Usage of Joblib Memory Class

Use case

The Memory class defines a context for lazy evaluation of function, by putting the results in a store, by default using a disk, and not re-running the function twice for the same arguments.

It works by explicitly saving the output to a file and it is designed to work with non-hashable and potentially large input and output data types such as numpy arrays.

A simple example:

First, define the cache directory:

>>> cachedir = 'your_cache_location_directory'

Then, instanciate a memory context that uses this cache directory:

>>> from joblib import Memory
>>> memory = Memory(cachedir, verbose=0)

After these initial steps, just decorate a function to cache its output in this context:

>>> @memory.cache
... def f(x):
...     print('Running f(%s)' % x)
...     return x

Calling this function twice with the same argument does not execute it the second time, the output is just reloaded from a pickle file in the cache directory:

>>> print(f(1))
Running f(1)
>>> print(f(1))

However, calling the function with a different parameter executes it and recomputes the output:

>>> print(f(2))
Running f(2)

Comparison with memoize

The memoize decorator ( caches in memory all the inputs and outputs of a function call. It can thus avoid running twice the same function, with a very small overhead. However, it compares input objects with those in cache on each call. As a result, for big objects there is a huge overhead. Moreover this approach does not work with numpy arrays, or other objects subject to non-significant fluctuations. Finally, using memoize with large objects will consume all the memory, where with Memory, objects are persisted to disk, using a persister optimized for speed and memory usage (joblib.dump()).

In short, memoize is best suited for functions with “small” input and output objects, whereas Memory is best suited for functions with complex input and output objects, and aggressive persistence to disk.

Using with numpy

The original motivation behind the Memory context was to have a memoize-like pattern on numpy arrays. Memory uses fast cryptographic hashing of the input arguments to check if they have been computed;

An example

Define two functions: the first with a number as an argument, outputting an array, used by the second one. Both functions are decorated with Memory.cache:

>>> import numpy as np

>>> @memory.cache
... def g(x):
...     print('A long-running calculation, with parameter %s' % x)
...     return np.hamming(x)

>>> @memory.cache
... def h(x):
...     print('A second long-running calculation, using g(x)')
...     return np.vander(x)

If the function h is called with the array created by the same call to g, h is not re-run:

>>> a = g(3)
A long-running calculation, with parameter 3
>>> a
array([0.08, 1.  , 0.08])
>>> g(3)
array([0.08, 1.  , 0.08])
>>> b = h(a)
A second long-running calculation, using g(x)
>>> b2 = h(a)
>>> b2
array([[0.0064, 0.08  , 1.    ],
       [1.    , 1.    , 1.    ],
       [0.0064, 0.08  , 1.    ]])
>>> np.allclose(b, b2)

Using memmapping

Memmapping (memory mapping) speeds up cache looking when reloading large numpy arrays:

>>> cachedir2 = 'your_cachedir2_location'
>>> memory2 = Memory(cachedir2, mmap_mode='r')
>>> square = memory2.cache(np.square)
>>> a = np.vander(np.arange(3)).astype(np.float)
>>> square(a)
[Memory] Calling square...
square(array([[0., 0., 1.],
       [1., 1., 1.],
       [4., 2., 1.]]))
___________________________________________________________square - 0.0s, 0.0min
memmap([[ 0.,  0.,  1.],
        [ 1.,  1.,  1.],
        [16.,  4.,  1.]])

Note: Notice the debug mode used in the above example. It is useful for tracing of what is being reexecuted, and where the time is spent.

If the square function is called with the same input argument, its return value is loaded from the disk using memmapping:

>>> res = square(a)
>>> print(repr(res))
memmap([[ 0.,  0.,  1.],
        [ 1.,  1.,  1.],
        [16.,  4.,  1.]])

The memmap file must be closed to avoid file locking on Windows; closing numpy.memmap objects is done with del, which flushes changes to the disk

>>> del res

Note: If the memory mapping mode used was ‘r’, as in the above example, the array will be read only, and will be impossible to modified in place.

On the other hand, using ‘r+’ or ‘w+’ will enable modification of the array, but will propagate these modification to the disk, which will corrupt the cache. If you want modification of the array in memory, we suggest you use the ‘c’ mode: copy on write.

Shelving: using references to cached values

In some cases, it can be useful to get a reference to the cached result, instead of having the result itself. A typical example of this is when a lot of large numpy arrays must be dispatched accross several workers: instead of sending the data themselves over the network, send a reference to the joblib cache, and let the workers read the data from a network filesystem, potentially taking advantage of some system-level caching too.

