I've been going through all the Stackoverflow answers on dynamic property setting, but for whatever reason I can't seem to get this to work.
I have a class, Evolution_Base, that in its init creates an instance of Value_Differences. Value_Differences should be dynamically creating properties, based on the list I pass, that returns the function value from _get_df_change:
from pandas import DataFrame
from dataclasses import dataclass
import pandas as pd
class Evolution_Base():
def __init__(self, res_date_0 : DataFrame , res_date_1 : DataFrame):
#dataclass
class Results_Data():
res_date_0_df : DataFrame
res_date_1_df : DataFrame
self.res = Results_Data(res_date_0_df= res_date_0,
res_date_1_df= res_date_1)
property_list = ['abc', 'xyz']
self.difference = Value_Differences(parent = self, property_list=property_list)
# Shared Functions
def _get_df_change(self, df_name, operator = '-'):
df_0 = getattr(self.res.res_date_0_df, df_name.lower())
df_1 = getattr(self.res.res_date_1_df, df_name.lower())
return self._df_change(df_1, df_0, operator=operator)
def _df_change(self, df_1 : pd.DataFrame, df_0 : pd.DataFrame, operator = '-') -> pd.DataFrame:
"""
Returns df_1 <operator | default = -> df_0
"""
# is_numeric mask
m_1 = df_1.select_dtypes('number')
m_0 = df_0.select_dtypes('number')
def label_me(x):
x.columns = ['t_1', 't_0']
return x
if operator == '-':
return label_me(df_1[m_1] - df_0[m_0])
elif operator == '+':
return label_me(df_1[m_1] + df_0[m_0])
class Value_Differences():
def __init__(self, parent : Evolution_Base, property_list = []):
self._parent = parent
for name in property_list:
def func(self, prop_name):
return self._parent._get_df_change(name)
# I've tried the following...
setattr(self, name, property(fget = lambda cls_self: func(cls_self, name)))
setattr(self, name, property(func(self, name)))
setattr(self, name, property(func))
Its driving me nuts... Any help appreciated!
My desired outcome is for:
evolution = Evolution_Base(df_1, df_2)
evolution.difference.abc == evolution._df_change('abc')
evolution.difference.xyz == evolution._df_change('xyz')
EDIT: The simple question is really, how do I setattr for a property function?
As asked
how do I setattr for a property function?
To be usable as a property, the accessor function needs to be wrapped as a property and then assigned as an attribute of the class, not the instance.
That function, meanwhile, needs to have a single unbound parameter - which will be an instance of the class, but is not necessarily the current self. Its logic needs to use the current value of name, but late binding will be an issue because of the desire to create lambdas in a loop.
A clear and simple way to work around this is to define a helper function accepting the Value_Differences instance and the name to use, and then bind the name value eagerly.
Naively:
from functools import partial
def _get_from_parent(name, instance):
return instance._parent._get_df_change(name)
class Value_Differences:
def __init__(self, parent: Evolution_Base, property_list = []):
self._parent = parent
for name in property_list:
setattr(Value_Differences, name, property(
fget = partial(_get_from_parent, name)
))
However, this of course has the issue that every instance of Value_Differences will set properties on the class, thus modifying what properties are available for each other instance. Further, in the case where there are many instances that should have the same properties, the setup work will be repeated at each instance creation.
The apparent goal
It seems that what is really sought, is the ability to create classes dynamically, such that a list of property names is provided and a corresponding class pops into existence, with code filled in for the properties implementing a certain logic.
There are multiple approaches to this.
Factory A: Adding properties to an instantiated template
Just like how functions can be nested within each other and the inner function will be an object that can be modified and returned (as is common when creating a decorator), a class body can appear within a function and a new class object (with the same name) is created every time the function runs. (The code in the OP already does this, for the Results_Data dataclass.)
def example():
class Template:
pass
return Template
>>> TemplateA, TemplateB = example(), example()
>>> TemplateA is TemplateB
False
>>> isinstance(TemplateA(), TemplateB)
False
>>> isinstance(TemplateB(), TemplateA)
False
So, a "factory" for value-difference classes could look like
from functools import partial
def _make_value_comparer(property_names, access_func):
class ValueDifferences:
def __init__(self, parent):
self._parent = parent
for name in property_names:
setattr(Value_Differences, name, property(
fget = partial(access_func, name)
))
return ValueDifferences
Notice that instead of hard-coding a helper, this factory expects to be provided with a function that implements the access logic. That function takes two parameters: a property name, and the ValueDifferences instance. (They're in that order because it's more convenient for functools.partial usage.)
Factory B: Using the type constructor directly
The built-in type in Python has two entirely separate functions.
With one argument, it discloses the type of an object.
