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
initvals are not a real property of the model, but an aspect of the sampler. We already have a initvals kwarg in pm.sample that can be used to customize initial values when needed.
We can also allow extra kwargs to be passed from model.initial_point to make_initial_point_fn to allow the equivalent behavior of initval=x used in conjunction with model.initial_point()
Removing them would make our distribution and model objects simpler, and provide one consistent way of defining initvals.
On purpose, I did not make initvals part of the fgraph representation of models, which raises NotImplementedError when custom ones are set. This forces an error in utilities that make use of it: model.copy(), pm.observe, pm.do , etc...
If people really object to not removing them from the model, we have to go and make it part of the fgraph representation.
If we accept it, we should deprecate them for a while since they are an old API and used many times in conjunction with ordered transforms (which are mostly used because we don't have #7297 yet)
The text was updated successfully, but these errors were encountered:
Description
initvals are not a real property of the model, but an aspect of the sampler. We already have a
initvals
kwarg inpm.sample
that can be used to customize initial values when needed.We can also allow extra kwargs to be passed from
model.initial_point
tomake_initial_point_fn
to allow the equivalent behavior ofinitval=x
used in conjunction withmodel.initial_point()
Removing them would make our distribution and model objects simpler, and provide one consistent way of defining initvals.
On purpose, I did not make initvals part of the fgraph representation of models, which raises NotImplementedError when custom ones are set. This forces an error in utilities that make use of it:
model.copy()
,pm.observe
,pm.do
, etc...If people really object to not removing them from the model, we have to go and make it part of the fgraph representation.
If we accept it, we should deprecate them for a while since they are an old API and used many times in conjunction with ordered transforms (which are mostly used because we don't have #7297 yet)
The text was updated successfully, but these errors were encountered: