Generators.

[Jorge Cardona]

Hi,

I was trying to create a function that receive a generator and return
a list but that each elements were computed in a diferent core of my
machine. I start using islice function in order to split the job in a
way that if there is "n" cores each "i" core will compute the elements
i,i+n,i+2n,..., but islice has a weird (to me) behavior, look:

from itertools import islice

def f(x):
print("eval: %d"%x)
return x

X = range(10)
g = (f(x) for x in X)

print(list(x for x in islice(g,0,None,2)))

$ python2.5 test.py
eval: 0
eval: 1
eval: 2
eval: 3
eval: 4
eval: 5
eval: 6
eval: 7
eval: 8
eval: 9
[0, 2, 4, 6, 8]
$ python2.7 test.py
eval: 0
eval: 1
eval: 2
eval: 3
eval: 4
eval: 5
eval: 6
eval: 7
eval: 8
eval: 9
[0, 2, 4, 6, 8]
$ python3.0 test.py
eval: 0
eval: 1
eval: 2
eval: 3
eval: 4
eval: 5
eval: 6
eval: 7
eval: 8
eval: 9
[0, 2, 4, 6, 8]

islice execute the function at the generator and drop the elements
that aren't in the slice. I found that pretty weird, the way that i
see generators is like an association between and indexing set (an
iterator or another generator) and a computation that is made indexed
by the indexing set, and islice is like a "transformation" on the
indexing set,it doesn't matter the result of the function, the slice
should act only on the indexing set, some other "transformation" like
takewhile act on the result so, the execution it has to be made, but
in the islice, or other "transformation" that act only in the indexing
set, the function shouldn't be executed at each element, but only on
that new set that result of the application of the "transformation" on
the original set.

I search a little bit and found that gi_frame.f_locals['.0'] holds the
inner indexing element of a generator, so i decide to define my own
islice like this:

from itertools import islice
import sys


if sys.version_info[0] != 3:
def next(it):
return it.next()


def f(x):
print("eval: %d"%x)
return x


def islice(iterable, *args):
s = slice(*args)

# search the deepest iter (Base indexing set)
it = iterable
while hasattr(it, 'gi_frame'):
it = it.gi_frame.f_locals['.0']

# Consume the base indexing set until the first element
for i in range(s.start):
it.next()

for e in iterable:
yield e

# Consume the base indexing set until the next element
for i in range(s.step-1):
next(it)


X = range(10)
g = (f(x) for x in X)

print(list(x for x in islice(g,0,None,2)))


jcardona@terminus:/tmp$ python2.5 test.py
eval: 0
eval: 2
eval: 4
eval: 6
eval: 8
[0, 2, 4, 6, 8]
jcardona@terminus:/tmp$ python2.7 test.py
eval: 0
eval: 2
eval: 4
eval: 6
eval: 8
[0, 2, 4, 6, 8]
jcardona@terminus:/tmp$ python3.0 test.py
eval: 0
eval: 2
eval: 4
eval: 6
eval: 8
[0, 2, 4, 6, 8]

Well, it works for what i need, but is not very neat, and i think that
there it should be a formal way to act on the base indexing iterator,
such way exists? Is there a better approach to get what i need?

Thanks.

 

[Lie Ryan]

Hi,

I was trying to create a function that receive a generator and return
a list but that each elements were computed in a diferent core of my
machine. I start using islice function in order to split the job in a
way that if there is "n" cores each "i" core will compute the elements
i,i+n,i+2n,..., but islice has a weird (to me) behavior, look:

it's nothing weird, python just do what you're telling it to:

transform all x in X with f(x)

g = (f(x) for x in X) then slice the result of that
print(list(x for x in islice(g,0,None,2)))

what you want to do is to slice before you transform:eval: 0
eval: 2
eval: 4
eval: 6
eval: 8
[0, 2, 4, 6, 8]

islice execute the function at the generator and drop the elements
that aren't in the slice. I found that pretty weird, the way that i
see generators is like an association between and indexing set (an
iterator or another generator) and a computation that is made indexed
by the indexing set, and islice is like a "transformation" on the
indexing set,it doesn't matter the result of the function, the slice
should act only on the indexing set, some other "transformation" like
takewhile act on the result so, the execution it has to be made, but
in the islice, or other "transformation" that act only in the indexing
set, the function shouldn't be executed at each element, but only on
that new set that result of the application of the "transformation" on
the original set.

that seems like an extremely lazy evaluation, I don't know if even a
true lazy language do that. Python is a strict language, with a few
laziness provided by generators, in the end it's still a strict language.

Well, it works for what i need, but is not very neat, and i think that
there it should be a formal way to act on the base indexing iterator,
such way exists? Is there a better approach to get what i need?

Reverse your operation.

 

[Jorge Cardona]

it's nothing weird, python just do what you're telling it to:

transform all x in X with f(x)

then slice the result of that

When i wrote that first line of code i thought that i was creating a
generator that will later compute the elements of X with function f,
if i will like to transform them immediately i would use [] instead
(), so, the result is not where i want to slice, even so, is exactly
the same to slice, after or before, the only difference is that after
the compute i'm losing 5 in unnecessary execution of the function f.

what you want to do is to slice before you transform:eval: 0
eval: 2
eval: 4
eval: 6
eval: 8
[0, 2, 4, 6, 8]

What i want to do is a function that receive any kind of generator and
execute it in several cores (after a fork) and return the data, so, i
can't slice the set X before create the generator.

that seems like an extremely lazy evaluation, I don't know if even a true
lazy language do that. Python is a strict language, with a few laziness
provided by generators, in the end it's still a strict language.

Yes, it looks like lazy evaluation, but, i don't see why there is not
a better control over the iterable associated to a generator, even
with Python that is a strict language, it will increase the
functionality of it, and the performance too, imagine that you pass a
function that takes 1 sec in run, and for several reason you can't
slice before (as the smp function that i want to create), the final
performance with the actual islice it gets really reduced
Just creating the separation between those transformation that act on
the index(islice, or tee) on those that act on the result(dropwhile,
takewhile, etc.) the control could be fine enough to increase
usability (that's the way i think right now), and you will be able to
combine generator without lose performance.

Reverse your operation.

 

[Taylor]

2009/12/7 Lie Ryan <[email protected]>:

On 12/7/2009 7:22 AM, Jorge Cardona wrote:

it's nothing weird, python just do what you're telling it to:

transform all x in X with f(x)

then slice the result of that

When i wrote that first line of code i thought that i was creating a
generator that will later compute  the elements of X with function f,
if i will like to transform them immediately i would use [] instead
(), so, the result is not where i want to slice, even so, is exactly
the same to slice, after or before, the only difference is that after
the compute i'm losing 5 in unnecessary execution of the function f.

what you want to do is to slice before you transform:

g = (x for x in islice(X, 0, None, 2))
print(list(f(x) for x in g))

eval: 0
eval: 2
eval: 4
eval: 6
eval: 8
[0, 2, 4, 6, 8]

What i want to do is a function that receive any kind of generator and
execute it in several cores (after a fork) and return the data, so, i
can't slice the set X before create the generator.

that seems like an extremely lazy evaluation, I don't know if even a true
lazy language do that. Python is a strict language, with a few laziness
provided by generators, in the end it's still a strict language.

Yes, it looks like lazy evaluation, but, i don't see why there is not
a better control over the iterable associated to a generator, even
with Python that is a strict language, it will increase the
functionality of it, and the performance too, imagine that you pass a
function that takes 1 sec in run, and for several reason you can't
slice before (as the smp function that i want to create), the final
performance with the actual islice it gets really reduced
Just creating the separation between those transformation that act on
the index(islice, or tee) on those that act on the result(dropwhile,
takewhile, etc.) the control could be fine enough to increase
usability (that's the way i think right now), and you will be able to
combine generator without lose performance.

Reverse your operation.

What would you have your islice do for the following generator?

def fib(n):
a,b=0,1
for x in range(n):
a,b=b,a+b
yield a

In order for some value it yields to be correct, it needs execute all
the previous times. If it happens that some of the results aren't
used, they are thrown out. As is, using your islice (on, say, fib(10))
gives a KeyError for the key '.0'.
Point is, generators don't work that way. You aren't guaranteed to be
able to jump around, only to find the next value. If you need to split
a particular generator up into parts that can be computed separately,
your best bet is probably to rewrite that generator.

 

[Lie Ryan]

First, I apologize for rearranging your message out of order.

Yes, it looks like lazy evaluation, but, i don't see why there is not
a better control over the iterable associated to a generator, even
with Python that is a strict language, it will increase the
functionality of it, and the performance too, imagine that you pass a
function that takes 1 sec in run, and for several reason you can't
slice before (as the smp function that i want to create), the final
performance with the actual islice it gets really reduced
Just creating the separation between those transformation that act on
the index(islice, or tee) on those that act on the result(dropwhile,
takewhile, etc.) the control could be fine enough to increase
usability (that's the way i think right now), and you will be able to
combine generator without lose performance.

Theoretically yes, but the semantic of generators in python is they work
on an Iterable (i.e. objects that have __iter__), instead of a Sequence
(i.e. objects that have __getitem__). That means semantically,
generators would call obj.__iter__() and call the iter.__next__() and do
its operation for each items returned by the iterator's iterable's
__next__().

The lazy semantic would be hard to fit the current generator model
without changing the semantics of generator to require a Sequence that
supports indexing.

Yes, it looks like lazy evaluation, but, i don't see why there is not
a better control over the iterable associated to a generator, even
with Python that is a strict language

You can control the laziness by making it explicitly lazy:

from functools import partial
def f(x):
print("eval: %d"%x)
return x

X = range(10)
g = (partial(f, x) for x in X)

print(list(x() for x in islice(g,0,None,2)))
# # or without partial:
# g = ((lambda: f(x)) for x in X)
# print(list(f() for f in islice(g,0,None,2)))

In a default-strict language, you have to explicitly say if you want
lazy execution.

What i want to do is a function that receive any kind of generator and
execute it in several cores (after a fork) and return the data, so, i
can't slice the set X before create the generator.

beware that a generator's contract is to return a valid iterator *once*
only. You can use itertools.tee() to create more generators, but tee
built a list of the results internally.

 

[Jorge Cardona]

2009/12/7 Lie Ryan <[email protected]>:

On 12/7/2009 7:22 AM, Jorge Cardona wrote:

I was trying to create a function that receive a generator and return
a list but that each elements were computed in a diferent core of my
machine. I start using islice function in order to split the job in a
way that if there is "n" cores each "i" core will compute the elements
i,i+n,i+2n,..., but islice has a weird (to me) behavior, look:

it's nothing weird, python just do what you're telling it to:

transform all x in X with f(x)

g = (f(x) for x in X)

then slice the result of that

print(list(x for x in islice(g,0,None,2)))

When i wrote that first line of code i thought that i was creating a
generator that will later compute  the elements of X with function f,
if i will like to transform them immediately i would use [] instead
(), so, the result is not where i want to slice, even so, is exactly
the same to slice, after or before, the only difference is that after
the compute i'm losing 5 in unnecessary execution of the function f.

what you want to do is to slice before you transform:
g = (x for x in islice(X, 0, None, 2))
print(list(f(x) for x in g))
eval: 0
eval: 2
eval: 4
eval: 6
eval: 8
[0, 2, 4, 6, 8]

What i want to do is a function that receive any kind of generator and
execute it in several cores (after a fork) and return the data, so, i
can't slice the set X before create the generator.

islice execute the function at the generator and drop the elements
that aren't in the slice. I found that pretty weird, the way that i
see generators is like an association between and indexing set (an
iterator or another generator) and a computation that is made indexed
by the indexing set, and islice is like a "transformation" on the
indexing set,it doesn't matter the result of the function, the slice
should act only on the indexing set, some other "transformation" like
takewhile act on the result so, the execution it has to be made, but
in the islice, or other "transformation" that act only in the indexing
set, the function shouldn't be executed at each element, but only on
that new set that result of the application of the "transformation" on
the original set.

that seems like an extremely lazy evaluation, I don't know if even a true
lazy language do that. Python is a strict language, with a few laziness
provided by generators, in the end it's still a strict language.

Yes, it looks like lazy evaluation, but, i don't see why there is not
a better control over the iterable associated to a generator, even
with Python that is a strict language, it will increase the
functionality of it, and the performance too, imagine that you pass a
function that takes 1 sec in run, and for several reason you can't
slice before (as the smp function that i want to create), the final
performance with the actual islice it gets really reduced
Just creating the separation between those transformation that act on
the index(islice, or tee) on those that act on the result(dropwhile,
takewhile, etc.) the control could be fine enough to increase
usability (that's the way i think right now), and you will be able to
combine generator without lose performance.

Well, it works for what i need, but is not very neat, and i think that
there it should be a formal way to act on the base indexing iterator,
such way exists? Is there a better approach to get what i need?

Reverse your operation.

What would you have your islice do for the following generator?

def fib(n):
   a,b=0,1
   for x in range(n):
       a,b=b,a+b
       yield a

In order for some value it yields to be correct, it needs execute all
the previous times. If it happens that some of the results aren't
used, they are thrown out. As is, using your islice (on, say, fib(10))
gives a KeyError for the key '.0'.
Point is, generators don't work that way. You aren't guaranteed to be
able to jump around, only to find the next value. If you need to split
a particular generator up into parts that can be computed separately,
your best bet is probably to rewrite that generator.

yes, it doesn't work with that example, f_locals has only the n
argument. Let me rewrite the islice like this:

def islice(iterable, *args):
s = slice(*args)

# search the deepest iter (Base indexing set)
it = iterable
while hasattr(it, 'gi_frame') and ('.0' in it.gi_frame.f_locals):
it = it.gi_frame.f_locals['.0']

# Consume the base indexing set until the first element
for i in range(s.start):
it.next()

for e in iterable:
yield e
# Consume the base indexing set until the next element
for i in range(s.step-1):
next(it)

And then:

def fib(n):
a,b=0,1
for x in range(n):
a,b=b,a+b
yield a

g = fib(10)
print(list(x for x in islice(g,0, None,2)))

Will result in:
[1, 3, 8, 21, 55]

How I see that particular example is that the fibonacci generator is a
base indexing set, look that you didn't define it in terms of an
association with another set and a particular function, even when
there is a range(n) and a sum operation, I can't think at this moment
in a way to represent fibonacci in generator expression terms.

The fact that there is a state that has to be preserved between each
called item of the generator defined with fib(n) (in order to avoid a
recursive implementation) makes impossible to define the underlying
function as fib(n) because it really depends on that state
(fib(state)), so, a unique indexing set couldn't be used here without
maintain an auxiliary set that holds that state at each compute.
Even with fibonacci there is only need to maintain the state variable,
and create a full range of numbers without known a priori the n number
that maps to the result, with factorial one can hold an state that
represent the past factorial and call the function with a a present n.
Maintain the state variables is permitted by the definition of the
generator as a function with yield, and in your fib it's maintained in
a and b.

(((Is there any "clean" way to define a (fib(x) for x in count())
without falls in recursive errors and high memory usage?, but as a
generator expression, i can't think in one right now.)))

A new element is shown here (well, or I just note that at this
moment), is that a generator defined as a function with yield could
not be seen as an associations between an indexing set and a function,
but just as an indexing set of new generators as in:

g = (f(x) for x in fib(10))
print(list(x for x in islice(g,0, None,2)))

that result in:

eval: 1
eval: 2
eval: 5
eval: 13
eval: 34
[1, 2, 5, 13, 34]

I just download the unittests for itertools from svn of python2.7:

test_islice (__main__.TestGC) ... ok

and adding islice above:

test_islice (__main__.TestGC) ... ok

 

[Jorge Cardona]

First, I apologize for rearranging your message out of order.



