Can Python pickle lambda functions?
Solution 1
Yes, python can pickle lambda functions… but only if you have something that uses copy_reg
to register how to pickle lambda functions -- the package dill
loads the copy_reg
you need into the pickle registry for you, when you import dill
.
Python 2.7.8 (default, Jul 13 2014, 02:29:54)
[GCC 4.2.1 Compatible Apple Clang 4.1 ((tags/Apple/clang-421.11.66))] on darwin
Type "help", "copyright", "credits" or "license" for more information.
>>>
>>> import dill # the code below will fail without this line
>>>
>>> import pickle
>>> s = pickle.dumps(lambda x, y: x+y)
>>> f = pickle.loads(s)
>>> assert f(3,4) == 7
>>> f
<function <lambda> at 0x10aebdaa0>
get dill here: https://github.com/uqfoundation
Solution 2
Python can pickle lambdas. We will cover Python 2 and 3 separately as implementation of pickle are different in different Python versions.
- Python 3.6
In Python 3, there is no module named cPickle
. We have pickle
instead which also doesn't support pickling of lambda
functions by default. Let's see it's dispatch table:
>> import pickle
>> pickle.Pickler.dispatch_table
<member 'dispatch_table' of '_pickle.Pickler' objects>
Wait. I tried looking up dispatch_table of pickle
not _pickle
. _pickle
is the alternative and faster C implementation of pickle. But we haven't imported it yet! This C implementation is imported automatically, if it is available, at the end of pure Python pickle
module.
# Use the faster _pickle if possible
try:
from _pickle import (
PickleError,
PicklingError,
UnpicklingError,
Pickler,
Unpickler,
dump,
dumps,
load,
loads
)
except ImportError:
Pickler, Unpickler = _Pickler, _Unpickler
dump, dumps, load, loads = _dump, _dumps, _load, _loads
We are still left with the question of pickling lambdas in Python 3. The answer is you CAN'T with the native pickle
or _pickle
. You will need to import dill
or cloudpickle and use that instead of the native pickle module.
>> import dill
>> dill.loads(dill.dumps(lambda x:x))
<function __main__.<lambda>>
- Python 2.7
pickle
uses pickle registry which is nothing but a mapping from type
to the function to use for serializing (pickling) objects of that type.
You can see pickle registry as:
>> pickle.Pickler.dispatch
{bool: <function pickle.save_bool>,
instance: <function pickle.save_inst>,
classobj: <function pickle.save_global>,
float: <function pickle.save_float>,
function: <function pickle.save_global>,
int: <function pickle.save_int>,
list: <function pickle.save_list>,
long: <function pickle.save_long>,
dict: <function pickle.save_dict>,
builtin_function_or_method: <function pickle.save_global>,
NoneType: <function pickle.save_none>,
str: <function pickle.save_string>,
tuple: <function pickle.save_tuple>,
type: <function pickle.save_global>,
unicode: <function pickle.save_unicode>}
To pickle custom types, Python provides copy_reg
module to register our functions. You can read more about it here. By default, copy_reg
module supports pickling of the following additional types:
>> import copy_reg
>> copy_reg.dispatch_table
{code: <function ipykernel.codeutil.reduce_code>,
complex: <function copy_reg.pickle_complex>,
_sre.SRE_Pattern: <function re._pickle>,
posix.statvfs_result: <function os._pickle_statvfs_result>,
posix.stat_result: <function os._pickle_stat_result>}
Now, type of lambda
functions is types.FunctionType
. However, the builtin function for this type function: <function pickle.save_global>
is not able to serialize lambda functions. Therefore, all third party libraries like dill
, cloudpickle
, etc override the inbuilt method to serialize lambda functions with some additional logic. Let's import dill
and see what it does.
