You are here: Home Dive Into Python 3

Difficulty level: ♦♦♦♦♦

Case Study: Porting chardet to Python 3

Words, words. They’re all we have to go on.
Rosencrantz and Guildenstern are Dead

 

Diving In

Question: what’s the #1 cause of gibberish text on the web, in your inbox, and across every computer system ever written? It’s character encoding. In the Strings chapter, I talked about the history of character encoding and the creation of Unicode, the “one encoding to rule them all.” I’d love it if I never had to see a gibberish character on a web page again, because all authoring systems stored accurate encoding information, all transfer protocols were Unicode-aware, and every system that handled text maintained perfect fidelity when converting between encodings.

I’d also like a pony.

A Unicode pony.

A Unipony, as it were.

I’ll settle for character encoding auto-detection.

What is Character Encoding Auto-Detection?

It means taking a sequence of bytes in an unknown character encoding, and attempting to determine the encoding so you can read the text. It’s like cracking a code when you don’t have the decryption key.

Isn’t That Impossible?

In general, yes. However, some encodings are optimized for specific languages, and languages are not random. Some character sequences pop up all the time, while other sequences make no sense. A person fluent in English who opens a newspaper and finds “txzqJv 2!dasd0a QqdKjvz” will instantly recognize that that isn’t English (even though it is composed entirely of English letters). By studying lots of “typical” text, a computer algorithm can simulate this kind of fluency and make an educated guess about a text’s language.

In other words, encoding detection is really language detection, combined with knowledge of which languages tend to use which character encodings.

Does Such An Algorithm Exist?

As it turns out, yes. All major browsers have character encoding auto-detection, because the web is full of pages that have no encoding information whatsoever. Mozilla Firefox contains an encoding auto-detection library which is open source. I ported the library to Python 2 and dubbed it the chardet module. This chapter will take you step-by-step through the process of porting the chardet module from Python 2 to Python 3.

Introducing The chardet Module

Before we set off porting the code, it would help if you understood how the code worked! This is a brief guide to navigating the code itself. The chardet library is too large to include inline here, but you can download it from chardet.feedparser.org.

The main entry point for the detection algorithm is universaldetector.py, which has one class, UniversalDetector. (You might think the main entry point is the detect function in chardet/__init__.py, but that’s really just a convenience function that creates a UniversalDetector object, calls it, and returns its result.)

There are 5 categories of encodings that UniversalDetector handles:

  1. UTF-n with a Byte Order Mark (BOM). This includes UTF-8, both Big-Endian and Little-Endian variants of UTF-16, and all 4 byte-order variants of UTF-32.
  2. Escaped encodings, which are entirely 7-bit ASCII compatible, where non-ASCII characters start with an escape sequence. Examples: ISO-2022-JP (Japanese) and HZ-GB-2312 (Chinese).
  3. Multi-byte encodings, where each character is represented by a variable number of bytes. Examples: Big5 (Chinese), SHIFT_JIS (Japanese), EUC-KR (Korean), and UTF-8 without a BOM.
  4. Single-byte encodings, where each character is represented by one byte. Examples: KOI8-R (Russian), windows-1255 (Hebrew), and TIS-620 (Thai).
  5. windows-1252, which is used primarily on Microsoft Windows by middle managers who wouldn’t know a character encoding from a hole in the ground.

UTF-n With A BOM

If the text starts with a BOM, we can reasonably assume that the text is encoded in UTF-8, UTF-16, or UTF-32. (The BOM will tell us exactly which one; that’s what it’s for.) This is handled inline in UniversalDetector, which returns the result immediately without any further processing.

Escaped Encodings

If the text contains a recognizable escape sequence that might indicate an escaped encoding, UniversalDetector creates an EscCharSetProber (defined in escprober.py) and feeds it the text.

EscCharSetProber creates a series of state machines, based on models of HZ-GB-2312, ISO-2022-CN, ISO-2022-JP, and ISO-2022-KR (defined in escsm.py). EscCharSetProber feeds the text to each of these state machines, one byte at a time. If any state machine ends up uniquely identifying the encoding, EscCharSetProber immediately returns the positive result to UniversalDetector, which returns it to the caller. If any state machine hits an illegal sequence, it is dropped and processing continues with the other state machines.

Multi-Byte Encodings

Assuming no BOM, UniversalDetector checks whether the text contains any high-bit characters. If so, it creates a series of “probers” for detecting multi-byte encodings, single-byte encodings, and as a last resort, windows-1252.

The multi-byte encoding prober, MBCSGroupProber (defined in mbcsgroupprober.py), is really just a shell that manages a group of other probers, one for each multi-byte encoding: Big5, GB2312, EUC-TW, EUC-KR, EUC-JP, SHIFT_JIS, and UTF-8. MBCSGroupProber feeds the text to each of these encoding-specific probers and checks the results. If a prober reports that it has found an illegal byte sequence, it is dropped from further processing (so that, for instance, any subsequent calls to UniversalDetector.feed() will skip that prober). If a prober reports that it is reasonably confident that it has detected the encoding, MBCSGroupProber reports this positive result to UniversalDetector, which reports the result to the caller.

Most of the multi-byte encoding probers are inherited from MultiByteCharSetProber (defined in mbcharsetprober.py), and simply hook up the appropriate state machine and distribution analyzer and let MultiByteCharSetProber do the rest of the work. MultiByteCharSetProber runs the text through the encoding-specific state machine, one byte at a time, to look for byte sequences that would indicate a conclusive positive or negative result. At the same time, MultiByteCharSetProber feeds the text to an encoding-specific distribution analyzer.

The distribution analyzers (each defined in chardistribution.py) use language-specific models of which characters are used most frequently. Once MultiByteCharSetProber has fed enough text to the distribution analyzer, it calculates a confidence rating based on the number of frequently-used characters, the total number of characters, and a language-specific distribution ratio. If the confidence is high enough, MultiByteCharSetProber returns the result to MBCSGroupProber, which returns it to UniversalDetector, which returns it to the caller.

