是否有一个装饰器来简单地缓存函数的返回值？

Consider the following:

@property
def name(self):

    if not hasattr(self, '_name'):

        # expensive calculation
        self._name = 1 + 1

    return self._name

I’m new, but I think the caching could be factored out into a decorator. Only I didn’t find one like it ;)

PS the real calculation doesn’t depend on mutable values

Starting from Python 3.2 there is a built-in decorator:

@functools.lru_cache(maxsize=100, typed=False)

Decorator to wrap a function with a memoizing callable that saves up to the maxsize most recent calls. It can save time when an expensive or I/O bound function is periodically called with the same arguments.

Example of an LRU cache for computing Fibonacci numbers:

@lru_cache(maxsize=None)
def fib(n):
    if n < 2:
        return n
    return fib(n-1) + fib(n-2)

>>> print([fib(n) for n in range(16)])
[0, 1, 1, 2, 3, 5, 8, 13, 21, 34, 55, 89, 144, 233, 377, 610]

>>> print(fib.cache_info())
CacheInfo(hits=28, misses=16, maxsize=None, currsize=16)

If you are stuck with Python 2.x, here’s a list of other compatible memoization libraries:

• functools32 | PyPI | Source code
• repoze.lru | PyPI | Source code
• pylru | PyPI | Source code
• backports.functools_lru_cache | PyPI | Source code

It sounds like you’re not asking for a general-purpose memoization decorator (i.e., you’re not interested in the general case where you want to cache return values for different argument values). That is, you’d like to have this:

x = obj.name  # expensive
y = obj.name  # cheap

while a general-purpose memoization decorator would give you this:

x = obj.name()  # expensive
y = obj.name()  # cheap

I submit that the method-call syntax is better style, because it suggests the possibility of expensive computation while the property syntax suggests a quick lookup.

[Update: The class-based memoization decorator I had linked to and quoted here previously doesn’t work for methods. I’ve replaced it with a decorator function.] If you’re willing to use a general-purpose memoization decorator, here’s a simple one:

def memoize(function):
  memo = {}
  def wrapper(*args):
    if args in memo:
      return memo[args]
    else:
      rv = function(*args)
      memo[args] = rv
      return rv
  return wrapper

Example usage:

@memoize
def fibonacci(n):
  if n < 2: return n
  return fibonacci(n - 1) + fibonacci(n - 2)

Another memoization decorator with a limit on the cache size can be found here.

class memorize(dict):
    def __init__(self, func):
        self.func = func

    def __call__(self, *args):
        return self[args]

    def __missing__(self, key):
        result = self[key] = self.func(*key)
        return result

Sample uses:

>>> @memorize
... def foo(a, b):
...     return a * b
>>> foo(2, 4)
8
>>> foo
{(2, 4): 8}
>>> foo('hi', 3)
'hihihi'
>>> foo
{(2, 4): 8, ('hi', 3): 'hihihi'}

Python 3.8 functools.cached_property decorator

https://docs.python.org/dev/library/functools.html#functools.cached_property

cached_property from Werkzeug was mentioned at: https://stackoverflow.com/a/5295190/895245 but a supposedly derived version will be merged into 3.8, which is awesome.

This decorator can be seen as caching @property, or as a cleaner @functools.lru_cache for when you don’t have any arguments.

The docs say:

@functools.cached_property(func)

Transform a method of a class into a property whose value is computed once and then cached as a normal attribute for the life of the instance. Similar to property(), with the addition of caching. Useful for expensive computed properties of instances that are otherwise effectively immutable.

Example:

class DataSet:
    def __init__(self, sequence_of_numbers):
        self._data = sequence_of_numbers

    @cached_property
    def stdev(self):
        return statistics.stdev(self._data)

    @cached_property
    def variance(self):
        return statistics.variance(self._data)

New in version 3.8.

Note This decorator requires that the dict attribute on each instance be a mutable mapping. This means it will not work with some types, such as metaclasses (since the dict attributes on type instances are read-only proxies for the class namespace), and those that specify slots without including dict as one of the defined slots (as such classes don’t provide a dict attribute at all).

Werkzeug has a cached_property decorator (docs, source)

I coded this simple decorator class to cache function responses. I find it VERY useful for my projects:

from datetime import datetime, timedelta 

class cached(object):
    def __init__(self, *args, **kwargs):
        self.cached_function_responses = {}
        self.default_max_age = kwargs.get("default_cache_max_age", timedelta(seconds=0))

    def __call__(self, func):
        def inner(*args, **kwargs):
            max_age = kwargs.get('max_age', self.default_max_age)
            if not max_age or func not in self.cached_function_responses or (datetime.now() - self.cached_function_responses[func]['fetch_time'] > max_age):
                if 'max_age' in kwargs: del kwargs['max_age']
                res = func(*args, **kwargs)
                self.cached_function_responses[func] = {'data': res, 'fetch_time': datetime.now()}
            return self.cached_function_responses[func]['data']
        return inner

The usage is straightforward:

import time

@cached
def myfunc(a):
    print "in func"
    return (a, datetime.now())

@cached(default_max_age = timedelta(seconds=6))
def cacheable_test(a):
    print "in cacheable test: "
    return (a, datetime.now())


print cacheable_test(1,max_age=timedelta(seconds=5))
print cacheable_test(2,max_age=timedelta(seconds=5))
time.sleep(7)
print cacheable_test(3,max_age=timedelta(seconds=5))

DISCLAIMER: I’m the author of kids.cache.

