sort order for strings of digits

[djc]

I learn lots of useful things from the list, some not always welcome. No
sooner had I found a solution to a minor inconvenience in my code, than
a recent thread here drew my attention to the fact that it will not work
for python 3. So suggestions please:

TODO 2012-10-22: sort order numbers first then alphanumeric('a', 'ab', 'acd', 'bcd', '1a', 'a1', '222 bb', 'b a 4')

sorted(n) ['1', '10', '101', '13', '2', '2000', '3', '31', '40']
sorted(s) ['1a', '222 bb', 'a', 'a1', 'ab', 'acd', 'b a 4', 'bcd']
sorted(n+s)

['1', '10', '101', '13', '1a', '2', '2000', '222 bb', '3', '31', '40',
'a', 'a1', 'ab', 'acd', 'b a 4', 'bcd']



Possibly there is a better way but for Python 2.7 this gives the
required result

Python 2.7.3 (default, Sep 26 2012, 21:51:14)
[1, 2, 3, 10, 13, 31, 40, 101, 2000, '1a', '222 bb', 'a', 'a1', 'ab',
'acd', 'b a 4', 'bcd']


[str(x) for x in sorted(int(x) if x.isdigit() else x for x in n+s)]
['1', '2', '3', '10', '13', '31', '40', '101', '2000', '1a', '222 bb',
'a', 'a1', 'ab', 'acd', 'b a 4', 'bcd']


But not for Python 3
Python 3.2.3 (default, Oct 19 2012, 19:53:16)
['1', '10', '101', '13', '1a', '2', '2000', '222 bb', '3', '31', '40',
'a', 'a1', 'ab', 'acd', 'b a 4', 'bcd']
Traceback (most recent call last):

The best I can think of is to split the input sequence into two lists,
sort each and then join them.

 

[Hans Mulder]

Python 3.2.3 (default, Oct 19 2012, 19:53:16)
['1', '10', '101', '13', '1a', '2', '2000', '222 bb', '3', '31', '40',
'a', 'a1', 'ab', 'acd', 'b a 4', 'bcd']
Traceback (most recent call last):

['1', '10', '101', '13', '1a', '2', '2000', '222 bb', '3', '31', '40',
'a', 'a1', 'ab', 'acd', 'b a 4', 'bcd']

The best I can think of is to split the input sequence into two lists,
sort each and then join them.

That might well be the most readable solution.


Hope this helps,

-- HansM

 

[Ian Kelly]

The best I can think of is to split the input sequence into two lists, sort
each and then join them.

In the example you have given they already seem to be split, so you
could just do:

sorted(n, key=int) + sorted(s)

If that's not really the case, then you could construct (str, int)
tuples as sort keys:

sorted(n+s, key=lambda x: ('', int(x)) if x.isdigit() else (x, -1))

Note that the empty string sorts before all numbers here, which may or
may not be desirable.

 

[Mark Lawrence]

Why -- I doubt Python 3.x .sort() and sorted() have removed the
optional key and cmp keywords.

Nope. I'm busy porting my own code from 2.7 to 3.3 and cmp seems to be
very dead.

This doesn't help either.

c:\Users\Mark\Cash\Python>2to3.py
Traceback (most recent call last):
File "C:\Python33\Tools\Scripts\2to3.py", line 3, in <module>
from lib2to3.main import main
ImportError: No module named main

 

[Dennis Lee Bieber]

Use functools.cmp_to_key for porting cmp functions. "sort(x, my_cmp)"
becomes "sort(x, key=cmp_to_key(my_cmp))"

The cmp builtin is also gone. If you need it, the suggested replacement
for "cmp(a, b)" is "(b < a) - (a < b)".

OUCH... Just another reason for my to hang onto the 2.x series as
long as possible (I only installed 2.7 this summer, I'd been using
2.5... And a project at my former employment was still running 2.3 and
having conflicts with a program with a binary extension written for 2.2
-- and we couldn't find the source for the extension to rebuild it!)