The case of Japanese is more difficult. Single-character distribution analysis is not always sufficient to distinguish between EUC-JP and SHIFT_JIS, so the SJISProber (defined in sjisprober.py) also uses 2-character distribution analysis. SJISContextAnalysis and EUCJPContextAnalysis (both defined in jpcntx.py and both inheriting from a common JapaneseContextAnalysis class) check the frequency of Hiragana syllabary characters within the text. Once enough text has been processed, they return a confidence level to SJISProber, which checks both analyzers and returns the higher confidence level to MBCSGroupProber.

Single-Byte Encodings

The single-byte encoding prober, SBCSGroupProber (defined in sbcsgroupprober.py), is also just a shell that manages a group of other probers, one for each combination of single-byte encoding and language: windows-1251, KOI8-R, ISO-8859-5, MacCyrillic, IBM855, and IBM866 (Russian); ISO-8859-7 and windows-1253 (Greek); ISO-8859-5 and windows-1251 (Bulgarian); ISO-8859-2 and windows-1250 (Hungarian); TIS-620 (Thai); windows-1255 and ISO-8859-8 (Hebrew).

SBCSGroupProber feeds the text to each of these encoding+language-specific probers and checks the results. These probers are all implemented as a single class, SingleByteCharSetProber (defined in sbcharsetprober.py), which takes a language model as an argument. The language model defines how frequently different 2-character sequences appear in typical text. SingleByteCharSetProber processes the text and tallies the most frequently used 2-character sequences. Once enough text has been processed, it calculates a confidence level based on the number of frequently-used sequences, the total number of characters, and a language-specific distribution ratio.

Hebrew is handled as a special case. If the text appears to be Hebrew based on 2-character distribution analysis, HebrewProber (defined in hebrewprober.py) tries to distinguish between Visual Hebrew (where the source text actually stored “backwards” line-by-line, and then displayed verbatim so it can be read from right to left) and Logical Hebrew (where the source text is stored in reading order and then rendered right-to-left by the client). Because certain characters are encoded differently based on whether they appear in the middle of or at the end of a word, we can make a reasonable guess about direction of the source text, and return the appropriate encoding (windows-1255 for Logical Hebrew, or ISO-8859-8 for Visual Hebrew).

windows-1252

If UniversalDetector detects a high-bit character in the text, but none of the other multi-byte or single-byte encoding probers return a confident result, it creates a Latin1Prober (defined in latin1prober.py) to try to detect English text in a windows-1252 encoding. This detection is inherently unreliable, because English letters are encoded in the same way in many different encodings. The only way to distinguish windows-1252 is through commonly used symbols like smart quotes, curly apostrophes, copyright symbols, and the like. Latin1Prober automatically reduces its confidence rating to allow more accurate probers to win if at all possible.

Running 2to3

We’re going to migrate the chardet module from Python 2 to Python 3. Python 3 comes with a utility script called 2to3, which takes your actual Python 2 source code as input and auto-converts as much as it can to Python 3. In some cases this is easy — a function was renamed or moved to a different module — but in other cases it can get pretty complex. To get a sense of all that it can do, refer to the appendix, Porting code to Python 3 with 2to3. In this chapter, we’ll start by running 2to3 on the chardet package, but as you’ll see, there will still be a lot of work to do after the automated tools have performed their magic.

The main chardet package is split across several different files, all in the same directory. The 2to3 script makes it easy to convert multiple files at once: just pass a directory as a command line argument, and 2to3 will convert each of the files in turn.

C:\home\chardet> python c:\Python30\Tools\Scripts\2to3.py -w chardet\
RefactoringTool: Skipping implicit fixer: buffer
RefactoringTool: Skipping implicit fixer: idioms
RefactoringTool: Skipping implicit fixer: set_literal
RefactoringTool: Skipping implicit fixer: ws_comma
--- chardet\__init__.py (original)
+++ chardet\__init__.py (refactored)
@@ -18,7 +18,7 @@
 __version__ = "1.0.1"

 def detect(aBuf):
-    import universaldetector
+    from . import universaldetector
     u = universaldetector.UniversalDetector()
     u.reset()
     u.feed(aBuf)
--- chardet\big5prober.py (original)
+++ chardet\big5prober.py (refactored)
@@ -25,10 +25,10 @@
 # 02110-1301  USA
 ######################### END LICENSE BLOCK #########################

-from mbcharsetprober import MultiByteCharSetProber
-from codingstatemachine import CodingStateMachine
-from chardistribution import Big5DistributionAnalysis
-from mbcssm import Big5SMModel
+from .mbcharsetprober import MultiByteCharSetProber
+from .codingstatemachine import CodingStateMachine
+from .chardistribution import Big5DistributionAnalysis
+from .mbcssm import Big5SMModel

 class Big5Prober(MultiByteCharSetProber):
     def __init__(self):
--- chardet\chardistribution.py (original)
+++ chardet\chardistribution.py (refactored)
@@ -25,12 +25,12 @@
 # 02110-1301  USA
 ######################### END LICENSE BLOCK #########################

-import constants
-from euctwfreq import EUCTWCharToFreqOrder, EUCTW_TABLE_SIZE, EUCTW_TYPICAL_DISTRIBUTION_RATIO
-from euckrfreq import EUCKRCharToFreqOrder, EUCKR_TABLE_SIZE, EUCKR_TYPICAL_DISTRIBUTION_RATIO
-from gb2312freq import GB2312CharToFreqOrder, GB2312_TABLE_SIZE, GB2312_TYPICAL_DISTRIBUTION_RATIO
-from big5freq import Big5CharToFreqOrder, BIG5_TABLE_SIZE, BIG5_TYPICAL_DISTRIBUTION_RATIO
-from jisfreq import JISCharToFreqOrder, JIS_TABLE_SIZE, JIS_TYPICAL_DISTRIBUTION_RATIO
+from . import constants
+from .euctwfreq import EUCTWCharToFreqOrder, EUCTW_TABLE_SIZE, EUCTW_TYPICAL_DISTRIBUTION_RATIO
+from .euckrfreq import EUCKRCharToFreqOrder, EUCKR_TABLE_SIZE, EUCKR_TYPICAL_DISTRIBUTION_RATIO
+from .gb2312freq import GB2312CharToFreqOrder, GB2312_TABLE_SIZE, GB2312_TYPICAL_DISTRIBUTION_RATIO
+from .big5freq import Big5CharToFreqOrder, BIG5_TABLE_SIZE, BIG5_TYPICAL_DISTRIBUTION_RATIO
+from .jisfreq import JISCharToFreqOrder, JIS_TABLE_SIZE, JIS_TYPICAL_DISTRIBUTION_RATIO