You should check kids.cache, it provides a @cache decorator that works on python 2 and python 3. No dependencies, ~100 lines of code. It’s very straightforward to use, for instance, with your code in mind, you could use it like this:

pip install kids.cache

Then

from kids.cache import cache
...
class MyClass(object):
    ...
    @cache            # <-- That's all you need to do
    @property
    def name(self):
        return 1 + 1  # supposedly expensive calculation

Or you could put the @cache decorator after the @property (same result).

Using cache on a property is called lazy evaluation, kids.cache can do much more (it works on function with any arguments, properties, any type of methods, and even classes…). For advanced users, kids.cache supports cachetools which provides fancy cache stores to python 2 and python 3 (LRU, LFU, TTL, RR cache).

IMPORTANT NOTE: the default cache store of kids.cache is a standard dict, which is not recommended for long running program with ever different queries as it would lead to an ever growing caching store. For this usage you can plugin other cache stores using for instance (@cache(use=cachetools.LRUCache(maxsize=2)) to decorate your function/property/class/method…)

Ah, just needed to find the right name for this: “Lazy property evaluation“.

I do this a lot too; maybe I’ll use that recipe in my code sometime.

There is fastcache, which is “C implementation of Python 3 functools.lru_cache. Provides speedup of 10-30x over standard library.”

Same as chosen answer, just different import:

from fastcache import lru_cache
@lru_cache(maxsize=128, typed=False)
def f(a, b):
    pass

Also, it comes installed in Anaconda, unlike functools which needs to be installed.

There is yet another example of a memoize decorator at Python Wiki:

http://wiki.python.org/moin/PythonDecoratorLibrary#Memoize

That example is a bit smart, because it won’t cache the results if the parameters are mutable. (check that code, it’s very simple and interesting!)

If you are using Django Framework, it has such a property to cache a view or response of API’s using @cache_page(time) and there can be other options as well.

Example:

@cache_page(60 * 15, cache="special_cache")
def my_view(request):
    ...

More details can be found here.

Along with the Memoize Example I found the following python packages:

• cachepy; It allows to set up ttl and\or the number of calls for cached functions; Also, one can use encrypted file-based cache…
• percache

I implemented something like this, using pickle for persistance and using sha1 for short almost-certainly-unique IDs. Basically the cache hashed the code of the function and the hist of arguments to get a sha1 then looked for a file with that sha1 in the name. If it existed, it opened it and returned the result; if not, it calls the function and saves the result (optionally only saving if it took a certain amount of time to process).

That said, I’d swear I found an existing module that did this and find myself here trying to find that module… The closest I can find is this, which looks about right: http://chase-seibert.github.io/blog/2011/11/23/pythondjango-disk-based-caching-decorator.html

The only problem I see with that is it wouldn’t work well for large inputs since it hashes str(arg), which isn’t unique for giant arrays.

It would be nice if there were a unique_hash() protocol that had a class return a secure hash of its contents. I basically manually implemented that for the types I cared about.

Try joblib http://pythonhosted.org/joblib/memory.html

from joblib import Memory
memory = Memory(cachedir=cachedir, verbose=0)
@memory.cache
    def f(x):
        print('Running f(%s)' % x)
        return x

If you are using Django and want to cache views, see Nikhil Kumar’s answer.

But if you want to cache ANY function results, you can use django-cache-utils.

It reuses Django caches and provides easy to use cached decorator:

from cache_utils.decorators import cached

@cached(60)
def foo(x, y=0):
    print 'foo is called'
    return x+y

@lru_cache is not perfect with default function values

my mem decorator:

import inspect


def get_default_args(f):
    signature = inspect.signature(f)
    return {
        k: v.default
        for k, v in signature.parameters.items()
        if v.default is not inspect.Parameter.empty
    }


def full_kwargs(f, kwargs):
    res = dict(get_default_args(f))
    res.update(kwargs)
    return res


def mem(func):
    cache = dict()

    def wrapper(*args, **kwargs):
        kwargs = full_kwargs(func, kwargs)
        key = list(args)
        key.extend(kwargs.values())
        key = hash(tuple(key))
        if key in cache:
            return cache[key]
        else:
            res = func(*args, **kwargs)
            cache[key] = res
            return res
    return wrapper

and code for testing:

from time import sleep


@mem
def count(a, *x, z=10):
    sleep(2)
    x = list(x)
    x.append(z)
    x.append(a)
    return sum(x)


def main():
    print(count(1,2,3,4,5))
    print(count(1,2,3,4,5))
    print(count(1,2,3,4,5, z=6))
    print(count(1,2,3,4,5, z=6))
    print(count(1))
    print(count(1, z=10))


if __name__ == '__main__':
    main()

result – only 3 times with sleep

but with @lru_cache it will be 4 times, because this:

print(count(1))
print(count(1, z=10))

will be calculated twice (bad working with defaults)