As reported on this stackoverflow question the current implementation of liblinear models is using a special memory layout for the coefficient (currently stored in the raw_coef_ array). Those coefficients are made read-only accessible to the user by means of the python properties coef_ and intercept_. However this solution prevents the user to change some of the coefficient values or to reassign a complete coef_ array on a the fitted estimator.

It would be much nicer to have coef_ / intercept_ as traditional array attributes instead.

Options:

• store parameters fitted by liblinear using the traditional coef_ / intercept_ layout and copy those parameters at prediction time to a temporary array suitable for prediction using liblinear
• implement predict and predict_proba in numpy instead of using liblinear at prediction time hence we could use standard numpy arrays with the usual layout for coef_ and intercept_