Okon Samuel

> A related question is whether instead of returning a _single_ value, in the case of `reports_each_observation=false` (eg, `auc`) we instead report a _constant_ vector. This would eliminate some bothersome...

What the status with integration with EvalMetrics.jl??

Now that you put it that way annotating with types won't be helpful as there as too many possibilities. For example a measures vector could consist of `confusion_matrix`, `true_positive` and...

> By the way, I seem to recall reading somewhere on Discourse that you need to use `@code_warntype optimize=true` in order to get the actual inferred return types By default...

@tlienart. What do you think? Since we are still going to call `matrix(X)` in the inner methods. Why don't we don't in some cases call `matrix(X)` earlier so we can...

In addition i noticed that internally all tables are converted to dense matrices. Are there not cases where we would like to have a sparse matrix.(Maybe if something like a...

> what code are you referring to here? Code at MLJModels > you can have a Table wrapping around a sparse matrix and recover said sparse matrix that's nice if...

> An implementation can already cache the matrix version of `data`, by returning this in `cache`. But now, assuming `model` has not changed (same hyper-parameters) `fit!` calls `update` which can...

I guess we can have two version of `fit!` function. - An internal one (maybe `fit_!`) used only by MLJ (which might have the implementation you suggested above) and -...

> Are you referring to the fact that a user might mutate data bound to a machine in between a call to `fit` (dispatched by their first call to `fit!`)...