 ENOUGH_DATA_THRESHOLD = 1024
 SURE_YES = 0.99
.
.
. (it goes on like this for a while)
.
.
RefactoringTool: Files that were modified:
RefactoringTool: chardet\__init__.py
RefactoringTool: chardet\big5prober.py
RefactoringTool: chardet\chardistribution.py
RefactoringTool: chardet\charsetgroupprober.py
RefactoringTool: chardet\codingstatemachine.py
RefactoringTool: chardet\constants.py
RefactoringTool: chardet\escprober.py
RefactoringTool: chardet\escsm.py
RefactoringTool: chardet\eucjpprober.py
RefactoringTool: chardet\euckrprober.py
RefactoringTool: chardet\euctwprober.py
RefactoringTool: chardet\gb2312prober.py
RefactoringTool: chardet\hebrewprober.py
RefactoringTool: chardet\jpcntx.py
RefactoringTool: chardet\langbulgarianmodel.py
RefactoringTool: chardet\langcyrillicmodel.py
RefactoringTool: chardet\langgreekmodel.py
RefactoringTool: chardet\langhebrewmodel.py
RefactoringTool: chardet\langhungarianmodel.py
RefactoringTool: chardet\langthaimodel.py
RefactoringTool: chardet\latin1prober.py
RefactoringTool: chardet\mbcharsetprober.py
RefactoringTool: chardet\mbcsgroupprober.py
RefactoringTool: chardet\mbcssm.py
RefactoringTool: chardet\sbcharsetprober.py
RefactoringTool: chardet\sbcsgroupprober.py
RefactoringTool: chardet\sjisprober.py
RefactoringTool: chardet\universaldetector.py
RefactoringTool: chardet\utf8prober.py

Now run the 2to3 script on the testing harness, test.py.

C:\home\chardet> python c:\Python30\Tools\Scripts\2to3.py -w test.py
RefactoringTool: Skipping implicit fixer: buffer
RefactoringTool: Skipping implicit fixer: idioms
RefactoringTool: Skipping implicit fixer: set_literal
RefactoringTool: Skipping implicit fixer: ws_comma
--- test.py (original)
+++ test.py (refactored)
@@ -4,7 +4,7 @@
 count = 0
 u = UniversalDetector()
 for f in glob.glob(sys.argv[1]):
-    print f.ljust(60),
+    print(f.ljust(60), end=' ')
     u.reset()
     for line in file(f, 'rb'):
         u.feed(line)
@@ -12,8 +12,8 @@
     u.close()
     result = u.result
     if result['encoding']:
-        print result['encoding'], 'with confidence', result['confidence']
+        print(result['encoding'], 'with confidence', result['confidence'])
     else:
-        print '******** no result'
+        print('******** no result')
     count += 1
-print count, 'tests'
+print(count, 'tests')
RefactoringTool: Files that were modified:
RefactoringTool: test.py

Well, that wasn’t so hard. Just a few imports and print statements to convert. Speaking of which, what was the problem with all those import statements? To answer that, you need to understand how the chardet module is split into multiple files.

A Short Digression Into Multi-File Modules

chardet is a multi-file module. I could have chosen to put all the code in one file (named chardet.py), but I didn’t. Instead, I made a directory (named chardet), then I made an __init__.py file in that directory. If Python sees an __init__.py file in a directory, it assumes that all of the files in that directory are part of the same module. The module’s name is the name of the directory. Files within the directory can reference other files within the same directory, or even within subdirectories. (More on that in a minute.) But the entire collection of files is presented to other Python code as a single module — as if all the functions and classes were in a single .py file.

What goes in the __init__.py file? Nothing. Everything. Something in between. The __init__.py file doesn’t need to define anything; it can literally be an empty file. Or you can use it to define your main entry point functions. Or you put all your functions in it. Or all but one.

A directory with an __init__.py file is always treated as a multi-file module. Without an __init__.py file, a directory is just a directory of unrelated .py files.

Let’s see how that works in practice.

>>> import chardet
>>> dir(chardet)             
['__builtins__', '__doc__', '__file__', '__name__',
 '__package__', '__path__', '__version__', 'detect']
>>> chardet                  
<module 'chardet' from 'C:\Python31\lib\site-packages\chardet\__init__.py'>
  1. Other than the usual class attributes, the only thing in the chardet module is a detect() function.
  2. Here’s your first clue that the chardet module is more than just a file: the “module” is listed as the __init__.py file within the chardet/ directory.

Let’s take a peek in that __init__.py file.

def detect(aBuf):                              
    from . import universaldetector            
    u = universaldetector.UniversalDetector()
    u.reset()
    u.feed(aBuf)
    u.close()
    return u.result
  1. The __init__.py file defines the detect() function, which is the main entry point into the chardet library.
  2. But the detect() function hardly has any code! In fact, all it really does is import the universaldetector module and start using it. But where is universaldetector defined?

The answer lies in that odd-looking import statement:

from . import universaldetector

Translated into English, that means “import the universaldetector module; that’s in the same directory I am,” where “I” is the chardet/__init__.py file. This is called a relative import. It’s a way for the files within a multi-file module to reference each other, without worrying about naming conflicts with other modules you may have installed in your import search path. This import statement will only look for the universaldetector module within the chardet/ directory itself.

These two concepts — __init__.py and relative imports — mean that you can break up your module into as many pieces as you like. The chardet module comprises 36 .py files — 36! Yet all you need to do to start using it is import chardet, then you can call the main chardet.detect() function. Unbeknownst to your code, the detect() function is actually defined in the chardet/__init__.py file. Also unbeknownst to you, the detect() function uses a relative import to reference a class defined in chardet/universaldetector.py, which in turn uses relative imports on five other files, all contained in the chardet/ directory.

