say how many steps "Solve" took

[jeisner]

We make solve run faster for shorter steps, so the user doesn't get a good idea of what slow convergence might look like.

So I wonder if we should provide a message at the end of Solve, saying how many steps it took (i.e., the number of evals of the LL function and of its gradient; the line searches don't need to evaluate the gradient). Maybe some kind of tooltip that fades away after a few seconds unless the user hovers over it. Or a pair of counters under the Solve button as Solve runs.

We might also want a checkbox to turn off the adaptive (Armijo) stepsize, so that the user can see first what goes wrong without it. But also the cost we pay for it in extra function evaluations per gradient step.

[fmof]

The first part about # of iterations is addressed; Armijo has been pulled out to #105

[jeisner]

Thanks. Comments:

• The word "iterations" has not been introduced anywhere. Our word is "step." So I suggest "18 steps taken" instead of "18 iterations done."
• Shouldn't pressing "Step" also increment the count in this message, rather than making the message vanish? After all, it is still taking steps on the same problem. I understand why you did it differently (what you really mean is "Solver took 18 steps"); but I think it may be nice for the user to see how many steps it takes by clicking manually with a fixed stepsize, too.
  • I think the message should vanish only if the user
    • manually changes the weights (including by zeroing them)
    • or changes the objective function (including by changing lesson # or regularization or C or N or new counts or new random challenge).
  • And I think it should appear (and reset the count to 0) if the user presses Step or Solve while it is currently invisible.
  • Note that if it maxes out at 200 iterations and the user presses solve again, the count should just continue at 201.
• I would put the message under the "Step/Solve buttons so that they don't change vertical position when it appears/disappears.
  • I wondered about whether the number of steps so far should instead be shown as part of the label on the step button itself. But I couldn't think of a good design.
• In the original issue, what I suggested was counting the number of function evals and gradient evals. (When we use Armijo, the number of steps is only the number of gradient evals but there may also be several function evals in there for the line search. Is this too complicated to show?)