MLJ model `BetaMLGMMRegressor` predicting row vectors instead of column vectors about betaml.jl HOT 5 CLOSED

Should have been fixed in master:

using Pkg
pkg"activate --temp"
pkg"add MLJBase, MLJModels, BetaML#master"
using MLJBase, MLJModels
import BetaML: GaussianMixtureRegressor, MultitargetGaussianMixtureRegressor

model = GaussianMixtureRegressor()
X, y = make_regression();
mach = machine(model, X, y) |> fit!
yhat = predict(mach, X);
l2(yhat, y) # ok

y2=hcat(y,y)
model2 = MultitargetGaussianMixtureRegressor()
mach2 = machine(model2, X, y2) |> fit!
yhat2 = predict(mach2, X);
l2(yhat2, y2) # ok

from betaml.jl.

Thanks, going check it...

from betaml.jl.

Hello, BetaMLGMMRegressor (that I have now renamed GaussianMixtureRegressor) is not predicting a row vector, but a matrix with a single column, as it can predict multiple targets, so currently always return matrices.
What do you suggest I change:

• the model registration metadata
• I automatically remove one dimension when y is a single variable
  ?

from betaml.jl.

For better or worse, the MLJ standard for returning multi-target predictions is to return a table. (This may not be actually documented but it is adopted in several places) . So there is a case distinction here. This means we generally have separate model types for regular and multitarget predictors. Which is not too bad, as it is not the main use-case.

I suppose you could have a single model that always predicts a table, but I think this will be totally unexpected for users. So, while I realize this is a pain, I think the best plan is a separate MultitargetGaussianMixtureRegressor for the MLJ interface.

a row vector, but a matrix with a single column

Same thing, in my vernacular 😉

from betaml.jl.

And if y2 is a table instead of a matrix, is that allowed? Elswhere, y2 and yhat2 would both be tables. See eg https://github.com/JuliaAI/NearestNeighborModels.jl/blob/f4362b3967dc1e4874cd9875f428c4d13316b761/src/models.jl#L312

Sorry this isn't documented anywhere.

from betaml.jl.