ProgressMeter

[cortner]

Reviving #84

printing the iteration history is useful when you are developing and experimenting, but as user you probably don't want to see it and still get some progress information?

Here are two variants I cooked up, which I both like, and both can probably be fine-tuned quite nicely.

using ProgressMeter
prog = ProgressThresh(1e-8, "Minimizing:")
df(x) = 4*x^3 - 2*x + 0.8
x = 1.0
h = 0.1
while df(x) > 1e-8
  x -= h * df(x)
  ProgressMeter.update!(prog, abs(df(x)))
  sleep(0.3)
end

using ProgressMeter
TOL = 1e-8
GRAN = 100
prog = Progress(floor(Int, GRAN * abs(log(TOL))), desc = "Minimizing:")
df(x) = 4*x^3 - 2*x + 0.8
x = 1.0
h = 0.1
while df(x) > TOL
  x -= h * df(x)
  ProgressMeter.update!(prog, ceil(Int, GRAN * abs(log(abs(df(x))))))
  sleep(0.3)
end

[antoine-levitt]

Very nice! I like the first solution a lot more, because it does not pretend that the norm of the gradient is an accurate representation of progress (which it is usually not, unless you do gradient descent or a similarly "nice" method with a tight tolerance). Maybe displaying the value of the function to be optimized would be nice also? show_trace could become a flag: True prints to STDOUT, :progress does this, and False does nothing.

[cortner]

yeah, the second one was more tongue in cheek.

I think the first version with a little more information will be great

[pkofod]

Looks great!

[aakhmetz]

looks indeed great!

[pkofod]

I think this will probably be a callback thing if anything and thus up to the user.