Getting a reference to the cache can be done using the call_and_shelve method on the wrapped function:

>>> result = g.call_and_shelve(4)
A long-running calculation, with parameter 4
>>> result  
MemorizedResult(location="...", func="...g...", argument_hash="...")

Once computed, the output of g is stored on disk, and deleted from memory. Reading the associated value can then be performed with the get method:

>>> result.get()
array([0.08, 0.77, 0.77, 0.08])

The cache for this particular value can be cleared using the clear method. Its invocation causes the stored value to be erased from disk. Any subsequent call to get will cause a KeyError exception to be raised:

>>> result.clear()
>>> result.get()  
Traceback (most recent call last):
KeyError: 'Non-existing cache value (may have been cleared).\nFile ... does not exist'

A MemorizedResult instance contains all that is necessary to read the cached value. It can be pickled for transmission or storage, and the printed representation can even be copy-pasted to a different python interpreter.

Shelving when cache is disabled

In the case where caching is disabled (e.g. Memory(None)), the call_and_shelvemethod returns a NotMemorizedResult instance, that stores the full function output, instead of just a reference (since there is nothing to point to). All the above remains valid though, except for the copy-pasting feature.


Across sessions, function cache is identified by the function’s name. Thus assigning the same name to different functions, their cache will override each-others (e.g. there are ‘name collisions’), and unwanted re-run will happen:

>>> @memory.cache
... def func(x):
...     print('Running func(%s)' % x)

>>> func2 = func

>>> @memory.cache
... def func(x):
...     print('Running a different func(%s)' % x)

As long as the same session is used, there are no collisions (in joblib 0.8 and above), altough joblib does warn you that you are doing something dangerous:

>>> func(1)
Running a different func(1)

>>> # FIXME: The next line should create a JolibCollisionWarning but does not
>>> # memory.rst:0: JobLibCollisionWarning: Possible name collisions between functions 'func' (<doctest memory.rst>:...) and 'func' (<doctest memory.rst>:...)
>>> func2(1)  
Running func(1)

>>> func(1) # No recomputation so far
>>> func2(1) # No recomputation so far

But suppose the interpreter is exited and then restarted, the cache will not be identified properly, and the functions will be rerun:

>>> # FIXME: The next line will should create a JoblibCollisionWarning but does not. Also it is skipped because it does not produce any output
>>> # memory.rst:0: JobLibCollisionWarning: Possible name collisions between functions 'func' (<doctest memory.rst>:...) and 'func' (<doctest memory.rst>:...)
>>> func(1) 
Running a different func(1)
>>> func2(1)  
Running func(1)

As long as the same session is used, there are no needless recomputation:

>>> func(1) # No recomputation now
>>> func2(1) # No recomputation now

lambda functions

Beware that with Python 2.7 lambda functions cannot be separated out:

>>> def my_print(x):
...     print(x)

>>> f = memory.cache(lambda : my_print(1))
>>> g = memory.cache(lambda : my_print(2))

>>> f()
>>> f()
>>> g() 
memory.rst:0: JobLibCollisionWarning: Cannot detect name collisions for function '<lambda>'
>>> g() 
>>> f() 

memory cannot be used on some complex objects, e.g. a callable object with a __call__ method.

However, it works on numpy ufuncs:

>>> sin = memory.cache(np.sin)
>>> print(sin(0))

caching methods: memory is designed for pure functions and it is not recommended to use it for methods. If one wants to use cache inside a class the recommended pattern is to cache a pure function and use the cached function inside your class, i.e. something like this:

def compute_func(arg1, arg2, arg3):
    # long computation
    return result

class Foo(object):
    def __init__(self, args): = None

    def compute(self): = compute_func(self.arg1, self.arg2, 40)

Using Memory for methods is not recommended and has some caveats that make it very fragile from a maintenance point of view because it is very easy to forget about these caveats when a software evolves. If this cannot be avoided (we would be interested about your use case by the way), here are a few known caveats:

1. a method cannot be decorated at class definition, because when the class is instantiated, the first argument (self) is bound, and no longer accessible to the Memory object. The following code won’t work:

class Foo(object):

    @mem.cache  # WRONG
    def method(self, args):

The right way to do this is to decorate at instantiation time:

class Foo(object):

    def __init__(self, args):
        self.method = mem.cache(self.method)

    def method(self, ...):

2. The cached method will have self as one of its arguments. That means that the result will be recomputed if anything with self changes. For example if self.attr has changed calling self.method will recompute the result even ifself.method does not use self.attr in its body. Another example is changing self inside the body of self.method. The consequence is that self.method will create cache that will not be reused in subsequent calls. To alleviate these problems and if you know that the result of self.method does not depend on self you can use self.method = mem.cache(self.method, ignore=['self']).

