Python UTF-16 CSV reader

Question

I have a UTF-16 CSV file which I have to read. Python csv module does not seem to support UTF-16.

I am using python 2.7.2. CSV files I need to parse are huge size running into several GBs of data.

Answers for John Machin questions below

print repr(open('test.csv', 'rb').read(100))

Output with test.csv having just abc as content

'\xff\xfea\x00b\x00c\x00'

I think csv file got created on windows machine in USA. I am using Mac OSX Lion.

If I use code provided by phihag and test.csv containing one record.

sample test.csv content used. Below is print repr(open('test.csv', 'rb').read(1000)) output

'\xff\xfe1\x00,\x002\x00,\x00G\x00,\x00S\x00,\x00H\x00 \x00f\x00\xfc\x00r\x00 \x00e\x00 \x00\x96\x00 \x00m\x00 \x00\x85\x00,\x00,\x00I\x00\r\x00\n\x00'

Code by phihag

import codecs
import csv
with open('test.csv','rb') as f:
      sr = codecs.StreamRecoder(f,codecs.getencoder('utf-8'),codecs.getdecoder('utf-8'),codecs.getreader('utf-16'),codecs.getwriter('utf-16'))      
      for row in csv.reader(sr):
         print row

Output of the above code

['1', '2', 'G', 'S', 'H f\xc3\xbcr e \xc2\x96 m \xc2\x85']
['', '', 'I']

expected output is

['1', '2', 'G', 'S', 'H f\xc3\xbcr e \xc2\x96 m \xc2\x85','','I']

Answer 1

At the moment, the csv module does not support UTF-16.

In Python 3.x, csv expects a text-mode file and you can simply use the encoding parameter of open to force another encoding:

# Python 3.x only
import csv
with open('utf16.csv', 'r', encoding='utf16') as csvf:
    for line in csv.reader(csvf):
        print(line) # do something with the line

In Python 2.x, you can recode the input:

# Python 2.x only
import codecs
import csv

class Recoder(object):
    def __init__(self, stream, decoder, encoder, eol='\r\n'):
        self._stream = stream
        self._decoder = decoder if isinstance(decoder, codecs.IncrementalDecoder) else codecs.getincrementaldecoder(decoder)()
        self._encoder = encoder if isinstance(encoder, codecs.IncrementalEncoder) else codecs.getincrementalencoder(encoder)()
        self._buf = ''
        self._eol = eol
        self._reachedEof = False

    def read(self, size=None):
        r = self._stream.read(size)
        raw = self._decoder.decode(r, size is None)
        return self._encoder.encode(raw)

    def __iter__(self):
        return self

    def __next__(self):
        if self._reachedEof:
            raise StopIteration()
        while True:
            line,eol,rest = self._buf.partition(self._eol)
            if eol == self._eol:
                self._buf = rest
                return self._encoder.encode(line + eol)
            raw = self._stream.read(1024)
            if raw == '':
                self._decoder.decode(b'', True)
                self._reachedEof = True
                return self._encoder.encode(self._buf)
            self._buf += self._decoder.decode(raw)
    next = __next__

    def close(self):
        return self._stream.close()

with open('test.csv','rb') as f:
    sr = Recoder(f, 'utf-16', 'utf-8')

    for row in csv.reader(sr):
        print (row)

open and codecs.open require the file to start with a BOM. If it doesn't (or you're on Python 2.x), you can still convert it in memory, like this:

try:
    from io import BytesIO
except ImportError: # Python < 2.6
    from StringIO import StringIO as BytesIO
import csv
with open('utf16.csv', 'rb') as binf:
    c = binf.read().decode('utf-16').encode('utf-8')
for line in csv.reader(BytesIO(c)):
    print(line) # do something with the line

Answer 2

The Python 2.x csv module documentation example shows how to handle other encodings.

Answer 3

I would strongly suggest that you recode your file(s) to UTF-8. Under the very likely condition that you don't have any Unicode characters outside the BMP, you can take advantage of the fact that UTF-16 is a fixed-length encoding to read fixed-length blocks from your input file without worrying about straddling block boundaries.