If you ever find yourself writing a large library in Python (or more likely, when you realize that your small library has grown into a large one), take the time to refactor it into a multi-file module. It’s one of the many things Python is good at, so take advantage of it.

Fixing What 2to3 Can’t

False is invalid syntax

Now for the real test: running the test harness against the test suite. Since the test suite is designed to cover all the possible code paths, it’s a good way to test our ported code to make sure there aren’t any bugs lurking anywhere.

C:\home\chardet> python test.py tests\*\*
Traceback (most recent call last):
  File "test.py", line 1, in <module>
    from chardet.universaldetector import UniversalDetector
  File "C:\home\chardet\chardet\universaldetector.py", line 51
    self.done = constants.False
                              ^
SyntaxError: invalid syntax

Hmm, a small snag. In Python 3, False is a reserved word, so you can’t use it as a variable name. Let’s look at constants.py to see where it’s defined. Here’s the original version from constants.py, before the 2to3 script changed it:

import __builtin__
if not hasattr(__builtin__, 'False'):
    False = 0
    True = 1
else:
    False = __builtin__.False
    True = __builtin__.True

This piece of code is designed to allow this library to run under older versions of Python 2. Prior to Python 2.3, Python had no built-in bool type. This code detects the absence of the built-in constants True and False, and defines them if necessary.

However, Python 3 will always have a bool type, so this entire code snippet is unnecessary. The simplest solution is to replace all instances of constants.True and constants.False with True and False, respectively, then delete this dead code from constants.py.

So this line in universaldetector.py:

self.done = constants.False

Becomes

self.done = False

Ah, wasn’t that satisfying? The code is shorter and more readable already.

No module named constants

Time to run test.py again and see how far it gets.

C:\home\chardet> python test.py tests\*\*
Traceback (most recent call last):
  File "test.py", line 1, in <module>
    from chardet.universaldetector import UniversalDetector
  File "C:\home\chardet\chardet\universaldetector.py", line 29, in <module>
    import constants, sys
ImportError: No module named constants

What’s that you say? No module named constants? Of course there’s a module named constants. It’s right there, in chardet/constants.py.

Remember when the 2to3 script fixed up all those import statements? This library has a lot of relative imports — that is, modules that import other modules within the same library — but the logic behind relative imports has changed in Python 3. In Python 2, you could just import constants and it would look in the chardet/ directory first. In Python 3, all import statements are absolute by default. If you want to do a relative import in Python 3, you need to be explicit about it:

from . import constants

But wait. Wasn’t the 2to3 script supposed to take care of these for you? Well, it did, but this particular import statement combines two different types of imports into one line: a relative import of the constants module within the library, and an absolute import of the sys module that is pre-installed in the Python standard library. In Python 2, you could combine these into one import statement. In Python 3, you can’t, and the 2to3 script is not smart enough to split the import statement into two.

The solution is to split the import statement manually. So this two-in-one import:

import constants, sys

Needs to become two separate imports:

from . import constants
import sys

There are variations of this problem scattered throughout the chardet library. In some places it’s “import constants, sys”; in other places, it’s “import constants, re”. The fix is the same: manually split the import statement into two lines, one for the relative import, the other for the absolute import.

Onward!

Name 'file' is not defined

And here we go again, running test.py to try to execute our test cases…

C:\home\chardet> python test.py tests\*\*
tests\ascii\howto.diveintomark.org.xml
Traceback (most recent call last):
  File "test.py", line 9, in <module>
    for line in file(f, 'rb'):
NameError: name 'file' is not defined

This one surprised me, because I’ve been using this idiom as long as I can remember. In Python 2, the global file() function was an alias for the open() function, which was the standard way of opening text files for reading. In Python 3, the global file() function no longer exists, but the open() function still exists.

Thus, the simplest solution to the problem of the missing file() is to call the open() function instead:

for line in open(f, 'rb'):

And that’s all I have to say about that.

Can’t use a string pattern on a bytes-like object

Now things are starting to get interesting. And by “interesting,” I mean “confusing as all hell.”

C:\home\chardet> python test.py tests\*\*
tests\ascii\howto.diveintomark.org.xml
Traceback (most recent call last):
  File "test.py", line 10, in <module>
    u.feed(line)
  File "C:\home\chardet\chardet\universaldetector.py", line 98, in feed
    if self._highBitDetector.search(aBuf):
TypeError: can't use a string pattern on a bytes-like object

To debug this, let’s see what self._highBitDetector is. It’s defined in the __init__ method of the UniversalDetector class:

class UniversalDetector:
    def __init__(self):
        self._highBitDetector = re.compile(r'[\x80-\xFF]')

This pre-compiles a regular expression designed to find non-ASCII characters in the range 128–255 (0x80–0xFF). Wait, that’s not quite right; I need to be more precise with my terminology. This pattern is designed to find non-ASCII bytes in the range 128-255.

And therein lies the problem.

In Python 2, a string was an array of bytes whose character encoding was tracked separately. If you wanted Python 2 to keep track of the character encoding, you had to use a Unicode string (u'') instead. But in Python 3, a string is always what Python 2 called a Unicode string — that is, an array of Unicode characters (of possibly varying byte lengths). Since this regular expression is defined by a string pattern, it can only be used to search a string — again, an array of characters. But what we’re searching is not a string, it’s a byte array. Looking at the traceback, this error occurred in universaldetector.py:

def feed(self, aBuf):
    .
    .
    .
    if self._mInputState == ePureAscii:
        if self._highBitDetector.search(aBuf):

And what is aBuf? Let’s backtrack further to a place that calls UniversalDetector.feed(). One place that calls it is the test harness, test.py.

u = UniversalDetector()
.
.
.
for line in open(f, 'rb'):
    u.feed(line)

And here we find our answer: in the UniversalDetector.feed() method, aBuf is a line read from a file on disk. Look carefully at the parameters used to open the file: 'rb'. 'r' is for “read”; OK, big deal, we’re reading the file. Ah, but 'b' is for “binary.” Without the 'b' flag, this for loop would read the file, line by line, and convert each line into a string — an array of Unicode characters — according to the system default character encoding. But with the 'b' flag, this for loop reads the file, line by line, and stores each line exactly as it appears in the file, as an array of bytes. That byte array gets passed to UniversalDetector.feed(), and eventually gets passed to the pre-compiled regular expression, self._highBitDetector, to search for high-bit… characters. But we don’t have characters; we have bytes. Oops.