Theoretically yes, but the semantic of generators in python is they work on
an Iterable (i.e. objects that have __iter__), instead of a Sequence (i.e..
objects that have __getitem__). That means semantically, generators would
call obj.__iter__() and call the iter.__next__() and do its operation for
each items returned by the iterator's iterable's __next__().

The lazy semantic would be hard to fit the current generator model without
changing the semantics of generator to require a Sequence that supports
indexing.

Why?

The goal is add a formal way to separate the transformation of a
generator in those that act on the indexing set and those that act on
the result set.

Well, a little (and not so elaborated) example could be:

from itertools import islice


class MyGenerator:
def __init__(self, function, indexing_set):
self._indexing_set = (x for x in indexing_set)
self._function = function

def indexing_set(self):
return (x for x in self._indexing_set)

def result_set(self):
return (self._function(x) for x in self.indexing_set())

def function(self):
return self._function

def f(x):
print("eval: %d"%x)
return x

def myslice(iterable, *args):
return MyGenerator(iterable.f, islice(iterable.indexing_set(),*args))


g = MyGenerator(f, xrange(10))
print(list(g.result_set()))

g = MyGenerator(f, xrange(10))
new_g = myslice(g,0,None,2)
print(list(new_g.result_set()))

that returns:

eval: 0
eval: 1
eval: 2
eval: 3
eval: 4
eval: 5
eval: 6
eval: 7
eval: 8
eval: 9
[0, 1, 2, 3, 4, 5, 6, 7, 8, 9]
eval: 0
eval: 2
eval: 4
eval: 6
eval: 8
[0, 2, 4, 6, 8]

I don't see why is needed add a sequence to support the indexing, but
some separation feature of the base components of the generator
(function, indexing_set).

You can control the laziness by making it explicitly lazy:

from functools import partial
def f(x):
   print("eval: %d"%x)
   return x

X = range(10)
g = (partial(f, x) for x in X)

print(list(x() for x in islice(g,0,None,2)))
# # or without partial:
# g = ((lambda: f(x)) for x in X)
# print(list(f() for f in islice(g,0,None,2)))

I keep here the problem in that i shouldn't be able to define the
original generator because the function receive the already defined
generator.

In a default-strict language, you have to explicitly say if you want lazy
execution.


beware that a generator's contract is to return a valid iterator *once*
only. You can use itertools.tee() to create more generators, but tee built a
list of the results internally.

Oh, yes, i used tee first, but i note then that I wasn't using the
same iterator in the same process, so, when the fork is made I can
use the initial generator in different processes without this problem,
so tee is not necessary in this case.

 

[Lie Ryan]

Why?

The goal is add a formal way to separate the transformation of a
generator in those that act on the indexing set and those that act on
the result set.

<snip code>

Well, that's just a pretty convoluted way to reverse the order of the
operation (hey, why not reverse it altogether?). Nevertheless, it still
requires a semantic change in that all generators must be able to
produce the underlying stream, which is not always the case. Take this
small example:

def foo(x):
i = 0
while True:
yield i
i += 1

What would you propose the .indexing_set to be? Surely not the same as
..result_set? How would you propose to skip over and islice such
generator? without executing the in-betweens?

I keep here the problem in that i shouldn't be able to define the
original generator because the function receive the already defined
generator.


Oh, yes, i used tee first, but i note then that I wasn't using the
same iterator in the same process, so, when the fork is made I can
use the initial generator in different processes without this problem,
so tee is not necessary in this case.

You would have to change tee as well:
.... print('eval: %s' % x)
.... return x + 1
....

>>> l = [1, 2, 3, 4, 5, 6, 7]
>>> it = iter(l)
>>> it = (foo(x) for x in it)
>>> a, b = itertools.tee(it)
>>> # now on to you
>>> it_a = itertools.islice(a, 0, None, 2)
>>> it_b = itertools.islice(b, 1, None, 2)
>>> next(it_b)

eval: 1
eval: 2
3eval: 3
eval: 4
5eval: 5
eval: 6
7