[Feature Request] Generalize pr_curves to visualize any 2D scalar data
alexlee-gk opened this issue · comments
The pr_curves plugin is great to visualize any 2D scalar data, but currently it's geared only towards precisio-recall curves. For instance, the summaries have PR-specific fields (e.g. Threshold, Precision, Recall, TP, FP, TN, FN) and the plots have hard-coded default ranges of [-0.05, 1.05]. This plugin can be generalized for any 2D data, by taking X and Y values, rather than Recall and Precision values. I have achieved this by filling in dummy values for some fields (Threshold, TP, FP, TN, FN) and commenting out _defaultXRange and _defaulYRange in tf-pr-curve-card.html. This is useful for plotting metrics of time-series data, i.e. per-time-step metrics vs prediction time steps.
Your feature request makes sense to me. A couple teams internal to Google have already used the PR curves dashboard to plot other curves, which further confirms your idea.
The one caveat is that we need the engineering hours to make this more generic dashboard that supports all sorts of 2D plots. Again, I like the idea.
Bump.
Has any work gone in this direction?