Lazy attributes

class sage.misc.lazy_attribute.lazy_attribute(f)

Bases: object

A lazy attribute for an object is like a usual attribute, except that, instead of being computed when the object is constructed (i.e. in __init__), it is computed on the fly the first time it is accessed.

For constant values attached to an object, lazy attributes provide a shorter syntax and automatic caching (unlike methods), while playing well with inheritance (like methods): a subclass can easily override a given attribute; you don’t need to call the super class constructor, etc.

Technically, a lazy_attribute is a non-data descriptor (see Invoking Descriptors in the Python reference manual).

EXAMPLES:

We create a class whose instances have a lazy attribute x:

sage: class A(object):
...       def __init__(self):
...           self.a=2 # just to have some data to calculate from
...
...       @lazy_attribute
...       def x(self):
...           print "calculating x in A"
...           return self.a + 1
...

For an instance a of A, a.x is calculated the first time it is accessed, and then stored as a usual attribute:

sage: a = A()
sage: a.x
calculating x in A
3
sage: a.x
3

Implementation details

We redo the same example, but opening the hood to see what happens to the internal dictionary of the object:

sage: a = A()
sage: a.__dict__
{'a': 2}
sage: a.x
calculating x in A
3
sage: a.__dict__
{'a': 2, 'x': 3}
sage: a.x
3
sage: timeit('a.x') # random
625 loops, best of 3: 89.6 ns per loop

This shows that, after the first calculation, the attribute x becomes a usual attribute; in particular, there is no time penalty to access it.

A lazy attribute may be set as usual, even before its first access, in which case the lazy calculation is completely ignored:

sage: a = A()
sage: a.x = 4
sage: a.x
4
sage: a.__dict__
{'a': 2, 'x': 4}

Class binding results in the lazy attribute itself:

sage: A.x
<sage.misc.lazy_attribute.lazy_attribute object at ...>

Conditional definitions

The function calculating the attribute may return NotImplemented to declare that, after all, it is not able to do it. In that case, the attribute lookup proceeds in the super class hierarchy:

sage: class B(A):
...       @lazy_attribute
...       def x(self):
...           if hasattr(self, "y"):
...               print "calculating x from y in B"
...               return self.y
...           else:
...               print "y not there; B does not define x"
...               return NotImplemented
...
sage: b = B()
sage: b.x
y not there; B does not define x
calculating x in A
3
sage: b = B()
sage: b.y = 1
sage: b.x
calculating x from y in B
1

Attribute existence testing

Testing for the existence of an attribute with hasattr currently always triggers its full calculation, which may not be desirable when the calculation is expensive:

sage: a = A()
sage: hasattr(a, "x")
calculating x in A
True

It would be great if we could take over the control somehow, if at all possible without a special implementation of hasattr, so as to allow for something like:

sage: class A (object):
...       @lazy_attribute
...       def x(self, existence_only=False):
...           if existence_only:
...               print "testing for x existence"
...               return True
...           else:
...               print "calculating x in A"
...               return 3
...
sage: a = A()
sage: hasattr(a, "x") # todo: not implemented
testing for x existence
sage: a.x
calculating x in A
3
sage: a.x
3

Here is a full featured example, with both conditional definition and existence testing:

sage: class B(A):
...       @lazy_attribute
...       def x(self, existence_only=False):
...           if hasattr(self, "y"):
...               if existence_only:
...                   print "testing for x existence in B"
...                   return True
...               else:
...                   print "calculating x from y in B"
...                   return self.y
...           else:
...               print "y not there; B does not define x"
...               return NotImplemented
...
sage: b = B()
sage: hasattr(b, "x") # todo: not implemented
y not there; B does not define x
testing for x existence
True
sage: b.x
y not there; B does not define x
calculating x in A
3
sage: b = B()
sage: b.y = 1
sage: hasattr(b, "x") # todo: not implemented
testing for x existence in B
True
sage: b.x
calculating x from y in B
1

lazy attributes and introspection

TODO: make the following work nicely:

sage: b.x?                # todo: not implemented
sage: b.x??               # todo: not implemented

Right now, the first one includes the doc of this class, and the second one brings up the code of this class, both being not very useful.

TESTS:

.. rubric:: Partial support for old style classes

Old style and new style classes play a bit differently with @property and attribute setting:

sage: class A:
...       @property
...       def x(self):
...           print "calculating x"
...           return 3
...
sage: a = A()
sage: a.x = 4
sage: a.__dict__
{'x': 4}
sage: a.x
4
sage: a.__dict__['x']=5
sage: a.x
5

sage: class A (object):
...       @property
...       def x(self):
...           print "calculating x"
...           return 3
...
sage: a = A()
sage: a.x = 4
...
AttributeError: can't set attribute
sage: a.__dict__
{}
sage: a.x
calculating x
3
sage: a.__dict__['x']=5
sage: a.x
calculating x
3

In particular, lazy_attributes need to be implemented as non-data descriptors for new style classes, so as to leave access to setattr. We now check that this implementation also works for old style classes (conditional definition does not work yet):

sage: class A:
...       def __init__(self):
...           self.a=2 # just to have some data to calculate from
...
...       @lazy_attribute
...       def x(self):
...           print "calculating x"
...           return self.a + 1
...
sage: a = A()
sage: a.__dict__
{'a': 2}
sage: a.x
calculating x
3
sage: a.__dict__
{'a': 2, 'x': 3}
sage: a.x
3
sage: timeit('a.x') # random
625 loops, best of 3: 115 ns per loop

sage: a = A()
sage: a.x = 4
sage: a.x
4
sage: a.__dict__
{'a': 2, 'x': 4}

sage: class B(A):
...       @lazy_attribute
...       def x(self):
...           if hasattr(self, "y"):
...               print "calculating x from y in B"
...               return self.y
...           else:
...               print "y not there; B does not define x"
...               return NotImplemented
...
sage: b = B()
sage: b.x                         # todo: not implemented
y not there; B does not define x
calculating x in A
3
sage: b = B()
sage: b.y = 1
sage: b.x
calculating x from y in B
1

lazy_attributes and cpdef functions

This attempts to check that lazy_attributes work with built-in functions like cpdef methods:

sage: class A:
...       def __len__(x):
...           return int(5)
...       len = lazy_attribute(len)
...
sage: A().len
5

About descriptor specifications

The specifications of descriptors (see 3.4.2.3 Invoking Descriptors in the Python reference manual) are incomplete w.r.t. inheritance, and maybe even ill-implemented. We illustrate this on a simple class hierarchy, with an instrumented descriptor:

sage: class descriptor(object):
...       def __get__(self, obj, cls):
...           print cls
...           return 1
...   
sage: class A(object):
...       x = descriptor()
...   
sage: class B(A):
...       pass
...

This is fine:

sage: A.x
<class '__main__.A'>
1

The behaviour for the following case is not specified (see Instance Binding) when x is not in the dictionary of B but in that of some super category:

sage: B().x
<class '__main__.B'>
1

It would seem more natural (and practical!) to get A rather than B.

From the specifications for Super Binding, it would be expected to get A and not B as cls parameter:

sage: super(B, B()).x
<class '__main__.B'>
1

Due to this, the natural implementation runs into an infinite loop in the following example:

sage: class A(object):
...       @lazy_attribute
...       def unimplemented_A(self):
...           return NotImplemented
...       @lazy_attribute
...       def unimplemented_AB(self):
...           return NotImplemented
...       @lazy_attribute
...       def unimplemented_B_implemented_A(self):
...           return 1
...
sage: class B(A):
...       @lazy_attribute
...       def unimplemented_B(self):
...           return NotImplemented
...       @lazy_attribute
...       def unimplemented_AB(self):
...           return NotImplemented
...       @lazy_attribute
...       def unimplemented_B_implemented_A(self):
...           return NotImplemented
...
sage: class C(B):
...       pass
...

This is the simplest case where, without workaround, we get an infinite loop:

sage: hasattr(B(), "unimplemented_A") # todo: not implemented
False

TODO: improve the error message:

sage: B().unimplemented_A # todo: not implemented
...
AttributeError: 'super' object has no attribute 'unimplemented_A'

We now make some systematic checks:

sage: B().unimplemented_A
...
AttributeError: '...' object has no attribute 'unimplemented_A'
sage: B().unimplemented_B
...
AttributeError: '...' object has no attribute 'unimplemented_B'
sage: B().unimplemented_AB
...
AttributeError: '...' object has no attribute 'unimplemented_AB'
sage: B().unimplemented_B_implemented_A
1

sage: C().unimplemented_A()
...
AttributeError: '...' object has no attribute 'unimplemented_A'
sage: C().unimplemented_B()
...
AttributeError: '...' object has no attribute 'unimplemented_B'
sage: C().unimplemented_AB()
...
AttributeError: '...' object has no attribute 'unimplemented_AB'
sage: C().unimplemented_B_implemented_A # todo: not implemented
1