Saving compiled DPP problem for reuse.

[kmonson]

Is your feature request related to a problem? Please describe.
Based on issue #1332 it would be useful to save already compiled DPP problems to reload later. The idea is to make it possible to perform a long compilation of a problem ahead of time to shorten future load times by a large margin. If possible this should be generic with respect to the solver while still overcoming the majority of the compile time.

Describe the solution you'd like
A dump, dumps, (implies compilation if it hasn't happened already) load, and loads methods on the cvxpy.Problem class.

A compile method on the cvxpy.Problem object to force compilation without solving first.

dump(str | Path, *args from cvxpy.Problem.solve as needed) - dumps compiled problem to str | Path
dumps(str | Path, *args from cvxpy.Problem.solve as needed) - returns str rep of compiled problem.

load and loads - inverse of dump, dumps

compile(*args from cvxpy.Problem.solve as needed) - forces compile of the problem without solving.

Additional context
Dumped problems would retain info on which solver they are for if needed, gp, gpc, and any other settings that would need to be baked into the compiled result.

Not sure what the best approach is for loading a problem for a specific solver and then calling solve with a different solver specified. What is done with DPP problems when solve it called twice with different parameters? Raise an exception?

[akshayka]

This is a very interesting idea, and one that I have been thinking about recently. What you are asking fits within the more general goal of a serialization format for CVXPY problems, which could make them more portable and may have other uses. Any solution will need to be designed very carefully.

It will take us some time before we have a properly spec'd solution for this. In the meantime, can you make do with Python pickle?

[akshayka]

One more comment: Python pickle is of course not as elegant of a solution as you propose, and does not provide the nice semantics & validation that you propose. But perhaps you could make do with a custom wrapper around pickle for now. We can leave this issue open until we have a proper solution.

[kmonson]

I hadn't thought of Pickle. I'd need to push this out to a repo and then eventually it would get packaged up and deployed to our users. The real hassle is making sure that the pickle works for all of our developers, some of whom are also users. I'm not sure if they trade off is worth it for our case. The current non DPP version isn't terrible, I'm just hoping to squeeze more performance out for the users in some cases. I'll run it by my cohorts.