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 arviz-plots we use stats_kwargs argument with datasets as values to indicate we don't want a computation to happen but instead we want the provided values to be used. I think we should make this possible here too as it can significantly reduce the computational cost, for example, computing multiple hdi intervals currently recomputes the kde for each interval; (in current arviz) we use psis log weights in multiple places, but the ability to provide log weights that have been already computed is quite limited...
The initial goal of this issue is to discuss what API we could use for that. I think it would be fine to consider this a somewhat advanced feature, so that manual precomputation is required instead of having the kde that was computed within hdi to be accessible to the user so it can be provided in a future call to hdi for example.
Some ideas:
Have a precomputed keyword argument that takes a dictionary of keywords representing intermediate results and their values as datasets/dataarrays/arrays
add explicit keyword arguments for each intermediate result where providing precomputed results is supported. We'd have to look into potential argument name clashes.
The text was updated successfully, but these errors were encountered:
In
arviz-plots
we usestats_kwargs
argument with datasets as values to indicate we don't want a computation to happen but instead we want the provided values to be used. I think we should make this possible here too as it can significantly reduce the computational cost, for example, computing multiple hdi intervals currently recomputes the kde for each interval; (in current arviz) we use psis log weights in multiple places, but the ability to provide log weights that have been already computed is quite limited...The initial goal of this issue is to discuss what API we could use for that. I think it would be fine to consider this a somewhat advanced feature, so that manual precomputation is required instead of having the kde that was computed within
hdi
to be accessible to the user so it can be provided in a future call tohdi
for example.Some ideas:
precomputed
keyword argument that takes a dictionary of keywords representing intermediate results and their values as datasets/dataarrays/arraysThe text was updated successfully, but these errors were encountered: