Better visualization with respect to hparams

[atishagarwala]

When doing hyperparameter tuning, i'll often have tensorboards that look like this:

which technically has all the information from my tuning run. However, it's hard to parse and use. Ideally, I would be able to:

• Label curves by hparam settings (where I can define which hparams to display).
• Let different hparam choices set the line formatting.

As an example for the latter point: suppose I had a hyperparameter sweep which included different learning rates and different batch sizes. I would love to be able to understand the effect of both by, say:

• Coloring curves with the same learning rates the same color.
• Choosing a different linestyle (solid, dotted, dashed, etc.) for each batch size.
  There may be different ways to provide multiple dimensions of curve sorting, IANAE. However I believe having more visualization options can help make tensorboards more useful.

[psybuzz]

Thanks for the report! These are great use cases for us to collect. Isolating the subset of curves that used a specific hparam value (or range) seems definitely valuable.

[ssgosh]

@atishagarwala As a workaround, I perform independent runs of my training scripts using a meta-script. I name training directories with the relevant hparams. Then, I open tensorboard with logdir set to the top-level directory. It is then easy to filter the runs with hparams. However, advanced filtering cannot be done (e.g. some param lambda > 0.1).