Performance of Scikit-learn parallelization

This repository contains a script to evaluate the speed up of training Scikit-learn algorithms leveraging parallelization, mainly using the flag n_jobs from joblib.parallel_backend. The script experiment.py times the training of a given set of models with a dummy dataset using different numbers of logical processors and parallelization backends. The measured times are logged as well as plotted for a convenient analysis.

Getting started

To run it in your machine, follow the next steps:

1. Install dependencies using pip install -r requirements.txt.
2. Select your experiment configuration in config.json.
   • List of number of logical processors to be tested, "n_jobs_to_test": [1, 3, 5, 10]. To test each number of logical processors in your machine, input "n_jobs_to_test": -1.
   • Set of models, e.g., "models": ["RandomForestClassifier", "KNeighborsClassifier", "LogisticRegression", "MLPClassifier", "SVC", "GaussianNB", "DecisionTreeClassifier", "SGDClassifier"].
   • Parallelization backends, e.g., "parallelization_backends": ["threading", "loky"].
   • Dummy dataset size, e.g., "n_samples": 100 and "n_features": 100.
   • Number of trials per combination, e.g., "n_trials": 3. Useful to take into account the variance, included in the plots with error bars.
   • Base name for the plot images, e.g., "image_name_base": "TrainingDurations".
3. Execute experiment.py. A log file (e.g. experiment_2023_03_12_19_20_48.log) can be used to track the execution.
4. Find your results as plots the automatically generated directory Results.

Please write me if you have any question understanding the code ([email protected]).