FAQ
Hello all

I am testing my code with list comprehensions against for loops.

the loop:

dipList=[float(val[1]) for val in datalist]
dip1=[]
for dp in dipList:
if dp == 90:
dip1.append(dp - 0.01)
else:
dip1.append(dp)

listcomp:

dipList=[float(val[1]) for val in datalist]
dip1=[(dp, dp-0.01)[dp=�.0] for dp in dipList]

Tenting the time spent by each approach (using time.clock()), with a
file with about 100,000 entries, I get 0.03s for the loop and 0.05s
for the listcomp.

thoughts?

TIA
Carlos

## Search Discussions

•  at Dec 17, 2009 at 5:42 pm ⇧

* Carlos Grohmann:
Hello all

I am testing my code with list comprehensions against for loops.

the loop:

dipList=[float(val[1]) for val in datalist]
dip1=[]
for dp in dipList:
if dp == 90:
dip1.append(dp - 0.01)
else:
dip1.append(dp)

listcomp:

dipList=[float(val[1]) for val in datalist]
dip1=[(dp, dp-0.01)[dp=�.0] for dp in dipList]

Tenting the time spent by each approach (using time.clock()), with a
file with about 100,000 entries, I get 0.03s for the loop and 0.05s
for the listcomp.

thoughts?
In the list comprehension you're constructing n tuples that you're not
constructing in the loop.

Have you tried this with

dip1 = [dp - 0.01 if dp == 90 else dp for dp in dipList]

?

Cheers & hth.,

- Alf
•  at Dec 17, 2009 at 6:04 pm ⇧

Have you tried this with

? ?dip1 = [dp - 0.01 if dp == 90 else dp for dp in dipList]
Yes that is better! many thanks!
•  at Dec 18, 2009 at 6:49 pm ⇧

On 17 Des, 18:42, "Alf P. Steinbach" wrote:

Have you tried this with

? ?dip1 = [dp - 0.01 if dp == 90 else dp for dp in dipList]
And for comparison with map:

map(lambda dp: dp - 0.01 if dp == 90 else dp, dipList)
•  at Dec 18, 2009 at 6:44 pm ⇧

On 17 Des, 18:37, Carlos Grohmann wrote:

Tenting the time spent by each approach (using time.clock()), with a
file with about 100,000 entries, I get 0.03s for the loop and 0.05s
for the listcomp.

thoughts?
Anything else being equal, list comprehensions will be the faster
becuase they incur fewer name and attribute lookups. It will be the
same as the difference between a for loop and a call to map. A list
comprehension is basically an enhancement of map.
•  at Dec 19, 2009 at 1:28 am ⇧

Tenting the time spent by each approach (using time.clock()), with a
file with about 100,000 entries, I get 0.03s for the loop and 0.05s
for the listcomp.
Anything else being equal, list comprehensions will be the faster
becuase they incur fewer name and attribute lookups. It will be the
same as the difference between a for loop and a call to map. A list
comprehension is basically an enhancement of map.
Not so. If you use the "dis" module to peek at the bytecode generated
for a list comprehension, you'll see it's very similar to that generated
for an explicit for-loop. The byte-code for a call to map is very
different.

Basically: both a for-loop and a list-comp do the looping in python
bytecode, while a call to map will do the actual looping in C.
def comper():
... return [i*2 for i in xrange(10)]
...
dis.dis(comper)
2 0 BUILD_LIST 0
3 DUP_TOP
4 STORE_FAST 0 (_[1])
13 CALL_FUNCTION 1
16 GET_ITER
17 FOR_ITER 17 (to 37)
20 STORE_FAST 1 (i)
32 BINARY_MULTIPLY
33 LIST_APPEND
34 JUMP_ABSOLUTE 17
37 DELETE_FAST 0 (_[1])
40 RETURN_VALUE

def maper():
... return map(lambda i: i*2,xrange(10))
...
dis.dis(maper)
3 LOAD_CONST 1 (<code object ...)
6 MAKE_FUNCTION 0
15 CALL_FUNCTION 1
18 CALL_FUNCTION 2
21 RETURN_VALUE
>>>

Cheers,

Ryan

--
Ryan Kelly
http://www.rfk.id.au | This message is digitally signed. Please visit
ryan at rfk.id.au | http://www.rfk.id.au/ramblings/gpg/ for details

-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 197 bytes
Desc: This is a digitally signed message part
URL: <http://mail.python.org/pipermail/python-list/attachments/20091219/32b992f8/attachment-0001.pgp>
•  at Dec 19, 2009 at 5:49 am ⇧

Ryan Kelly wrote:
Someone else wrote:
It will be the
same as the difference between a for loop and a call to map.
Not so. If you use the "dis" module to peek at the bytecode generated
for a list comprehension, you'll see it's very similar to that generated
for an explicit for-loop.
The usual advice is that if you have a built-in function that
does what you want done for each element, then using map() is
probably the fastest way.

However, if you need to create a Python function to pass to
map(), the list comprehension may well be faster, because it
avoids the cost of a Python function call per element.

