__autoinit__ (Was: Proposal: reducing self.x=x; self.y=y;self.z=z boilerplate code)

[Ralf W. Grosse-Kunstleve]

My initial proposal
(http://cci.lbl.gov/~rwgk/python/adopt_init_args_2005_07_02.html) didn't
exactly get a warm welcome...

And Now for Something Completely Different:

class autoinit(object):

def __init__(self, *args, **keyword_args):
self.__dict__.update(
zip(self.__autoinit__.im_func.func_code.co_varnames[1:], args))
self.__dict__.update(keyword_args)
self.__autoinit__(*args, **keyword_args)

class grouping(autoinit):

def __autoinit__(self, x, y, z):
print self.x, self.y, self.z

group = grouping(1,2,z=3)
group = grouping(z=1,x=2,y=3)
try: grouping(1)
except TypeError, e: print e
try: grouping(1,2,3,a=0)
except TypeError, e: print e


Almost like my original favorite solution, only better, and it doesn't require
a syntax change.

Under a hypothetical new proposal __autoinit__ would become a standard feature
of object.

Any takers?

Cheers,
Ralf


__________________________________________________
Do You Yahoo!?
Tired of spam? Yahoo! Mail has the best spam protection around
http://mail.yahoo.com

 

[Kay Schluehr]

My initial proposal
(http://cci.lbl.gov/~rwgk/python/adopt_init_args_2005_07_02.html) didn't
exactly get a warm welcome...

Well ... yes

Ralf, if you want to modify the class instantiation behaviour you
should have a look on metaclasses. That's what they are for. It is not
a particular good idea to integrate a new method into the object base
class for each accidental idea and write a PEP for it.

I provide you an example which is actually your use case. It doesn't
change the class hierarchy : the metaclass semantics is not "is a" as
for inheritance but "customizes" as one would expect also for
decorators.

class autoattr(type):
'''
The autoattr metaclass is used to extract auto_xxx parameters from
the argument-tuple or the keyword arguments of an object
constructor __init__
and create object attributes mit name xxx and the value of auto_xxx
passed
into __init__
'''
def __init__(cls,name, bases, dct):
super(autoattr,cls).__init__(name,bases,dct)
old_init = cls.__init__
defaults = cls.__init__.im_func.func_defaults
varnames = cls.__init__.im_func.func_code.co_varnames[1:]

def new_init(self,*args,**kwd):
for var,default in zip(varnames[-len(defaults):],defaults):
if var.startswith("auto_"):
self.__dict__[var[5:]] = default
for var,arg in zip(varnames,args):
if var.startswith("auto_"):
self.__dict__[var[5:]] = arg
for (key,val) in kwd.items():
if key.startswith("auto_"):
self.__dict__[key[5:]] = val
old_init(self,*args,**kwd)
cls.__init__ = new_init

class app:
__metaclass__ = autoattr
def __init__(self, auto_x, y, auto_z = 9):
pass
-> AttributeError


Kay

 

[Scott David Daniels]

My initial proposal
(http://cci.lbl.gov/~rwgk/python/adopt_init_args_2005_07_02.html) didn't
exactly get a warm welcome...

And Now for Something Completely Different:

class autoinit(object):

def __init__(self, *args, **keyword_args):
self.__dict__.update(
zip(self.__autoinit__.im_func.func_code.co_varnames[1:], args))
self.__dict__.update(keyword_args)
self.__autoinit__(*args, **keyword_args)

Should be:
class autoinit(object):
def __init__(self, *args, **keyword_args):
for name, value in zip(self.__autoinit__.im_func.func_code.
co_varnames[1:], args):
setattr(self, name, value)
for name, value in keyword_args.items():
setattr(self, name, value)
self.__autoinit__(*args, **keyword_args)

Since using setattr will take care of any slots used in other classes.
Not all data is stored in the __dict__.

For example:

class Example(autoinit):
__slots__ = 'abc',
def __autoinit__(self, a=1, abc=1):
print a, abc

a = Example(1,2)
print a.__dict__
print a.a
print a.abc


--Scott David Daniels
(e-mail address removed)

 

[Kay Schluehr]

I stripped your code down to the essence. See attachment.

