Keyword arguments - strange behaviour?

[brian.bird]

Can anyone explain the behaviour of python when running this script?

def method(n, bits=[]):

.... bits.append(n)
.... print bits
....

method(1) [1]
method(2)

[1, 2]

It's the same in python 1.5, 2.3 and 2.4 so it's not a bug. But I
expected the variable "bits" to be re-initialised to an empty list as
each method was called. Whenever I explain optional keyword arguments
to someone I have usually (wrongly as it turns out) said it is
equivalent to:
.... if bits is None:
.... bits=[]
.... bits.append(n)
.... print bits
....

method(1) [1]
method(2)

[2]

Is there a good reason why these scripts are not the same? I can
understand how/why they are different, it's just not what I expected.
(It seems strange to me that the result of the first method can only be
determined if you know how many times it has already been called)

Is this behaviour what you would (should?) intuitively expect?
Thanks,

Brian

 

[deelan]

Can anyone explain the behaviour of python when running this script? (...)

Is there a good reason why these scripts are not the same? I can
understand how/why they are different, it's just not what I expected.
(It seems strange to me that the result of the first method can only be
determined if you know how many times it has already been called)

see "python gotchas":
<http://www.ferg.org/projects/python_gotchas.html#bct_sec_5>

bye.

 

[Fredrik Lundh]

Can anyone explain the behaviour of python when running this script?

the language reference has the full story:

http://www.python.org/doc/2.4/ref/function.html

"Default parameter values are evaluated when the function
definition is executed"

</F>

 

[Duncan Booth]

wrote:

Can anyone explain the behaviour of python when running this script?

def method(n, bits=[]):

... bits.append(n)
... print bits
...

method(1) [1]
method(2)

[1, 2]

It's the same in python 1.5, 2.3 and 2.4 so it's not a bug. But I
expected the variable "bits" to be re-initialised to an empty list as
each method was called. Whenever I explain optional keyword arguments
to someone I have usually (wrongly as it turns out) said it is
equivalent to:

<snipped erroneous comparison>

No, it is closer to:

# Assume you have done this earlier:
import new
def _realmethod(n, bits):
bits.append(n)
print bits

# The def is roughly equivalent to this:
_defaultArgs = ([], )
method = new.function(_realmethod.func_code, globals(), 'method',
_defaultArgs)

Each time you re-execute the def you get a new set of default arguments
evaluated, but the result of the evaluation is simply a value passed in to
the function constructor.

The code object is compiled earlier then def creates a new function object
from the code object, the global dictionary, the function name, and the
default arguments (and any closure as well, but its a bit harder to
illustrate that this way).

 

[brian.bird]

Thanks, this makes perfect sense. The phrase which sums it up neatly is
"Default parameter values are evaluated when the function definition is
executed"

However, is there a good reason why default parameters aren't evaluated
as the function is called? (apart from efficiency and backwards
compatibility)? Is this something that's likely to stay the same in
python3.0?

I'm really looking for a neat way to do the following:

def method(a,b,opt1=None,opt2=None,opt3="",opt4=None):
if opt1 is None: opt1=[]
if opt2 is None: opt2={}
if opt4 is None: opt4=[]

Python syntax is normally so neat but this just looks a mess if there
are lots of parameters.

 

[Fredrik Lundh]

However, is there a good reason why default parameters aren't evaluated
as the function is called? (apart from efficiency and backwards compatibility)?

how would you handle this case:

def function(arg=otherfunction(value)):
return arg

</F>

 

[brian.bird]

def function(arg=otherfunction(value)):
return arg

My expectation would have been that otherfunction(value) would be
called if (and only if) the arg keyword parameter was missing from the
function() call (ie. the optional value is evaluated the lazy way).
Also, otherfunction would be called each and every time this function()
is called without the arg keyword. (At least, I would have assumed this
before today)

Still, I can see why it's been implemented the way it has, it just
seems a shame there isn't a neat shortcut to default lots of optional
arguments to new mutable objects. And since I'm not the only one to
fall into this trap it makes me wonder why the default behaviour isn't
made to be what most people seem to expect?

 

[Fredrik Lundh]

def function(arg=otherfunction(value)):
return arg

My expectation would have been that otherfunction(value) would be
called if (and only if) the arg keyword parameter was missing from the
function() call (ie. the optional value is evaluated the lazy way).

what otherfunction? what value?

</F>

 

[harold fellermann]

Hi,

I cannot see any strange behavior. this code works exacly as you and I
suspect:
.... return x
........ return arg
....5

Or is this not what you excepted?

- harold -

def function(arg=otherfunction(value)):
return arg

My expectation would have been that otherfunction(value) would be
called if (and only if) the arg keyword parameter was missing from the
function() call (ie. the optional value is evaluated the lazy way).
Also, otherfunction would be called each and every time this function()
is called without the arg keyword. (At least, I would have assumed this
before today)

Still, I can see why it's been implemented the way it has, it just
seems a shame there isn't a neat shortcut to default lots of optional
arguments to new mutable objects. And since I'm not the only one to
fall into this trap it makes me wonder why the default behaviour isn't
made to be what most people seem to expect?

