How to efficiently parse fixed width files?

I’m not really sure if this is efficient, but it should be readable (as opposed to do the slicing manually). I defined a function slices that gets a string and column lengths, and returns the substrings. I made it a generator, so for really long lines, it doesn’t build a temporary list of substrings.

def slices(s, *args):
    position = 0
    for length in args:
        yield s[position:position + length]
        position += length

Example

In [32]: list(slices('abcdefghijklmnopqrstuvwxyz0123456789', 2))
Out[32]: ['ab']

In [33]: list(slices('abcdefghijklmnopqrstuvwxyz0123456789', 2, 10, 50))
Out[33]: ['ab', 'cdefghijkl', 'mnopqrstuvwxyz0123456789']

In [51]: d,c,h = slices('dogcathouse', 3, 3, 5)
In [52]: d,c,h
Out[52]: ('dog', 'cat', 'house')

But I think the advantage of a generator is lost if you need all columns at once. Where one could benefit from is when you want to process columns one by one, say in a loop.

The code below gives a sketch of what you might want to do if you have some serious fixed-column-width file handling to do.

“Serious” = multiple record types in each of multiple file types, records up to 1000 bytes, the layout-definer and “opposing” producer/consumer is a government department with attitude, layout changes result in unused columns, up to a million records in a file, …

Features: Precompiles the struct formats. Ignores unwanted columns. Converts input strings to required data types (sketch omits error handling). Converts records to object instances (or dicts, or named tuples if you prefer).

Code:

import struct, datetime, io, pprint

# functions for converting input fields to usable data
cnv_text = rstrip
cnv_int = int
cnv_date_dmy = lambda s: datetime.datetime.strptime(s, "%d%m%Y") # ddmmyyyy
# etc

# field specs (field name, start pos (1-relative), len, converter func)
fieldspecs = [
    ('surname', 11, 20, cnv_text),
    ('given_names', 31, 20, cnv_text),
    ('birth_date', 51, 8, cnv_date_dmy),
    ('start_date', 71, 8, cnv_date_dmy),
    ]

fieldspecs.sort(key=lambda x: x[1]) # just in case

# build the format for struct.unpack
unpack_len = 0
unpack_fmt = ""
for fieldspec in fieldspecs:
    start = fieldspec[1] - 1
    end = start + fieldspec[2]
    if start > unpack_len:
        unpack_fmt += str(start - unpack_len) + "x"
    unpack_fmt += str(end - start) + "s"
    unpack_len = end
field_indices = range(len(fieldspecs))
print unpack_len, unpack_fmt
unpacker = struct.Struct(unpack_fmt).unpack_from

class Record(object):
    pass
    # or use named tuples

raw_data = """
....v....1....v....2....v....3....v....4....v....5....v....6....v....7....v....8
          Featherstonehaugh   Algernon Marmaduke  31121969            01012005XX
"""

f = cStringIO.StringIO(raw_data)
headings = f.next()
for line in f:
    # The guts of this loop would of course be hidden away in a function/method
    # and could be made less ugly
    raw_fields = unpacker(line)
    r = Record()
    for x in field_indices:
        setattr(r, fieldspecs[x][0], fieldspecs[x][3](raw_fields[x]))
    pprint.pprint(r.__dict__)
    print "Customer name:", r.given_names, r.surname

Output:

78 10x20s20s8s12x8s
{'birth_date': datetime.datetime(1969, 12, 31, 0, 0),
 'given_names': 'Algernon Marmaduke',
 'start_date': datetime.datetime(2005, 1, 1, 0, 0),
 'surname': 'Featherstonehaugh'}
Customer name: Algernon Marmaduke Featherstonehaugh

> str = '1234567890'
> w = [0,2,5,7,10]
> [ str[ w[i-1] : w[i] ] for i in range(1,len(w)) ]
['12', '345', '67', '890']

Two more options that are easier and prettier than already mentioned solutions:

The first is using pandas:

import pandas as pd

path = 'filename.txt'

#inferred - as suggested in the comments by James Paul Mason
data = pd.read_fwf(path, colspecs='infer')

# Or using Pandas with a column specification
col_specification = [(0, 20), (21, 30), (31, 50), (51, 100)]
data = pd.read_fwf(path, colspecs=col_specification)

And the second option using numpy.loadtxt:

import numpy as np

# Using NumPy and letting it figure it out automagically
data_also = np.loadtxt(path)

It really depends on in what way you want to use your data.

