I’m still worried about our pm.glm that is not good design. If we add more models we’ll need more such files in scope. I think we should refactor that in a single module with different models.
Also GLM should be refactored into tensor variable subclass. We did similar things in TransformedRV
, Minibatch
and ObservedRV
with clever theano.compile.view_op
usage. We’ll get intuitive design that way.
I agree with this (maybe not with Zoo
as the name tho). To me, I prefer the specific functional module in a submodule:
- zoo
- GLM
- GP
- ODE?
- Mixture models?
- ...
1 Like
I also have the following proposal
- models
- base (moved
pm.model
)
- GLM
- GP
- ODE?
- Mixture models?
- …
Hmm, I would prefer pm.model being a more general class.
Also, for each zoo module, we can have some classical and model specific algorithm for model fitting. It would be cool to then compare with general inference method such as NUTS and VI
pm.Model
will be the base class for all the rest. That’s why I want to move it to base.py