I've refactored how the merged-dictionary (all_classes) below is created, but I'm wondering if it can be more efficient.

I have a dictionary of dictionaries, like this:

groups_and_classes = {'group_1': {'class_A': [1, 2, 3],
                                  'class_B': [1, 3, 5, 7], 
                                  'class_c': [1, 2], # ...many more items like this
                                 },
                      'group_2': {'class_A': [11, 12, 13],
                                  'class_C': [5, 6, 7, 8, 9]
                                 }, # ...and many more items like this
                     }

A function creates a new object from groups_and_classes like this (the function to create this is called often):

all_classes = {'class_A': [1, 2, 3, 11, 12, 13],
               'class_B': [1, 3, 5, 7, 9],
               'class_C': [1, 2, 5, 6, 7, 8, 9]
              }

Right now, there is a loop that does this:

all_classes = {}
for group in groups_and_classes.values():
    for c, vals in group.iteritems():
        for v in vals:
            if all_classes.has_key(c):
                if v not in all_classes[c]:
                    all_classes[c].append(v)
            else:
                all_classes[c] = [v]

So far, I changed the code to use a set instead of a list since the order of the list doesn't matter and the values need to be unique:

all_classes = {}
for group in groups_and_classes.values():
    for c, vals in group.iteritems():
        try:
            all_classes[c].update(set(vals))
        except KeyError:
            all_classes[c] = set(vals)

This is a little nicer, and I didn't have to convert the sets to lists because of how all_classes is used in the code.

Question: Is there a more efficient way of creating all_classes (aside from building it at the same time groups_and_classes is built, and changing everywhere this function is called)?