For the user your approach then becomes:

class grouping:
__metaclass__ = autoattr
def __init__(self, x, y, z):
pass

No. This is clearly NOT what I had in mind. I translated your original
proposal which introduced a punctuation syntax '.x' for constructor
parameters forcing the interpreter to create equally named object
attributes into a naming convention that can be handled by a metaclass
customizer. The grouping.__init__ above does exacly nothing according
to my implementation. I would never accept dropping fine-tuning
capabilities. The "auto_" prefix is all the declarative magic.

My __autoinit__ suggestion would result in (assuming object supports
this by default):

class grouping(object):
def __autoinit__(self, x, y, z):
pass

I think that's far more intuitive.

Being intuitive is relative to someones intuition.

Kay

 

[Terry Reedy]

I have a suggestion I don't remember seeing for flagging which vars to
autoinit without new syntax: use '_' instead of '.'. I have never seen
local vars with a leading '_'. So, in combination with whatever other
mechanism (metaclass, __init__ decorator?)

def __init__(self, _x, y, _z) :

would automatically do self.x = _x, self.z = _z, but not self.y = y.

Terry J. Reedy

 

[Bengt Richter]

I have a suggestion I don't remember seeing for flagging which vars to
autoinit without new syntax: use '_' instead of '.'. I have never seen
local vars with a leading '_'. So, in combination with whatever other
mechanism (metaclass, __init__ decorator?)

def __init__(self, _x, y, _z) :

would automatically do self.x = _x, self.z = _z, but not self.y = y.

Of course, if it works for __init__, it should work for an arbitrary
method or even function (if a decorator is used).

I think I could write a byte-code-modifying decorator that would insert
code at the beginning of a function to do the
self.x = _x; self.z = _z # but not self.y = y
in your example, and similarly for others.
But I am not sure it's not a waste of time to do these byte-code things
except to see how using such decorators feels in practice, or whether any
byte-code munging after generation is a good idea.

What would be the next step? To implement the same in C as a built in decorator?
I'm wondering if the effort of munging byte code isn't misplaced, and would create
a growing maintenance problem as more built-in decorators like that accumulated.

So, if we want alternate code generated, perhaps we need a way of talking to the
original code generation process? I'm thinking some kind of source-processing-time
decorator (maybe marked with double-@, like
@@auto_init_self_attributes
def __init__(self, _x, y, _z) : print 'self.y=%r not set'%y

The idea is for @@deco to be intercepted as part of the AST of the unit being compiled,
and be converted at that time to call with the AST and its own location in the tree being
passed to it so it can do mods to the AST (and cut itself out), and then let the resulting
AST be compiled as normal. Multiple @@decorators would each get their chance to mod the AST
before compilation. This still would have some version dependency, but it wouldn't be
byte code hacking ;-) One of these days I'll think some more on @@. As an experiment, I think
the tokenizer could translate @@deco to _AT_AT_deco and let the AST be formed with that
apparent function call at that place in the code. Then the AST could be walked to find the _AT_AT_<deconame>
calls and extract them and compile them and execute them (the _AT_AT_deco functions would
obviously have to be defined already, so they might be looked up in some standard place
like trying to do place = __import__(_AT_AT_MODULE_) wher _AT_AT_MODULE_ gets defined sort
of like __METACLASS__.), passing the AST and the _AT_AT_deco call location therein, and the rest
of the parameters.

AST decoration would introduce macro-like capabilities, but restricted to transforming valid ASTs,
so it isn't like source text rewriting macros.

Sorry for the run-on stream of consciousness ;-/

Regards,
Bengt Richter

 

[Dan Sommers]

news:[email protected]...

I have a suggestion I don't remember seeing for flagging which vars to
autoinit without new syntax: use '_' instead of '.'. I have never seen
local vars with a leading '_'. So, in combination with whatever other
mechanism (metaclass, __init__ decorator?)

def __init__(self, _x, y, _z) :

would automatically do self.x = _x, self.z = _z, but not self.y = y.

