See comment: https://github.com/metric-learn/metric-learn/pull/159#discussion_r250234204 We should: - Allow tuples, lists, numpy arrays, (any array like, TODO: see if there is other array-like than these 3 types) as arguments for `bounds` for ITML - I think a unique type of return for `bounds_` would be better, say a numpy array - Do tests to ensure that this works as expected - Do the right advertising in the dosctring