Shared variables

The documentation defines shared variables as variables that "may be shared between different functions". Well it turns out they have a much deeper meaning that was not explicited in Aesara, and was not made obvious by this name. Let's explore what shared variables are, what they can be used for, where they occur in the library and where they make things difficult.

Example taken from the documentation:

from aesara import function, shared
import aesara.tensor as at

state = shared(0)
inc = at.iscalar('inc')
accumulate = function([inc], state, updates=[(state, state+inc)])


decrement = function([inc], state, updates=[(state, state-inc)])


shared variables are not completely symbolic as they carry some internal state. They interact with the graph in two ways:

  1. get_value gets the current value of the variable, stored in a container.
  2. set_value sets the value of the shared variable for all the functions it is used in.

They also allow sharing state between compiled graphs.

updates parameter in function tells the compiler that when the function is going to be run the value of the shared variable is going to be modified in X way. They are necessary because the value of the shared variables needs to be propagated globally?

TODO Conceptual understanding: what is a shared variable? Why is it not a Variable?

Variables are placeholders, mere step in a computation. They are variables in the mathematical sense. Shared variables are variables in the computing sense: a value store. The problem is that shared variables introduce a notion of imperative programming (namely "update this register in memory with this value"). Shared variables allow us to:

  • Interact with our graphs once they've been computed;
  • Link different compiled graphs;
  • SOmething with scan.

We can read the following aesara.function's docstring:

  1. RemoveShared: shared variables are just an abstraction to make things more convenient for the user. The shared variables are

transformed into implicit inputs and implicit outputs. The optimizations don't see which variables are shared or not.

TODO How does scan work with updates?

Well, scan returns a graph that says how the shared variable within should be treated.

import aesara

a = aesara.shared(1)
values, updates = aesara.scan(lambda: {a: a+1}, n_steps=10)
b = a + 1
c = updates[a] + 1
f = aesara.function([], [b, c], updates=updates)

It has an interesting behavior. If we run the function we get something expected:


b is the original value, plus one. c includes the increments that we built by scan. If we run the function again we see that a value was incremented by 10:


What if we don't pass the updates to the function?

from aesara import function, shared

a = shared(1)
values, updates = aesara.scan(lambda: {a: a+1}, n_steps=10)
b = a + 1
c = updates[a] + 1
f = aesara.function([], [b, c])

The updates were not applied to the shared variable. What passing the updates to the function means is "also update the shared variables' value". Does it otherwise? Let's take a simple example:

a = shared(0)
a = a + 1
fn = aesara.function([], a)

So Variables are placeholders, they're not variables in the sense of the host language (python) as they don't store any value. SharedVariables however are variables in the sense of the host language; we can modify them in the host process runtime, or in the graph runtime using the updates kwarg in the function.

TODO What happens to shared variables when the graph is copied?

If the updates still refer to the original shared variables then nothing is going to happen; so using aesara.graph.basic.clone_get_equiv and passing the origin updates will likely lead to buggy code since the original updates won't be applied to the cloned variable.

Instead we need to use functions like aesara.compile.function.pfunc.rebuild_collec_shared to also get the updated updates.

TODO Why do we use shared variables for the RandomStream?

TODO Shared variables don't have a sense of scope (give example)


Links to this note