 

[Steven D'Aprano]

The best I can think of is to split the input sequence into two lists,
sort each and then join them.

According to your example code, you don't have to split the input because
you already have two lists, one filled with numbers and one filled with
strings.

But I think that what you actually have is a single list of strings, and
you are supposed to sort the strings such that they come in numeric order
first, then alphanumerical. E.g.:

['9', '1000', 'abc2', '55', '1', 'abc', '55a', '1a']
=> ['1', '1a', '9', '55', '55a', '1000', 'abc', 'abc2']

At least that is what I would expect as the useful thing to do when
sorting.

The trick is to take each string and split it into a leading number and a
trailing alphanumeric string. Either part may be "empty". Here's a pure
Python solution:

from sys import maxsize # use maxint in Python 2
def split(s):
for i, c in enumerate(s):
if not c.isdigit():
break
else: # aligned with the FOR, not the IF
return (int(s), '')
return (int(s[:i] or maxsize), s[i:])

Now sort using this as a key function:

py> L = ['9', '1000', 'abc2', '55', '1', 'abc', '55a', '1a']
py> sorted(L, key=split)
['1', '1a', '9', '55', '55a', '1000', 'abc', 'abc2']


The above solution is not quite general:

* it doesn't handle negative numbers or numbers with a decimal point;

* it doesn't handle the empty string in any meaningful way;

* in practice, you may or may not want to ignore leading whitespace,
or trailing whitespace after the number part;

* there's a subtle bug if a string contains a very large numeric prefix,
finding and fixing that is left as an exercise.

 

[Steven D'Aprano]

OUCH... Just another reason for my to hang onto the 2.x series as
long as possible

On the contrary. If you are using cmp with sort, your sorts are slow, and
you should upgrade to using a key function as soon as possible.

For small lists, you may not notice, but for large lists using a
comparison function is a BAD IDEA.

Here's an example: sorting a list of numbers by absolute value.

py> L = [5, -6, 1, -2, 9, -8, 4, 3, -7, 2, -3]
py> sorted(L, key=abs)
[1, -2, 2, 3, -3, 4, 5, -6, -7, -8, 9]
py> sorted(L, lambda a, b: cmp(abs(a), abs(b)))
[1, -2, 2, 3, -3, 4, 5, -6, -7, -8, 9]

But the amount of work done is radically different. Let's temporarily
shadow the built-ins with patched versions:

py> _abs = abs
py> _abs, _cmp = abs, cmp
py> c1 = c2 = 0
py> def abs(x):
.... global c1
.... c1 += 1
.... return _abs(x)
....
py> def cmp(a, b):
.... global c2
.... c2 += 1
.... return _cmp(a, b)
....

Now we can see just how much work is done under the hood using a key
function vs a comparison function:

py> sorted(L, key=abs)
[1, -2, 2, 3, -3, 4, 5, -6, -7, -8, 9]
py> c1
11

So the key function is called once for each item in the list. But:


py> c1 = 0 # reset the count
py> sorted(L, lambda a, b: cmp(abs(a), abs(b)))
[1, -2, 2, 3, -3, 4, 5, -6, -7, -8, 9]
py> c1, c2
(54, 27)

The comparison function is called 27 times for a list of nine items (a
average of 2.5 calls to cmp per item), and abs is called twice for each
call to cmp. (Well, duh.)

If the list is bigger, it gets worse:

py> c2 = 0
py> x = sorted(L*10, lambda a, b: cmp(abs(a), abs(b)))
py> c2
592

That's an average of 5.4 calls to cmp per item. And it gets even worse as
the list gets bigger.

As your lists get bigger, the amount of work done calling the comparison
function gets ever bigger still. Sorting large lists with a comparison
function is SLOOOW.

py> del abs, cmp # remove the monkey-patched versions
py> L = L*1000000
py> with Timer():
.... x = sorted(L, key=abs)
....
time taken: 9.165448 seconds
py> with Timer():
.... x = sorted(L, lambda a, b: cmp(abs(a), abs(b)))
....
time taken: 63.579679 seconds


The Timer() context manager used can be found here:

http://code.activestate.com/recipes/577896

 

[DJC]

According to your example code, you don't have to split the input because
you already have two lists, one filled with numbers and one filled with
strings.

