The shape of the RandomVariables is hard to reason about

[rlouf]

The RandomVariables intialized with different values for size have a concrete shape when the size is broadcastable, and symbolic otherwise:

import aesara.tensor as at

srng = at.random.randomstream(0)

a_rv = srng.normal(0, 1)
# breakpoint inside `make_node`
# print(size, shape, bcast)
# tensorconstant{[]} tensorconstant{[]} ()
print(a_rv.shape)
# tensorconstant{[]}

a_rv = srng.normal(0, 1, size=(0,))
# breakpoint inside `make_node`
# print(size, shape, bcast)
# tensorconstant{(1,) of 0} tensorconstant{(1,) of 0} (false,)
print(a_rv.shape)
# shape.0


a_rv = srng.normal(0, 1, size=(1,))
# breakpoint inside `make_node`
# print(size, shape, bcast)
# tensorconstant{(1,) of 1} tensorconstant{(1,) of 1} (true,)
print(a_rv.shape)
# tensorconstant{(1,) of 1}


a_rv = srng.normal(0, 1, size=(2,))
# breakpoint inside `make_node`
# print(size, shape, bcast)
# tensorconstant{(1,) of 2} tensorconstant{(1,) of 2} (false,)
print(a_rv.shape)
# shape.0

This is problematic for some downstream applications. In AeHMC we create a scalar/vector inverse_mass_matrix parameter based on the number of dimensions of the variable, which thus does not have a definite shape. When a_rv.shape is a TensorConstant the shape checks performed in Scan will raise an error.

I would expect all of the above shapes to be TensorConstants, or at the very least to be all symbolic. The shapes are determined when computing the output type with TensorType(dtype=dtype, shape=bcast).

I have marked this as a question since I might also be initializing the inverse_mass_matrix incorrectly (as at.scalar or at.vector without specifying the shape).

[ricardoV94]

We can output better static shape info beyond broadcastable/not but otherwise I don't se a problem.

You should be able to use at.specify_shape(input, rv.shape) if you are having issues with replacing RVs by values or something of that sort. Or preemptively constant fold shapes if you need them.

Also you might want to use var.type.shape instead of var.shape depending on what you are trying to achieve.

[rlouf]

Turning this into a bug after an offline discussion with @brandonwillard, a_rv.shape should return a TensorConstant in every example given in the original comment.

[brandonwillard]

Turning this into a bug after an offline discussion with @brandonwillard, a_rv.shape should return a TensorConstant in every example given in the original comment.

More specifically, it looks like aesara.tensor.basic.infer_broadcastable should be returning any and all constant shapes produced by the relatively costly rewrite_graph call, instead of discarding that information via the deprecated broadcastable encoding.

[rlouf]

Indeed, if I insert a breakpoint right before

aesara/aesara/tensor/basic.py

Line 1366 in 739bd49

bcast = tuple(getattr(s, "data", s) == 1 for s in folded_shape)

folded_shape contains the correct type information (constant shapes for the cases size=(0,) and size=(2,) in the original post).

[brandonwillard]

Indeed, if I insert a breakpoint right before

aesara/aesara/tensor/basic.py

Line 1366 in 739bd49

bcast = tuple(getattr(s, "data", s) == 1 for s in folded_shape)

folded_shape contains the correct type information (constant shapes for the cases size=(0,) and size=(2,) in the original post).

Yeah, I have a fix in #1253, but I should split that off from some of the other experimental stuff in that PR.