Preventing a crash for LSML in a verbose mode

[searchivarius]

If LSML is created with verbose=True, the code crashes here https://github.com/searchivarius/metric-learn/blob/fac5ed5d4f0ae24d274e0956f4b7d2d6d09d2841/metric_learn/lsml.py#L95 due to uninitialized variable s_best. This is I believe a way to fix the problem.

[Callidior]

It's an interesting case if the program gets there with l_best uninitialized, because it means that no single gradient descent step could improve the loss. Thus, maybe l_best = 0 is a more appropriate initialization?

[searchivarius]

@Callidior I am torturing your software using substantially non-metric data :-) Whatever initialization is best in this case is up to you.

[perimosocordiae]

l_best is recording the length of the best step size, so I agree that it makes sense to initialize it to zero.

[searchivarius]

Hi guys, sure, fell free to make this modification.

[perimosocordiae]

Just pushed a fix. Thanks, @searchivarius!