Add tests for distributions moments #5087
Conversation
Codecov Report
@@ Coverage Diff @@
## main #5087 +/- ##
==========================================
- Coverage 77.81% 77.78% -0.04%
==========================================
Files 87 88 +1
Lines 14008 14115 +107
==========================================
+ Hits 10900 10979 +79
- Misses 3108 3136 +28
|
Definitely
None yet. We should probably test that it respects the RV size, while also working with broadcasted params. |
ricardoV94
force-pushed
the
more-moments
branch
from
d76c132 to
2be8cdc
Compare
ricardoV94
changed the title
ricardoV94
force-pushed
the
more-moments
branch
from
2be8cdc to
1583962
Compare
| assert_moment_is_expected(model, expected) | ||
|
|
||
|
|
||
| @pytest.mark.skip(reason="aeppl interval transform fails when both edges are None") |
ricardoV94
force-pushed
the
more-moments
branch
from
1583962 to
e098857
Compare
There was a problem hiding this comment.
I can't approve it, even though @ricardoV94 did 98 % of the changes.
ricardoV94
force-pushed
the
more-moments
branch
from
e098857 to
490ac22
Compare
Fixes some pre-existing moments Adds moments for HalfNormal and TruncatedNormal distributions Adds helper rv_size_is_none function
ricardoV94
force-pushed
the
more-moments
branch
from
490ac22 to
7ddc9d4
Compare
|
Do we still need more moments? |
Yes, see #5078 |
| @@ -828,6 +850,12 @@ def dist(cls, sigma=None, tau=None, sd=None, *args, **kwargs): | |||
|
|
|||
| return super().dist([0.0, sigma], **kwargs) | |||
|
|
|||
| def get_moment(rv, size, loc, sigma): | |||
| moment = loc + sigma | |||
There was a problem hiding this comment.
I have a doubt here.
Current implementation: loc + sigma
Wikipedia version: loc + \sqrt{\frac{2}{\pi}}sigma
There was a problem hiding this comment.
I think you are right, we should probably be using that. Do you want to open a PR to fix it?
There was a problem hiding this comment.
A bit more involved, we should probably also use the truncated normal mean as the moment. We might even be able to simplify the switch statement, if the normal logcdf behaves well with +- infinity values: https://en.wikipedia.org/wiki/Truncated_normal_distribution
Just checking the CI already.
Do we want to systematically test the moments?
Any ideas how to do that without lots of copy-paste?
Related to #5078