iterparse() lets you do stuff while building the tree, that means that unless you remove what you don't need anymore, you'll still end up with the whole tree in the end.
For more information: read this by the author of the original ElementTree implementation (but it's also applicable to lxml)