To Do

Set up a single command for multiple imputation: a) pick m random subsets of data, b) choose model of rank k, regularization constant α for each subset, c) impute missing data from each of the m selected models.
- (nandana will do this)
Documentation!
- how to think about mpca
- imputation
- error metrics for cross validation
- new syntax for fitting data frame
  - how to specify loss function(s)
- parallel fitting
- full rank model / prisma
Poisson loss
- scaling?
- to log or not to log? that is the interpretative issue

Bugs

init_nndsvd! doesn't work (probably an upgrade-to-1.0 bug)
M_estimator doesn't work (losses.jl); bug in Optim?
sample doesn't work
lots of bugs in fit_dataframe_w_type_imputation; deprecated for now. (also it's an odd thing to do.)
imputation doesn't return correct type (for dataframes)