I organized some Python built-in containers that show up in HuggingFace and coding tests.
collections
UserDict
Docs Identical to a regular Python dictionary. It’s a simple wrapper made to be easier to subclass or handle as an object.
HuggingFace BatchEncoding Github When you call HuggingFace’s tokenizer it returns a BatchEncoding type. I came across code that called pop on BatchEncoding like below and didn’t understand it at first. Turns out BatchEncoding is a subclass of UserDict, so it just uses the same pop behavior as a Python dictionary.
sample_mapping = tokenized_examples.pop("overflow_to_sample_mapping")OrderedDict
ref blog Before Python 3.6, dictionary ordering wasn’t guaranteed, so OrderedDict was necessary. From 3.6 onward, the standard dictionary also preserves insertion order during iteration.
The difference remains in comparison: a regular dictionary ignores order when comparing, while OrderedDict considers it.
deque
Double-ended queue. The ref blog has a better summary than the official docs.
- deque.append(item): Insert item at the right end of the deque.
- deque.appendleft(item): Insert item at the left end of the deque.
- deque.pop(): Remove and return the rightmost element.
- deque.popleft(): Remove and return the leftmost element.
- deque.extend(array): Iterate over the given array and append each element to the right.
- deque.extendleft(array): Iterate over the given array and append each element to the left.
- deque.remove(item): Find and remove item from the deque.
- deque.rotate(num): Rotate the deque by num steps (positive = right, negative = left).