You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
In some cases, one may need to specify inits for Stan models (and potentially for others as well). In the current design it is a bit of a hard question where inits should be stored. One can currently specify them when building a backend, but that's not always convenient, because:
Inits may depend on unmodelled data (e.g. the size of an array), so one would need to change the backend to correspond with the data, which is inelegant and clunky
One may plausibly want to share the same inits across multiple backends
On the other hand, inits are - strictly speaking - an argument for the backend and different backends may require inits in different formats (e.g. sampling backends can use different inits for different chains) or even different inits (e.g. some backends may be able to fit the model well with default initialization, others may need carefully tuned inits).
So it is a bit of an unclear conceptual problem. I'll try to think about it a bit more - maybe there is some clever solution to get best of both worlds.
The text was updated successfully, but these errors were encountered:
In some cases, one may need to specify inits for Stan models (and potentially for others as well). In the current design it is a bit of a hard question where inits should be stored. One can currently specify them when building a backend, but that's not always convenient, because:
On the other hand, inits are - strictly speaking - an argument for the backend and different backends may require inits in different formats (e.g. sampling backends can use different inits for different chains) or even different inits (e.g. some backends may be able to fit the model well with default initialization, others may need carefully tuned inits).
So it is a bit of an unclear conceptual problem. I'll try to think about it a bit more - maybe there is some clever solution to get best of both worlds.
The text was updated successfully, but these errors were encountered: