[python] Accessing dict keys like an attribute?

I find it more convenient to access dict keys as obj.foo instead of obj['foo'], so I wrote this snippet:

class AttributeDict(dict):
    def __getattr__(self, attr):
        return self[attr]
    def __setattr__(self, attr, value):
        self[attr] = value

However, I assume that there must be some reason that Python doesn't provide this functionality out of the box. What would be the caveats and pitfalls of accessing dict keys in this manner?

The best way to do this is:

class AttrDict(dict):
    def __init__(self, *args, **kwargs):
        super(AttrDict, self).__init__(*args, **kwargs)
        self.__dict__ = self

Some pros:

• It actually works!
• No dictionary class methods are shadowed (e.g. .keys() work just fine. Unless - of course - you assign some value to them, see below)
• Attributes and items are always in sync
• Trying to access non-existent key as an attribute correctly raises AttributeError instead of KeyError
• Supports [Tab] autocompletion (e.g. in jupyter & ipython)

Cons:

• Methods like .keys() will not work just fine if they get overwritten by incoming data
• Each AttrDict instance actually stores 2 dictionaries, one inherited and another one in __dict__
• Causes a memory leak in Python < 2.7.4 / Python3 < 3.2.3
• Pylint goes bananas with E1123(unexpected-keyword-arg) and E1103(maybe-no-member)
• For the uninitiated it seems like pure magic.

A short explanation on how this works

• All python objects internally store their attributes in a dictionary that is named __dict__.
• There is no requirement that the internal dictionary __dict__ would need to be "just a plain dict", so we can assign any subclass of dict() to the internal dictionary.
• In our case we simply assign the AttrDict() instance we are instantiating (as we are in __init__).
• By calling super()'s __init__() method we made sure that it (already) behaves exactly like a dictionary, since that function calls all the dictionary instantiation code.

One reason why Python doesn't provide this functionality out of the box

As noted in the "cons" list, this combines the namespace of stored keys (which may come from arbitrary and/or untrusted data!) with the namespace of builtin dict method attributes. For example:

d = AttrDict()
d.update({'items':["jacket", "necktie", "trousers"]})
for k, v in d.items():    # TypeError: 'list' object is not callable
    print "Never reached!"

Update - 2020

Since this question was asked almost ten years ago, quite a bit has changed in Python itself since then.

While this approach is still valid for some cases, e.g. legacy projects stuck to older versions of Python and cases where you really need to handle dictionaries with very dynamic string keys - I think that in general the dataclasses introduced in Python 3.7 are the obvious/correct solution to vast majority of the use cases of AttrDict.

tuples can be used dict keys. How would you access tuple in your construct?

Also, namedtuple is a convenient structure which can provide values via the attribute access.

This isn't a 'good' answer, but I thought this was nifty (it doesn't handle nested dicts in current form). Simply wrap your dict in a function:

def make_funcdict(d=None, **kwargs)
    def funcdict(d=None, **kwargs):
        if d is not None:
            funcdict.__dict__.update(d)
        funcdict.__dict__.update(kwargs)
        return funcdict.__dict__
    funcdict(d, **kwargs)
    return funcdict

Now you have slightly different syntax. To acces the dict items as attributes do f.key. To access the dict items (and other dict methods) in the usual manner do f()['key'] and we can conveniently update the dict by calling f with keyword arguments and/or a dictionary

Example

d = {'name':'Henry', 'age':31}
d = make_funcdict(d)
>>> for key in d():
...     print key
... 
age
name
>>> print d.name
... Henry
>>> print d.age
... 31
>>> d({'Height':'5-11'}, Job='Carpenter')
... {'age': 31, 'name': 'Henry', 'Job': 'Carpenter', 'Height': '5-11'}

And there it is. I'll be happy if anyone suggests benefits and drawbacks of this method.

What if you wanted a key which was a method, such as __eq__ or __getattr__?

And you wouldn't be able to have an entry that didn't start with a letter, so using 0343853 as a key is out.

And what if you didn't want to use a string?

Just to add some variety to the answer, sci-kit learn has this implemented as a Bunch:

class Bunch(dict):                                                              
    """ Scikit Learn's container object                                         

    Dictionary-like object that exposes its keys as attributes.                 
    >>> b = Bunch(a=1, b=2)                                                     
    >>> b['b']                                                                  
    2                                                                           
    >>> b.b                                                                     
    2                                                                           
    >>> b.c = 6                                                                 
    >>> b['c']                                                                  
    6                                                                           
    """                                                                         

    def __init__(self, **kwargs):                                               
        super(Bunch, self).__init__(kwargs)                                     

    def __setattr__(self, key, value):                                          
        self[key] = value                                                       

    def __dir__(self):                                                          
        return self.keys()                                                      

    def __getattr__(self, key):                                                 
        try:                                                                    
            return self[key]                                                    
        except KeyError:                                                        
            raise AttributeError(key)                                           

    def __setstate__(self, state):                                              
        pass                       

All you need is to get the setattr and getattr methods - the getattr checks for dict keys and the moves on to checking for actual attributes. The setstaet is a fix for fix for pickling/unpickling "bunches" - if inerested check https://github.com/scikit-learn/scikit-learn/issues/6196

From This other SO question there's a great implementation example that simplifies your existing code. How about:

class AttributeDict(dict):
    __slots__ = () 
    __getattr__ = dict.__getitem__
    __setattr__ = dict.__setitem__

Much more concise and doesn't leave any room for extra cruft getting into your __getattr__ and __setattr__ functions in the future.

Let me post another implementation, which builds upon the answer of Kinvais, but integrates ideas from the AttributeDict proposed in http://databio.org/posts/python_AttributeDict.html.

The advantage of this version is that it also works for nested dictionaries:

class AttrDict(dict):
    """
    A class to convert a nested Dictionary into an object with key-values
    that are accessible using attribute notation (AttrDict.attribute) instead of
    key notation (Dict["key"]). This class recursively sets Dicts to objects,
    allowing you to recurse down nested dicts (like: AttrDict.attr.attr)
    """

    # Inspired by:
    # http://stackoverflow.com/a/14620633/1551810
    # http://databio.org/posts/python_AttributeDict.html

    def __init__(self, iterable, **kwargs):
        super(AttrDict, self).__init__(iterable, **kwargs)
        for key, value in iterable.items():
            if isinstance(value, dict):
                self.__dict__[key] = AttrDict(value)
            else:
                self.__dict__[key] = value

No need to write your own as setattr() and getattr() already exist.

The advantage of class objects probably comes into play in class definition and inheritance.

I found myself wondering what the current state of "dict keys as attr" in the python ecosystem. As several commenters have pointed out, this is probably not something you want to roll your own from scratch, as there are several pitfalls and footguns, some of them very subtle. Also, I would not recommend using Namespace as a base class, I've been down that road, it isn't pretty.

Fortunately, there are several open source packages providing this functionality, ready to pip install! Unfortunately, there are several packages. Here is a synopsis, as of Dec 2019.

Contenders (most recent commit to master|#commits|#contribs|coverage%):

• addict (2019-04-28 | 217 | 22 | 100%)
• munch (2019-12-16 | 160 | 17 | ?%)
• easydict (2018-10-18 | 51 | 6 | ?%)
• attrdict (2019-02-01 | 108 | 5 | 100%)
• prodict (2019-10-01 | 65 | 1 | ?%)

No longer maintained or under-maintained:

• treedict (2014-03-28 | 95 | 2 | ?%)
• bunch (2012-03-12 | 20 | 2 | ?%)
• NeoBunch

I currently recommend munch or addict. They have the most commits, contributors, and releases, suggesting a healthy open-source codebase for each. They have the cleanest-looking readme.md, 100% coverage, and good looking set of tests.

I do not have a dog in this race (for now!), besides having rolled my own dict/attr code and wasted a ton of time because I was not aware of all these options :). I may contribute to addict/munch in the future as I would rather see one solid package than a bunch of fragmented ones. If you like them, contribute! In particular, looks like munch could use a codecov badge and addict could use a python version badge.

addict pros:

• recursive initialization (foo.a.b.c = 'bar'), dict-like arguments become addict.Dict

addict cons:

• shadows typing.Dict if you from addict import Dict
• No key checking. Due to allowing recursive init, if you misspell a key, you just create a new attribute, rather than KeyError (thanks AljoSt)

munch pros:

• unique naming
• built-in ser/de functions for JSON and YAML

munch cons:

• no recursive init / only can init one attr at a time

Wherein I Editorialize

Many moons ago, when I used text editors to write python, on projects with only myself or one other dev, I liked the style of dict-attrs, the ability to insert keys by just declaring foo.bar.spam = eggs. Now I work on teams, and use an IDE for everything, and I have drifted away from these sorts of data structures and dynamic typing in general, in favor of static analysis, functional techniques and type hints. I've started experimenting with this technique, subclassing Pstruct with objects of my own design:

class  BasePstruct(dict):
    def __getattr__(self, name):
        if name in self.__slots__:
            return self[name]
        return self.__getattribute__(name)

    def __setattr__(self, key, value):
        if key in self.__slots__:
            self[key] = value
            return
        if key in type(self).__dict__:
            self[key] = value
            return
        raise AttributeError(
            "type object '{}' has no attribute '{}'".format(type(self).__name__, key))


class FooPstruct(BasePstruct):
    __slots__ = ['foo', 'bar']

This gives you an object which still behaves like a dict, but also lets you access keys like attributes, in a much more rigid fashion. The advantage here is I (or the hapless consumers of your code) know exactly what fields can and can't exist, and the IDE can autocomplete fields. Also subclassing vanilla dict means json serialization is easy. I think the next evolution in this idea would be a custom protobuf generator which emits these interfaces, and a nice knock-on is you get cross-language data structures and IPC via gRPC for nearly free.

If you do decide to go with attr-dicts, it's essential to document what fields are expected, for your own (and your teammates') sanity.

Feel free to edit/update this post to keep it recent!

This is what I use

args = {
        'batch_size': 32,
        'workers': 4,
        'train_dir': 'train',
        'val_dir': 'val',
        'lr': 1e-3,
        'momentum': 0.9,
        'weight_decay': 1e-4
    }
args = namedtuple('Args', ' '.join(list(args.keys())))(**args)

print (args.lr)

Here's a short example of immutable records using built-in collections.namedtuple:

def record(name, d):
    return namedtuple(name, d.keys())(**d)

and a usage example:

rec = record('Model', {
    'train_op': train_op,
    'loss': loss,
})

print rec.loss(..)

class AttrDict(dict):

     def __init__(self):
           self.__dict__ = self

if __name__ == '____main__':

     d = AttrDict()
     d['ray'] = 'hope'
     d.sun = 'shine'  >>> Now we can use this . notation
     print d['ray']
     print d.sun

You can do it using this class I just made. With this class you can use the Map object like another dictionary(including json serialization) or with the dot notation. I hope help you:

class Map(dict):
    """
    Example:
    m = Map({'first_name': 'Eduardo'}, last_name='Pool', age=24, sports=['Soccer'])
    """
    def __init__(self, *args, **kwargs):
        super(Map, self).__init__(*args, **kwargs)
        for arg in args:
            if isinstance(arg, dict):
                for k, v in arg.iteritems():
                    self[k] = v

        if kwargs:
            for k, v in kwargs.iteritems():
                self[k] = v

    def __getattr__(self, attr):
        return self.get(attr)

    def __setattr__(self, key, value):
        self.__setitem__(key, value)

    def __setitem__(self, key, value):
        super(Map, self).__setitem__(key, value)
        self.__dict__.update({key: value})

    def __delattr__(self, item):
        self.__delitem__(item)

    def __delitem__(self, key):
        super(Map, self).__delitem__(key)
        del self.__dict__[key]

Usage examples:

m = Map({'first_name': 'Eduardo'}, last_name='Pool', age=24, sports=['Soccer'])
# Add new key
m.new_key = 'Hello world!'
print m.new_key
print m['new_key']
# Update values
m.new_key = 'Yay!'
# Or
m['new_key'] = 'Yay!'
# Delete key
del m.new_key
# Or
del m['new_key']

This doesn't address the original question, but should be useful for people that, like me, end up here when looking for a lib that provides this functionality.

Addict it's a great lib for this: https://github.com/mewwts/addict it takes care of many concerns mentioned in previous answers.

An example from the docs:

body = {
    'query': {
        'filtered': {
            'query': {
                'match': {'description': 'addictive'}
            },
            'filter': {
                'term': {'created_by': 'Mats'}
            }
        }
    }
}

With addict:

from addict import Dict
body = Dict()
body.query.filtered.query.match.description = 'addictive'
body.query.filtered.filter.term.created_by = 'Mats'

Apparently there is now a library for this - https://pypi.python.org/pypi/attrdict - which implements this exact functionality plus recursive merging and json loading. Might be worth a look.

Caveat emptor: For some reasons classes like this seem to break the multiprocessing package. I just struggled with this bug for awhile before finding this SO: Finding exception in python multiprocessing

Wherein I Answer the Question That Was Asked

Why doesn't Python offer it out of the box?

I suspect that it has to do with the Zen of Python: "There should be one -- and preferably only one -- obvious way to do it." This would create two obvious ways to access values from dictionaries: obj['key'] and obj.key.

Caveats and Pitfalls

These include possible lack of clarity and confusion in the code. i.e., the following could be confusing to someone else who is going in to maintain your code at a later date, or even to you, if you're not going back into it for awhile. Again, from Zen: "Readability counts!"

>>> KEY = 'spam'
>>> d[KEY] = 1
>>> # Several lines of miscellaneous code here...
... assert d.spam == 1

If d is instantiated or KEY is defined or d[KEY] is assigned far away from where d.spam is being used, it can easily lead to confusion about what's being done, since this isn't a commonly-used idiom. I know it would have the potential to confuse me.

Additonally, if you change the value of KEY as follows (but miss changing d.spam), you now get:

>>> KEY = 'foo'
>>> d[KEY] = 1
>>> # Several lines of miscellaneous code here...
... assert d.spam == 1
Traceback (most recent call last):
  File "<stdin>", line 2, in <module>
AttributeError: 'C' object has no attribute 'spam'

IMO, not worth the effort.

Other Items

As others have noted, you can use any hashable object (not just a string) as a dict key. For example,

>>> d = {(2, 3): True,}
>>> assert d[(2, 3)] is True
>>> 

is legal, but

>>> C = type('C', (object,), {(2, 3): True})
>>> d = C()
>>> assert d.(2, 3) is True
  File "<stdin>", line 1
  d.(2, 3)
    ^
SyntaxError: invalid syntax
>>> getattr(d, (2, 3))
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
TypeError: getattr(): attribute name must be string
>>> 

is not. This gives you access to the entire range of printable characters or other hashable objects for your dictionary keys, which you do not have when accessing an object attribute. This makes possible such magic as a cached object metaclass, like the recipe from the Python Cookbook (Ch. 9).

Wherein I Editorialize

I prefer the aesthetics of spam.eggs over spam['eggs'] (I think it looks cleaner), and I really started craving this functionality when I met the namedtuple. But the convenience of being able to do the following trumps it.

>>> KEYS = 'spam eggs ham'
>>> VALS = [1, 2, 3]
>>> d = {k: v for k, v in zip(KEYS.split(' '), VALS)}
>>> assert d == {'spam': 1, 'eggs': 2, 'ham': 3}
>>>

This is a simple example, but I frequently find myself using dicts in different situations than I'd use obj.key notation (i.e., when I need to read prefs in from an XML file). In other cases, where I'm tempted to instantiate a dynamic class and slap some attributes on it for aesthetic reasons, I continue to use a dict for consistency in order to enhance readability.

I'm sure the OP has long-since resolved this to his satisfaction, but if he still wants this functionality, then I suggest he download one of the packages from pypi that provides it:

• Bunch is the one I'm more familiar with. Subclass of dict, so you have all that functionality.
• AttrDict also looks like it's also pretty good, but I'm not as familiar with it and haven't looked through the source in as much detail as I have Bunch.
• Addict Is actively maintained and provides attr-like access and more.
• As noted in the comments by Rotareti, Bunch has been deprecated, but there is an active fork called Munch.

