In a blog post, Facebook AI announced its newly open sourced tool HiPlot for data visualization tasks. It is designed to be lightweight and should help machine learning researchers discover patterns and correlations in their high-dimensional data.
As the blog states, HiPlot is being used at Facebook AI “to explore and efficiently analyze hyperparameter tuning of deep neural networks with dozens of hyperparameters.”
HiPlot can be used either with Jupyter Notebook for Python data or as a webserver when data is stored in a CSV file. It is available on PyPI and can be installed via
pip install hiplot, or as npm package.
Interactivity, extendability and more
HiPlot is designed to offer some advantages compared to similiar visualization tools, one of which is interactivity. This is made possible by displaying parallel plots, which allow researchers to customize the visualization as needed.
For example, parallel plots let you reorder or remove axes or extract a certain data selection. See the interactive demo to try this for yourself.
While HiPlot’s web server can parse CSV or JSON files, a custom Python parser can be used as well. In addition, HiPlot is compatible with the logs of other Facebook AI libraries including [email protected], Nevergrad and fairseq.
Furthermore, HiPlot provides visualization for population-based training, a genetic algorithm.
On GitHub, there is an example for using HiPlot with Jupyter Notebook. The following input creates the HiPlot visualization as seen below: