I had seen some documentation on contributing new functions to Stan, but they seemed more on the development than the modeling side. This appears to be more modeling but still require a little more than just writing the model, but it may be what I am looking for after all. Is there a full example of this in the CmdStan documentation? That would be really helpful for the general audience (in which I include myself). Also, if it is implemented as a Stan/Math header and called via CmdStan it means it should be available to any of the interfaces, is that right?
If I manage maybe I can try to write a tutorial for it myself. Thanks.
Section 4.4
No, there is an additional step where you have to modify the Stan library so that the new function is exposed in the Stan language.
Thanks for the pointers. Iāll give it a go. I think this can be a powerful way of extending Stan, but I think between the CmdStan reference and online wiki the documentation is probably lacking for non-developers, so at this point I still find it easier to implement a simpler version of the model and sampler from scratch than implementing a new Stan function.
The thread mentioned above is helpful, too. I understand the basic usage of templates, but I normally donāt use them and I donāt understand why they are needed here, so all in all Iām still at a loss here. Iāll write an example function and maybe open a new thread for that (or maybe extend that one mentioned). Thanks.
They are needed because if you provide them then the Stan math library automatically calculates the gradient of the function for you, which is required for HMC, ADVI, and L/BFGS.
Thatās an understatement. Weāre working on developer-facing doc. Itās always the last to get attention. The cobblerās children have no shoes and all that.
One way to see how to write a templated function for Stan in full generality is to look at the C++ generated by Stan for a user-defined function. Usually, you wonāt need to template that generally for ordinary use because youāll know the types.
1 Like