Local variables in nested functions

My understanding is that cage is looked for in the parent function namespace when the yielded pet_function is actually called, not before.

So when you do

funs = list(get_petters())

You generate 3 functions which will find the lastly created cage.

If you replace your last loop with :

for name, f in get_petters():
    print name + ":", 
    f()

You will actually get :

cow: Mary pets the cow.
dog: Mary pets the dog.
cat: Mary pets the cat.

The nested function looks up variables from the parent scope when executed, not when defined.

The function body is compiled, and the 'free' variables (not defined in the function itself by assignment), are verified, then bound as closure cells to the function, with the code using an index to reference each cell. pet_function thus has one free variable (cage) which is then referenced via a closure cell, index 0. The closure itself points to the local variable cage in the get_petters function.

When you actually call the function, that closure is then used to look at the value of cage in the surrounding scope at the time you call the function. Here lies the problem. By the time you call your functions, the get_petters function is already done computing it's results. The cage local variable at some point during that execution was assigned each of the 'cow', 'dog', and 'cat' strings, but at the end of the function, cage contains that last value 'cat'. Thus, when you call each of the dynamically returned functions, you get the value 'cat' printed.

The work-around is to not rely on closures. You can use a partial function instead, create a new function scope, or bind the variable as a default value for a keyword parameter.

• Partial function example, using functools.partial():

  from functools import partial
  
  def pet_function(cage=None):
      print "Mary pets the " + cage.animal + "."
  
  yield (animal, partial(gotimes, partial(pet_function, cage=cage)))
  

• Creating a new scope example:

  def scoped_cage(cage=None):
      def pet_function():
          print "Mary pets the " + cage.animal + "."
      return pet_function
  
  yield (animal, partial(gotimes, scoped_cage(cage)))
  

• Binding the variable as a default value for a keyword parameter:

  def pet_function(cage=cage):
      print "Mary pets the " + cage.animal + "."
  
  yield (animal, partial(gotimes, pet_function))
  

There is no need to define the scoped_cage function in the loop, compilation only takes place once, not on each iteration of the loop.

Let's simplify the question. Define:

def get_petters():
    for animal in ['cow', 'dog', 'cat']:
        def pet_function():
            return "Mary pets the " + animal + "."

        yield (animal, pet_function)

Then, just like in the question, we get:

>>> for name, f in list(get_petters()):
...     print(name + ":", f())

cow: Mary pets the cat.
dog: Mary pets the cat.
cat: Mary pets the cat.

But if we avoid creating a list() first:

>>> for name, f in get_petters():
...     print(name + ":", f())

cow: Mary pets the cow.
dog: Mary pets the dog.
cat: Mary pets the cat.

What's going on? Why does this subtle difference completely change our results?

---

If we look at list(get_petters()), it's clear from the changing memory addresses that we do indeed yield three different functions:

>>> list(get_petters())

[('cow', <function get_petters.<locals>.pet_function at 0x7ff2b988d790>),
 ('dog', <function get_petters.<locals>.pet_function at 0x7ff2c18f51f0>),
 ('cat', <function get_petters.<locals>.pet_function at 0x7ff2c14a9f70>)]

However, take a look at the cells that these functions are bound to:

>>> for _, f in list(get_petters()):
...     print(f(), f.__closure__)

Mary pets the cat. (<cell at 0x7ff2c112a9d0: str object at 0x7ff2c3f437f0>,)
Mary pets the cat. (<cell at 0x7ff2c112a9d0: str object at 0x7ff2c3f437f0>,)
Mary pets the cat. (<cell at 0x7ff2c112a9d0: str object at 0x7ff2c3f437f0>,)

>>> for _, f in get_petters():
...     print(f(), f.__closure__)

Mary pets the cow. (<cell at 0x7ff2b86b5d00: str object at 0x7ff2c1a95670>,)
Mary pets the dog. (<cell at 0x7ff2b86b5d00: str object at 0x7ff2c1a952f0>,)
Mary pets the cat. (<cell at 0x7ff2b86b5d00: str object at 0x7ff2c3f437f0>,)

For both loops, the cell object remains the same throughout the iterations. However, as expected, the specific str it references varies in the second loop. The cell object refers to animal, which is created when get_petters() is called. However, animal changes what str object it refers to as the generator function runs.

In the first loop, during each iteration, we create all the fs, but we only call them after the generator get_petters() is completely exhausted and a list of functions is already created.

In the second loop, during each iteration, we are pausing the get_petters() generator and calling f after each pause. Thus, we end up retrieving the value of animal at that moment in time that the generator function is paused.

As @Claudiu puts in an answer to a similar question:

Three separate functions are created, but they each have the closure of the environment they're defined in - in this case, the global environment (or the outer function's environment if the loop is placed inside another function). This is exactly the problem, though -- in this environment, animal is mutated, and the closures all refer to the same animal.

[Editor note: i has been changed to animal.]

This stems from the following

for i in range(2): 
    pass

print(i)  # prints 1

after iterating the value of i is lazily stored as its final value.

As a generator the function would work (i.e. printing each value in turn), but when transforming to a list it runs over the generator, hence all calls to cage (cage.animal) return cats.