Terry J. Reedy

That's a pretty big change; now all formal parameters beginning with an
underscore have a brand new meaning.

How about this:

def __init__(self, self.x, y, self.z):
# self.x, self.z from first and third explicit parameters
do_something_with_y()

where "self" in "self.x" and "self.y" would have to match the first
parameter (so that the pathological among us could write this:

def __init__(this, this.x, y, this.z):
do_something_with_y()

instead).

(Sorry if this posts multiple times; gnus and/or my news server were not
happy when I was composing this reply.)

Regards,
Dan

 

[Paul Rubin]

def __init__(self, self.x, y, self.z):
# self.x, self.z from first and third explicit parameters
do_something_with_y()

Hey, I like that.

 

[Terry Reedy]

That's a pretty big change; now all formal parameters beginning with an
underscore have a brand new meaning.

As I said, 'in combination with whatever other mechanism', meaning one that
one has to intentionally invoke. So there would be no code breaking: the
suggestion is for a convention used with currently available mechanism
(assuming that such is possible) that would give the fine-grained control
not so easily possible with the current update_dict_with_locals idiom.

It would be easier to write the decorator if it were passed the (quoted)
names of the parameters to be 'attributed'. But then the user would have
to write and keep in synchrony two lists, one a (quoted) subset of the
other. So I thought it better to have the decorator writer and decorator
function work a little harder to introspect and interpret one list with
some items marked somehow in a currently legal but unusual manner.

How about this:

def __init__(self, self.x, y, self.z):
# self.x, self.z from first and third explicit parameters

This is new syntax that is not currently legal. My suggestion was for a
solution that avoided that difficulty and that could, I believe, be
implemented now, in less time that this thread has been going on, rather
than maybe some years from now.

Terry J. Reedy

 

[Dan Sommers]

As I said, 'in combination with whatever other mechanism', meaning one
that one has to intentionally invoke ...

My mistake; I didn't read carefully enough.

Regards,
Dan

 

[Bengt Richter]

Hey, I like that.

Me too. I liked the leading _, but on second thought it is a weird language change
re names in a special context. Probably not so good.

OTOH, I see in the above a generalizable principle of formal parameter expressions as
automatic assignment targets at the beginning of function execution that could allow
anything normally allowed locally on the left of an assignment. This is contrived,
but illustrates:

def setem(key, adict[key]): pass # note that assignment is left-right,
# so key is available for adict[key]
# and adict is global here
adict = {}
setem('k', 'value')
adict -> {'k':'value'}

Regards,
Bengt Richter

 

[Kay Schluehr]

How about this:

def __init__(self, self.x, y, self.z):
# self.x, self.z from first and third explicit parameters
do_something_with_y()

Can you tell me in which way it is anyhow better than the original
proposal

def __init__(self, .x, y, .z):
# self.x, self.z from first and third explicit parameters
do_something_with_y()

besides that it is more verbose?

Kay

 

[Terry Reedy]

Me too. I liked the leading _, but on second thought it is a weird
language change
re names in a special context. Probably not so good.

To repeat: while my '_' proposal could have been for a language change (in
3.0), it was actually for a convention recognized by a metaclass or, more
likely, decorator (and implementable now, I think). *Any* flag would serve
the purpose, but I picked one that was a single char while being visually
striking and, as far as I know, rarely used in current practice even though
quite legal. A 'self_' prefix would do as well except for being more to
type. The prefix could even be an argument to the decorator!

Terry J. Reedy

 

[Reinhold Birkenfeld]

Can you tell me in which way it is anyhow better than the original
proposal

def __init__(self, .x, y, .z):
# self.x, self.z from first and third explicit parameters
do_something_with_y()

besides that it is more verbose?

It is more explicit. Explicit is better than implicit.

But as with many proposals, this raises consequential questions, for
example, how "self.x" parameters are handled in other methods, or even
functions, as __init__ is not special-cased by the parser.

