how to get the time of one boost round?

[showkeyjar]

I use gridsearch and optuna to search the best hyper parameters of xgboost.

but I found when the tree is deep and big,

then the boost round will very slow, and cost too much time

and the big tree parameters usually not use for productive system.

so the big tree search usually waste time.

I want to find a way that can measure the boost round time,

if the time is too long, then stop train, and this strategy can also avoid OOM.

please help how can I do this? thanks!

[trivialfis]

Hi, you can define a callback function. Please see Python doc and demo. https://xgboost.readthedocs.io/en/stable/python/callbacks.html

[showkeyjar]

I read the doc and demo,

I can use callback measure time, but

there is no information for how to stop train in callback functions

[trivialfis]

From the demo, there's a after_iteration function defined in the callback class, which returns a boolean value. According to the document of its parent class TrainingCallback https://github.com/dmlc/xgboost/blob/b0015fda9658d44c1e7ac3eb0361ba8d1f44bcdc/python-package/xgboost/callback.py#L44 , it should return True when the training should stop. Also, in the demo https://github.com/dmlc/xgboost/blob/b0015fda9658d44c1e7ac3eb0361ba8d1f44bcdc/demo/guide-python/callbacks.py#L50 it should return False when the training should continue.

It might not be intuitive enough, could you please open a PR for suggested changes to the document so that it can be more friendly to new users?