Early stopping features for calibrate and optimize?

[djinnome]

For calibrate, a nice intermediate result is the current loss.

For optimize, would a duality gap or optimal value be a useful intermediate value? @anirban-chaudhuri what would make sense as input to the progress hook like what we do for calibrate.

Right now, Terarium kills the process if we want to stop early, but if there is a more hygenic way to do this?

[SamWitty]

@djinnome , could you please break this out into smaller atomic issues for each feature request. This could be decomposed into intermediate results for calibrate, intermediate results for optimize, early stopping features for calibrate, and early stopping features for optimize.

Also, could you clarify how this request differs from #477 and #362 .

[djinnome]

Hi @SamWitty

It sounds like progress hooks for calibrate and optimize are already handled by #362 and #477, respectively, so this post can reduce to answering the question about whether we want to control early stopping ourselves, or let Terarium handle it for us.