>> import dill
>> pickle.Pickler.dispatch
{_pyio.BufferedReader: <function dill.dill.save_file>,
_pyio.TextIOWrapper: <function dill.dill.save_file>,
_pyio.BufferedWriter: <function dill.dill.save_file>,
_pyio.BufferedRandom: <function dill.dill.save_file>,
functools.partial: <function dill.dill.save_functor>,
operator.attrgetter: <function dill.dill.save_attrgetter>,
operator.itemgetter: <function dill.dill.save_itemgetter>,
cStringIO.StringI: <function dill.dill.save_stringi>,
cStringIO.StringO: <function dill.dill.save_stringo>,
bool: <function pickle.save_bool>,
cell: <function dill.dill.save_cell>,
instancemethod: <function dill.dill.save_instancemethod0>,
instance: <function pickle.save_inst>,
classobj: <function dill.dill.save_classobj>,
code: <function dill.dill.save_code>,
property: <function dill.dill.save_property>,
method-wrapper: <function dill.dill.save_instancemethod>,
dictproxy: <function dill.dill.save_dictproxy>,
wrapper_descriptor: <function dill.dill.save_wrapper_descriptor>,
getset_descriptor: <function dill.dill.save_wrapper_descriptor>,
member_descriptor: <function dill.dill.save_wrapper_descriptor>,
method_descriptor: <function dill.dill.save_wrapper_descriptor>,
file: <function dill.dill.save_file>,
float: <function pickle.save_float>,
staticmethod: <function dill.dill.save_classmethod>,
classmethod: <function dill.dill.save_classmethod>,
function: <function dill.dill.save_function>,
int: <function pickle.save_int>,
list: <function pickle.save_list>,
long: <function pickle.save_long>,
dict: <function dill.dill.save_module_dict>,
builtin_function_or_method: <function dill.dill.save_builtin_method>,
module: <function dill.dill.save_module>,
NotImplementedType: <function dill.dill.save_singleton>,
NoneType: <function pickle.save_none>,
xrange: <function dill.dill.save_singleton>,
slice: <function dill.dill.save_slice>,
ellipsis: <function dill.dill.save_singleton>,
str: <function pickle.save_string>,
tuple: <function pickle.save_tuple>,
super: <function dill.dill.save_functor>,
type: <function dill.dill.save_type>,
weakcallableproxy: <function dill.dill.save_weakproxy>,
weakproxy: <function dill.dill.save_weakproxy>,
weakref: <function dill.dill.save_weakref>,
unicode: <function pickle.save_unicode>,
thread.lock: <function dill.dill.save_lock>}
Now, let's try to pickle lambda function.
>> pickle.loads(pickle.dumps(lambda x:x))
<function __main__.<lambda>>
It WORKS!!
In Python 2 we have two versions of pickle
-
import pickle # pure Python version
pickle.__file__ # <install directory>/python-2.7/lib64/python2.7/pickle.py
import cPickle # C extension
cPickle.__file__ # <install directory>/python-2.7/lib64/python2.7/lib-dynload/cPickle.so
Now, let's try to pickle lambda with C implementation cPickle
.
>> import cPickle
>> cPickle.loads(cPickle.dumps(lambda x:x))
TypeError: can't pickle function objects
What went wrong? Let's see the dispatch table of cPickle
.
>> cPickle.Pickler.dispatch_table
AttributeError: 'builtin_function_or_method' object has no attribute 'dispatch_table'
The implementation of pickle
and cPickle
is different. Importing
dill makes only Python version of pickle
work. The disadvantage of using pickle
instead of cPickle
is that it can be as much as 1000 times slower than cPickle.
I hope this clears all the doubts.
Solution 3
No, Python can't pickle lambda functions:
>>> import cPickle as pickle
>>> s = pickle.dumps(lambda x,y: x+y)
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
File "/System/Library/Frameworks/Python.framework/Versions/2.7/lib/python2.7/copy_reg.py", line 70, in _reduce_ex
raise TypeError, "can't pickle %s objects" % base.__name__
TypeError: can't pickle function objects
Not sure what you did that succeeded...