--
Greg
•  at Dec 19, 2009 at 8:04 am ⇧

On Sat, 19 Dec 2009 12:28:32 +1100, Ryan Kelly wrote:

Anything else being equal, list comprehensions will be the faster
becuase they incur fewer name and attribute lookups. It will be the
same as the difference between a for loop and a call to map. A list
comprehension is basically an enhancement of map.
Not so. If you use the "dis" module to peek at the bytecode generated
for a list comprehension, you'll see it's very similar to that generated
for an explicit for-loop. The byte-code for a call to map is very
different.
"Very similar" and "very different" byte-code mean very little regarding
speed.

Basically: both a for-loop and a list-comp do the looping in python
bytecode, while a call to map will do the actual looping in C.
This is a classic example of the confirmation fallacy -- if you say that
for-loops and list-comps are very similar, you need to actually check the
byte-code of both. You don't. You need to compare the byte-code of all
three operations, not just two of them, e.g.:

dis.dis(compile("map(f, seq)", '', 'exec'))
dis.dis(compile("[f(x) for x in seq]", '', 'exec'))
dis.dis(compile("L = []\nfor x in seq: L.append(f(x))", '', 'exec'))

But in fact just looking at the byte-code isn't helpful, because it tells
you nothing about the relative speed of each operation. You need to
actually time the operations.
from timeit import Timer
t1 = Timer("map(len, 'abcdefgh')", setup='')
t2 = Timer("[len(c) for c in 'abcdefgh']", setup='')
t3 = Timer("""L = []
... for c in 'abcdefgh':
... L.append(len(c))
... """, setup='')
min(t1.repeat())
3.9076540470123291
min(t2.repeat())
4.5931642055511475
min(t3.repeat())
7.4744069576263428

So, on my PC, with Python 2.5, with this example, a for-loop is about 60%
slower than a list comp and about 90% slower than map; the list comp is

But that only holds for *that* example. Here's another one:

def f(x):
... return 1+2*x+3*x**2
...
values = [1,2,3,4,5,6]
t1 = Timer("map(f, values)", setup='from __main__ import f, values')
t2 = Timer("[f(x) for x in values]",
... setup='from __main__ import f, values')
t3 = Timer("""L = []
... for x in values:
... L.append(f(x))
... """, setup='from __main__ import f, values')
min(t1.repeat())
7.0339860916137695
min(t2.repeat())
6.8053178787231445
min(t3.repeat())
9.1957418918609619

For this example map and the list comp are nearly the same speed, with
map slightly slower; but the for-loop is still significantly worse.

Of course, none of these timing tests are terribly significant. The
actual difference in time is of the order of a millionth of a second per
call to map compared to the list comp or the for-loop, for these small
examples. Most of the time you shouldn't care about time differences of
that magnitude, and write whatever is easiest.

--
Steven
•  at Dec 19, 2009 at 2:18 pm ⇧

On 19 Des, 02:28, Ryan Kelly wrote:

Not so. ?If you use the "dis" module to peek at the bytecode generated
for a list comprehension, you'll see it's very similar to that generated
for an explicit for-loop. ?The byte-code for a call to map is very
different.
First, you failed to realize that the bytecode is different because
map is doing the work in C.

Second, you did not provide bytecode for the for-loop.
•  at Dec 18, 2009 at 6:50 pm ⇧

On Dec 17, 9:37?am, Carlos Grohmann wrote:
Tenting the time spent by each approach (using time.clock()), with a
file with about 100,000 entries, I get 0.03s for the loop and 0.05s
for the listcomp.

thoughts?
You shouldn't trust your intuition in things like this. Some features
were added to Python to make writing easier, not to make it run
faster. This time your intuition was correct. Next time, who knows?

Carl Banks
•  at Dec 18, 2009 at 6:55 pm ⇧

On 17 Des, 18:37, Carlos Grohmann wrote:

Tenting the time spent by each approach (using time.clock()), with a
file with about 100,000 entries, I get 0.03s for the loop and 0.05s
for the listcomp.

thoughts?
Let me ask a retoric question:

- How much do you really value 20 ms of CPU time?
•  at Dec 18, 2009 at 8:44 pm ⇧

On Fri, Dec 18, 2009 at 11:55 AM, sturlamolden wrote:
On 17 Des, 18:37, Carlos Grohmann wrote:

Tenting the time spent by each approach (using time.clock()), with a
file with about 100,000 entries, I get 0.03s for the loop and 0.05s
for the listcomp.

thoughts?
Let me ask a retoric question:

- How much do you really value 20 ms of CPU time?
If it takes 1 nanosecond to execute a single instruction then 20
milliseconds represents 20 million instructions. Therefore I value 20ms of
CPU time very much indeed.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.python.org/pipermail/python-list/attachments/20091218/137c15ce/attachment-0001.htm>

## Related Discussions

Discussion Overview
 group python-list categories python posted Dec 17, '09 at 5:37p active Dec 19, '09 at 2:18p posts 12 users 8 website python.org

### 8 users in discussion

Content

People

Support

Translate

site design / logo © 2022 Grokbase