What we need this regular expression to search is not an array of characters, but an array of bytes.

Once you realize that, the solution is not difficult. Regular expressions defined with strings can search strings. Regular expressions defined with byte arrays can search byte arrays. To define a byte array pattern, we simply change the type of the argument we use to define the regular expression to a byte array. (There is one other case of this same problem, on the very next line.)

  class UniversalDetector:
      def __init__(self):
-         self._highBitDetector = re.compile(r'[\x80-\xFF]')
-         self._escDetector = re.compile(r'(\033|~{)')
+         self._highBitDetector = re.compile(b'[\x80-\xFF]')
+         self._escDetector = re.compile(b'(\033|~{)')
          self._mEscCharSetProber = None
          self._mCharSetProbers = []
          self.reset()

Searching the entire codebase for other uses of the re module turns up two more instances, in charsetprober.py. Again, the code is defining regular expressions as strings but executing them on aBuf, which is a byte array. The solution is the same: define the regular expression patterns as byte arrays.

  class CharSetProber:
      .
      .
      .
      def filter_high_bit_only(self, aBuf):
-         aBuf = re.sub(r'([\x00-\x7F])+', ' ', aBuf)
+         aBuf = re.sub(b'([\x00-\x7F])+', b' ', aBuf)
          return aBuf
    
      def filter_without_english_letters(self, aBuf):
-         aBuf = re.sub(r'([A-Za-z])+', ' ', aBuf)
+         aBuf = re.sub(b'([A-Za-z])+', b' ', aBuf)
          return aBuf

Can't convert 'bytes' object to str implicitly

Curiouser and curiouser…

C:\home\chardet> python test.py tests\*\*
tests\ascii\howto.diveintomark.org.xml
Traceback (most recent call last):
  File "test.py", line 10, in <module>
    u.feed(line)
  File "C:\home\chardet\chardet\universaldetector.py", line 100, in feed
    elif (self._mInputState == ePureAscii) and self._escDetector.search(self._mLastChar + aBuf):
TypeError: Can't convert 'bytes' object to str implicitly

There’s an unfortunate clash of coding style and Python interpreter here. The TypeError could be anywhere on that line, but the traceback doesn’t tell you exactly where it is. It could be in the first conditional or the second, and the traceback would look the same. To narrow it down, you should split the line in half, like this:

elif (self._mInputState == ePureAscii) and \
    self._escDetector.search(self._mLastChar + aBuf):

And re-run the test:

C:\home\chardet> python test.py tests\*\*
tests\ascii\howto.diveintomark.org.xml
Traceback (most recent call last):
  File "test.py", line 10, in <module>
    u.feed(line)
  File "C:\home\chardet\chardet\universaldetector.py", line 101, in feed
    self._escDetector.search(self._mLastChar + aBuf):
TypeError: Can't convert 'bytes' object to str implicitly

Aha! The problem was not in the first conditional (self._mInputState == ePureAscii) but in the second one. So what could cause a TypeError there? Perhaps you’re thinking that the search() method is expecting a value of a different type, but that wouldn’t generate this traceback. Python functions can take any value; if you pass the right number of arguments, the function will execute. It may crash if you pass it a value of a different type than it’s expecting, but if that happened, the traceback would point to somewhere inside the function. But this traceback says it never got as far as calling the search() method. So the problem must be in that + operation, as it’s trying to construct the value that it will eventually pass to the search() method.

We know from previous debugging that aBuf is a byte array. So what is self._mLastChar? It’s an instance variable, defined in the reset() method, which is actually called from the __init__() method.

class UniversalDetector:
    def __init__(self):
        self._highBitDetector = re.compile(b'[\x80-\xFF]')
        self._escDetector = re.compile(b'(\033|~{)')
        self._mEscCharSetProber = None
        self._mCharSetProbers = []
        self.reset()

    def reset(self):
        self.result = {'encoding': None, 'confidence': 0.0}
        self.done = False
        self._mStart = True
        self._mGotData = False
        self._mInputState = ePureAscii
        self._mLastChar = ''

And now we have our answer. Do you see it? self._mLastChar is a string, but aBuf is a byte array. And you can’t concatenate a string to a byte array — not even a zero-length string.

So what is self._mLastChar anyway? In the feed() method, just a few lines down from where the trackback occurred.

if self._mInputState == ePureAscii:
    if self._highBitDetector.search(aBuf):
        self._mInputState = eHighbyte
    elif (self._mInputState == ePureAscii) and \
            self._escDetector.search(self._mLastChar + aBuf):
        self._mInputState = eEscAscii

self._mLastChar = aBuf[-1]

The calling function calls this feed() method over and over again with a few bytes at a time. The method processes the bytes it was given (passed in as aBuf), then stores the last byte in self._mLastChar in case it’s needed during the next call. (In a multi-byte encoding, the feed() method might get called with half of a character, then called again with the other half.) But because aBuf is now a byte array instead of a string, self._mLastChar needs to be a byte array as well. Thus:

  def reset(self):
      .
      .
      .
-     self._mLastChar = ''
+     self._mLastChar = b''

Searching the entire codebase for “mLastChar” turns up a similar problem in mbcharsetprober.py, but instead of tracking the last character, it tracks the last two characters. The MultiByteCharSetProber class uses a list of 1-character strings to track the last two characters. In Python 3, it needs to use a list of integers, because it’s not really tracking characters, it’s tracking bytes. (Bytes are just integers from 0-255.)

