Python Rocks! and other rants

Weblog of Kent S Johnson

#72 2006-04-06 15:18:37

lxml for Windows!

lxml now has a Windows binary available. I'm thrilled! lxml is a fast, Pythonic XML library with full XPath support. I don't know of any other Python XML library that has all three of these attributes:

• ElementTree is fast and Pythonic but only supports a limited subset of XPath.
• Amara is Pythonic and supports XPath but in my limited experience it seems quite slow.
• PyXML uses w3c DOM models which are clunky and un-Pythonic.

I have raved many times about dom4j, an XML library for Java which is easy to use and includes XPath support. I have long wished for a Python package as powerful and easy-to-use. I'm optimistic that lxml will fill the bill.

Categories: Python

#70 2005-12-28 10:49:20

Why I love Python 5

// The XPathAPI is in different packages in Java 1.4 and 1.5.
// Use introspection to find the right one
private static Method __selectSingleNode;

static {
    // Look for the XPathAPI class in two places
    Class XPathAPI = null;
    try {
        XPathAPI = Class.forName("org.apache.xpath.XPathAPI");
    } catch (ClassNotFoundException e) {
        try {
            XPathAPI = Class.forName("com.sun.org.apache.xpath.internal.XPathAPI");
        } catch (ClassNotFoundException e1) {
        }
    }

    // Get the methods we support
    try {
        __selectSingleNode =
          XPathAPI.getMethod("selectSingleNode",
                             new Class[] { Node.class, String.class} );
    } catch (SecurityException e) {
    } catch (NoSuchMethodException e) {
    }
}

/** XPathAPI.selectSingleNode */
public static Node selectSingleNode(Node node, String xpath) {
    try {
        return (Node)__selectSingleNode.invoke(null, new Object[] { node, xpath });
    } catch (IllegalArgumentException e) {
        e.printStackTrace();
    } catch (IllegalAccessException e) {
        e.printStackTrace();
    } catch (InvocationTargetException e) {
        e.printStackTrace();
    }
    return null;
}

try:
  import org.apache.xpath.XPathAPI as XPathAPI
except ImportError:
  import com.sun.org.apache.xpath.internal.XPathAPI as XPathAPI

__selectSingleNode = getattr(XPathAPI, 'selectSingleNode')

try:
  __selectSingleNode = ...
except (SecurityException, NoSuchMethodException), e:
  e.printStackTrace()

Categories: Java, Python

#69 2005-12-06 22:37:36

Simple itertools.groupby() example

Suppose you have a (sorted) list of dicts containing the names of cities and states, and you want to print them out with headings by state:

>>> cities = [
...     { 'city' : 'Harford', 'state' : 'Connecticut' },
...     { 'city' : 'Boston', 'state' : 'Massachusetts' },
...     { 'city' : 'Worcester', 'state' : 'Massachusetts' },
...     { 'city' : 'Albany', 'state' : 'New York' },
...     { 'city' : 'New York City', 'state' : 'New York' },
...     { 'city' : 'Yonkers', 'state' : 'New York' },
... ]

First let me explain operator.itemgetter(). This function is a factory for new functions. It creates functions that access items using a key. In this case I will use it to create a function to access the 'state' item of each record:

>>> from operator import itemgetter
>>> getState = itemgetter('state')
>>> getState
<operator.itemgetter object at 0x00A31D90>
>>> getState(cities[0])
'Connecticut'
>>> [ getState(record) for record in cities ]
['Connecticut', 'Massachusetts', 'Massachusetts', 'New York', 'New York', 'New York']

So the value returned by itemgetter('state') is a function that accepts a dict as an argument and returns the 'state' item of the dict. Calling getState(d) is the same as writing d['state'].

What does this have to do with itertool.groupby()?

>>> from itertools import groupby
>>> help(groupby)
Help on class groupby in module itertools:

class groupby(__builtin__.object)
|  groupby(iterable[, keyfunc]) -> create an iterator which returns
|  (key, sub-iterator) grouped by each value of key(value).

groupby() takes an optional second argument which is a function to extract keys from the data. getState() is just the function we need.

>>> groups = groupby(cities, getState)
>>> groups
<itertools.groupby object at 0x00A88300>

Hmm. That's a bit opaque. groupby() returns an iterator. Each item in the iterator is a pair of (key, group). Let's take a look:

>>> for key, group in groups:
...   print key, group
...
Connecticut <itertools._grouper object at 0x0089D0F0>
Massachusetts <itertools._grouper object at 0x0089D0C0>
New York <itertools._grouper object at 0x0089D0F0>

Hmm. Still a bit opaque :-) The key part is clear - that's the state, extracted with getState - but group is another iterator. One way to look at it's contents is to use a nested loop. Note that I have to call groupby() again, the old iterator was consumed by the last loop:

>>> for key, group in groupby(cities, getState):
...   print key
...   for record in group:
...     print record
...
Connecticut
{'city': 'Harford', 'state': 'Connecticut'}
Massachusetts
{'city': 'Boston', 'state': 'Massachusetts'}
{'city': 'Worcester', 'state': 'Massachusetts'}
New York
{'city': 'Albany', 'state': 'New York'}
{'city': 'New York City', 'state': 'New York'}
{'city': 'Yonkers', 'state': 'New York'}

Well, that makes more sense! And it's not too far from the original requirement, we just need to pretty up the output a bit. How about this:

>>> for key, group in groupby(cities, getState):
...   print 'State:', key
...   for record in group:
...     print '   ', record['city']
...
State: Connecticut
     Harford
State: Massachusetts
     Boston
     Worcester
State: New York
     Albany
     New York City
     Yonkers

Other than misspelling Hartford (sheesh, and I grew up in Connecticut!) that's not too bad!

Categories: Python

#66 2005-10-24 21:14:40

What's so great about Ruby?

I'm reading Bruce Tate's latest book, Beyond Java. In it, he argues that Java has become overgrown, unwieldy and vulnerable for replacement in many applications. Prime candidates to replace it are the dynamic languages, particularly Ruby.

As a staunch Python advocate I read his description of Ruby with interest. Most Ruby features that he thinks are cool are available in Python in some form. Some are considered wizard-level tricks in Python instead of the mainstream practices they seem to be in Ruby.

For example, in Ruby you can easily add to a class definition. You just declare the class again and extend the definition. This works even for built-in classes. The Ruby approach is conceptually very simple--it reuses the class definition syntax. In Python you can add methods to a class after it is defined by adding attributes to the class. Python's approach is fairly obscure - getting it right can take a few tries.

Ruby allows mixins--class fragments that can be added to a class definition to extend it. Python can do the same with multiple inheritance, at the time a class is defined, or by appending to __bases__, which might be considered a hack.

Ruby has support for creating aliases of methods and replacing them, and this is considered a good thing. In Python this is called monkeypatching and is generally frowned on.

So there is not that much difference in capability. In Ruby some of these things are easier, and I don't discount that. But the main difference seems to be philosophical or cultural. In Python classes are thought of as fairly static--once you create it, it's done. Metaprogramming tricks are used during class creation to get some special effect, or to meet some unusual need.

In Ruby, though, classes are thought of as malleable. A class definition is just a starting point for the full functionality of the class. It's an interesting way of looking at it.

Categories: Python

#64 2005-06-10 09:24:16

Why I love Python 4

Sometimes Java almost seems to go out of its way to make coding difficult, while Python goes out of its way to make it easy. Here is a case in point.

I needed a Java method that accepts a duration expressed as a number of minutes in a string, and returns the same duration formatted as HH:MM:SS. In Python, this is trivial:

def formatDuration(durStr):
  hours, minutes = divmod(int(durStr), 60)
  return '%02d:%02d:00' % (hours, minutes)

Java, on the other hand, makes me jump through hoops to do the same thing:

static String formatDuration(String spcsfDur) {
    int duration = Integer.parseInt(spcsfDur);
    int minutes = duration % 60;
    int hours = (duration - minutes) / 60;

    Object[] args = new Object[] {
            new Integer(hours),
            new Integer(minutes)
    };

    String durStr = MessageFormat.format("{0,number,00}:{1,number,00}:00", args);

    return durStr;
}

Ouch.

Categories: Python

#61 2004-10-22 20:40:32

When should I use classes?

A question beginners sometimes ask is, "Should I use classes in my program?" This article gives some ideas of when it is appropriate to introduce classes into a design.

