Variationally Informed Parameterization #276
Conversation
ferrine
force-pushed
the
vip-reparameterization
branch
from
fa8f763 to
9443b63
Compare
|
Anything else I should do in this PR? |
|
Content looks good. Perhaps add an entry on the docs (we have to start introducing headers and sections though) |
|
Oh and a copy pastable example in the docstrings? |
|
Ah, sure, docs are needed |
| vip_rv = model_free_rv( | ||
| vip_rv_, | ||
| vip_rv_.clone(), | ||
| transform, |
There was a problem hiding this comment.
@ricardoV94 should I worry about this?
There was a problem hiding this comment.
Might be safer to raise if transform is not None, like you do with the size. I can only imagine people using OrderedTransform but you never know the things people try :D
| ) | ||
|
|
||
|
|
||
| @singledispatch |
There was a problem hiding this comment.
Seems a bit overkill at this point?
There was a problem hiding this comment.
I'd extend it to student T, Gamma, Logit Normal, etc
|
@ricardoV94 this is failing for some unknown reason |
I think a rewrite from the most recent PyTensor is leading to an error in the R2D2 prior tests. Either a rewrite bug or an issue with the R2D2. Will have to investigate |
|
Ah seems to be just a warning probably by numpy when doing something with -inf... Have to check if it's inoffensive and just ignore/catch it in the test |
|
@ricardoV94 what are the next steps? |
|
Need to check why the warning is being triggered and see if it can be safely ignored |
ferrine
force-pushed
the
vip-reparameterization
branch
from
b49a13a to
7a563c9
Compare
|
it should be all good now |
This PR adds a tool to autoparameterize your model in the following way: https://arxiv.org/abs/1906.03028
The intended api is like this
You need to convert tour model to change its structure
Then you use your samplers
For stability, you might want to turn on clipping for lambdas