Add Possibility to Load XGBoost Models as PySpark Models

[ayoub317]

Closes #11400

The default values for device, use_gpu and tree_method were added due to this function call, and eventually this one which assumes they already exist in _paramMap or _defaultParamMap.

[trivialfis]

cc @WeichenXu123 @wbo4958

[wbo4958]

will this function be exposed to users?

[ayoub317]

I think it is good to expose it to users so that if they have already loaded or trained an sklearn model, and maybe want to make predictions on a large dataset they can do it without the need to save the model to disk and then load it back with the load_model method.

[trivialfis]

Thank you for working on this. But what's the difference between the approach taken here and the following snippet?

from xgboost.spark import SparkXGBRegressorModel
from xgboost import XGBRegressor

reg = XGBRegressor()

SparkXGBRegressorModel(reg, None)

[ayoub317]

Hello @trivialfis ,
The code snippet you provided doesn't seem to load an XGBoost model and make predictions. I expect the loading part to work, but an error to occur during the prediction without this commit. It also I think packages the loading as an sklearn model and the conversion to a Spark model in a nice way.

[wbo4958]

LGTM.