This went to spam for some reason. I'm not sure I understand the problem well, specifically the second point about trained_model_serialization. You mean when you change the parameters without using the parameter framework, it is sometimes not considered when serializing because it's not registered? So the problem boils down to detecting unregistered parameters, is that right?
On Wed, Jul 10, 2019 at 12:33 PM Heiko Strathmann < heiko.strathm...@gmail.com> wrote: > While writing tests to check certain functionality for all machines, I > constantly come across random segfaults (bugs, missing inits), crashes, > missing registered parameters (so clone/equals doesnt work), etc. All those > are use-case motivated tests so quite bad. > > We have some tests for all objects that check clone/equals consistency, > but the problem is that this is only for the default ctor construction of > objects. If for example, parameters are not registered (and so clone doesnt > work), this will not be caught be the test. > > Another example is trained_model_serialization: if I set a parameter of > some of the machines to the non-default value, the test fails simply > because the parameter is not considered when serializing. > > The problem is that it is very hard to write non-default initializations > for all machines by hand (default is done in ctor so that is fine). It > would be nice to be able to instantiate a random combination of parameters > for machines. So Any::random or something, at least for non-pointer > parameters. > > Any thoughts on that? > > H >