With three arguments, it creates a new type. The class syntax is in fact syntactic sugar for a call to this builtin. The arguments are:
a string name (will be set as the __name__ attribute)
a list of classes to use as superclasses (will be set as __bases__)
a dict mapping attribute names to their values (including methods and properties - will become the __dict__, roughly)
In this style, the same factory could look something like:
from functools import partial
def _make_value_comparer(property_names, access_func):
methods = {
name: property(fget = partial(access_func, name)
for name in property_names
}
methods['__init__'] = lambda self, parent: setattr(self, '_parent', parent)
return type('ValueDifferences', [], methods)
Using the factory
In either of the above cases, EvolutionBase would be modified in the same way.
Presumably, every EvolutionBase should use the same ValueDifferences class (i.e., the one that specifically defines abc and xyz properties), so the EvolutionBase class can cache that class as a class attribute, and use it later:
class Evolution_Base():
def _get_from_parent(name, mvd):
# mvd._parent will be an instance of Evolution_Base.
return mvd._parent._get_df_change(name)
_MyValueDifferences = _make_value_comparer(['abc', 'xyz'], _get_from_parent)
def __init__(self, res_date_0 : DataFrame , res_date_1 : DataFrame):
#dataclass
class Results_Data():
res_date_0_df : DataFrame
res_date_1_df : DataFrame
self.res = Results_Data(res_date_0_df= res_date_0,
res_date_1_df= res_date_1)
self.difference = _MyValueDifferences(parent = self)
Notice that the cached _MyValueDifferences class no longer requires a list of property names to be constructed. That's because it was already provided when the class was created. The actual thing that varies per instance of _MyValueDifferences, is the parent, so that's all that gets passed.
Simpler approaches
It seems that the goal is to have a class whose instances are tightly associated with instances of Evolution_Base, providing properties specifically named abc and xyz that are computed using the Evolution_Base's data.
That could just be hard-coded as a nested class:
class Evolution_Base:
class EBValueDifferences:
def __init__(self, parent):
self._parent = parent
#property
def abc(self):
return self._parent._get_df_change('abc')
#property
def xyz(self):
return self._parent._get_df_change('xyz')
def __init__(self, res_date_0 : DataFrame , res_date_1 : DataFrame):
#dataclass
class Results_Data():
res_date_0_df : DataFrame
res_date_1_df : DataFrame
self.res = Results_Data(res_date_0_df = res_date_0,
res_date_1_df = res_date_1)
self.difference = EBValueDifferences(self)
# _get_df_change etc. as before
Even simpler, provide corresponding properties directly on Evolution_Base:
class Evolution_Base:
#property
def abc_difference(self):
return self._get_df_change('abc')
#property
def xyz_difference(self):
return self._get_df_change('xyz')
def __init__(self, res_date_0 : DataFrame , res_date_1 : DataFrame):
#dataclass
class Results_Data():
res_date_0_df : DataFrame
res_date_1_df : DataFrame
self.res = Results_Data(res_date_0_df = res_date_0,
res_date_1_df = res_date_1)
# _get_df_change etc. as before
# client code now calls my_evolution_base.abc_difference
# instead of my_evolution_base.difference.abc
If there are a lot of such properties, they could be attached using a much simpler dynamic approach (that would still be reusable for other classes that define a _get_df_change):
def add_df_change_property(name, cls):
setattr(
cls, f'{name}_difference',
property(fget = lambda instance: instance._get_df_change(name))
)
which can also be adapted for use as a decorator:
from functools import partial
def exposes_df_change(name):
return partial(add_df_change_property, name)
#exposes_df_change('abc')
#exposes_df_change('def')
class Evolution_Base:
# `self.difference` can be removed, no other changes needed
This is quite the rabbit hole. Impossible is a big call, but I will say this: they don't intend you to do this. The 'Pythonic' way of achieving your example use case is the __getattr__ method. You could also override the __dir__ method to insert your custom attributes for discoverability.
This is the code for that:
class Value_Differences():
def __init__(self, parent : Evolution_Base, property_list = []):
self._parent = parent
self._property_list = property_list
def __dir__(self):
return sorted(set(
dir(super(Value_Differences, self)) + \
list(self.__dict__.keys()) + self._property_list))
def __getattr__(self, __name: str):
if __name in self._property_list:
return self._parent._get_df_change(__name)
But that wasn't the question, and respect for a really, really interesting question. This is one of those things that you look at and say 'hmm, should be possible' and can get almost to a solution. I initially thought what you asked for was technically possible, just very hacky to achieve. But it turns out that it would be very, very weird hackery if it was possible.
Two small foundational things to start with:
Remind ourselves of the hierarchy of Python objects that the runtime is working with when defining and instantiating classes:
The metaclass (defaulting to type), which is used to build classes. I'm going to refer to this as the Metaclass Type Object (MTO).
The class definition, which is used to build objects. I'm going to refer to this as the Class Type Object (CTO).
And the class instance or object, which I'll refer to as the Class Instance Object (CIO).
MTOs are subclasses of type. CTOs are subclasses of object. CIOs are instances of CTOs, but instantiated by MTOs.