Ignoring some arguments

It may be useful not to recalculate a function when certain arguments change, for instance a debug flag. Memory provides the ignore list:

>>> @memory.cache(ignore=['debug'])
... def my_func(x, debug=True):
...     print('Called with x = %s' % x)
>>> my_func(0)
Called with x = 0
>>> my_func(0, debug=False)
>>> my_func(0, debug=True)
>>> # my_func was not reevaluated

Reference documentation of the Memory class

class joblib.memory.Memory(location=None, backend='local', cachedir=None, mmap_mode=None, compress=False, verbose=1, bytes_limit=None, backend_options={})

A context object for caching a function’s return value each time it is called with the same input arguments.

All values are cached on the filesystem, in a deep directory structure.

Read more in the User Guide.

Parameter NameDescription
location: str or NoneThe path of the base directory to use as a data store or None. If None is given, no caching is done and the Memory object is completely transparent. This option replaces cachedir since version 0.12.
backend: str, optionalType of store backend for reading/writing cache files. Default: ‘local’. The ‘local’ backend is using regular filesystem operations to manipulate data (open, mv, etc) in the backend.
cachedir: str or None, optional 
mmap_mode: {None, ‘r+’, ‘r’, ‘w+’, ‘c’}, optionalThe memmapping mode used when loading from cache numpy arrays. See numpy.load for the meaning of the arguments.
compress: boolean, or integer, optionalWhether to zip the stored data on disk. If an integer is given, it should be between 1 and 9, and sets the amount of compression. Note that compressed arrays cannot be read by memmapping.
verbose: int, optionalVerbosity flag, controls the debug messages that are issued as functions are evaluated.
bytes_limit: int, optionalLimit in bytes of the size of the cache.
backend_options: dict, optionalContains a dictionnary of named parameters used to configure the store backend.

Methods in Memory class:

Method NameParametersDescription
cache(func=None, ignore=None, verbose=None, mmap_mode=False)func: callable, optionalThe function to be decoratedignore: list of stringsA list of arguments name to ignore in the hashingverbose: integer, optionalThe verbosity mode of the function. By default that of the memory object is used.mmap_mode: {None, ‘r+’, ‘r’, ‘w+’, ‘c’}, optionalThe memmapping mode used when loading from cache numpy arrays. See numpy.load for the meaning of the arguments. By default that of the memory object is used.Decorates the given function func to only compute its return value for input arguments not cached on disk.
clear(warn=True) Erase the complete cache directory.
eval(func, *args, **kwargs) Eval function func with arguments *args and **kwargs, in the context of the memory.This method works similarly to the builtin apply, except that the function is called only if the cache is not up to date.

Useful methods of decorated functions

Function decorated by Memory.cache() are MemorizedFunc objects that, in addition of behaving like normal functions, expose methods useful for cache exploration and management.

class joblib.memory.MemorizedFunc(func, location, backend='local', ignore=None, mmap_mode=None, compress=False, verbose=1, timestamp=None)

Callable object decorating a function for caching its return value each time it is called.

Methods are provided to inspect the cache or clean it.

Attribute NameDescription
func: callableThe original, undecorated, function.
location: stringThe location of joblib cache. Depends on the store backend used.
backend: strType of store backend for reading/writing cache files. Default is ‘local’, in which case the location is the path to a disk storage.
ignore: list or NoneList of variable names to ignore when choosing whether to recompute.
mmap_mode: {None, ‘r+’, ‘r’, ‘w+’, ‘c’}The memmapping mode used when loading from cache numpy arrays. See numpy.load for the meaning of the different values.
compress: boolean, or integerWhether to zip the stored data on disk. If an integer is given, it should be between 1 and 9, and sets the amount of compression. Note that compressed arrays cannot be read by memmapping.
verbose: int, optionalThe verbosity flag, controls messages that are issued as the function is evaluated.

Methods in MemorizedFunc class:

Method NameDescription
call(*args, **kwargs)Force the execution of the function with the given arguments and persist the output values.
clear(warn=True)Empty the function’s cache.

Leave a Comment

Your email address will not be published.