[dask] warn if attempting to use tree_learner other than data parallel #3848
Conversation
| @@ -244,6 +244,7 @@ def _train(client, data, label, params, model_factory, sample_weight=None, group | |||
| for tree_learner_param in _ConfigAliases.get('tree_learner'): | |||
| tree_learner = params.get(tree_learner_param) | |||
| if tree_learner is not None: | |||
| params['tree_learner'] = tree_learner | |||
There was a problem hiding this comment.
I know this is a little awkward in the case where aliases are used, because multiple aliases for tree_learner might exist in params.
That change will only live on master for a short time, because I'm cleaning it up in #3813 (comment) right now
There was a problem hiding this comment.
LGTM!
I know you recommended an error in #3834 (comment), but in my opinion it would be better to make this only a warning so that expert users (who know how to do the Dask partitioning for the different methods) are still able to experiment with the other methods.
OK, but please be prepared to that it is very common to use
import warnings
warnings.filterwarnings("ignore")
in Python scripts with LightGBM 😃
|
This pull request has been automatically locked since there has not been any recent activity since it was closed. To start a new related discussion, open a new issue at https://github.com/microsoft/LightGBM/issues including a reference to this. |
Based on #3834 (comment).
Today, LightGBM's distributed learning supports three methods (feature parallel, data parallel, and voting parallel) --> https://lightgbm.readthedocs.io/en/latest/Features.html#optimization-in-parallel-learning
The Dask interface in
lightgbm.daskis currently only tested with data parallel, where each chunk of the dataset is a row-wise chunk (all features, subset of rows).The other methods MIGHT work with
lightgbm.dask, but significant testing and documentation is needed. Supporting those methods might include adding extra validation to the Dask package as well.Changes in this PR
The other parallel learning methods are optimizations, and I don't think the work to support them (#3834) needs to make it into the next LightGBM release (3.2.0). So that testing and documentation won't happen in the next few weeks.
This PR proposes adding a warning if you use one of these methods, to tell users that support for methods other than "data" is experimental and might break.
@StrikerRUS I know you recommended an error in #3834 (comment), but in my opinion it would be better to make this only a warning so that expert users (who know how to do the Dask partitioning for the different methods) are still able to experiment with the other methods.