Control caching of composite models

Hi,

Problem description

I have just had the late realization that setting cache=false for a composite model will not transfer to all sub machines built at fit time.

using MLJBase
using MLJLinearModels

n = 100
X = MLJBase.table(rand(n, 3))
y = rand(n)
stack = Stack(metalearner=LinearRegressor(),
    model1 = LinearRegressor())

mach = machine(stack, X, y, cache=false)
fit!(mach, verbosity=0)

# Top level machine
@assert !isdefined(mach, :data) # not defined : ok
mach.cache # contains data: ?

# Any submachine
submachines = report(mach).machines
for submach in submachines
    @assert isdefined(submach, :data)
end

I am currently working with very big Stacks and I think this is the main reason for which I run out of memory.

Ideas:

• Of course I guess there is the possibility of adding a hyperparameter to the composite model that can be transferred after to the machine definitions. This would probably solve my personnal issue (and would be a short term solution) however I don't think this is ideal in the long term because any user defined composite model will have to add this hyperparameter.

• More generally I think this issue arises from the current impossibility(?) of communication between the machine and its submachines in the current design. If I am not mistaken we will have the same problem with computational resources as briefly raised here.

A vague idea would be to define the learning graph in a method like machine(m::MyComposite, args...) instead of the current fit.

What are your thoughts?

Start by tracing machine(...), fit!, and report(mach).machines in the composite-model path, using the supplied Stack example to reproduce cache=false behavior. Determine how caching should propagate to submachines, including user-defined composite models; done requires an agreed design and verification that submachines do not retain unintended training data.