  class MultiByteCharSetProber(CharSetProber):
      def __init__(self):
          CharSetProber.__init__(self)
          self._mDistributionAnalyzer = None
          self._mCodingSM = None
-         self._mLastChar = ['\x00', '\x00']
+         self._mLastChar = [0, 0]

      def reset(self):
          CharSetProber.reset(self)
          if self._mCodingSM:
              self._mCodingSM.reset()
          if self._mDistributionAnalyzer:
              self._mDistributionAnalyzer.reset()
-         self._mLastChar = ['\x00', '\x00']
+         self._mLastChar = [0, 0]

Unsupported operand type(s) for +: 'int' and 'bytes'

I have good news, and I have bad news. The good news is we’re making progress…

C:\home\chardet> python test.py tests\*\*
tests\ascii\howto.diveintomark.org.xml
Traceback (most recent call last):
  File "test.py", line 10, in <module>
    u.feed(line)
  File "C:\home\chardet\chardet\universaldetector.py", line 101, in feed
    self._escDetector.search(self._mLastChar + aBuf):
TypeError: unsupported operand type(s) for +: 'int' and 'bytes'

…The bad news is it doesn’t always feel like progress.

But this is progress! Really! Even though the traceback calls out the same line of code, it’s a different error than it used to be. Progress! So what’s the problem now? The last time I checked, this line of code didn’t try to concatenate an int with a byte array (bytes). In fact, you just spent a lot of time ensuring that self._mLastChar was a byte array. How did it turn into an int?

The answer lies not in the previous lines of code, but in the following lines.

if self._mInputState == ePureAscii:
    if self._highBitDetector.search(aBuf):
        self._mInputState = eHighbyte
    elif (self._mInputState == ePureAscii) and \
            self._escDetector.search(self._mLastChar + aBuf):
        self._mInputState = eEscAscii

self._mLastChar = aBuf[-1]

This error doesn’t occur the first time the feed() method gets called; it occurs the second time, after self._mLastChar has been set to the last byte of aBuf. Well, what’s the problem with that? Getting a single element from a byte array yields an integer, not a byte array. To see the difference, follow me to the interactive shell:

>>> aBuf = b'\xEF\xBB\xBF'         
>>> len(aBuf)
3
>>> mLastChar = aBuf[-1]
>>> mLastChar                      
191
>>> type(mLastChar)                
<class 'int'>
>>> mLastChar + aBuf               
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
TypeError: unsupported operand type(s) for +: 'int' and 'bytes'
>>> mLastChar = aBuf[-1:]          
>>> mLastChar
b'\xbf'
>>> mLastChar + aBuf               
b'\xbf\xef\xbb\xbf'
  1. Define a byte array of length 3.
  2. The last element of the byte array is 191.
  3. That’s an integer.
  4. Concatenating an integer with a byte array doesn’t work. You’ve now replicated the error you just found in universaldetector.py.
  5. Ah, here’s the fix. Instead of taking the last element of the byte array, use list slicing to create a new byte array containing just the last element. That is, start with the last element and continue the slice until the end of the byte array. Now mLastChar is a byte array of length 1.
  6. Concatenating a byte array of length 1 with a byte array of length 3 returns a new byte array of length 4.

So, to ensure that the feed() method in universaldetector.py continues to work no matter how often it’s called, you need to initialize self._mLastChar as a 0-length byte array, then make sure it stays a byte array.

              self._escDetector.search(self._mLastChar + aBuf):
          self._mInputState = eEscAscii

- self._mLastChar = aBuf[-1]
+ self._mLastChar = aBuf[-1:]

ord() expected string of length 1, but int found

Tired yet? You’re almost there…

C:\home\chardet> python test.py tests\*\*
tests\ascii\howto.diveintomark.org.xml                       ascii with confidence 1.0
tests\Big5\0804.blogspot.com.xml
Traceback (most recent call last):
  File "test.py", line 10, in <module>
    u.feed(line)
  File "C:\home\chardet\chardet\universaldetector.py", line 116, in feed
    if prober.feed(aBuf) == constants.eFoundIt:
  File "C:\home\chardet\chardet\charsetgroupprober.py", line 60, in feed
    st = prober.feed(aBuf)
  File "C:\home\chardet\chardet\utf8prober.py", line 53, in feed
    codingState = self._mCodingSM.next_state(c)
  File "C:\home\chardet\chardet\codingstatemachine.py", line 43, in next_state
    byteCls = self._mModel['classTable'][ord(c)]
TypeError: ord() expected string of length 1, but int found

OK, so c is an int, but the ord() function was expecting a 1-character string. Fair enough. Where is c defined?

# codingstatemachine.py
def next_state(self, c):
    # for each byte we get its class
    # if it is first byte, we also get byte length
    byteCls = self._mModel['classTable'][ord(c)]

That’s no help; it’s just passed into the function. Let’s pop the stack.

# utf8prober.py
def feed(self, aBuf):
    for c in aBuf:
        codingState = self._mCodingSM.next_state(c)

Do you see it? In Python 2, aBuf was a string, so c was a 1-character string. (That’s what you get when you iterate over a string — all the characters, one by one.) But now, aBuf is a byte array, so c is an int, not a 1-character string. In other words, there’s no need to call the ord() function because c is already an int!

Thus:

  def next_state(self, c):
      # for each byte we get its class
      # if it is first byte, we also get byte length
-     byteCls = self._mModel['classTable'][ord(c)]
+     byteCls = self._mModel['classTable'][c]

Searching the entire codebase for instances of “ord(c)” uncovers similar problems in sbcharsetprober.py

# sbcharsetprober.py
def feed(self, aBuf):
    if not self._mModel['keepEnglishLetter']:
        aBuf = self.filter_without_english_letters(aBuf)
    aLen = len(aBuf)
    if not aLen:
        return self.get_state()
    for c in aBuf:
        order = self._mModel['charToOrderMap'][ord(c)]

…and latin1prober.py

# latin1prober.py
def feed(self, aBuf):
    aBuf = self.filter_with_english_letters(aBuf)
    for c in aBuf:
        charClass = Latin1_CharToClass[ord(c)]

c is iterating over aBuf, which means it is an integer, not a 1-character string. The solution is the same: change ord(c) to just plain c.

