Community-maintained fork: This is a community-maintained fork of Facebook Research’s HiPlot, which has been archived. We aim to keep the project alive with bug fixes, security updates, and new features.
HiPlot is a lightweight interactive visualization tool to help AI researchers discover correlations and patterns in high-dimensional data using parallel plots and other graphical ways to represent information.
Note
HiPlot is now maintained as a community fork at mindthemath/hiplot, after the original facebookresearch/hiplot repository was archived.
HiPlot demonstration¶
Given about 7000 experimental datapoints, we want to understand which parameters influence the metric we want to optimize: valid ppl. How can HiPlot help?
On the parallel plot, each line represents one datapoint. Slicing on the
valid pplaxis reveals that higher values forlrlead to better models.We will focus on higher values for the
lrthen. Un-slice thevalid pplaxis by clicking on the axis, but outside of the current slice. Slice on thelraxis values above1e-2, then click theKeepbutton.Let’s see now how the training goes by adding a line plot. Right click the
epochaxis title and selectSet as X axis. Similarly, setvalid pplas the Y axis. Once you have done both, an XY line plot should appear below the parallel plot.Slicing through the
dropout,embedding_sizeandlraxis reveals how they can affect the training dynamics: convergence speed and maximum performance.