--
Freunde, nur Mut,
Lächelt und sprecht:
Die Menschen sind gut --
Bloß die Leute sind schlecht.
-- Erich Kästner

 

[Steve Holden]

Hi,

I cannot see any strange behavior. this code works exacly as you and I
suspect:

.... return x
....
.... return arg
....
5

Or is this not what you excepted?

Channelling the effbot, I think he was asking what namespace context you
expected the expression "arg=otherfunction(x)" to be evaluated in when
it's used at the time of a function call to dynamically create a new
default value for arg.

regards
Steve

 

[Fredrik Lundh]

I cannot see any strange behavior. this code works exacly as you and I suspect:

you seem to have missed some of the posts that led up to the one you
replied to. (most importantly, the first one).

</F>

 

[Steven Bethard]

However, is there a good reason why default parameters aren't evaluated
as the function is called? (apart from efficiency and backwards
compatibility)?

So, one of my really common use cases that takes advantage of the fact
that default parameters are evaluated at function definition time:

def foo(bar, baz, matcher=re.compile(r'...')):
...
text = matcher.sub(r'...', text)
...

If default parameters were evaluated when the function was called, my
regular expression would get re-compiled every time foo() was called.
This would be inefficient, especially if foo() got called a lot. If
Python 3000 changed the evaluation time of default parameters, I could
rewrite this code as:

class foo(object):
matcher=re.compile(r'...')
def __new__(self, bar, baz, matcher=None):
if matcher is None:
matcher = self.matcher
...
text = matcher.sub(r'...', text)
...

But that seems like a lot of work to do something that used to be pretty
simple...

Steve

 

[brian.bird]

Channelling the effbot, I think he was asking what namespace context
you

expected the expression "arg=otherfunction(x)" to be evaluated in when
it's used at the time of a function call to dynamically create a new
default value for arg.

Thanks, I realise now that's what was meant. I think I was just
assuming otherfunction() would be globally available - but that's not
something I want to go into since I can see now there will be problems
trying to define what python should do if it evaluated defaults at the
function call

I think I can now explain to someone why there are good reasons that
the default params are evaluated at the definition. However, I still
can't give a nice looking solution on how to re-write a function to
have empty mutable values as default arguments: eg.

def method(a,b,opt1=[],opt2=None,opt3="",opt4={})

How could I re-write this (especially if there are perhaps 20 optional
parameters,any number of which may have mutable defaults) without
writing 20 "if opt1 is None: opt1=[]" statements?

Brian

 

[Fuzzyman]

Channelling the effbot, I think he was asking what namespace

context

you
expected the expression "arg=otherfunction(x)" to be evaluated in when
it's used at the time of a function call to dynamically create a new
default value for arg.

Thanks, I realise now that's what was meant. I think I was just
assuming otherfunction() would be globally available - but that's not
something I want to go into since I can see now there will be problems
trying to define what python should do if it evaluated defaults at the
function call

I think I can now explain to someone why there are good reasons that
the default params are evaluated at the definition. However, I still
can't give a nice looking solution on how to re-write a function to
have empty mutable values as default arguments: eg.

def method(a,b,opt1=[],opt2=None,opt3="",opt4={})

How could I re-write this (especially if there are perhaps 20 optional
parameters,any number of which may have mutable defaults) without
writing 20 "if opt1 is None: opt1=[]" statements?

Brian

One way that is *slightly neater* is to simply collect all keyword
arguments as a dictionary. Then compare with a dictionary of defaults
for any missing keywords.

def afunction(**keywargs):
defaults = {'param1' : [], 'param2' : {}, 'param3' : []}
for entry in defaults:
if not keywargs.has_key(entry):
keywargs[entry] = defaults[entry]

This keeps all your defaults in a single dictionary and avoids a really
long function definition.
Regards,

Fuzzy
http://www.voidspace.org.uk/python/index.shtml

 

[Fuzzyman]

Channelling the effbot, I think he was asking what namespace

context

you
expected the expression "arg=otherfunction(x)" to be evaluated in when
it's used at the time of a function call to dynamically create a new
default value for arg.

Thanks, I realise now that's what was meant. I think I was just
assuming otherfunction() would be globally available - but that's not
something I want to go into since I can see now there will be problems
trying to define what python should do if it evaluated defaults at the
function call

I think I can now explain to someone why there are good reasons that
the default params are evaluated at the definition. However, I still
can't give a nice looking solution on how to re-write a function to
have empty mutable values as default arguments: eg.

def method(a,b,opt1=[],opt2=None,opt3="",opt4={})

How could I re-write this (especially if there are perhaps 20 optional
parameters,any number of which may have mutable defaults) without
writing 20 "if opt1 is None: opt1=[]" statements?