Here’s a simple module for Python 3, based on John Machin’s answer – adapt as needed 🙂

"""
fixedwidth

Parse and iterate through a fixedwidth text file, returning record objects.

Adapted from https://stackoverflow.com/a/4916375/243392


USAGE

    import fixedwidth, pprint

    # define the fixed width fields we want
    # fieldspecs is a list of [name, description, start, width, type] arrays.
    fieldspecs = [
        ["FILEID", "File Identification", 1, 6, "A/N"],
        ["STUSAB", "State/U.S. Abbreviation (USPS)", 7, 2, "A"],
        ["SUMLEV", "Summary Level", 9, 3, "A/N"],
        ["LOGRECNO", "Logical Record Number", 19, 7, "N"],
        ["POP100", "Population Count (100%)", 30, 9, "N"],
    ]

    # define the fieldtype conversion functions
    fieldtype_fns = {
        'A': str.rstrip,
        'A/N': str.rstrip,
        'N': int,
    }

    # iterate over record objects in the file
    with open(f, 'rb'):
        for record in fixedwidth.reader(f, fieldspecs, fieldtype_fns):
            pprint.pprint(record.__dict__)

    # output:
    {'FILEID': 'SF1ST', 'LOGRECNO': 2, 'POP100': 1, 'STUSAB': 'TX', 'SUMLEV': '040'}
    {'FILEID': 'SF1ST', 'LOGRECNO': 3, 'POP100': 2, 'STUSAB': 'TX', 'SUMLEV': '040'}    
    ...

"""

import struct, io


# fieldspec columns
iName, iDescription, iStart, iWidth, iType = range(5)


def get_struct_unpacker(fieldspecs):
    """
    Build the format string for struct.unpack to use, based on the fieldspecs.
    fieldspecs is a list of [name, description, start, width, type] arrays.
    Returns a string like "6s2s3s7x7s4x9s".
    """
    unpack_len = 0
    unpack_fmt = ""
    for fieldspec in fieldspecs:
        start = fieldspec[iStart] - 1
        end = start + fieldspec[iWidth]
        if start > unpack_len:
            unpack_fmt += str(start - unpack_len) + "x"
        unpack_fmt += str(end - start) + "s"
        unpack_len = end
    struct_unpacker = struct.Struct(unpack_fmt).unpack_from
    return struct_unpacker


class Record(object):
    pass
    # or use named tuples


def reader(f, fieldspecs, fieldtype_fns):
    """
    Wrap a fixedwidth file and return records according to the given fieldspecs.
    fieldspecs is a list of [name, description, start, width, type] arrays.
    fieldtype_fns is a dictionary of functions used to transform the raw string values, 
    one for each type.
    """

    # make sure fieldspecs are sorted properly
    fieldspecs.sort(key=lambda fieldspec: fieldspec[iStart])

    struct_unpacker = get_struct_unpacker(fieldspecs)

    field_indices = range(len(fieldspecs))

    for line in f:
        raw_fields = struct_unpacker(line) # split line into field values
        record = Record()
        for i in field_indices:
            fieldspec = fieldspecs[i]
            fieldname = fieldspec[iName]
            s = raw_fields[i].decode() # convert raw bytes to a string
            fn = fieldtype_fns[fieldspec[iType]] # get conversion function
            value = fn(s) # convert string to value (eg to an int)
            setattr(record, fieldname, value)
        yield record


if __name__=='__main__':

