Overflow handling

[jlsuarezdiaz]

Some algorithms (NCA, NCMML, NCMC, LMNN) can overflow with some datasets if learning rates are too big. They may return matrices with NAs which cause errors in subsequent algorithms, like classifiers or a tune function. During the tuning process the function should be able to not abort if there is a failure in any of the cases, and keep evaluating the other cases specified.