Use HiPlot integration to analyse multiple experiments¶
Parallel coordinates plot is a powerful tool that allows AI researchers analyse correlations and patterns between experiments’ metrics, parameters and properties.
Parallel plots are especially useful when inspecting hyper-parameter optimization jobs that usually consists of hundreds of experiments. Neptune allows you to very easily generate such plot in a Notebook or Python script.
Visualization is build with HiPlot, lightweight interactive visualization tool published by the Facebook AI group.
This feature makes use of the HiPlot library and is implemented as a part of the neptune-contrib. Make sure that you have all dependencies installed:
Use this command to install all dependencies:
pip install neptune-client neptune-contrib[viz] hiplot
Generate parallel coordinates plot¶
Make sure you have your project set:
import neptune from neptunecontrib.viz.parallel_coordinates_plot import make_parallel_coordinates_plot neptune.init('USERNAME/example-project') make_parallel_coordinates_plot(html_file_path='my_visual.html', metrics= ['epoch_accuracy', 'epoch_loss', 'eval_accuracy', 'eval_loss'], params = ['activation', 'batch_size', 'dense_units', 'dropout', 'learning_rate', 'optimizer'], tag='optuna')
Customize visualization to your need¶
Set axes order,
Drop unused axes,
Apply coloring to axis,
Sort by clicking on axis,
Select range in axis & slide.
Inspect experiments lineage¶
Right-click on the axis name,
Use options ‘Set as X axis’ and ‘Set as Y axis’ (in the menu XY group at the bottom),
When both are selected, you will see lineage plot below parallel coordinates plot.
Check example notebooks in Neptune¶
These notebooks are tracked in Neptune public projects. Feel free to play with the plots - they are interactive.