Brian

One way that is *slightly neater* is to simply collect all keyword
arguments as a dictionary. Then compare with a dictionary of defaults
for any missing keywords.

def afunction(**keywargs):
defaults = {'param1' : [], 'param2' : {}, 'param3' : []}
for entry in defaults:
if not keywargs.has_key(entry):
keywargs[entry] = defaults[entry]

This keeps all your defaults in a single dictionary and avoids a really
long function definition.
Regards,

Fuzzy
http://www.voidspace.org.uk/python/index.shtml

 

[Fuzzyman]

So, one of my really common use cases that takes advantage of the fact
that default parameters are evaluated at function definition time:

def foo(bar, baz, matcher=re.compile(r'...')):
...
text = matcher.sub(r'...', text)
...

Surely "re.compile(r'...')" is effectively a constant ? So your above
code is equivalent to :

aConst = re.compile(r'...')
def foo(bar, baz, matcher=aConst):
....
text = matcher.sub(r'...', text)
....

I agree that dynamic evaluation of default arguments seems much more
pythonic, as well as getting rid of a common python gotcha.

Regards,

Fuzzy
http://www.voidspace.org.uk/python/index.shmtl

If default parameters were evaluated when the function was called, my

 

[Steven Bethard]

Surely "re.compile(r'...')" is effectively a constant ? So your above
code is equivalent to :

aConst = re.compile(r'...')
def foo(bar, baz, matcher=aConst):
...
text = matcher.sub(r'...', text)
...

Basically, yes. Though you add an extra name to the module or class
namespace by defining 'aConst'. I consider this a minor pollution of
the namespace since 'matcher' is only used in the function, not the
module or class.

But if we're arguing for equivalencies, the oft-asked for

def foo(lst=[]):
...

where [] is evaluated for each function call, is equivalent for most
purposes to:

def foo(lst=None):
if lst is None:
lst = []

There's a minor pollution of the range of values 'lst' can take on -- if
you need lst to be able to take on the value None, you'll want to
rewrite this something like:

no_value = object()
def foo(lst=no_value):
if lst is no_value:
lst = []

My point here is that there's always going to be some equivalent
expression (turing completeness etc.) I was just pointing out a quite
reasonable use of the current default parameter evaluation system.

Steve

 

[Fuzzyman]

Basically, yes. Though you add an extra name to the module or class
namespace by defining 'aConst'. I consider this a minor pollution of

the namespace since 'matcher' is only used in the function, not the
module or class.

But if we're arguing for equivalencies, the oft-asked for

def foo(lst=[]):
...

where [] is evaluated for each function call, is equivalent for most
purposes to:

def foo(lst=None):
if lst is None:
lst = []

There's a minor pollution of the range of values 'lst' can take on -- if
you need lst to be able to take on the value None, you'll want to
rewrite this something like:

no_value = object()
def foo(lst=no_value):
if lst is no_value:
lst = []

My point here is that there's always going to be some equivalent
expression (turing completeness etc.) I was just pointing out a quite
reasonable use of the current default parameter evaluation system.

Steve

Sure.. but you also gave an example of an alternative that was complex,
and used it's complexity as an argument against having default
arguments dynamically evaluated. What I was saying is that there is an
alternative that is at least as readable (if not more) than your
example.

Having default arguments evaluated when the function is defined is
unintuitive and unpythonic.
Regards,


Fuzzy
http://www.voidspace.org.uk/python/index.shtml

 

[Steve Holden]

Fuzzyman wrote:

[...]

Having default arguments evaluated when the function is defined is
unintuitive and unpythonic.

With due respect that's a matter of opinion.

Having default arguments evaluated when the function is defined is the
result of several design decisions in Python, not the least of which is
the decision to have "def" be an executable statement. Ever wondered how
come you can def the same function name several times in the same
module? It's simply because on of def's side-effects is the binding of
its name in the current namespace.

Unfortunately there's absolutely no guarantee that the namespace context
when a function is called will bear any relation to the namespace
context when it was defined (particularly given that definition and call
can take place in entirely separate modules). So, how would one specify
a "deferred expression" to be evaluated when the function is called
rather than when it is defined? There presumably has to be some
technique that's different from the current ones, since presumably you
would also want to retain the option of having the default evaluated at
definition time.

I still haven't seen any reasonable suggestions as to how one might
exactly specify the execution context for such a deferred expression,
let alone how to spell the deferred expression itself. Would you have us
provide a code object that can be eval()'d?

regards
Steve

 

[Steven Bethard]

Sure.. but you also gave an example of an alternative that was complex,

Interesting. I would have thought that my example was pretty simple.
Maybe it would be helpful to generalize it to:

def foo(bar, baz, spam=badger(x, y, z)):
...

All it does is use a default value that was produced by a function call.
I'm surprised you haven't run into this situation before...

Of course, what is complex or simple is a matter of personal opinion. I
use this pattern so often that it's quite simple to me, but I guess I
can understand that if you don't use such a pattern, it might seem
foreign to you.

Steve