Python runs code inside class definitions as if it was running a function:
class Class1:
print("1")
def __init__(self, v1):
print("4")
print("2")
print("3")
c1 = Class1("x")
print("5")
gives 1, 2, 3, 4, 5
Put these two things together with:
class Class1:
def attr1_get(self):
return 'attr1 value'
attr1 = property(attr1_get)
we are defining a function attr1_get as part of the class definition. We are then running an inline piece of code that creates an object of type property. Note that this is just the name of the object's type - it isn't a property as you would describe it. Just an object with some attributes, being references to various functions. We then assign that object to an attribute in the class we are defining.
In the terms I used above, once that code is run we have a CTO instantiated as an object in memory that contains an attribute attr1 of type property (an object subclass, containing a bunch of attributes itself - one of which is a reference to the function attr1_get).
That can be used to instantiate an object, the CIO.
This is where the MTO comes in. You instantiate the property object while defining the CTO so that when the runtime applies the MTO to create the CIO from the CTO, an attribute on the CIO will be formed with a custom getter function for that attribute rather than the 'standard' getter function the runtime would use. The property object means something to the type object when it is building a new object.
So when we run:
c1 = Class1()
we don't get a CIO c1 with an attribute attr1 that is an object of type property. The metaclass of type type formed a set of references against the attribute's internal state to all the functions we stored in the property object. Note that this is happening inside the runtime, and you can't call this directly from your code - you just tell the type metaclass to do it by using the property wrapper object.
So if you directly assign a property() result to an attribute of a CIO, you have a Pythonic object assigned that references some functions, but the internal state for the runtime to use to reference the getter, setter, etc. is not set up. The getter of an attribute that contains a property object is the standard getter and so returns the object instance, and not the result of the functions it wraps,
This next bit of code demonstrates how this flows:
print("Let's begin")
class MetaClass1(type):
print("Starting to define MetaClass1")
def __new__(cls, name, bases, dct):
x = super().__new__(cls, name, bases, dct)
print("Metaclass1 __new__({})".format(str(cls)))
return x
print("__new__ of MetaClass1 is defined")
def __init__(cls, name, bases, dct):
print("Metaclass1 __init__({})".format(str(cls)))
print("__init__ of MetaClass1 is defined")
print("Metaclass is defined")
class Class1(object,metaclass=MetaClass1):
print("Starting to define Class1")
def __new__(cls, *args, **kwargs):
print("Class1 __new__({})".format(str(cls)))
return super(Class1, cls).__new__(cls, *args, **kwargs)
print("__new__ of Class1 is defined")
def __init__(self):
print("Class1 __init__({})".format(str(self)))
print("__init__ of Class1 is defined")
def g1(self):
return 'attr1 value'
print("g1 of Class1 is defined")
attr1 = property(g1)
print("Class1.attr1 = ", attr1)
print("attr1 of Class1 is defined")
def addProperty(self, name, getter):
setattr(self, name, property(getter))
print("self.", name, " = ", getattr(self, name))
print("addProperty of Class1 is defined")
print("Class is defined")
c1 = Class1()
print("Instance is created")
print(c1.attr1)
def g2(cls):
return 'attr2 value'
c1.addProperty('attr2', g2)
print(c1.attr2)
I have put all those print statements there to demonstrate the order in which things happen very clearly.
In the middle, you see:
g1 of Class1 is defined
Class1.attr1 = <property object at 0x105115c10>
attr1 of Class1 is defined
We have created an object of type property and assigned it to a class attribute.
Continuing:
addProperty of Class1 is defined
Metaclass1 __new__(<class '__main__.MetaClass1'>)
Metaclass1 __init__(<class '__main__.Class1'>)
Class is defined
The metaclass got instantiated, being passed first itself (__new__) and then the class it will work on (__init__). This happened right as we stepped out of the class definition. I have only included the metaclass to show what will happen with the type metaclass by default.
Then:
Class1 __new__(<class '__main__.Class1'>)
Class1 __init__(<__main__.Class1 object at 0x105124c10>)
Instance is created
attr1 value
self. attr2 = <property object at 0x105115cb0>
<property object at 0x105115cb0>
Class1 is instantiated, providing first its type to __new__ and then its instance to __init__.
We see that attr1 is instantiated properly, but attr2 is not. That is because setattr is being called once the class instance is already constructed and is just saying attr2 is an instance of the class property and not defining attr2 as the actual runtime construct of a property.
Which is made more clear if we run:
print(c1.attr2.fget(c1))
print(c1.attr1.fget(c1))
attr2 (a property object) isn't aware of the class or instance of the containing attribute's parent. The function it wraps still needs to be given the instance to work on.
attr1 doesn't know what to do with that, because as far as it is concerned it is a string object, and has no concept of how the runtime is mapping its getter.
The fundamental reason why what you tried doesn't work is that a property, a use case of a descriptor, by design must be stored as a class variable, not as an instance attribute.
Excerpt from the documentation of descriptor:
To use the descriptor, it must be stored as a class variable in
another class:
To create a class with dynamically named properties that has access to a parent class, one elegant approach is to create the class within a method of the main class, and use setattr to create class attributes with dynamic names and property objects. A class created in the closure of a method automatically has access to the self object of the parent instance, avoiding having to manage a clunky _parent attribute like you do in your attempt:
class Evolution_Base:
def __init__(self, property_list):
self.property_list = property_list
self._difference = None
#property
def difference(self):
if not self._difference:
class Value_Differences:
pass
for name in self.property_list:
# use default value to store the value of name in each iteration
def func(obj, prop_name=name):
return self._get_df_change(prop_name) # access self via closure
setattr(Value_Differences, name, property(func))
self._difference = Value_Differences()
return self._difference
def _get_df_change(self, df_name):
return f'df change of {df_name}' # simplified return value for demo purposes
so that:
evolution = Evolution_Base(['abc', 'xyz'])
print(evolution.difference.abc)
print(evolution.difference.xyz)
would output:
df change of abc
df change of xyz
Demo: https://replit.com/#blhsing/ExtralargeNaturalCoordinate
Responding directly to your question, you can create a class:
class FooBar:
def __init__(self, props):
def make_prop(name):
return property(lambda accessor_self: self._prop_impl(name))
self.accessor = type(
'Accessor',
tuple(),
{p: make_prop(p) for p in props}
)()
def _prop_impl(self, arg):
return arg
o = FooBar(['foo', 'bar'])
assert o.accessor.foo == o._prop_impl('foo')
assert o.accessor.bar == o._prop_impl('bar')
Further, it would be beneficiary to cache created class to make equivalent objects more similar and eliminate potential issues with equality comparison.
That said, I am not sure if this is desired. There's little benefit of replacing method call syntax (o.f('a')) with property access (o.a). I believe it can be detrimental on multiple accounts: dynamic properties are confusing, harder to document, etc., finally while none of this is strictly guaranteed in crazy world of dynamic python -- they kind of communicate wrong message: that the access is cheap and does not involve computation and that perhaps you can attempt to write to it.
I think that when you define the function func in the loop, it closes over the current value of the name variable, not the value of the name variable at the time the property is accessed. To fix this, you can use a lambda function to create a closure that captures the value of name at the time the property is defined.
class Value_Differences():
def __init__(self, parent : Evolution_Base, property_list = []):
self._parent = parent
for name in property_list:
setattr(self, name, property(fget = lambda self, name=name: self._parent._get_df_change(name)))
Does this help you ?
The simple question is really, how do I setattr for a property function?
In python we can set dynamic attributes like this:
class DynamicProperties():
def __init__(self, property_list):
self.property_list = property_list
def add_properties(self):
for name in self.property_list:
setattr(self.__class__, name, property(fget=lambda self: 1))
dync = DynamicProperties(['a', 'b'])
dync.add_properties()
print(dync.a) # prints 1
print(dync.b) # prints 1
Correct me if I am wrong but from reviewing your code, you want to create a dynamic attributes then set their value to a specific function call within the same class, where the passed in data is passed in attributes in the constructor " init " this is achievable, an example:
class DynamicProperties():
def __init__(self, property_list, data1, data2):
self.property_list = property_list
self.data1 = data1
self.data2 = data2
def add_properties(self):
for name in self.property_list:
setattr(self.__class__, name, property(fget=lambda self: self.change(self.data1, self.data2) ))
def change(self, data1, data2):
return data1 - data2
dync = DynamicProperties(['a', 'b'], 1, 2)
dync.add_properties()
print(dync.a == dync.change(1, 2)) # prints true
print(dync.b == dync.change(1,2)) # prints true
You just have to add more complexity to the member, __getattr__ / __setattr__ gives you the string, so it can be interpreted as needed. The biggest "problem" doing this is that the return might no be consistent and piping it back to a library that expect an object to have a specific behavior can cause soft errors.
This example is not the same as yours, but it has the same concept, manipulate columns with members. To get a copy with changes a set is not needed, with a copy, modify and return, the new instance can be created with whatever needed.
For example, the __getattr__ in this line will:
Check and interpret the string xyz_mull_0
Validate that the members and the operand exists
Make a copy of data_a
Modify the copy and return it
var = data_a.xyz_mull_0()
This looks more complex that it actually is, with the same instance members its clear what it is doing, but the _of modifier needs a callback, this is because the __getattr__ can only have one parameter, so it needs to save the attr and return a callback to be called with the other instance that then will call back to the __getattr__ and complete the rest of the function.