Solution 4
Even though it might be obvious I would like to add an other possible solution. As you probably know lambda functions are just anonymous function declarations. If you don't have many lambdas that are used only once and it wouldn't add much noise to your code you could just name your lambda and pass the name of it (without the parentheses) like this:
import cPickle as pickle
def addition(x, y):
return x+y
if __name__ == "__main__":
s = pickle.dumps(addition)
f = pickle.loads(s)
assert f(3,4) == 7
The name also adds more semantic and you wouldn't need an additional dependency like Dill. But only do that if that outweighs the added noise of the additional function(s).
Solution 5
what worked for me (windows 10, python 3.7) was to pass a function instead of a lambda function:
def merge(x):
return Image.merge("RGB", x.split()[::-1])
transforms.Lambda(merge)
instead of:
transforms.Lambda(lambda x: Image.merge("RGB", x.split()[::-1]))
no dill or cPickel needed.
Lars
Updated on July 05, 2022Comments
-
Lars almost 2 years
I have read in a number of threads that Python
pickle
/cPickle
cannot pickle lambda functions. However the following code works, using Python 2.7.6:import cPickle as pickle if __name__ == "__main__": s = pickle.dumps(lambda x, y: x+y) f = pickle.loads(s) assert f(3,4) == 7
So what is going on? Or, rather, what is the limit of pickling lambdas?
[EDIT] I think i know why this code runs. I forgot (sorry!) i am running stackless python, which has a form of micro-threads called tasklets executing a function. These tasklets can be halted, pickled, unpickled and continued, so i guess (asked on the stackless mailing list) that it also provides a way to pickle function bodies.
-
Ramast about 8 yearsI've tried on python3 In [1]: import dill In [2]: import pickle In [3]: pickle.dumps(lambda x: (x+1, x+2)) --------------------------------------------------------------------------- PicklingError Traceback (most recent call last) <ipython-input-3-924e2f4cc7e0> in <module>() ----> 1 pickle.dumps(lambda x: (x+1, x+2)) PicklingError: Can't pickle <function <lambda> at 0x7f08ee40ca60>: attribute lookup <lambda> on main failed. It only works if you import dill as pickle
-
Ramast about 8 yearsI don't know why this comment is down voted. pickle can't serialize lambdas only dill package can
-
Mike McKerns about 8 years@Ramast: you are correct -- in python3, you currently have to
import dill as pickle
. In python2, what I have above works either way you do it. -
Mike McKerns over 7 years@CharlieParker: Can you elaborate? The above should work for most "arbitrary" functions.
-
naught101 over 6 yearsWhy can't python pickle lambdas?
-
Pavel Vlasov over 5 yearsThe answer is incorrect for Python 3.6 as it is - use
dill.dumps()
anddill.loads()
instead. -
Mike McKerns over 5 years@PavelVlasov: your comment is a duplicate of the initial comment on the answer.
-
Ciprian Tomoiagă almost 5 years@MikeMcKerns can you please expand on the differences between Py3 and Py2 version of
dill
? Isdill
not able to hijack thepickle
module anymore? I was hoping that importingdill
beforerq
(which internally usespickle
), I can makerq
work with lambdas -
Mike McKerns almost 5 years@CiprianTomoiagă: The primary difference is that in python 2, there was
pickle
(python) andcPickle
(C) -- and in python 3, it's nowpickle
(python) and_pickle
(C)... however,pickle.dump
(anddumps
) uses_pickle
(i.e. C)... anddill
can currently only inject new methods into the python pickle registry. So, just importingdill
doesn't work like it does in python 2. Note that there ispickle._dump
(and_dumps
), which does use the python registry... so that works as in python 2. Unfortunately most packages don't fall back to_dump
, whendump
fails. -
EliadL almost 5 yearsThis answer should be accepted. It nicely explains the usability and limitations with the
dill
package in each python version. Well done! -
Charlie Parker about 4 years@MikeMcKerns does it work if the code that has arbitrary pytorch (and data science) code? I am trying to pickle a pytorch neural net work pointer to lambda functions.
-
Charlie Parker about 4 yearslike in this case: stackoverflow.com/questions/61510810/…
-
Mike McKerns about 4 yearsit should. I answered your question directly at the above link