Using functools.lru_cache on functions with constant but non-hashable objects

Question

Is it possible to use functools.lru_cache for caching a partial function created by functools.partial?

My problem is a function that takes hashable parameters and contant, non-hashable objects such as NumPy arrays.

Consider this toy example:

import numpy as np
from functools import lru_cache, partial

def foo(key, array):
    print('%s:' % key, array)
a = np.array([1,2,3])

Since NumPy arrays are not hashable, this will not work:

@lru_cache(maxsize=None)
def foo(key, array):
    print('%s:' % key, array)
foo(1, a)

As expected you get following error:

/Users/ch/miniconda/envs/sci34/lib/python3.4/functools.py in __init__(self, tup, hash)
    349     def __init__(self, tup, hash=hash):
    350         self[:] = tup
--> 351         self.hashvalue = hash(tup)
    352 
    353     def __hash__(self):

TypeError: unhashable type: 'numpy.ndarray'

So my next idea was to use functools.partial to get rid of the NumPy array (which is constant anyway)

pfoo = partial(foo, array=a)
pfoo(2)

So now I have a function that only takes hashable arguments, and should be perfect for lru_cache. But is it possible to use lru_cache in this situation? I cannot use it as a wrapping function instead of the @lru_cache decorator, can I?

Is there a clever way to solve this?

Answer 1

As the array is constant you can use a wrapper around the actual lru cached function and simply pass the key value to it:

from functools import lru_cache, partial
import numpy as np


def lru_wrapper(array=None):
    @lru_cache(maxsize=None)
    def foo(key):
        return '%s:' % key, array
    return foo


arr = np.array([1, 2, 3])
func = lru_wrapper(array=arr)

for x in [0, 0, 1, 2, 2, 1, 2, 0]:
    print (func(x))

print (func.cache_info())

Outputs:

('0:', array([1, 2, 3]))
('0:', array([1, 2, 3]))
('1:', array([1, 2, 3]))
('2:', array([1, 2, 3]))
('2:', array([1, 2, 3]))
('1:', array([1, 2, 3]))
('2:', array([1, 2, 3]))
('0:', array([1, 2, 3]))
CacheInfo(hits=5, misses=3, maxsize=None, currsize=3)