  # sbcharsetprober.py
  def feed(self, aBuf):
      if not self._mModel['keepEnglishLetter']:
          aBuf = self.filter_without_english_letters(aBuf)
      aLen = len(aBuf)
      if not aLen:
          return self.get_state()
      for c in aBuf:
-         order = self._mModel['charToOrderMap'][ord(c)]
+         order = self._mModel['charToOrderMap'][c]

  # latin1prober.py
  def feed(self, aBuf):
      aBuf = self.filter_with_english_letters(aBuf)
      for c in aBuf:
-         charClass = Latin1_CharToClass[ord(c)]
+         charClass = Latin1_CharToClass[c]

Unorderable types: int() >= str()

Let’s go again.

C:\home\chardet> python test.py tests\*\*
tests\ascii\howto.diveintomark.org.xml                       ascii with confidence 1.0
tests\Big5\0804.blogspot.com.xml
Traceback (most recent call last):
  File "test.py", line 10, in <module>
    u.feed(line)
  File "C:\home\chardet\chardet\universaldetector.py", line 116, in feed
    if prober.feed(aBuf) == constants.eFoundIt:
  File "C:\home\chardet\chardet\charsetgroupprober.py", line 60, in feed
    st = prober.feed(aBuf)
  File "C:\home\chardet\chardet\sjisprober.py", line 68, in feed
    self._mContextAnalyzer.feed(self._mLastChar[2 - charLen :], charLen)
  File "C:\home\chardet\chardet\jpcntx.py", line 145, in feed
    order, charLen = self.get_order(aBuf[i:i+2])
  File "C:\home\chardet\chardet\jpcntx.py", line 176, in get_order
    if ((aStr[0] >= '\x81') and (aStr[0] <= '\x9F')) or \
TypeError: unorderable types: int() >= str()

So what’s this all about? “Unorderable types”? Once again, the difference between byte arrays and strings is rearing its ugly head. Take a look at the code:

class SJISContextAnalysis(JapaneseContextAnalysis):
    def get_order(self, aStr):
        if not aStr: return -1, 1
        # find out current char's byte length
        if ((aStr[0] >= '\x81') and (aStr[0] <= '\x9F')) or \
           ((aStr[0] >= '\xE0') and (aStr[0] <= '\xFC')):
            charLen = 2
        else:
            charLen = 1

And where does aStr come from? Let’s pop the stack:

def feed(self, aBuf, aLen):
    .
    .
    .
    i = self._mNeedToSkipCharNum
    while i < aLen:
        order, charLen = self.get_order(aBuf[i:i+2])

Oh look, it’s our old friend, aBuf. As you might have guessed from every other issue we’ve encountered in this chapter, aBuf is a byte array. Here, the feed() method isn’t just passing it on wholesale; it’s slicing it. But as you saw earlier in this chapter, slicing a byte array returns a byte array, so the aStr parameter that gets passed to the get_order() method is still a byte array.

And what is this code trying to do with aStr? It’s taking the first element of the byte array and comparing it to a string of length 1. In Python 2, that worked, because aStr and aBuf were strings, and aStr[0] would be a string, and you can compare strings for inequality. But in Python 3, aStr and aBuf are byte arrays, aStr[0] is an integer, and you can’t compare integers and strings for inequality without explicitly coercing one of them.

In this case, there’s no need to make the code more complicated by adding an explicit coercion. aStr[0] yields an integer; the things you’re comparing to are all constants. Let’s change them from 1-character strings to integers. And while we’re at it, let’s change aStr to aBuf, since it’s not actually a string.

  class SJISContextAnalysis(JapaneseContextAnalysis):
-     def get_order(self, aStr):
-      if not aStr: return -1, 1
+     def get_order(self, aBuf):
+      if not aBuf: return -1, 1
          # find out current char's byte length
-         if ((aStr[0] >= '\x81') and (aStr[0] <= '\x9F')) or \
-            ((aBuf[0] >= '\xE0') and (aBuf[0] <= '\xFC')):
+         if ((aBuf[0] >= 0x81) and (aBuf[0] <= 0x9F)) or \
+            ((aBuf[0] >= 0xE0) and (aBuf[0] <= 0xFC)):
              charLen = 2
          else:
              charLen = 1

          # return its order if it is hiragana
-      if len(aStr) > 1:
-             if (aStr[0] == '\202') and \
-                (aStr[1] >= '\x9F') and \
-                (aStr[1] <= '\xF1'):
-                return ord(aStr[1]) - 0x9F, charLen
+      if len(aBuf) > 1:
+             if (aBuf[0] == 202) and \
+                (aBuf[1] >= 0x9F) and \
+                (aBuf[1] <= 0xF1):
+                return aBuf[1] - 0x9F, charLen

          return -1, charLen

  class EUCJPContextAnalysis(JapaneseContextAnalysis):
-     def get_order(self, aStr):
-      if not aStr: return -1, 1
+     def get_order(self, aBuf):
+      if not aBuf: return -1, 1
          # find out current char's byte length
-         if (aStr[0] == '\x8E') or \
-           ((aStr[0] >= '\xA1') and (aStr[0] <= '\xFE')):
+         if (aBuf[0] == 0x8E) or \
+           ((aBuf[0] >= 0xA1) and (aBuf[0] <= 0xFE)):
              charLen = 2
-         elif aStr[0] == '\x8F':
+         elif aBuf[0] == 0x8F:
              charLen = 3
          else:
              charLen = 1

        # return its order if it is hiragana
-    if len(aStr) > 1:
-           if (aStr[0] == '\xA4') and \
-              (aStr[1] >= '\xA1') and \
-              (aStr[1] <= '\xF3'):
-                 return ord(aStr[1]) - 0xA1, charLen
+    if len(aBuf) > 1:
+           if (aBuf[0] == 0xA4) and \
+              (aBuf[1] >= 0xA1) and \
+              (aBuf[1] <= 0xF3):
+               return aBuf[1] - 0xA1, charLen

        return -1, charLen

Searching the entire codebase for occurrences of the ord() function uncovers the same problem in chardistribution.py (specifically, in the EUCTWDistributionAnalysis, EUCKRDistributionAnalysis, GB2312DistributionAnalysis, Big5DistributionAnalysis, SJISDistributionAnalysis, and EUCJPDistributionAnalysis classes. In each case, the fix is similar to the change we made to the EUCJPContextAnalysis and SJISContextAnalysis classes in jpcntx.py.

