Cuml: [FEA] dask train test split

[gumdropsteve]

Is your feature request related to a problem? Please describe.
I wish I could use cuML to train_test_split a dask_cudf.DataFrame.

Describe the solution you'd like

from cuml.dask.preprocessing.model_selection import train_test_split

https://ml.dask.org/modules/generated/dask_ml.model_selection.train_test_split.html

Describe alternatives you've considered
Running normal train_test_split then making dask_cudf.DataFrames from the results.

[JohnZed]

Thank you for submitting this! train_test_split for Dask is certainly a logical feature for cuML to add and is on our priority list for 0.15. Adding this issue to that release target.

[cjnolet]

In the meantime, if your data is already shuffled randomly across the cluster, Dask’s random_split function does work with Dask_cudf.

Here’s an example of where this is used on TPCxBB: https://github.com/rapidsai/tpcx-bb/blob/master/tpcx_bb/queries/q28/tpcx_bb_query_28.py#L339

[pseudotensor]

Yes, I would like this feature as well, simplifies use of scikit-learn like cuml functions. Currently I hit:

2020-10-21 19:45:33,533 C: 25% D:367.0GB M:47.6GB  NODE:LOCAL1      4613   DATA   |   File "/home/jon/miniconda/lib/python3.6/site-packages/cuml/preprocessing/model_selection.py", line 130, in train_test_split
2020-10-21 19:45:33,534 C: 25% D:367.0GB M:47.6GB  NODE:LOCAL1      4613   DATA   |     if X.shape[0] != y.shape[0]:
2020-10-21 19:45:33,535 C: 25% D:367.0GB M:47.6GB  NODE:LOCAL1      4613   DATA   |   File "/home/jon/miniconda/lib/python3.6/site-packages/dask/delayed.py", line 563, in __bool__
2020-10-21 19:45:33,535 C: 25% D:367.0GB M:47.6GB  NODE:LOCAL1      4613   DATA   |     raise TypeError("Truth of Delayed objects is not supported")
2020-10-21 19:45:33,535 C: 25% D:367.0GB M:47.6GB  NODE:LOCAL1      4613   DATA   | TypeError: Truth of Delayed objects is not supported
2020-10-21 19:45:33,536 C: 25% D:367.0GB M:47.6GB  NODE:LOCAL1      4613   DATA   | ].