import re
class FlexibleFrame:
operand_mod = {
'sub': lambda a, b: a - b,
'add': lambda a, b: a + b,
'div': lambda a, b: a / b,
'mod': lambda a, b: a % b,
'mull': lambda a, b: a * b,
}
#staticmethod
def add_operand(name, func):
if name not in FlexibleFrame.operand_mod.keys():
FlexibleFrame.operand_mod[name] = func
# This makes this class subscriptable
def __getitem__(self, item):
return self.__dict__[item]
# Uses:
# -> object.value
# -> object.member()
# -> object.<name>_<operand>_<name|int>()
# -> object.<name>_<operand>_<name|int>_<flow>()
def __getattr__(self, attr):
if re.match(r'^[a-zA-Z]+_[a-zA-Z]+_[a-zA-Z0-9]+(_of)?$', attr):
seg = attr.split('_')
var_a, operand, var_b = seg[0:3]
# If there is a _of: the second operand is from the other
# instance, the _of is removed and a callback is returned
if len(seg) == 4:
self.__attr_ref = '_'.join(seg[0:3])
return self.__getattr_of
# Checks if this was a _of attribute and resets it
if self.__back_ref is not None:
other = self.__back_ref
self.__back_ref = None
self.__attr_ref = None
else:
other = self
if var_a not in self.__dict__:
raise AttributeError(
f'No match of {var_a} in (primary) {__class__.__name__}'
)
if operand not in FlexibleFrame.operand_mod.keys():
raise AttributeError(
f'No match of operand {operand}'
)
# The return is a copy of self, if not the instance
# is getting modified making x = a.b() useless
ret = FlexibleFrame(**self.__dict__)
# Checks if the second operand is a int
if re.match(r'^\d+$', var_b) :
ref_b_num = int(var_b)
for i in range(len(self[var_a])):
ret[var_a][i] = FlexibleFrame.operand_mod[operand](
self[var_a][i], ref_b_num
)
elif var_b in other.__dict__:
for i in range(len(self[var_a])):
# out_index = operand[type](in_a_index, in_b_index)
ret[var_a][i] = FlexibleFrame.operand_mod[operand](
self[var_a][i], other[var_b][i]
)
else:
raise AttributeError(
f'No match of {var_b} in (secondary) {__class__.__name__}'
)
# This swaps the .member to a .member()
# it also adds and extra () in __getattr_of
return lambda: ret
# return ret
if attr in self.__dict__:
return self[attr]
raise AttributeError(
f'No match of {attr} in {__class__.__name__}'
)
def __getattr_of(self, other):
self.__back_ref = other
return self.__getattr__(self.__attr_ref)()
def __init__(self, **kwargs):
self.__back_ref = None
self.__attr_ref = None
#TODO: Check if data columns match in size
# if not, implement column_<name>_filler=<default>
for i in kwargs:
self.__dict__[i] = kwargs[i]
if __name__ == '__main__':
data_a = FlexibleFrame(**{
'abc': [i for i in range(10)],
'nmv': [i for i in range(10)],
'xyz': [i for i in range(10)],
})
data_b = FlexibleFrame(**{
'fee': [i + 10 for i in range(10)],
'foo': [i + 10 for i in range(10)],
})
FlexibleFrame.add_operand('set', lambda a, b: b)
var = data_a.xyz_mull_0()
var = var.abc_set_xyz()
var = var.xyz_add_fee_of(data_b)
As a extra thing, lambdas in python have this thing, so it can make difficult using them when self changes.
It seems you're bending the language to do weird things. I'd take it as a smell that your code is probably getting convoluted but I'm not saying there would never be a use-case for it so here is a minimal example of how to do it:
class Obj:
def _df_change(self, arg):
print('change', arg)
class DynAttributes(Obj):
def __getattr__(self, name):
return self._df_change(name)
class Something:
difference = DynAttributes()
a = Something()
b = Obj()
assert a.difference.hello == b._df_change('hello')
When calling setattr , use self.__class__ instead of self
Code sample:
class A:
def __init__(self,names : List[str]):
for name in names:
setattr(self.__class__,name,property(fget=self.__create_getter(name)))
def __create_getter(self,name: str):
def inner(self):
print(f"invoking {name}")
return 10
return inner
a = A(['x','y'])
print(a.x + 1)
print(a.y + 2)
I follow the tutorial out of the docs and an example by fluent Python. In the book they teach me how to avoid the AttributeError by get, (e.g., when you do z = Testing.x) and I wanted to do something simliar with the set method. But it seems like, it lead to a broken class with no error.
To be more specific about the issue:
With outcommented line Testing.x = 1 it invoke the __set__ methods.
With uncommented line #Testing.x = 1 it does not invoke the __set__ methods.
Can someone teach me why it behaves this way?
import abc
class Descriptor:
def __init__(self):
cls = self.__class__
self.storage_name = cls.__name__
def __get__(self, instance, owner):
if instance is None:
return self
else:
return getattr(instance, self.storage_name)
def __set__(self, instance, value):
print(instance,self.storage_name)
setattr(instance, self.storage_name, value)
class Validator(Descriptor):
def __set__(self, instance, value):
value = self.validate(instance, value)
super().__set__(instance, value)
#abc.abstractmethod
def validate(self, instance, value):
"""return validated value or raise ValueError"""
class NonNegative(Validator):
def validate(self, instance, value):
if value <= 0:
raise ValueError(f'{value!r} must be > 0')
return value
class Testing:
x = NonNegative()
def __init__(self,number):
self.x = number
#Testing.x = 1
t = Testing(1)
t.x = 1
Attribute access is generally handled by object.__getattribute__ and type.__getattribute__ (for instances of type, i.e. classes). When an attribute lookup of the form a.x involves a descriptor as x, then various binding rules come into effect, based on what x is:
Instance binding: If binding to an object instance, a.x is transformed into the call: type(a).__dict__['x'].__get__(a, type(a)).
