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 ppl axis reveals that higher values for
lr lead to better models.
We will focus on higher values for the
lr then. Un-slice the
valid ppl axis by clicking on the axis, but outside of the current slice. Slice on the
lr axis values above
1e-2, then click the
Let’s see now how the training goes by adding a line plot. Right click the
epoch axis title and select
Set as X axis. Similarly, set
valid ppl as the Y axis. Once you have done both, an XY line plot should appear below the parallel plot.
Slicing through the
lr axis reveals how they can affect the training dynamics: convergence speed and maximum performance.