GPU Support

[eladmw]

Considering the number of aggregations from each model in each layer, it would be nice to train the models faster

[xuyxu]

Hi @eladmw, thanks for your suggestions.

Do you have any good material on the GPU forest? It would be impossible to implement the GPU version from scratch with our current workforce.

[eladmw]

Looks like there's many references to XGBRegressor/XGBClassifier, which has compatibility with GPUs. I skimmed so correct me if I'm wrong, but I think that you can do XGBRegressor(tree_method='gpu_hist', gpu_id=0), or feed the params into the training interface with a dict setting :
param['gpu_id'] = 0
param['tree_method'] = 'gpu_hist'

With regards to Lightgbm, it is sort of a pain to enable GPUs.

[xuyxu]

Thanks @eladmw!

I think the first step we need to take is to support customized base estimators in the layer (see issue #29). After then, we can make it compatiable with the GPU version of XGB model. Let's keep this issue open for a while ;-)

[xuyxu]

Closed via #14