Add single component samplers. #450
Conversation
|
This also fixes the travis issue currently on master. |
| @@ -61,7 +61,7 @@ def rmap(self, apt): | |||
| dpt = self.dpt.copy() | |||
|
|
|||
| for var, slc, shp in self.ordering.vmap: | |||
| dpt[var] = np.reshape(apt[slc], shp) | |||
| dpt[var] = np.reshape(np.atleast_1d(apt)[slc], shp) | |||
There was a problem hiding this comment.
this means we won't distinguish between, 0-d and 1-d values, is that okay? That makes me a little nervous.
There was a problem hiding this comment.
Well 0d is a float which you can't index into which is what produced an exception here.
There was a problem hiding this comment.
I think they actually can: array(1.0)[()] gives 1.0 and array(1.0).ravel() gives array([1.])
There was a problem hiding this comment.
Hm, so for some reason a float got passed in here. Alternatively we could wrap it with np.asarray().
There was a problem hiding this comment.
Using asarray here seems good. Though I would slightly prefer figuring out why a float was passed.
|
This will be a good addition. Thanks for fixing the problem in master too! |
|
This is the wrong approach I realized now. We should just add a kwarg |
This PR adds
SingleComponentSliceandSingleComponentMetropolis(derived from base classSingleComponentSampler) and appropriate unittests.It was confusing before that in PyMC 2 the normal usage of samplers mostly produced single component updates while instantiating
Slice()orMetropolis()in PyMC 3 does block updating by default.To have an easy option to get to single component samplers I added these.
I had to skip the unittest for Metropolis as it was converging to the correct value, I think related to #358. (FWIW I did review the code but couldn't find any logic bug).