Categories: Python

#60 2004-10-19 15:18:24

Relief

It's such a relief to be coding in Jython again after working with Java. In Java I feel like I'm fighting the language and the APIs all the time. It's way too hard to get anything done. Python just does what I want with much less straining and fuss.

Aaahhh!

Categories: Java, Python

#58 2004-10-06 19:53:36

Spring Free

For my current project, a distributed application with a database back-end, I tried Hibernate and Spring. I really wanted to like them! They have lots of cool features and make some things incredibly easy. But in the end I have gone back to tried-and-true simple tools - Jython, Velocity, Jetty and not much else.

Hibernate felt like it got in the way as much as it helped. I was constantly having to figure out the Hibernate way of things. My data is largely a tree structure flattened into a database table. Hibernate's lazy loading was sometimes great and sometimes exactly wrong.

Spring just felt too big. Every time I needed a new piece I would have to add another Spring jar to my lib directory and usually a few more jars that it depended on. It felt like using a sledgehammer to squash an ant.

Java also feels quite cumbersome now. I have been working primarily in Jython for about a year now and I hate the hoops that Java makes me jump though to get anything done.

In the end it all felt too confining. I was Mech Warrior, high in a robot programming vehicle, directing awesome power from from my command post. But I longed to put my feet on the ground, pick up a light pack and run.

So I have chucked it all, salvaging what I can, starting over for the rest. What a relief it is!

Categories: Java, Python

#57 2004-09-14 21:29:36

Python and Unicode

Python has extensive support for Unicode data. Two issues that are not well documented elsewhere are the handling of non-Ascii characters in the Python interpreter, and use of the default system encoding. I cover those here.

Categories: Python

#56 2004-08-27 23:39:44

concat vs join - followup

A couple of people have made good points about my last post comparing string concatenation and join.

Marilyn Davis pointed out that in my data, the crossover point where join beats concatenation is always around 500 total characters in the final string. Hans Nowak pointed out that for much longer strings such the lines of a file or parts of a web page, the crossover point comes very quickly.

So here is ConcatTimer version 2 :-) This version dispenses with the fancy graphics and just looks for the crossover point. (It's not too smart about it, either.) It also looks at much larger text chunks - up to 80 characters. Here is the program:

import timeit

reps = 100 # How many reps to try?
unit = '    ' # Concat this string

# Naive concatenation using string +
def concatPlus(count):
   s=''
   for i in range(count):
       s += unit
   return s

# Concatention with string.join
def concatJoin(count):
   s=[]
   for i in range(count):
       s.append(unit)
   return ''.join(s)


# Time one test case
def timeOne(fn, count):
   setup = "from __main__ import " + fn.__name__
   stmt = '%s(%d)' % (fn.__name__, count)

   t = timeit.Timer(stmt, setup)
   secs = min(t.repeat(3, reps))
   return secs


# For strings of length unitLen, find the crossover point where appending
# takes the same amount of time as joining
def findOne(unitLen):
   global unit
   unit = ' ' * unitLen
   t = 2

   while 1:
       tPlus = timeOne(concatPlus, t)
       tJoin = timeOne(concatJoin, t)
       if tPlus > tJoin:
           break
       t += 1

   return t, tPlus, tJoin


for unitLen in range(1,80):
   t, tPlus, tJoin = findOne(unitLen)
   print '%2d %3d %3d %1.5f %1.5f' % (unitLen, t, t*unitLen, tPlus, tJoin)

And here is an elided list of results. The columns are the length of the pieces, the number of pieces where concat becomes more expensive than join, the total number of characters in the string at the crossover point, and the actual times. (I cut the number of reps down to keep this from taking too long to run.)