Class binding: If binding to a class, A.x is transformed into the call: A.__dict__['x'].__get__(None, A).
Super binding: [...]
For the scope of this question, only (2) is relevant. Here, Testing.x invokes the descriptor via __get__(None, Testing). Now one might ask why this is done instead of simply returning the descriptor object itself (as if it was any other object, say an int). This behavior is useful to implement the classmethod decorator. The descriptor HowTo guide provides an example implementation:
class ClassMethod:
def __init__(self, f):
self.f = f
def __get__(self, obj, cls=None):
print(f'{obj = }, {cls = }')
return self.f.__get__(cls, cls) # simplified version
class Test:
#ClassMethod
def func(cls, x):
pass
Test().func(2) # call from instance
Test.func(1) # this requires binding without any instance
We can observe that for the second case Test.func(1) there is no instance involved, but the ClassMethod descriptor can still bind to the cls.
Given that __get__ is used for both, instance and class binding, one might ask why this isn't the case for __set__. Specifically, for x.y = z, if y is a data descriptor, why doesn't it invoke y.__set__(None, z)? I guess the reason is that there is no good use case for that and it unnecessarily complicates the descriptor API. What would the descriptor do with that information anyway? Typically, managing how attributes are set is done by the class (or metaclass for types), via object.__setattr__ or type.__setattr__.
So to prevent Testing.x from being replaced by a user, you could use a custom metaclass:
class ProtectDataDescriptors(type):
def __setattr__(self, name, value):
if hasattr(getattr(self, name, None), '__set__'):
raise AttributeError(f'Cannot override data descriptor {name!r}')
super().__setattr__(name, value)
class Testing(metaclass=ProtectDataDescriptors):
x = NonNegative()
def __init__(self, number):
self.x = number
Testing.x = 1 # now this raises AttributeError
However, this is not an absolute guarantee as users can still use type.__setattr__ directly to override that attribute:
type.__setattr__(Testing, 'x', 1) # this will bypass ProtectDataDescriptors.__setattr__
The line
Testing.x = 1
replaces the descriptor you've set as a class attribute for Testing with an integer.
Since the descriptor is no more, self.x = ... or t.x = ... is just an assignment that doesn't involve a descriptor.
As an aside, surely you've noticed there is no true x attribute anymore with your descriptor, and you can't use more than one instance of the same descriptor without conflicts?
class Testing:
x = NonNegative()
y = NonNegative()
def __init__(self, number):
self.x = number
t = Testing(2345)
t.x = 1234
t.y = 5678
print(vars(t))
prints out
{'NonNegative': 5678}
I have a state object that represents a system. Properties within the state object are populated from [huge] text files. As not every property is accessed every time a state instance, is created, it makes sense to lazily load them.:
class State:
def import_positions(self):
self._positions = {}
# Code which populates self._positions
#property
def positions(self):
try:
return self._positions
except AttributeError:
self.import_positions()
return self._positions
def import_forces(self):
self._forces = {}
# Code which populates self._forces
#property
def forces(self):
try:
return self._forces
except AttributeError:
self.import_forces()
return self._forces
There's a lot of repetitive boilerplate code here. Moreover, sometimes an import_abc can populate a few variables (i.e. import a few variables from a small data file if its already open).
It makes sense to overload #property such that it accepts a function to "provide" that variable, viz:
class State:
def import_positions(self):
self._positions = {}
# Code which populates self._positions
#lazyproperty(import_positions)
def positions(self):
pass
def import_forces(self):
self._forces = {}
# Code which populates self._forces and self._strain
#lazyproperty(import_forces)
def forces(self):
pass
#lazyproperty(import_forces)
def strain(self):
pass
However, I cannot seem to find a way to trace exactly what method are being called in the #property decorator. As such, I don't know how to approach overloading #property into my own #lazyproperty.
Any thoughts?