Sorry for the confusion, the pair of strings was just a way of testing
variations on the input. So a sequence with any combination of strings
that can be read as numbers and strings of chars that don't look like
numbers (even if that string includes digits) is the expected input

But I think that what you actually have is a single list of strings, and
you are supposed to sort the strings such that they come in numeric order
first, then alphanumerical. E.g.:

['9', '1000', 'abc2', '55', '1', 'abc', '55a', '1a']
=> ['1', '1a', '9', '55', '55a', '1000', 'abc', 'abc2']

Not quite, what I want is to ensure that if the strings look like
numbers they are placed in numerical order. ie 1 2 3 10 100 not 1 10 100
2 3. Cases where a string has some leading digits can be treated as
strings like any other.

At least that is what I would expect as the useful thing to do when
sorting.

Well it depends on the use case. In my case the strings are column and
row labels for a report. I want them to be presented in a convenient to
read sequence. Which the lexical sorting of the strings that look like
numbers is not. I want a reasonable do-what-i-mean default sort order
that can handle whatever strings are used.

The trick is to take each string and split it into a leading number and a
trailing alphanumeric string. Either part may be "empty". Here's a pure
Python solution:

from sys import maxsize # use maxint in Python 2
def split(s):
for i, c in enumerate(s):
if not c.isdigit():
break
else: # aligned with the FOR, not the IF
return (int(s), '')
return (int(s[:i] or maxsize), s[i:])

Now sort using this as a key function:

py> L = ['9', '1000', 'abc2', '55', '1', 'abc', '55a', '1a']
py> sorted(L, key=split)
['1', '1a', '9', '55', '55a', '1000', 'abc', 'abc2']


The above solution is not quite general:

* it doesn't handle negative numbers or numbers with a decimal point;

* it doesn't handle the empty string in any meaningful way;

* in practice, you may or may not want to ignore leading whitespace,
or trailing whitespace after the number part;

* there's a subtle bug if a string contains a very large numeric prefix,
finding and fixing that is left as an exercise.

That looks more than general enough for my purposes! I will experiment
along those lines, thank you.

 

[Mark Lawrence]

Perhaps you have a sys.path conflict?

Correct, now fixed, thanks.

Use functools.cmp_to_key for porting cmp functions. "sort(x, my_cmp)"
becomes "sort(x, key=cmp_to_key(my_cmp))"

The cmp builtin is also gone. If you need it, the suggested replacement
for "cmp(a, b)" is "(b < a) - (a < b)".

As it's my own small code base I've blown away all references to cmp,
it's rich comparisons all the way.

 

[Chris Angelico]

On the contrary. If you are using cmp with sort, your sorts are slow, and
you should upgrade to using a key function as soon as possible.

But cmp_to_key doesn't actually improve anything. So I'm not sure how
Py3 has achieved anything; Py2 supported key-based sorting already.

ChrisA

 

[Arnaud Delobelle]

The trick is to take each string and split it into a leading number and a
trailing alphanumeric string. Either part may be "empty". Here's a pure
Python solution:

from sys import maxsize # use maxint in Python 2
def split(s):
for i, c in enumerate(s):
if not c.isdigit():
break
else: # aligned with the FOR, not the IF
return (int(s), '')
return (int(s[:i] or maxsize), s[i:])

Now sort using this as a key function:

py> L = ['9', '1000', 'abc2', '55', '1', 'abc', '55a', '1a']
py> sorted(L, key=split)
['1', '1a', '9', '55', '55a', '1000', 'abc', 'abc2']

You don't actually need to split the string, it's enough to return a
pair consisting of the number of leading digits followed by the string
as the key. Here's an implementation using takewhile:
.... return sum(1 for c in takewhile(str.isdigit, s)) or 1000, s
....