Step 1: Determine what encoding you actually have. Examine the first few bytes of your file:

print repr(open('thefile.csv', 'rb').read(100))

Four possible ways of encoding u'abc'

\xfe\xff\x00a\x00b\x00c -> utf_16
\xff\xfea\x00b\x00c\x00 -> utf_16
\x00a\x00b\x00c -> utf_16_be
a\x00b\x00c\x00 -> utf_16_le

If you have any trouble with this step, edit your question to include the results of the above print repr()

Step 2: Here's a Python 2.X recode-UTF-16*-to-UTF-8 script:

import sys
infname, outfname, enc = sys.argv[1:4]
fi = open(infname, 'rb')
fo = open(outfname, 'wb')
BUFSIZ = 64 * 1024 * 1024
first = True
while 1:
    buf = fi.read(BUFSIZ)
    if not buf: break
    if first and enc == 'utf_16':
        bom = buf[:2]
        buf = buf[2:]
        enc = {'\xfe\xff': 'utf_16_be', '\xff\xfe': 'utf_16_le'}[bom]
        # KeyError means file doesn't start with a valid BOM
    first = False
    fo.write(buf.decode(enc).encode('utf8'))
fi.close()
fo.close()

Other matters:

You say that your files are too big to read the whole file, recode and rewrite, yet you can open it in vi. Please explain.

The <85> being treated as end of record is a bit of a worry. Looks like 0x85 is being recognised as NEL (C1 control code, NEWLINE). There is a strong possibility that the data was originally encoded in some legacy single-byte encoding where 0x85 has a meaning but has been transcoded to UTF-16 under the false assumption that the original encoding was ISO-8859-1 aka latin1. Where did the file originate? An IBM mainframe? Windows/Unix/classic Mac? What country, locale, language? You obviously think that the <85> is not meant to be a newline; what do you think that it means?

Please feel free to send a copy of a cut-down file (that includes some of the <85> stuff) to sjmachin at lexicon dot net

Update based on 1-line sample data provided.

This confirms my suspicions. Read this. Here's a quote from it:

... the C1 control characters ... are rarely used directly, except on specific platforms such as OpenVMS. When they turn up in documents, Web pages, e-mail messages, etc., which are ostensibly in an ISO-8859-n encoding, their code positions generally refer instead to the characters at that position in a proprietary, system-specific encoding such as Windows-1252 or the Apple Macintosh ("MacRoman") character set that use the codes provided for representation of the C1 set with a single 8-bit byte to instead provide additional graphic characters

This code:

s1 = '\xff\xfe1\x00,\x002\x00,\x00G\x00,\x00S\x00,\x00H\x00 \x00f\x00\xfc\x00r\x00 \x00e\x00 \x00\x96\x00 \x00m\x00 \x00\x85\x00,\x00,\x00I\x00\r\x00\n\x00'
s2 = s1.decode('utf16')
print 's2 repr:', repr(s2)
from unicodedata import name
from collections import Counter
non_ascii = Counter(c for c in s2 if c >= u'\x80')
print 'non_ascii:', non_ascii
for c in non_ascii:
    print "from: U+%04X %s" % (ord(c), name(c, "<no name>"))
    c2 = c.encode('latin1').decode('cp1252')
    print "to:   U+%04X %s" % (ord(c2), name(c2, "<no name>"))

s3 = u''.join(
    c.encode('latin1').decode('1252') if u'\x80' <= c < u'\xA0' else c
    for c in s2
    )
print 's3 repr:', repr(s3)
print 's3:', s3

produces the following (Python 2.7.2 IDLE, Windows 7):

s2 repr: u'1,2,G,S,H f\xfcr e \x96 m \x85,,I\r\n'
non_ascii: Counter({u'\x85': 1, u'\xfc': 1, u'\x96': 1})
from: U+0085 <no name>
to:   U+2026 HORIZONTAL ELLIPSIS
from: U+00FC LATIN SMALL LETTER U WITH DIAERESIS
to:   U+00FC LATIN SMALL LETTER U WITH DIAERESIS
from: U+0096 <no name>
to:   U+2013 EN DASH
s3 repr: u'1,2,G,S,H f\xfcr e \u2013 m \u2026,,I\r\n'
s3: 1,2,G,S,H für e – m …,,I

Which do you think is a more reasonable interpretation of \x96:

SPA i.e. Start of Protected Area (Used by block-oriented terminals.)
or
EN DASH
?

Looks like a thorough analysis of a much larger data sample is warranted. Happy to help.

Answer 4

Just open your file with codecs.open like in

import codecs, csv

stream = codecs.open(<yourfile.csv>, encoding="utf-16")
reader = csv.reader(stream)

And work through your program with unicode strings, as you should do anyway if you are processing text