Global name 'reduce' is not defined

Once more into the breach…

C:\home\chardet> python test.py tests\*\*
tests\ascii\howto.diveintomark.org.xml                       ascii with confidence 1.0
tests\Big5\0804.blogspot.com.xml
Traceback (most recent call last):
  File "test.py", line 12, in <module>
    u.close()
  File "C:\home\chardet\chardet\universaldetector.py", line 141, in close
    proberConfidence = prober.get_confidence()
  File "C:\home\chardet\chardet\latin1prober.py", line 126, in get_confidence
    total = reduce(operator.add, self._mFreqCounter)
NameError: global name 'reduce' is not defined

According to the official What’s New In Python 3.0 guide, the reduce() function has been moved out of the global namespace and into the functools module. Quoting the guide: “Use functools.reduce() if you really need it; however, 99 percent of the time an explicit for loop is more readable.” You can read more about the decision from Guido van Rossum’s weblog: The fate of reduce() in Python 3000.

def get_confidence(self):
    if self.get_state() == constants.eNotMe:
        return 0.01
  
    total = reduce(operator.add, self._mFreqCounter)

The reduce() function takes two arguments — a function and a list (strictly speaking, any iterable object will do) — and applies the function cumulatively to each item of the list. In other words, this is a fancy and roundabout way of adding up all the items in a list and returning the result.

This monstrosity was so common that Python added a global sum() function.

  def get_confidence(self):
      if self.get_state() == constants.eNotMe:
          return 0.01
  
-     total = reduce(operator.add, self._mFreqCounter)
+     total = sum(self._mFreqCounter)

Since you’re no longer using the operator module, you can remove that import from the top of the file as well.

  from .charsetprober import CharSetProber
  from . import constants
- import operator

I CAN HAZ TESTZ?

C:\home\chardet> python test.py tests\*\*
tests\ascii\howto.diveintomark.org.xml                       ascii with confidence 1.0
tests\Big5\0804.blogspot.com.xml                             Big5 with confidence 0.99
tests\Big5\blog.worren.net.xml                               Big5 with confidence 0.99
tests\Big5\carbonxiv.blogspot.com.xml                        Big5 with confidence 0.99
tests\Big5\catshadow.blogspot.com.xml                        Big5 with confidence 0.99
tests\Big5\coolloud.org.tw.xml                               Big5 with confidence 0.99
tests\Big5\digitalwall.com.xml                               Big5 with confidence 0.99
tests\Big5\ebao.us.xml                                       Big5 with confidence 0.99
tests\Big5\fudesign.blogspot.com.xml                         Big5 with confidence 0.99
tests\Big5\kafkatseng.blogspot.com.xml                       Big5 with confidence 0.99
tests\Big5\ke207.blogspot.com.xml                            Big5 with confidence 0.99
tests\Big5\leavesth.blogspot.com.xml                         Big5 with confidence 0.99
tests\Big5\letterlego.blogspot.com.xml                       Big5 with confidence 0.99
tests\Big5\linyijen.blogspot.com.xml                         Big5 with confidence 0.99
tests\Big5\marilynwu.blogspot.com.xml                        Big5 with confidence 0.99
tests\Big5\myblog.pchome.com.tw.xml                          Big5 with confidence 0.99
tests\Big5\oui-design.com.xml                                Big5 with confidence 0.99
tests\Big5\sanwenji.blogspot.com.xml                         Big5 with confidence 0.99
tests\Big5\sinica.edu.tw.xml                                 Big5 with confidence 0.99
tests\Big5\sylvia1976.blogspot.com.xml                       Big5 with confidence 0.99
tests\Big5\tlkkuo.blogspot.com.xml                           Big5 with confidence 0.99
tests\Big5\tw.blog.xubg.com.xml                              Big5 with confidence 0.99
tests\Big5\unoriginalblog.com.xml                            Big5 with confidence 0.99
tests\Big5\upsaid.com.xml                                    Big5 with confidence 0.99
tests\Big5\willythecop.blogspot.com.xml                      Big5 with confidence 0.99
tests\Big5\ytc.blogspot.com.xml                              Big5 with confidence 0.99
tests\EUC-JP\aivy.co.jp.xml                                  EUC-JP with confidence 0.99
tests\EUC-JP\akaname.main.jp.xml                             EUC-JP with confidence 0.99
tests\EUC-JP\arclamp.jp.xml                                  EUC-JP with confidence 0.99
.
.
.
316 tests

Holy crap, it actually works! /me does a little dance

Summary

What have we learned?

  1. Porting any non-trivial amount of code from Python 2 to Python 3 is going to be a pain. There’s no way around it. It’s hard.
  2. The automated 2to3 tool is helpful as far as it goes, but it will only do the easy parts — function renames, module renames, syntax changes. It’s an impressive piece of engineering, but in the end it’s just an intelligent search-and-replace bot.
  3. The #1 porting problem in this library was the difference between strings and bytes. In this case that seems obvious, since the whole point of the chardet library is to convert a stream of bytes into a string. But “a stream of bytes” comes up more often than you might think. Reading a file in “binary” mode? You’ll get a stream of bytes. Fetching a web page? Calling a web API? They return a stream of bytes, too.
  4. You need to understand your program. Thoroughly. Preferably because you wrote it, but at the very least, you need to be comfortable with all its quirks and musty corners. The bugs are everywhere.
  5. Test cases are essential. Don’t port anything without them. The only reason I have any confidence that chardet works in Python 3 is that I started with a test suite that exercised all major code paths. If you don’t have any tests, write some tests before you start porting to Python 3. If you have a few tests, write more. If you have a lot of tests, then the real fun can begin.

© 2001–10 Mark Pilgrim