    # test module

    import pprint, io

    # define the fields we want
    # fieldspecs are [name, description, start, width, type]
    fieldspecs = [
        ["FILEID", "File Identification", 1, 6, "A/N"],
        ["STUSAB", "State/U.S. Abbreviation (USPS)", 7, 2, "A"],
        ["SUMLEV", "Summary Level", 9, 3, "A/N"],
        ["LOGRECNO", "Logical Record Number", 19, 7, "N"],
        ["POP100", "Population Count (100%)", 30, 9, "N"],
    ]

    # define a conversion function for integers
    def to_int(s):
        """
        Convert a numeric string to an integer.
        Allows a leading ! as an indicator of missing or uncertain data.
        Returns None if no data.
        """
        try:
            return int(s)
        except:
            try:
                return int(s[1:]) # ignore a leading !
            except:
                return None # assume has a leading ! and no value

    # define the conversion fns
    fieldtype_fns = {
        'A': str.rstrip,
        'A/N': str.rstrip,
        'N': to_int,
        # 'N': int,
        # 'D': lambda s: datetime.datetime.strptime(s, "%d%m%Y"), # ddmmyyyy
        # etc
    }

    # define a fixedwidth sample
    sample = """
SF1ST TX04089000  00000023748        1 
SF1ST TX04090000  00000033748!       2
SF1ST TX04091000  00000043748!        
"""
    sample_data = sample.encode() # convert string to bytes
    file_like = io.BytesIO(sample_data) # create a file-like wrapper around bytes

    # iterate over record objects in the file
    for record in reader(file_like, fieldspecs, fieldtype_fns):
        # print(record)
        pprint.pprint(record.__dict__)

Here is what NumPy uses under the hood (much much simplified, but still – this code is found in the LineSplitter class within the _iotools module):

import numpy as np

DELIMITER = (20, 10, 10, 20, 10, 10, 20)

idx = np.cumsum([0] + list(DELIMITER))
slices = [slice(i, j) for (i, j) in zip(idx[:-1], idx[1:])]

def parse(line):
    return [line[s] for s in slices]

It does not handle negative delimiters for ignoring column so it is not as versatile as struct, but it is faster.

String slicing doesn’t have to be ugly as long as you keep it organized. Consider storing your field widths in a dictionary and then using the associated names to create an object:

from collections import OrderedDict

class Entry:
    def __init__(self, line):

        name2width = OrderedDict()
        name2width['foo'] = 2
        name2width['bar'] = 3
        name2width['baz'] = 2

        pos = 0
        for name, width in name2width.items():

            val = line[pos : pos + width]
            if len(val) != width:
                raise ValueError("not enough characters: '{}'".format(line))

            setattr(self, name, val)
            pos += width

file = "ab789yzncd987wxnef555uv"

entry = []

for line in file.split('n'):
    entry.append(Entry(line))

print(entry[1].bar) # output: 987

Because my old work often handles 1 million lines of fixwidth data, I did research on this issue when I started using Python.

There are 2 types of FixedWidth

1. ASCII FixedWidth (ascii character length = 1, double-byte encoded character length = 2)
2. Unicode FixedWidth (ascii character & double-byte encoded character length = 1)

If the resource string is all composed of ascii characters, then ASCII FixedWidth = Unicode FixedWidth

Fortunately, string and byte are different in py3, which reduces a lot of confusion when dealing with double-byte encoded characters (e.g.gbk, big5, euc-jp, shift-jis, etc.).
For the processing of “ASCII FixedWidth”, the String is usually converted to Bytes and then split.

Without importing third-party modules
totalLineCount = 1 million, lineLength = 800 byte , FixedWidthArgs=(10,25,4,….), I split the Line in about 5 ways and get the following conclusion:

1. struct is the fastest (1x)
2. Loop only, not pre-processing FixedWidthArgs is the slowest (5x+)
3. slice(bytes) is faster than slice(string)
4. The source string is the bytes test result: struct(1x) , operator.itemgetter(1.7x) , precompiled sliceObject & list comprehensions(2.8x), re.patten object (2.9x)

When dealing with large files, we often use with open ( file, "rb") as f:.
The method traverses one of the above files, about 2.4 second.
I think the appropriate handler, which processes 1 million rows of data, splits each row into 20 fields and takes less than 2.4 seconds.

