Minor issues in RidgeRegression_SmoothedL1TV

[duboism]

Hello,

Fouad and I have found the following issues in RidgeRegression_SmoothedL1TV:

• the get_params method try to access self.A (in estimators.py line 892) which doesn't exist; it should instead return self.Atv and self.Al1
• the __init__ function calls the parent __init__ (i.e. RegressionEstimator.__init__) without passing start_vector (therefore the default value in RegressionEstimator.__init__ (which is a random vector) will always be used); if it makes sense to use a different start_vector it should be added in __init__ and passed to the parent __init__

[duboism]

By the way is the goal of get_params is to be able to reconstruct an identical object (by passing the dictionary to the constructor)? In this case, it should return the same parameters than in __init__.

[tomlof]

Great!

However, the start_vector is actually never used in ExcessiveGapMethod, though is could be used in other algorithms in the future...

Note that in Excessive gap the start vector is given by the algorithm. It is possible to override this and force a given start vector (though, the algorithm is not guaranteed to converge in this case), but in that case the start vector is given explicitly to the fit method (but we do not allow that currently ;-)).

Yes, precisely, the idea of the get_params was to obtain all fields of the object so that it can be reconstructed using set_params. This is not implemented, nor tested, properly yet, though... Essentially, anything passed to init should be returned by get_params, but sometimes other parameters (computed ones) may be returned as well.