(De)serialization of graphs

[casadibot]

It would be very handy to be able to save MX and SX graphs to a filesystem and load back.
In a related effort, Python pickling should become possible.

Created by jgillis at: 2011-12-22T13:59:40
Last updated at: 2012-08-14T11:01:15

---

Comment 1 by jgillis at 2012-08-14T01:15:19

Pointers: http://stackoverflow.com/questions/9310053/how-to-make-my-swig-extension-module-work-with-pickle

---

Comment 2 by jgillis at 2012-08-14T11:01:15

Another pointer: http://www.picklingtools.com/

[jgillis]

Some serialization formats: yaml, json, bson

[jaeandersson]

Storing an SX graph is trivial. All information is contained in the algorithm. One problem is that if you reconstruct the graph, you might create common subexpressions.

MX is much harder. And you cannot really do it without also doing FX and derived classes. For that you would need some systematic approach.

I'm not sure I think this is the right way to go. So far, we have done a pretty good job making the symbolics fast enough that people aren't really complaining about regenerating all the expression graphs all the time. Codegen also makes this less important.

[jaeandersson]

About serialization of SX, my proposal would be to create a .nl exporter.

[jaeandersson]

Created two issues for serialization of NLPs, #755 and #756.

[jaeandersson]

I start to think that this would really be possible also for MX. It's a major effort, but it would offer a really nice platform-independent way of passing symbolic expressions.

[ghorn]

great!!!

probably need a version scheme also

[jaeandersson]

What do you mean by version scheme?

[ghorn]

If you saved it with a different casadi version and try to load it, and the serialization format changes, an assertion would be nice. Maybe that's too much complexity to worry about

[jaeandersson]

I think the first step has to be to get any working serialization. Before there is a stable serialization, versioning is of little use.