Metrics

[justusschock]

Scikit-Learn metrics often don't support enough dimension (i.e. they often don't support 3 dimensional data (NxHxW) or even 4 dimensional data (NxCxHxW) and thus we either need to provide custom implementations or any kind of mapping here.

[justusschock]

I'm not sure anymore if we really should do this. IMO we could assume that the user is able to rewrite the metrics him/herself, which is, why I'd also remove the metrics file and move the focal loss implementations to vision. What do you think on that @mibaumgartner ?