EXA speed up ensemble/tests/test_gradient_boosting

[glennfrutiz]

…g.py::test_gradient_boosting_early_stopping.

Reference Issues/PRs

Meta-issue: accelerate the slowest running tests #21407

What does this implement/fix? Explain your changes.

These changes halve the magnitude of n_estimators used in various functions called within the test.
As a result, the run time of the test has sped up to under 1 second on my machine.
Before changes it took 2.05 seconds and decreased to 0.6 seconds.

Any other comments?

[ogrisel]

Thanks for the PR, however the proposed changes break the intended meaning of the test that is to check that the actual number of trees learned without the early stopping option is significantly larger than the actual number of trees learned with the early stopping option enabled for the same learning rate value.

Please consider the proposed change below that will however increase back the test duration a bit.

It might be possible to further speed-up this test by decreasing the training set size, for instance with:

X, y = make_classification(n_samples=300, random_state=0)

however this might require adjusting the expected number of trees when early stopping is enabled to preserve the intent of the test.

[glennfrutiz]

@ogrisel Thank you so much for the detailed suggestions you have left. They have definitely clarified my understanding of this test. I have applied the changes you suggested. I attempted to decrease the training set size and adjust the number of trees when early stopping but I did not have any luck. With your suggestions the run time for this test on my machine is now 0.86 seconds.

[ogrisel]

I attempted to decrease the training set size and adjust the number of trees when early stopping but I did not have any luck.

Thanks for trying! A negative experiment result can be a positive contribution in itself :)

LGTM!

[lorentzenchr]

LGTM

[lorentzenchr]

A negative experiment result can be a positive contribution in itself :)

@ogrisel I couldn't agree more!