However, in order to improve readability of his code I strongly recommend that he not mix his notation styles. If he prefers this notation then he should simply instantiate a dynamic object, add his desired attributes to it, and call it a day:

>>> C = type('C', (object,), {})
>>> d = C()
>>> d.spam = 1
>>> d.eggs = 2
>>> d.ham = 3
>>> assert d.__dict__ == {'spam': 1, 'eggs': 2, 'ham': 3}

Wherein I Update, to Answer a Follow-Up Question in the Comments

In the comments (below), Elmo asks:

What if you want to go one deeper? ( referring to type(...) )

While I've never used this use case (again, I tend to use nested dict, for consistency), the following code works:

>>> C = type('C', (object,), {})
>>> d = C()
>>> for x in 'spam eggs ham'.split():
...     setattr(d, x, C())
...     i = 1
...     for y in 'one two three'.split():
...         setattr(getattr(d, x), y, i)
...         i += 1
...
>>> assert d.spam.__dict__ == {'one': 1, 'two': 2, 'three': 3}

The easiest way is to define a class let's call it Namespace. which uses the object dict.update() on the dict. Then, the dict will be treated as an object.

class Namespace(object):
    '''
    helps referencing object in a dictionary as dict.key instead of dict['key']
    '''
    def __init__(self, adict):
        self.__dict__.update(adict)



Person = Namespace({'name': 'ahmed',
                     'age': 30}) #--> added for edge_cls


print(Person.name)

How about Prodict, the little Python class that I wrote to rule them all:)

Plus, you get auto code completion, recursive object instantiations and auto type conversion!

You can do exactly what you asked for:

p = Prodict()
p.foo = 1
p.bar = "baz"

Example 1: Type hinting

class Country(Prodict):
    name: str
    population: int

turkey = Country()
turkey.name = 'Turkey'
turkey.population = 79814871

Example 2: Auto type conversion

germany = Country(name='Germany', population='82175700', flag_colors=['black', 'red', 'yellow'])

print(germany.population)  # 82175700
print(type(germany.population))  # <class 'int'>

print(germany.flag_colors)  # ['black', 'red', 'yellow']
print(type(germany.flag_colors))  # <class 'list'>

You can have all legal string characters as part of the key if you use array notation. For example, obj['!#$%^&*()_']

What would be the caveats and pitfalls of accessing dict keys in this manner?

As @Henry suggests, one reason dotted-access may not be used in dicts is that it limits dict key names to python-valid variables, thereby restricting all possible names.

The following are examples on why dotted-access would not be helpful in general, given a dict, d:

Validity

The following attributes would be invalid in Python:

d.1_foo                           # enumerated names
d./bar                            # path names
d.21.7, d.12:30                   # decimals, time
d.""                              # empty strings
d.john doe, d.denny's             # spaces, misc punctuation 
d.3 * x                           # expressions  

Style

PEP8 conventions would impose a soft constraint on attribute naming:

A. Reserved keyword (or builtin function) names:

d.in
d.False, d.True
d.max, d.min
d.sum
d.id

If a function argument's name clashes with a reserved keyword, it is generally better to append a single trailing underscore ...

B. The case rule on methods and variable names:

Variable names follow the same convention as function names.

d.Firstname
d.Country

Use the function naming rules: lowercase with words separated by underscores as necessary to improve readability.

Sometimes these concerns are raised in libraries like pandas, which permits dotted-access of DataFrame columns by name. The default mechanism to resolve naming restrictions is also array-notation - a string within brackets.

If these constraints do not apply to your use case, there are several options on dotted-access data structures.

You can use dict_to_obj https://pypi.org/project/dict-to-obj/ It does exactly what you asked for

From dict_to_obj import DictToObj
a = {
'foo': True
}
b = DictToObj(a)
b.foo
True

It doesn't work in generality. Not all valid dict keys make addressable attributes ("the key"). So, you'll need to be careful.

Python objects are all basically dictionaries. So I doubt there is much performance or other penalty.

