AutoSklearnClassifier seems to exit application on finishing fit

[jrs53]

AutoML(1)_fe4182f8674c69a5018ac777ba60abec.log
autosklearn.log

Describe the bug

When I run automl.fit, it seems to take the expected amount of time (specified by time_left_for_this_task), but my application immediately quits with no apparent error messages.

To Reproduce

automl = AutoSklearnClassifier(time_left_for_this_task = 2*60)
logging.info("Fitting classifier")
automl.fit(X_train, y_train)
logging.info("Finished fitting classifier")

This runs for ~2 mins, I see "Fitting Classifier" in my output, but never "Finished fitting classifier".

This only seems to happen with my data; e.g. load_digits from the Start page is fine.

Expected behavior

I expect the classifier to fit and execution to continue

Actual behavior, stacktrace or logfile

No stacktrace, logs attached

Environment and installation:

Please give details about your installation:

Running inside Docker container (mfeurer/auto-sklearn both master and development) on Amazon Linux

[jrs53]

Ahh, might automl.fit do something that means subsequent logging statements don't actually do what I expect?
If I use plain print instead of logging.info then I see output and it is clear execution is continuing

[mfeurer]

Ahh, might automl.fit do something that means subsequent logging statements don't actually do what I expect?

Yes, indeed it does. We should make sure to not mess with the default logger. @franchuterivera would it be possible to add this requirement to #1017

[franchuterivera]

Hello! I think this issue was solved with #1077, but let me clarify the expectation here:

• The API has support to provide a logging configuration similar to the default one used by autosklearn .

Notice that autosklearn by default makes the root logs to console be warning level. If you do logger.warning('Finisher') you will be able to see the desired msg. If you do want to see info level, kindly pass a logging yaml file to autosklearn.

Thanks a lot for originally reporting the issue! I am going to close it now, as it seems to be fixed.