Feature Request: Insights embedded in TensorBoard

TensorBoard is great for logging data during an ML training run. As the model author, you can output scalars, histograms, embeddings, images and others.

However, data is presented in the specific order that it was written (by epoch / step values) and doesn’t surface any specific metrics that might be out of range or anomalous. We’d like to propose a few features that might help this:

1. Users can specify valid ranges for data logged
   1. Data falling outside these validation values would be flagged or presented first
   2. This could potentially complement hparams for hyperparameter ranges
   3. Specification is TBD
2. Users can specify epoch / step based milestones
   1. Milestones would be shown on all time series plots (scalars, etc)
3. Others?

We might bring this up at the TensorBoard SIG as well. Comments and thoughts welcome!

cc @nfelt @manivaradarajan @GalOshri @natalialunova @lanpa @sanekmelnikov @caraya10 @jspisak

Start by reviewing TensorBoard’s existing scalar, histogram, embedding, image, and hparams views to understand how logged data and time-series plots are presented. The TensorBoard SIG discussion should clarify validation-range and milestone specifications; done means an agreed scope and implemented insights that surface out-of-range data or show epoch/step milestones.