max_iter limit argument inconsistencies between solvers.

This question likely stems from my poor understanding of sklearn and apologies if this has been answered but I was wondering why for some estimators, such as SVC(), the max_iter argument can be set to -1 for removing the limit but for other estimators such as SGDClassifier() there needs to be a hard limit, otherwise there is the error ValueError: max_iter must be > zero. Got -1.000000.

[rth]

Now that tol>0 by default, we could add max_iter=-1

Are we confident SGD will always converge within the given tolerance? If not this will produce an infinite loop.

+1 to support maxi_ter=-1 but I’m not so sure about making it a default.

[boris-shabash]

Is this still an open issue? I’d like to chime in with the fact I actually think estimators should have a pre-specified limit on max_iter for models. One example for why is the sklearn.svm.SVC object which when instantiated with max_iter=-1 can get stuck if it cannot converge. This makes it difficult to employ in Jupyter notebooks where sending an interrupt to the kernel isn’t enough to regain control.