[scikit-learn] MLPClassifier/Regressor and Kernel Processes when Multiprocessing

[Gael Varoquaux]

Hi,

I cannot look too much in details. However, I would advice you to try
using loky or joblib instead of multiprocessing, as a lot of work has
been put in them to protect against problems that can arise in
multi-process parallel computing (for instance the underlying numerical
libraries may not be fork safe, or they may have parallel computing
abilities themselves).

Hope this helps,

Gaël

On Tue, Apr 28, 2020 at 02:06:00PM -0500, Taylor J Keding wrote:
> Hi SciKit-Learn folks,

> I am building a stacked generalization classifier using the multilayer
> perceptron classifier as one of it's submodels. All data have been preprocessed
> appropriately and I am tuning each submodel's hyperparameters with a customized
> randomized search protocol (very similar to sklearn's RandomizedSearchCV).
> Importantly, I am using Python's Multiprocessing.Pool() to parallelize this
> search.

> When I start the hyperparameter search, jobs/threads do indeed spawn
> appropriately. Tuning other submodels (RandomForestClassifier, SVC,
> GradientBoostingClassifier, SDGClassifier) works perfectly, which each job
> (model with particular randomized parameters) being scored with cross_val_score
> and returning when the Pool of workers is complete. All is well until I reach
> the MLPClassifier model. Jobs spawn as with the other models, however, System
> CPU (Linux Kernel) processes surge and overwhelm my server. Approximately 20%
> of the CPUs are running User processes, while the other 80% of CPUS are running
> System/Kernel processes, causing immense slow-down. Again, this only happens
> with the MLPClassifier - all other models run appropriately with ~98% User
> processes and ~2% System/Kernel processes.

> Is there something unique in the MLPClassifier/Regressor models that causes
> increased System/Kernel processes compared to other models? In an attempt to
> troubleshoot, I used sklearn's RandomizedSearchCV instead of my custom
> implementation and the same problems happen (with n_jobs specified in the same
> way).

> Any help with why the MLP models are behaving this way during multiprocessing
> is much appreciated.
> Best,
> Taylor Keding

> _______________________________________________
> scikit-learn mailing list
> scikit-learn at python.org
> https://mail.python.org/mailman/listinfo/scikit-learn


-- 
    Gael Varoquaux
    Research Director, INRIA		  Visiting professor, McGill 
    http://gael-varoquaux.info            http://twitter.com/GaelVaroquaux