[QST] How to perform saving and loading of multi-GPU cuml dask RandomForestClassifier models?

[Santyk1993]

There aren't clear documentation examples for saving and loading the 'cuml.dask.ensemble.RandomForestClassifier' trained models.

I have trained the model using 4 GPUs.
After training, I have used the 'get_combined_model()' function on the distributed model to get a single GPU model for pickling, as mentioned in:
https://docs.rapids.ai/api/cuml/stable/pickling_cuml_models.html#Distributed-Model-Pickling

When I try to load this pickled model and use it for prediction, I get an error stating:
" AttributeError: 'NoneType' object has no attribute 'predict' "

I have tried performing the save using pickle and joblib libraries, and I have tried the save file formats: .sav, .pkl, and .model.
All approaches lead to the same error mentioned above.

Can someone advise on this issue?

[dantegd]

Hi @Santyk1993 I believe what you ran into is a bug in the 0.17 version, particularly described by this issue:

#3331

This PR should fix it (once it is merged) for the upcoming versions: #3388

Thanks!

[drobison00]

@Santyk1993 did this resolve the issues you were seeing?

[Santyk1993]

@drobison00 I read in issue #3388 that the issue with pickling the 'cuml.Dask' version of Random Forest Classification is now fixed. That's great to hear.

I had worked around the issue at that time, by simply not saving the model, and using it in the same run.
Currently, I am not running the Dask version of cuml RF, but I shall update in this thread as soon as I try it.

Thank you.

[github-actions]

This issue has been labeled inactive-30d due to no recent activity in the past 30 days. Please close this issue if no further response or action is needed. Otherwise, please respond with a comment indicating any updates or changes to the original issue and/or confirm this issue still needs to be addressed. This issue will be labeled inactive-90d if there is no activity in the next 60 days.