PDHG and new callbacks about cil HOT 2 OPEN

Thanks @epapoutsellis some initial thoughts but I can take another look next week -

• PDHG with verbose=2 does not print primal-dual gap as before

A callback to print the primal-dual gap would be useful for CIL and shouldn't be difficult to write - I can take a look next week.

• Different precision in objective values between ProgressCallback and TextProgressCallback

I think this is a question for @casperdcl. But are you referring to the different values for it/s or the 2dp vs3dp for ProgressCallback and TextProgressCallback?

• There is no timing attribute in PDHG or Algorithm class. This used to return a list with times (per update_objective_interval) without measuring computational time of the objectives.

Again, it should be possible to write a callback for this. What timings are you interested in? Being able to time each iteration, excluding the objective calculation?

• What does it mean 13.57it/s? For some stochastic algs, it would be nice to get the time per one iteration when a stochastic gradient is used and the time per one iteration when full gradient iscomputed.

13.57it/s is the number of iterations per second so the reciprocal of time per iteration.

from cil.

Since we compute primal-dual gap by default, I think is better to print it (if one wants) as we did before.

Also, it is strange to see a continuous representation for iterations, e.g., 13.57 iterations. Seconds per iteration is more reasonable which is equal to Seconds per execution time for the update method.
Why in ProgressCallback we get 11.57it/s and with TextProgressCallback 13.01 it/s?

We can do many things with callbacks but previous default behaviour was much better. Also, printing in jupyter is very strange.

from cil.