Reinhold

 

[Kay Schluehr]

It is more explicit. Explicit is better than implicit.

The punctuation syntax makes it explicit too. But maybe a point is a
more tiny and less explicit symbol than an @ that makes a decorator
explicit

But as with many proposals, this raises consequential questions, for
example, how "self.x" parameters are handled in other methods, or even
functions, as __init__ is not special-cased by the parser.

Reinhold

Yes. My argument against the syntax is more that of a language lawyer:
how a class uses the argument parameters of a constructor is an
implementation detail of a class and should not be published in the
constructor interface.

One may assign special attributes to the classes ( e.g. tagging it with
a metaclass ) or a qualifier. I had recently a look on Scala an
object-functional language running on top of the JVM. Scala introduces
the concept of a "case class" to represent object trees. All arguments
passed into a case class constructor become automagically object
attributes. This is for convenience and let the tree grow if the passed
arguments are case class instances again. Here it is the class type
that determines how it's construction is handled. I think this is a
reasonable approach.

Kay

 

[Dan Sommers]

It is more explicit. Explicit is better than implicit.

But as with many proposals, this raises consequential questions, for
example, how "self.x" parameters are handled in other methods, or even
functions, as __init__ is not special-cased by the parser.

So why limit it to __init__? Bengt Richter's idea of generalizing it is
a good one. Currently, when this method:

def f(self, x, y, z):
pass

is called, Python binds self to a reference to the object, x to the
first argument, y to the second argument, z to and the third. By
extension, hypothetically, this method:

def new_f(self, self.x, y, self.z):
do_something_with_y()

would be semantically identical to:

def new_f(self, __anonymous_argument_1, y, __anonymous_argument_2):
self.x = __anonymous_argument_1
del __anonymous_argument_1 # look: a use case for del! <wink>
self.z = __anonymous_argument_2
del __anonymous_argument_2
do_something_with_y()

It's not too far from the tuple unpacking that happens now:

def g(x, (a, b)):
pass

q = (3, 4)
g(1, q) # inside g, x = 1, a = q[0] = 3, b = q[1] = 4

and it's certainly not less explicit than properties.

Without thinking it all the way through, I suppose these:

def method_1(self, *self.l):
pass
def method_2(self, **self.d):
pass

could act as if they were these:

def method_1(self, *args):
self.l = args
del args
def method_2(self, **kw):
self.d = kw
del kw

Regards,
Dan

 

[Reinhold Birkenfeld]

Without thinking it all the way through, I suppose these:

def method_1(self, *self.l):
pass
def method_2(self, **self.d):
pass

could act as if they were these:

def method_1(self, *args):
self.l = args
del args
def method_2(self, **kw):
self.d = kw
del kw

I still think it's too specialized. What would, hypothetically, this do?

class Bar: pass

class Foo:
x = Bar()
def method_1(self, x.y):
pass

It's hard to explain that you can autoassign self.y but not x.y.

Reinhold

 

[Bengt Richter]

To repeat: while my '_' proposal could have been for a language change (in
3.0), it was actually for a convention recognized by a metaclass or, more
likely, decorator (and implementable now, I think). *Any* flag would serve
the purpose, but I picked one that was a single char while being visually
striking and, as far as I know, rarely used in current practice even though
quite legal. A 'self_' prefix would do as well except for being more to
type. The prefix could even be an argument to the decorator!

Yes, sorry, I should have read more carefully. Yours is the high ground ;-)

Regards,
Bengt Richter

 

[Bengt Richter]

I still think it's too specialized. What would, hypothetically, this do?

class Bar: pass

class Foo:
x = Bar()
def method_1(self, x.y):
pass

It's hard to explain that you can autoassign self.y but not x.y.

No, that limitation wouldn't exist, so you wouldn't have to explain it ;-)
I.e., the above would act like

class Foo:
x = Bar()
def method_1(self, _anonymous_arg_1):
x.y = _anonymous_arg_1

and would do whatever it would do now (probably look for a global x or a closure cell x, but
it wouldn't find the class variable in a normal method call)

Regards,
Bengt Richter