Ignore constant shape dependencies in MarginalModel #226
Conversation
| @@ -251,9 +253,24 @@ class FiniteDiscreteMarginalRV(MarginalRV): | |||
| """Base class for Finite Discrete Marginalized RVs""" | |||
|
|
|||
|
|
|||
| def static_shape_ancestors(vars): | |||
There was a problem hiding this comment.
So before pymc-devs/pytensor@0cb5fdb, these Shape Ops would have been caught by a the simple isinstance(var, Constant) check on line 276?
Now a "constant" shape is one which 1) has no Nones, and 2) has only Shape Op ancestors?
There was a problem hiding this comment.
Before there wouldn't be any dependency between x -> shape(x) -> y(shape(x)), because shape(x) would just be a constant (so no further edges to x)
| """Test that we don't consider dependencies through "constant" shape Ops""" | ||
| x1 = pt.matrix("x1", shape=(None, 5)) | ||
| y1 = pt.random.normal(size=pt.shape(x1)) | ||
| assert is_conditional_dependent(y1, x1, [x1, y1]) |
There was a problem hiding this comment.
I would have guessed this case would fail, because there is a None in the shape (line 256)?
There was a problem hiding this comment.
There being a None is what makes the dependency. We need to see x1 to know what y1 looks like. For the x2/y2 we already know everything there is to know, so there's no dependency
CI started to fail in latest version of PyTensor due to pymc-devs/pytensor@0cb5fdb
Fix by ignoring dependecies when Shape ought to be constant (all static type shape of entries of input are fixed).
Changes should be backwards compatible