 1 475 475 0.02733 0.02732
 2 263 526 0.01581 0.01581
 3 169 507 0.01024 0.01022
 4 129 516 0.00782 0.00778
 5 100 500 0.00622 0.00604
 6  85 510 0.00517 0.00515
 7  73 511 0.00447 0.00446
 8  63 504 0.00386 0.00385
 9  57 513 0.00354 0.00353
10  53 530 0.00333 0.00333
11  47 517 0.00294 0.00292
12  45 540 0.00287 0.00285
13  41 533 0.00262 0.00260
14  38 532 0.00246 0.00244
15  36 540 0.00232 0.00230
16  34 544 0.00222 0.00222
17  31 527 0.00200 0.00199
18  29 522 0.00189 0.00188
19  30 570 0.00199 0.00194
20  28 560 0.00188 0.00186
21  28 588 0.00190 0.00185
22  26 572 0.00177 0.00174
23  25 575 0.00170 0.00168
24  24 576 0.00165 0.00163
25  23 575 0.00158 0.00156
26  22 572 0.00153 0.00151
27  21 567 0.00146 0.00144
28  21 588 0.00146 0.00146
29  21 609 0.00147 0.00144
30  20 600 0.00142 0.00139
31  19 589 0.00134 0.00134
32  20 640 0.00143 0.00139
33  19 627 0.00137 0.00136
34  18 612 0.00130 0.00129
35  18 630 0.00131 0.00130
36  18 648 0.00133 0.00130
37  17 629 0.00126 0.00126
38  17 646 0.00126 0.00124
39  15 585 0.00112 0.00111
43  15 645 0.00113 0.00110
44  14 616 0.00106 0.00105
45  15 675 0.00114 0.00110
46  14 644 0.00106 0.00105
48  14 672 0.00109 0.00105
49  13 637 0.00100 0.00099
58  13 754 0.00104 0.00100
59  12 708 0.00098 0.00096
69  12 828 0.00102 0.00098
70  11 770 0.00093 0.00092
77  11 847 0.00094 0.00091
78  10 780 0.00086 0.00086
79  10 790 0.00087 0.00085

So, for anyone still reading, you can see that Hans is right and Marilyn is close:

• For longer strings and more than a few appends, join is clearly a win
• The total number of characters at the crossover isn't quite constant, but it grows slowly.

Based on this experiment I would say that if the total number of characters is less than 500-1000, concatenation is fine. For anything bigger, use join.

Of course the total amount of time involved in any case is pretty small. Unless you have a lot of characters or you are building a lot of strings, I don't think it really matters too much.

I started this experiment because I have been telling people on the Tutor mailing list to use join, and I wondered how much it really mattered. Does it make enough of a difference to bring it up to beginners? I'm not sure. It's good to teach best practices, but maybe it's a poor use of time to teach this to beginners. I won't be so quick to bring it up next time.

Update: Alan Gauld points out that this is an optimization, and the first rule of optimization is don't until you know you need it. That's a useful way to think about it. Thanks for the reminder!

Categories: Python

#55 2004-08-27 19:49:20

Which is really faster - concatenation or join?

A couple of times recently I have given out the conventional advice that for concatenating strings, it is better to build a list of the pieces and join it together than to use string concatenation to build the list. The reasoning is that string concatenation requires copying the entire string for each addition, while the list is designed to make concatenation efficient.

Because of all the copying, the time for string concatenation is proportional to the square of the number of additions - it is O(n^2). List append, however, happens in constant time, so building the list takes time proportional to the number of appends - it is O(n).

The trick with this, though, is that there is a proportionality constant here, and for small n, string concatenation may be faster. I decided to find out.

Here is a program that compares the time for string concatenation using the two methods. It varies both the number of append operations and the length of the appended strings. It prints the results in a series of tables, and it uses VPython to graph the results:

import timeit
from visual.graph import *

reps = 1000 # How many reps to try?
unit = '    ' # Concat this string

# Naive concatenation using string +
def concatPlus(count):
   s=''
   for i in range(count):
       s += unit
   return s

# Concatention the way the big boys do it, with string.join
def concatJoin(count):
   s=[]
   for i in range(count):
       s.append(unit)
   return ''.join(s)