L = ['9', '1000', 'abc2', '55', '1', 'abc', '55a', '1a']
sorted(L, key=prefix)

['1', '1a', '9', '55', '55a', '1000', 'abc', 'abc2']

Here's why it works:
[(1, '9'), (4, '1000'), (1000, 'abc2'), (2, '55'), (1, '1'), (1000,
'abc'), (2, '55a'), (1, '1a')]

 

[wxjmfauth]

Le mercredi 31 octobre 2012 16:17:19 UTC+1, djc a écrit :

I learn lots of useful things from the list, some not always welcome. No

sooner had I found a solution to a minor inconvenience in my code, than

a recent thread here drew my attention to the fact that it will not work

for python 3. So suggestions please:



TODO 2012-10-22: sort order numbers first then alphanumeric

('1', '10', '101', '3', '40', '31', '13', '2', '2000')

('a', 'ab', 'acd', 'bcd', '1a', 'a1', '222 bb', 'b a 4')



['1', '10', '101', '13', '2', '2000', '3', '31', '40']

['1a', '222 bb', 'a', 'a1', 'ab', 'acd', 'b a 4', 'bcd']

['1', '10', '101', '13', '1a', '2', '2000', '222 bb', '3', '31', '40',

'a', 'a1', 'ab', 'acd', 'b a 4', 'bcd']







Possibly there is a better way but for Python 2.7 this gives the

required result



Python 2.7.3 (default, Sep 26 2012, 21:51:14)



[1, 2, 3, 10, 13, 31, 40, 101, 2000, '1a', '222 bb', 'a', 'a1', 'ab',

'acd', 'b a 4', 'bcd']





[str(x) for x in sorted(int(x) if x.isdigit() else x for x in n+s)]

['1', '2', '3', '10', '13', '31', '40', '101', '2000', '1a', '222 bb',

'a', 'a1', 'ab', 'acd', 'b a 4', 'bcd']





But not for Python 3

Python 3.2.3 (default, Oct 19 2012, 19:53:16)



['1', '10', '101', '13', '1a', '2', '2000', '222 bb', '3', '31', '40',

'a', 'a1', 'ab', 'acd', 'b a 4', 'bcd']



Traceback (most recent call last):

File "<stdin>", line 1, in <module>

TypeError: unorderable types: str() < int()



The best I can think of is to split the input sequence into two lists,

sort each and then join them.





--

djc

# Py 3.2.3
z = ['1', '10', '101', '13', '1a', '2', '2000',

.... '222 bb', '3', '31', '40', 'a', 'a1', 'ab',
.... 'acd', 'b a 4', 'bcd'
.... ]

n, s = [], []
for e in z:

.... if e.isdigit():
.... n.append(int(e))
.... else:
.... s.append(e)
....

n.sort()
s.sort()
ns = [str(e) for e in n]
ns.extend(s)
ns

['1', '2', '3', '10', '13', '31', '40', '101', '2000', '1a',
'222 bb', 'a', 'a1', 'ab', 'acd', 'b a 4', 'bcd']

jmf

 

[Steven D'Aprano]

But cmp_to_key doesn't actually improve anything. So I'm not sure how
Py3 has achieved anything; Py2 supported key-based sorting already.

Yes, but there is a lot of old code pre-dating key-based sorts. There's
also some examples of code where it isn't obvious how to write it as a
key-based sort, but a comparison function is simple. And people coming
from other languages that only support comparison-based sorts (C?) will
probably continue with what they know.

Even though key-based sorting is better, there's a lot of comparison
sorting that falls under "if it ain't broke, don't fix it".

So even though key-based sorts are better, there are still comparison-
based sorts in the wild. Python 2 has to support them. Python 3, which is
allowed to break backwards compatibility, does not. So when porting to 3,
you have to change the sorts.

Most of the time it is simple to convert a comparison-based sort to a key-
based sort. For the cases where you either can't come up with a good key
function yourself, or were you want to do so mechanically, Python
provides cmp_to_key.