Multioutput gp.Latent #4764
Multioutput gp.Latent #4764
Conversation
|
Nothing wrong with doing marginal GP first, and then latent in a separate PR, but this isnt core functionality, so it can certainly wait if you'd prefer. |
pymc3/glm/families.py
Outdated
| @@ -36,7 +36,7 @@ def __call__(self, x): | |||
|
|
|||
|
|
|||
| identity = Identity() | |||
| logit = at.nnet.sigmoid | |||
| logit = at.sigmoid | |||
There was a problem hiding this comment.
Confused why this file is here after #4682.
There was a problem hiding this comment.
weird, I'll double check where exactly my main is at
pymc3/gp/gp.py
Outdated
| @@ -15,7 +15,7 @@ | |||
| import functools | |||
| import warnings | |||
|
|
|||
| import aesara.tensor as at | |||
| import aesara.tensor as aet | |||
There was a problem hiding this comment.
gotcha, happy to flip it back
|
@fonnesbeck sounds good. I think I'll try and keep them separate then if that's alright so the PRs are smaller (fixed the PR title). |
|
Looking through some old PRs... @bwengals do you want to get back to this one? |
|
@michaelosthege thanks for the nudge. Yes, I'll prioritize #5035 and this for 4.0. It's important to get this in |
Codecov Report
@@ Coverage Diff @@
## main #4764 +/- ##
===========================================
- Coverage 78.96% 68.33% -10.63%
===========================================
Files 88 88
Lines 14248 14254 +6
===========================================
- Hits 11251 9741 -1510
- Misses 2997 4513 +1516
|
|
@bwengals Is this PR still relevant / being pursued? |
|
Ah thanks for the nudge @ricardoV94, yes I need to finish this up. Can we keep it open? It shouldn't be too much work either I just need to do it |
Of course :) |
|
Gonna close, @danhphan is working on improvements to this |
This PR adds support for multi-output GPs, or matrix instead of just vector
yinputs. There have been a few requests for this on discourse, most recently this one.Not sure yet, but I think that for some GP implementations, vector
ywill need have shape(n, 1)instead of(n, ), in the same way theXinputs do, which would be a breaking change (if there's no way to broadcast it automatically in aesara?).At this point, the very basics of
gp.Marginalwork with matrixY, but none of the other implementations do.mainorv3orv4?