Skip to main content

Traversal over Python's objects subtree and calculate the total size of the subtree in bytes (deep size).

Project description


Coverage Status

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.


  • 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.

  1. 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.
  2. objsize has a simple and robust implementation with significantly fewer lines of code, compared to pympler. 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.


pip install objsize==0.6.1

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'))

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'})

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)

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])

Or simply let objsize detect that automatically:

>>> objsize.get_exclusive_deep_size(my_obj)

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)

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)


  • The default filter function is objsize.shared_object_or_function_filter.
  • When using objsize.shared_object_filter, shared functions and lambdas are also counted, but builtin functions are still excluded.

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))

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(
        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
... )

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 type(o) == torch.Tensor:

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, (
        *objsize.SharedObjectOrFunctionType,, 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
... )

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)
>>> o_ref = weakref.proxy(o)
>>> objsize.get_deep_size(o_ref)

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:
                yield o.__repr__.__self__
            except ReferenceError:

Then use it as follows:

>>> objsize.get_deep_size(o_ref, get_referents_func=get_weakref_referents)

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(o_ref, get_referents_func=get_weakref_referents)


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)
{'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}

Similar to before, non-exclusive objects can be ignored.

>>> for o in objsize.traverse_exclusive_bfs(my_obj):
...     print(o)
{'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}



Project details

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

objsize-0.6.1.tar.gz (8.2 kB view hashes)

Uploaded source

Built Distribution

objsize-0.6.1-py3-none-any.whl (9.3 kB view hashes)

Uploaded py3

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page