dask scikit-learn API: TypeError: fit() got an unexpected keyword argument 'callbacks'

[pseudotensor]

Is there planned support for dask to have callbacks? When using things like Optuna I hit:

  File "/home/jon/h2oai.fullcondatest/h2oaicore/models.py", line 1512, in objective
    model.fit(X, y, sample_weight=sample_weight, callbacks=[pruning_callback], **kwargs_fit)
  File "/home/jon/minicondadai/lib/python3.6/site-packages/xgboost/core.py", line 421, in inner_f
    return f(**kwargs)
TypeError: fit() got an unexpected keyword argument 'callbacks'

@trivialfis

[hcho3]

It is already supported:

xgboost/tests/python/test_with_dask.py

Lines 670 to 671 in fb56da5

	xgb.dask.train(client, {}, dtrain, num_boost_round=4, callbacks=[TestCallback()])[
	'booster']

[pseudotensor]

Is the scikit-learn API not supporting it? I'm using latest 1.3.0 master.

[hcho3]

Not yet, currently, you'd need to use xgb.dask.train() to use the callbacks argument.

[hcho3]

Also, Optuna's pruning callback function doesn't yet work with xgboost.dask: #6323. We recently overhauled the callback interface, specifically to support Dask interface, and the Optuna's pruning callback should be updated to use the new callback interface (it's using the old interface).

[pseudotensor]

Ok, so even if I tried to use native API for Optuna, it would fail some other way.

[hcho3]

That's right, until #6323 is resolved.

[pseudotensor]

Is the Optuna calback code complicated? I'll have to check it out.

[trivialfis]

I will submit a PR on optuna next week. We haven't done any test on using dask interface with optuna.

[pseudotensor]

Awesome, thanks @trivialfis !

I suppose no changes are required for the dask-optuna package https://github.com/jrbourbeau/dask-optuna

[trivialfis]

Closed by #6471