn package as it can
reduce the number of times you need to run a classifier to find the 'best'
parameters. This saves a hell of a lot of time and should be part of the
standard toolkit for any machine learnist - in my opinion.
I've started this thread to discuss the topic and see if its worth
proceeding with developing a OptunitySearchCV.
Many thanks,
Josh
I've created a GridSearchCV like class, but instead of blindly calculating
the score at all parameter combinations, OptunitySearchCV uses cleaver
optimisation algorithms to find the best parameters. As you could guess it
leverages a library called Optunity:
http://optunity.readthedocs.org/en/latest/. The optimisation algorithms
used are specifically designed for time-expensive non-smooth functions.
I've crated a pull request which demonstrates the proof-of-concept of the
idea: https://github.com/scikit-learn/scikit-learn/pull/6662.
I think this should be included in the Scikit learn package as it can
reduce the number of times you need to run a classifier to find the 'best'
parameters. This saves a hell of a lot of time and should be part of the
standard toolkit for any machine learnist - in my opinion.
I've started this thread to discuss the topic and see if its worth
proceeding with developing a OptunitySearchCV.
Many thanks,
Josh