How can I train use grow_gpu_hist but predict on cpu?

#1

Hi,I havd trained a model based on XGBRegressor,
set
'updater':'grow_gpu_hist','predictor': 'cpu_predictor'
,and pickle.dump success.
When I migrate the model to other machines on no-GPU,
pickle.load fail.

xgboost.core.XGBoostError: [20:26:52] /Users/travis/build/dmlc/xgboost/src/tree/tree_updater.cc:20: Unknown tree updater grow_gpu_hist

I has review https://github.com/dmlc/xgboost/issues/5291,but the function .get_booster() except error

need to call fit or load_model beforehand

Can I migrate to the CPU after GPU trained?

#2

To run training on GPU, please use tree_method="gpu_hist".

When loading a pickle on a CPU only machine, XGBoost should automatically goes back to hist tree method, with a warning saying no GPU is found.

#3

But when I tried to deploy the model on a server without GPU, XGBRegressor did not automatically switch to the hist tree method as expected, but threw the above exception.

#4

Hi, please use tree_method instead of updater.