[train] Fold v2.XGBoostTrainer API into the public trainer class as an alternate constructor
#50045
Conversation
Signed-off-by: Justin Yu <[email protected]>
Signed-off-by: Justin Yu <[email protected]>
Signed-off-by: Justin Yu <[email protected]>
justinvyu
requested review from
hongpeng-guo,
matthewdeng,
raulchen and
woshiyyya
as code owners
Signed-off-by: Justin Yu <[email protected]>
| @@ -67,7 +67,7 @@ def train_fn_per_worker(config: dict): | |||
| train_ds = ray.data.from_items([{"x": x, "y": x + 1} for x in range(32)]) | |||
| eval_ds = ray.data.from_items([{"x": x, "y": x + 1} for x in range(16)]) | |||
| trainer = XGBoostTrainer( | |||
| train_fn_per_worker, | |||
| train_loop_per_worker=train_fn_per_worker, | |||
There was a problem hiding this comment.
In the long term this is the main API change, right? That this needs to be a kwarg.
There was a problem hiding this comment.
Actually can we just keep this as a required argument since the V1 will always populate it? Are there any problems if we do?
| # TODO(justinvyu): [Deprecated] Legacy XGBoostTrainer API | ||
| label_column: Optional[str] = None, | ||
| params: Optional[Dict[str, Any]] = None, | ||
| num_boost_round: Optional[int] = None, |
There was a problem hiding this comment.
Are these needed for the V2 API? These are not passed in from the V1 API.
|
|
||
| num_boost_round = num_boost_round or 10 | ||
|
|
||
| _log_deprecation_warning(LEGACY_XGBOOST_TRAINER_DEPRECATION_MESSAGE) |
There was a problem hiding this comment.
Let's hold off on explicitly logging that it's deprecated until we have the GH issue and documentation for V2 published?
| scaling_config=ray.train.ScalingConfig(num_workers=4), | ||
| ) | ||
| result = trainer.fit() | ||
| booster = RayTrainReportCallback.get_model(result.checkpoint) |
There was a problem hiding this comment.
Is this a staticmethod of the RayTrainReportCallback? A side question of mine: does this get_model function best to be defined under a callback. Why not make it a utility under ray.train.xgboost if it is not an instance method of RayTrainReportCallback that will use instance attributes.
Summary
Currently, the new
XGBoostTrainerAPI is only accessible with a separate importray.train.xgboost.v2.XGBoostTrainer.To avoid unnecessary import changes, this PR folds the new API, which accepts new arguments
(train_loop_per_worker, train_loop_config, xgboost_config), into the publicray.train.xgboost.XGBoostTrainerclass.This also makes some changes in the Ray Train v2
XGBoostTrainerclass to improve the migration UX, since it does not support the legacyXGBoostTrainerAPI at all.TODO
XGBoostTrainerandLightGBMTrainerAPI revamps #50042