Singeleton pattern for the task API

Currently, in AutoGluon, we will do the following, in which TabularPrediction is used as a singeleton.

from autogluon import TabularPrediction as task
predictor = task.fit(...)

A more natural design choice is to create a new class object and call fit as a method.

from autogluon import TabularPrediction
task = TabularPrediction(SOME_ARGS)
predictor = task.fit()

@Jerryzcn

[sxjscience]

@Innixma I think .fit() will still take the data as the input parameter. The difference is that we can pass some specific requirements like time budget and inference latency when constructing the task, e.g.

from autogluon import TabularPrediction
task = TabularPrediction(time_limit=1000, latency='low')
predictor = task.fit(train_data, label='...')

Then, we will automatically learn a predictor that obeys the budget constraint.

[Innixma]

Closing as resolved, we have implemented the singleton pattern for v0.1.