Maybe you want something like this. It's a sort of simple memoization function combined with #property.
def lazyproperty(func):
values = {}
def wrapper(self):
if not self in values:
values[self] = func(self)
return values[self]
wrapper.__name__ = func.__name__
return property(wrapper)
class State:
#lazyproperty
def positions(self):
print 'loading positions'
return {1, 2, 3}
s = State()
print s.positions
print s.positions
Which prints:
loading positions
set([1, 2, 3])
set([1, 2, 3])
Caveat: entries in the values dictionary won't be garbage collected, so it's not suitable for long-running programs. If the loaded value is immutable across all classes, it can be stored on the function object itself for better speed and memory use:
try:
return func.value
except AttributeError:
func.value = func(self)
return func.value
I think you can remove even more boilerplate by writing a custom descriptor class that decorates the loader method. The idea is to have the descriptor itself encode the lazy-loading logic, meaning that the only thing you define in an actual method is the loader itself (which is the only thing that, apparently, really does have to vary for different values). Here's an example:
class LazyDesc(object):
def __init__(self, func):
self.loader = func
self.secretAttr = '_' + func.__name__
def __get__(self, obj, cls):
try:
return getattr(obj, self.secretAttr)
except AttributeError:
print("Lazily loading", self.secretAttr)
self.loader(obj)
return getattr(obj, self.secretAttr)
class State(object):
#LazyDesc
def positions(self):
self._positions = {'some': 'positions'}
#LazyDesc
def forces(self):
self._forces = {'some': 'forces'}
Then:
>>> x = State()
>>> x.forces
Lazily loading _forces
{'some': 'forces'}
>>> x.forces
{'some': 'forces'}
>>> x.positions
Lazily loading _positions
{'some': 'positions'}
>>> x.positions
{'some': 'positions'}
Notice that the "lazy loading" message was printed only on the first access for each attribute. This version also auto-creates the "secret" attribute to hold the real data by prepending an underscore to the method name (i.e., data for positions is stored in _positions. In this example, there's no setter, so you can't do x.positions = blah (although you can still mutate the positions with x.positions['key'] = val), but the approach could be extended to allow setting as well.
The nice thing about this approach is that your lazy logic is transparently encoded in the descriptor __get__, meaning that it easily generalizes to other kinds of boilerplate that you might want to abstract away in a similar manner.
However, I cannot seem to find a way to trace exactly what method are
being called in the #property decorator.
property is actually a type (whether you use it with the decorator syntax of not is orthogonal), which implements the descriptor protocol (https://docs.python.org/2/howto/descriptor.html). An overly simplified (I skipped the deleter, doc and quite a few other things...) pure-python implementation would look like this:
class property(object):
def __init__(self, fget=None, fset=None):
self.fget = fget
self.fset = fset
def setter(self, func):
self.fset = func
return func
def __get__(self, obj, type=None):
return self.fget(obj)
def __set__(self, obj, value):
if self.fset:
self.fset(obj, value)
else:
raise AttributeError("Attribute is read-only")
Now overloading property is not necessarily the simplest solution. In fact there are actually quite a couple existing implementations out there, including Django's "cached_property" (cf http://ericplumb.com/blog/understanding-djangos-cached_property-decorator.html for more about it) and pydanny's "cached-property" package (https://pypi.python.org/pypi/cached-property/0.1.5)
As the title says. It seems no matter what I do, __getattr__ will not be called. I also tried it for instance (absurd, I know), with predictably no response. As if __getattr__ was banned in meta classes.
I'd appreciate any pointer to documentation about this.
The code:
class PreinsertMeta(type):
def resolvedField(self):
if isinstance(self.field, basestring):
tbl, fld = self.field.split(".")
self.field = (tbl, fld)
return self.field
Field = property(resolvedField)
def __getattr__(self, attrname):
if attrname == "field":
if isinstance(self.field, basestring):
tbl, fld = self.field.split(".")
self.field = (tbl, fld)
return self.field
else:
return super(PreinsertMeta, self).__getattr__(attrname)
def __setattr__(self, attrname, value):
super(PreinsertMeta, self).__setattr__(attrname, value)
class Test(object):
__metaclass__ = PreinsertMeta
field = "test.field"
print Test.field # Should already print the tuple
Test.field = "another.field" # __setattr__ gets called nicely
print Test.field # Again with the string?
print Test.Field # note the capital 'F', this actually calls resolvedField() and prints the tuple
Thanks to BrenBarn, here's the final working implementation:
class PreinsertMeta(type):
def __getattribute__(self, attrname):
if attrname == "field" and isinstance(object.__getattribute__(self, attrname), basestring):
tbl, fld = object.__getattribute__(self, attrname).split(".")
self.field = (tbl, fld)
return object.__getattribute__(self, attrname)
As documented, __getattr__ is only called if the attribute does not exist. Since your class has a field attribute, that blocks __getattr__. You can use __getattribute__ if you really want to intercept all attribute access, although it's not clear from your example why you need to do this. Note that this has nothing to do with metaclasses; you would see the same behavior if you created an instance of an ordinary class and gave it some attribute.
Even assuming you used __getattribute__, so it was called when the attribute exists, your implementation doesn't make much sense. Inside __getattr__ you try to get a value for self.field. But if __getattribute__ was called in the first place, it will be called again for this access, creating an infinite recursion: in order to get self.field, it has to call __getattribute__, which again tries to get self.field, which again calls __getattribute__, etc. See the documentation for __getattribute__ for how to get around this.
