Memoize function result based on selected parameters

Question

I have a function that has the following signature:

def spectrogram(signal: numpy.ndarray, sampling_frequency=16000, win_len=512, hop=256, win_type='hanning')

The function expects a numpy array on input (signal), plus number of other parameters, and outputs a numpy array. It computes spectrogram for the given audio file in order to obtain certain acoustic features. Per file, I am going to call this function multiple times, often with the same parameters, but not always. For certain features I might change the hop or win_type. I was thinking to cache the results, so that I don't run the same computations more than once. The results are valid per file. The files are going to be processes in parallel with joblib.

I was thinking to memoize the results based on the file name (which isn't a parameter that I normally would have on the function) and fields sampling_frequency, win_len, hop and win_type (i.e. NOT signal - this can be a large array and it is much more efficient to look at filename, which is unique).

How can I best memoize the results? All solutions that I have seen, cache results based on provided input; in my case I'd like to memoize based on selected fields. I am on Python 3.6.

Answer 1

Here is my implementation, I had to mock up some data, so if it doesn’t do quite what you need I’m happy to tweak it a bit.

def memo(hashTable, fileName, signal: np.ndarray, sampling_frequency=16000, win_len=512, hop=256, win_type='hanning'):  
    new_hash = hash(fileName + str(sampling_frequency) + str(win_len) + str(hop) + win_type)
    if new_hash in hashTable.keys():
        return hashTable[new_hash]
    else:
        answer = spectrogram(signal, sampling_frequency, win_len, hop, win_type)
        hashTable[new_hash] = answer
        return answer


def spectrogram(signal: np.ndarray, sampling_frequency=16000, win_len=512, hop=256, win_type='hanning'):
    makeArrayUnique = hop - 256
    return np.arange(makeArrayUnique, 24 + makeArrayUnique).reshape(2,12)

def testHash():
    hashTable = {}
    dummySignal = np.zeros(10)
    print('First call', memo(hashTable, 'file1', signal=dummySignal))
    print('Second Call', memo(hashTable, 'file1', signal=dummySignal, hop=260))
    print('First call again', memo(hashTable, 'file1', signal=dummySignal))

    print('Hash Table', hashTable)

Output showing 3 calls but only two entries in hash table:

>>> testHash()
First call [[ 0  1  2  3  4  5  6  7  8  9 10 11]
 [12 13 14 15 16 17 18 19 20 21 22 23]]
Second Call [[ 4  5  6  7  8  9 10 11 12 13 14 15]
 [16 17 18 19 20 21 22 23 24 25 26 27]]
First call again [[ 0  1  2  3  4  5  6  7  8  9 10 11]
 [12 13 14 15 16 17 18 19 20 21 22 23]]
Hash Table {-4316472197502448580: array([[ 0,  1,  2,  3,  4,  5,  6,  7,  8,  9, 10, 11],
       [12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23]]), 6772234510013844540: array([[ 4,  5,  6,  7,  8,  9, 10, 11, 12, 13, 14, 15],
       [16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27]])}

Answer 2

I decided to post my own answer that builds on @Ethan answer (+1 vote) by adding two elements:

1. Bounding of the cache. That was one of my prerequisites, which is also why I could not accept Ethan's answer. The latter is unbounded and would quickly exhaust my memory.

2. I feel it is more elegant; it uses a decorator and a module designed for caching (and catching some corner cases). It's more reusable and therefore more friendly for others.

from cachetools.keys import hashkey
from cachetools import cached, LRUCache

def mykey(signal, *args, **kwargs):
    key = hashkey(*args, **kwargs)
    return key

@cached(LRUCache(maxsize=6), key=mykey)
def spectrogram(signal: numpy.ndarray, filename, sampling_frequency=16000, win_len=512, hop=256, win_type='hanning')

In essence, I am simply ignoring the signal and instead taking filename as an extra parameter for caching. Under certain circumstances even the filename would not be needed. If a separate process is spawned per file, there's no need for this safeguard, as the cache cannot be shared between processes anyway.

Bonus

I also decided to try Memory from joblib and it also performed well. Here's a snippet:

from joblib import Memory

memory = Memory('cachedir', verbose=0, bytes_limit=100000)

@memory.cache
def spectrogram(signal: numpy.ndarray, sampling_frequency=16000, win_len=512, hop=256, win_type='hanning')

It performed on average 25% worse than the first solution as:

• it writes to disk
• it computes a hash over complete numpy.ndarray

Considering the above, it's still a great score.