par-coordinates 0.3

Create parallel coordinates plots of the hyperparameter search

Introduction

Parallel coordinates plots are a convenient way of visualizing the hyperparameter search for machine learning and deep learning models. While many platforms for model development (see wandb) already include this kind of visualization, reproducing it in matplotlib for more advanced customization or publishing can be tricky. This package solves the issue by offering compatibility with the major frameworks for hyperparameter tuning. The current version supports keras_tuner, sklearn and optuna.

Installation

The package is available at par-coordinates and can be installed via:

pip install par-coordinates

Tutorial

The jupyter notebook analyses/tutorial.ipynb contains detailed examples of the package being used:

• random search in keras_tuner for a multilayer perceptron
• grid search and random search in sklearn for a random forest classifier
• Bayesian optimization in optuna for a XGBoost regressor

The workflow is always the same:

# import package
from par_coordinates import get_results
from par_coordinates import plot_par_coordinates
# get results dataframe for specific tuner
results = get_results.sklearn(random_search, "f1")
# display parallel coordinates plot
fig = plot_par_coordinates(results, labels=["maximum tree depth", "minimum samples in leaf node", 
                           "number of trees", "F1 score"], figsize=(8, 4), curves=True, 
                           linewidth=0.8, alpha=0.8, cmap=plt.get_cmap("copper"))

resulting in the following visualization (with room for customization):

For additional information the official documentation is available here.

License

The package is freely available under MIT license. The code in src/par_coordinates/plot.py is based on pcp.