Sorry, badly worded title. I hope a simple example will make it clear. Here's the easiest way to do what I want to do:
class Lemon(object):
headers = ['ripeness', 'colour', 'juiciness', 'seeds?']
def to_row(self):
return [self.ripeness, self.colour, self.juiciness, self.seeds > 0]
def save_lemons(lemonset):
f = open('lemons.csv', 'w')
out = csv.writer(f)
out.write(Lemon.headers)
for lemon in lemonset:
out.writerow(lemon.to_row())
This works alright for this small example, but I feel like I'm "repeating myself" in the Lemon class. And in the actual code I'm trying to write (where the number of variables I'm exporting is ~50 rather than 4, and where to_row calls a number of private methods that do a bunch of weird calculations), it becomes awkward.
As I write the code to generate a row, I need to constantly refer to the "headers" variable to make sure I'm building my list in the correct order. If I want to change the variables being outputted, I need to make sure to_row and headers are being changed in parallel (exactly the kind of thing that DRY is meant to prevent, right?).
Is there a better way I could design this code? I've been playing with function decorators, but nothing has stuck. Ideally I should still be able to get at the headers without having a particular lemon instance (i.e. it should be a class variable or class method), and I don't want to have a separate method for each variable.
In this case, getattr() is your friend: it allows you to get a variable based on a string name. For example:
def to_row(self):
return [getattr(self, head) for head in self.headers]
EDIT: to properly use the header seeds?, you would need to set the attribute seeds? for the objects. setattr(self, 'seeds?', self.seeds > 0) right above the return statement.
We could use some metaclass shenanegans to do this...
In python 2, attributes are passed to the metaclass in a dict, without
preserving order, we'll also want a base class to work with so we can
distinguish class attributes that should be mapped into the row. In python3, we could dispense with just about all of this base descriptor class.
import itertools
import functools
#functools.total_ordering
class DryDescriptor(object):
_order_gen = itertools.count()
def __init__(self, alias=None):
self.alias = alias
self.order = next(self._order_gen)
def __lt__(self, other):
return self.order < other.order
We will want a python descriptor for every attribute we wish to map into the
row. slots are a nice way to get data descriptors without much work. One
caveat, though, we'll have to manually remove the helper instance to make the
real slot descriptor visible.
class slot(DryDescriptor):
def annotate(self, attr, attrs):
del attrs[attr]
self.attr = attr
slots = attrs.setdefault('__slots__', []).append(attr)
def annotate_class(self, cls):
if self.alias is not None:
setattr(cls, self.alias, getattr(self.attr))
For computed fields, we can memoize results. Memoizing off of the annotated
instance is tricky without a memory leak, we need weakref. alternatively, we
could have arranged for another slot just to store the cached value. This also isn't quite thread safe, but pretty close.
import weakref
class memo(DryDescriptor):
_memo = None
def __call__(self, method):
self.getter = method
return self
def annotate(self, attr, attrs):
if self.alias is not None:
attrs[self.alias] = self
def annotate_class(self, cls): pass
def __get__(self, instance, owner):
if instance is None:
return self
if self._memo is None:
self._memo = weakref.WeakKeyDictionary()
try:
return self._memo[instance]
except KeyError:
return self._memo.setdefault(instance, self.getter(instance))
On the metaclass, all of the descriptors we created above are found, sorted by
creation order, and instructed to annotate the new, created class. This does
not correctly treat derived classes and could use some other conveniences like
an __init__ for all the slots.
class DryMeta(type):
def __new__(mcls, name, bases, attrs):
descriptors = sorted((value, key)
for key, value
in attrs.iteritems()
if isinstance(value, DryDescriptor))
for descriptor, attr in descriptors:
descriptor.annotate(attr, attrs)
cls = type.__new__(mcls, name, bases, attrs)
for descriptor, attr in descriptors:
descriptor.annotate_class(cls)
cls._header_descriptors = [getattr(cls, attr) for descriptor, attr in descriptors]
return cls
Finally, we want a base class to inherit from so that we can have a to_row
method. this just invokes all of the __get__s for all of the respective
descriptors, in order.
class DryBase(object):
__metaclass__ = DryMeta
def to_row(self):
cls = type(self)
return [desc.__get__(self, cls) for desc in cls._header_descriptors]
Assuming all of that is tucked away, out of sight, the definition of a class
that uses this feature is mostly free of repitition. The only short coming is
that to be practical, every field needs a python friendly name, thus we had the
alias key to associate 'seeds?' to has_seeds
class ADryRow(DryBase):
__slots__ = ['seeds']
ripeness = slot()
colour = slot()
juiciness = slot()
#memo(alias='seeds?')
def has_seeds(self):
print "Expensive!!!"
return self.seeds > 0
>>> my_row = ADryRow()
>>> my_row.ripeness = "tart"
>>> my_row.colour = "#8C2"
>>> my_row.juiciness = 0.3479
>>> my_row.seeds = 19
>>>
>>> print my_row.to_row()
Expensive!!!
['tart', '#8C2', 0.3479, True]
>>> print my_row.to_row()
['tart', '#8C2', 0.3479, True]