Solution is:

DICT_RESERVED_KEYS = vars(dict).keys()


class SmartDict(dict):
    """
    A Dict which is accessible via attribute dot notation
    """
    def __init__(self, *args, **kwargs):
        """
        :param args: multiple dicts ({}, {}, ..)
        :param kwargs: arbitrary keys='value'

        If ``keyerror=False`` is passed then not found attributes will
        always return None.
        """
        super(SmartDict, self).__init__()
        self['__keyerror'] = kwargs.pop('keyerror', True)
        [self.update(arg) for arg in args if isinstance(arg, dict)]
        self.update(kwargs)

    def __getattr__(self, attr):
        if attr not in DICT_RESERVED_KEYS:
            if self['__keyerror']:
                return self[attr]
            else:
                return self.get(attr)
        return getattr(self, attr)

    def __setattr__(self, key, value):
        if key in DICT_RESERVED_KEYS:
            raise AttributeError("You cannot set a reserved name as attribute")
        self.__setitem__(key, value)

    def __copy__(self):
        return self.__class__(self)

    def copy(self):
        return self.__copy__()

As noted by Doug there's a Bunch package which you can use to achieve the obj.key functionality. Actually there's a newer version called

NeoBunch

It has though a great feature converting your dict to a NeoBunch object through its neobunchify function. I use Mako templates a lot and passing data as NeoBunch objects makes them far more readable, so if you happen to end up using a normal dict in your Python program but want the dot notation in a Mako template you can use it that way:

from mako.template import Template
from neobunch import neobunchify

mako_template = Template(filename='mako.tmpl', strict_undefined=True)
data = {'tmpl_data': [{'key1': 'value1', 'key2': 'value2'}]}
with open('out.txt', 'w') as out_file:
    out_file.write(mako_template.render(**neobunchify(data)))

And the Mako template could look like:

% for d in tmpl_data:
Column1     Column2
${d.key1}   ${d.key2}
% endfor

I created this based on the input from this thread. I need to use odict though, so I had to override get and set attr. I think this should work for the majority of special uses.

Usage looks like this:

# Create an ordered dict normally...
>>> od = OrderedAttrDict()
>>> od["a"] = 1
>>> od["b"] = 2
>>> od
OrderedAttrDict([('a', 1), ('b', 2)])

# Get and set data using attribute access...
>>> od.a
1
>>> od.b = 20
>>> od
OrderedAttrDict([('a', 1), ('b', 20)])

# Setting a NEW attribute only creates it on the instance, not the dict...
>>> od.c = 8
>>> od
OrderedAttrDict([('a', 1), ('b', 20)])
>>> od.c
8

The class:

class OrderedAttrDict(odict.OrderedDict):
    """
    Constructs an odict.OrderedDict with attribute access to data.

    Setting a NEW attribute only creates it on the instance, not the dict.
    Setting an attribute that is a key in the data will set the dict data but 
    will not create a new instance attribute
    """
    def __getattr__(self, attr):
        """
        Try to get the data. If attr is not a key, fall-back and get the attr
        """
        if self.has_key(attr):
            return super(OrderedAttrDict, self).__getitem__(attr)
        else:
            return super(OrderedAttrDict, self).__getattr__(attr)


    def __setattr__(self, attr, value):
        """
        Try to set the data. If attr is not a key, fall-back and set the attr
        """
        if self.has_key(attr):
            super(OrderedAttrDict, self).__setitem__(attr, value)
        else:
            super(OrderedAttrDict, self).__setattr__(attr, value)

This is a pretty cool pattern already mentioned in the thread, but if you just want to take a dict and convert it to an object that works with auto-complete in an IDE, etc:

class ObjectFromDict(object):
    def __init__(self, d):
        self.__dict__ = d

You can pull a convenient container class from the standard library:

from argparse import Namespace

to avoid having to copy around code bits. No standard dictionary access, but easy to get one back if you really want it. The code in argparse is simple,

class Namespace(_AttributeHolder):
    """Simple object for storing attributes.

    Implements equality by attribute names and values, and provides a simple
    string representation.
    """

    def __init__(self, **kwargs):
        for name in kwargs:
            setattr(self, name, kwargs[name])

    __hash__ = None

    def __eq__(self, other):
        return vars(self) == vars(other)

    def __ne__(self, other):
        return not (self == other)

    def __contains__(self, key):
        return key in self.__dict__