Traversal over Python's objects subtree and calculate the total size of the subtree in bytes (deep size).
Project description
objsize
Traversal over Python's objects subtree and calculate the total size of the subtree in bytes (deep size).
This module traverses all child objects using Python's internal GC implementation.
It attempts to ignore shared objects (i.e., None
, types, modules, classes, functions, lambdas), as they are common
among all objects.
It is implemented without recursive calls for high performance.
Features
- Traverse objects' subtree
- Calculate objects' (deep) size in bytes
- Exclude non-exclusive objects
- Exclude specified objects subtree
- Allow the user to specify unique handlers for:
- Object's size calculation
- Object's referents (i.e., its children)
- Object filter (skip specific objects)
Pympler also supports determining an object deep size via pympler.asizeof()
.
There are two main differences between objsize
and pympler
.
objsize
has additional features:- Traversing the object subtree: iterating all the object's descendants one by one.
- Excluding non-exclusive objects. That is, objects that are also referenced from somewhere else in the program. This is true for calculating the object's deep size and for traversing its descendants.
objsize
has a simple and robust implementation with significantly fewer lines of code, compared topympler
. The Pympler implementation uses recursion, and thus have to use a maximal depth argument to avoid reaching Python's max depth.objsize
, however, uses BFS which is more efficient and simple to follow. Moreover, the Pympler implementation carefully takes care of any object type.objsize
archives the same goal with a simple and generic implementation, which has fewer lines of code.
Install
pip install objsize==0.6.0
Basic Usage
Calculate the size of the object including all its members in bytes.
>>> import objsize
>>> objsize.get_deep_size(dict(arg1='hello', arg2='world'))
340
It is possible to calculate the deep size of multiple objects by passing multiple arguments:
>>> objsize.get_deep_size(['hello', 'world'], dict(arg1='hello', arg2='world'), {'hello', 'world'})
628
Complex Data
objsize
can calculate the size of an object's entire subtree in bytes regardless of the type of objects in it, and its
depth.
Here is a complex data structure, for example, that include a self reference:
my_data = (list(range(3)), list(range(3, 6)))
class MyClass:
def __init__(self, x, y):
self.x = x
self.y = y
self.d = {'x': x, 'y': y, 'self': self}
def __repr__(self):
return "MyClass"
my_obj = MyClass(*my_data)
We can calculate my_obj
deep size, including its stored data.
>>> objsize.get_deep_size(my_obj)
708
We might want to ignore non-exclusive objects such as the ones stored in my_data
.
>>> objsize.get_deep_size(my_obj, exclude=[my_data])
384
Or simply let objsize
detect that automatically:
>>> objsize.get_exclusive_deep_size(my_obj)
384
Non Shared Functions or Classes
objsize
filters functions, lambdas, and classes by default since they are usually shared among many objects.
For example:
>>> method_dict = {"identity": lambda x: x, "double": lambda x: x*2}
>>> objsize.get_deep_size(method_dict)
232
Some objects, however, as illustrated in the above example, have unique functions not shared by other objects. Due to this, it may be useful to count their sizes. You can achieve this by providing an alternative filter function.
>>> objsize.get_deep_size(method_dict, filter_func=objsize.shared_object_filter)
986
- Note that using this filter function (
objsize.shared_object_filter_with_functions
) will also count shared functions and lambdas.
Special Cases
Some objects handle their data in a way that prevents Python's GC from detecting it. The user can supply a special way to calculate the actual size of these objects.
Case 1: torch
Using a simple calculation of the object size won't work for torch.Tensor
.
>>> import torch
>>> objsize.get_deep_size(torch.rand(200))
72
So the user can define its own size calculation handler for such cases:
import objsize
import sys
import torch
def get_size_of_torch(o):
# `objsize.safe_is_instance` catches `ReferenceError` caused by `weakref` objects
if objsize.safe_is_instance(o, torch.Tensor):
return sys.getsizeof(o.storage())
else:
return sys.getsizeof(o)
Then use it as follows:
>>> import torch
>>> objsize.get_deep_size(
... torch.rand(200),
... get_size_func=get_size_of_torch
... )
848
However, this neglects the object's internal structure.
The user can help objsize
to find the object's hidden storage by supplying it with its own referent and filter
functions:
import objsize
import gc
import torch
def get_referents_torch(*objs):
# Yield all native referents
yield from gc.get_referents(*objs)
for o in objs:
# If the object is a torch tensor, then also yield its storage
if objsize.safe_is_instance(o, torch.Tensor):
yield o.storage()
def filter_func(o):
# Torch storage points to another meta storage that is
# already included in the outer storage calculation,
# so we need to filter it.
# Also, `torch.dtype` is a common object like Python's types.
return not objsize.safe_is_instance(o, (type, torch.storage._UntypedStorage, torch.dtype))
Then use these as follows:
>>> import torch
>>> objsize.get_deep_size(
... torch.rand(200),
... get_referents_func=get_referents_torch,
... filter_func=filter_func
... )
1024
Case 2: weakref
Using a simple calculation of the object size won't work for weakref.proxy
.
>>> import weakref
>>> class Foo(list):
... pass
...
>>> o = Foo([0]*100)
>>> objsize.get_deep_size(o)
896
>>> o_ref = weakref.proxy(o)
>>> objsize.get_deep_size(oref)
72
To mitigate this, you can provide a method that attempts to fetch the proxy's referents:
import weakref
import gc
def get_weakref_referents(*objs):
yield from gc.get_referents(*objs)
for o in objs:
if type(o) in weakref.ProxyTypes:
try:
yield o.__repr__.__self__
except ReferenceError:
pass
Then use it as follows:
>>> objsize.get_deep_size(oref, get_referents_func=get_weakref_referents)
968
After the referenced object will be collected, then the size of the proxy object will be reduced.
>>> del o
>>> gc.collect()
>>> # Wait for the object to be collected
>>> objsize.get_deep_size(oref, get_referents_func=get_weakref_referents)
72
Traversal
A user can implement its own function over the entire subtree using the traversal method, which traverses all the objects in the subtree.
>>> for o in objsize.traverse_bfs(my_obj):
... print(o)
...
MyClass
{'x': [0, 1, 2], 'y': [3, 4, 5], 'd': {'x': [0, 1, 2], 'y': [3, 4, 5], 'self': MyClass}}
[0, 1, 2]
[3, 4, 5]
{'x': [0, 1, 2], 'y': [3, 4, 5], 'self': MyClass}
2
1
0
5
4
3
Similar to before, non-exclusive objects can be ignored.
>>> for o in objsize.traverse_exclusive_bfs(my_obj):
... print(o)
...
MyClass
{'x': [0, 1, 2], 'y': [3, 4, 5], 'd': {'x': [0, 1, 2], 'y': [3, 4, 5], 'self': MyClass}}
{'x': [0, 1, 2], 'y': [3, 4, 5], 'self': MyClass}
License
Project details
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.