# Time one test case
def timeOne(fn, count):
   setup = "from __main__ import " + fn.__name__
   stmt = '%s(%d)' % (fn.__name__, count)

   t = timeit.Timer(stmt, setup)
   secs = min(t.repeat(3, reps))
   return secs


# Draw the curves for a single length of the appended string
def graphOne(unitLen):
   global unit
   unit = ' ' * unitLen

   title = 'Unit length is %d' % len(unit)
   funct1 = gcurve(color=color.cyan)
   funct2 = gdots(color=color.yellow)

   print
   print title
   print '       tPlus  tJoin'

   for t in range(10, 100, 10) + range(100, 600, 50):
       tPlus = timeOne(concatPlus, t)
       tJoin = timeOne(concatJoin, t)
       print '%5d  %2.3f  %2.3f' % (t, tPlus, tJoin)

       funct1.plot( pos=(t, tPlus) )
       funct2.plot( pos=(t, tJoin) )


graph = gdisplay(title='Append speed', xtitle='count', ytitle='time')
for unitLen in [1,2,3,4,5]:
   graphOne(unitLen)

Here is the graph - the yellow dots are for concatJoin, the blue curves are concatPlus:

A couple of things stand out from this:

• For every string length, concatPlus is faster when the number of appends is relatively small - up to 80 appends in my tests
• For larger numbers of appends, not only does concatPlus show O(n^2) behavior, it gets worse as the size of the appended strings grows. concatJoin is O(n) and it doesn't really matter how long the appended string is. In fact, I think concatPlus is O(m*n) where m is the total length of the final string and n is the number of appends.

Based on these results, I think I will stop spreading the conventional wisdom. I think most uses of string concatenation are for small strings with a small number of concatenations. String join only pays off when there are a lot of appends.

Here is the raw data from the program:

D:\Personal\Tutor>python concattimer.py
Visual-2003-10-05

Unit length is 1
      tPlus  tJoin
  10  0.005  0.008
  20  0.008  0.014
  30  0.012  0.020
  40  0.015  0.025
  50  0.018  0.032
  60  0.022  0.037
  70  0.026  0.044
  80  0.029  0.049
  90  0.033  0.057
 100  0.038  0.062
 150  0.059  0.092
 200  0.082  0.122
 250  0.109  0.149
 300  0.145  0.178
 350  0.184  0.208
 400  0.222  0.237
 450  0.262  0.264
 500  0.307  0.292
 550  0.349  0.325

Unit length is 2
      tPlus  tJoin
  10  0.005  0.008
  20  0.008  0.014
  30  0.012  0.019
  40  0.015  0.026
  50  0.019  0.033
  60  0.023  0.039
  70  0.027  0.044
  80  0.033  0.051
  90  0.038  0.057
 100  0.042  0.062
 150  0.075  0.095
 200  0.114  0.125
 250  0.155  0.154
 300  0.200  0.185
 350  0.247  0.215
 400  0.295  0.243
 450  0.349  0.272
 500  0.404  0.305
 550  0.463  0.332

Unit length is 3
      tPlus  tJoin
  10  0.005  0.008
  20  0.008  0.014
  30  0.012  0.019
  40  0.016  0.026
  50  0.020  0.033
  60  0.025  0.039
  70  0.031  0.045
  80  0.036  0.051
  90  0.043  0.057
 100  0.050  0.064
 150  0.090  0.095
 200  0.134  0.127
 250  0.184  0.159
 300  0.236  0.187
 350  0.293  0.217
 400  0.360  0.247
 450  0.427  0.278
 500  0.499  0.306
 550  0.576  0.335

Unit length is 4
      tPlus  tJoin
  10  0.005  0.008
  20  0.008  0.014
  30  0.012  0.019
  40  0.017  0.025
  50  0.021  0.033
  60  0.026  0.039
  70  0.035  0.045
  80  0.042  0.051
  90  0.050  0.058
 100  0.057  0.063
 150  0.101  0.096
 200  0.153  0.129
 250  0.207  0.161
 300  0.271  0.192
 350  0.341  0.222
 400  0.416  0.249
 450  0.501  0.280
 500  0.580  0.310
 550  0.682  0.338

Unit length is 5
      tPlus  tJoin
  10  0.005  0.008
  20  0.008  0.014
  30  0.013  0.019
  40  0.017  0.025
  50  0.022  0.034
  60  0.029  0.040
  70  0.040  0.046
  80  0.047  0.052
  90  0.055  0.058
 100  0.063  0.064
 150  0.114  0.097
 200  0.169  0.130
 250  0.232  0.163
 300  0.306  0.195
 350  0.385  0.223
 400  0.470  0.252
 450  0.567  0.283
 500  0.668  0.313
 550  0.779  0.344

Categories: Python