I only find that stuct and itemgetter meet the requirements

ps: For normal display, I converted unicode str to bytes.
If you are in a double-byte environment, you don’t need to do this.

from itertools import accumulate
from operator import itemgetter

def oprt_parser(sArgs):
    sum_arg = tuple(accumulate(abs(i) for i in sArgs))
    # Negative parameter field index
    cuts = tuple(i for i,num in enumerate(sArgs) if num < 0)
    # Get slice args and Ignore fields of negative length
    ig_Args = tuple(item for i, item in enumerate(zip((0,)+sum_arg,sum_arg)) if i not in cuts)
    # Generate `operator.itemgetter` object
    oprtObj =itemgetter(*[slice(s,e) for s,e in ig_Args])
    return oprtObj

lineb = b'abcdefghijklmnopqrstuvwxyzxb0xa1xb2xbbxb4xd3xb5xc4xb6xeexb7xa2xb8xf6xbaxcd0123456789'
line = lineb.decode("GBK")

# Unicode Fixed Width
fieldwidthsU = (13, -13, 4, -4, 5,-5) # Negative width fields is ignored
# ASCII Fixed Width
fieldwidths = (13, -13, 8, -8, 5,-5) # Negative width fields is ignored
# Unicode FixedWidth processing
parse = oprt_parser(fieldwidthsU)
fields = parse(line)
print('Unicode FixedWidth','fields: {}'.format(tuple(map(lambda s: s.encode("GBK"), fields))))
# ASCII FixedWidth processing
parse = oprt_parser(fieldwidths)
fields = parse(lineb)
print('ASCII FixedWidth','fields: {}'.format(fields))
line = 'ABCDEFGHIJKLMNOPQRSTUVWXYZ0123456789n'
fieldwidths = (2, -10, 24)
parse = oprt_parser(fieldwidths)
fields = parse(line)
print(f"fields: {fields}")

Output:

Unicode FixedWidth fields: (b'abcdefghijklm', b'xb0xa1xb2xbbxb4xd3xb5xc4', b'01234')
ASCII FixedWidth fields: (b'abcdefghijklm', b'xb0xa1xb2xbbxb4xd3xb5xc4', b'01234')
fields: ('AB', 'MNOPQRSTUVWXYZ0123456789')

oprt_parser is 4x make_parser(list comprehensions + slice)

---

During the research, it was found that when the cpu speed is faster, it seems that the efficiency of the re method increases faster.
Since I don’t have more and better computers to test, provide my test code, if anyone is interested, you can test it with a faster computer.

Run Environment:

• os:win10
• python: 3.7.2
• CPU:amd athlon x3 450
• HD:seagate 1T

import timeit
import time
import re
from itertools import accumulate
from operator import itemgetter

def eff2(stmt,onlyNum= False,showResult=False):
    '''test function'''
    if onlyNum:
        rl = timeit.repeat(stmt=stmt,repeat=roundI,number=timesI,globals=globals())
        avg = sum(rl) / len(rl)
        return f"{avg * (10 ** 6)/timesI:0.4f}"
    else:
        rl = timeit.repeat(stmt=stmt,repeat=10,number=1000,globals=globals())
        avg = sum(rl) / len(rl)
        print(f"【{stmt}】")
        print(f"tquick avg = {avg * (10 ** 6)/1000:0.4f} s/million")
        if showResult:
            print(f"t  Result = {eval(stmt)}nt  timelist = {rl}n")
        else:
            print("")

def upDouble(argList,argRate):
    return [c*argRate for c in argList]

tbStr = "000000001111000002222真2233333333000000004444444QAZ55555555000000006666666ABC这些事中文字abcdefghijk"
tbBytes = tbStr.encode("GBK")
a20 = (4,4,2,2,2,3,2,2, 2 ,2,8,8,7,3,8,8,7,3, 12 ,11)
a20U = (4,4,2,2,2,3,2,2, 1 ,2,8,8,7,3,8,8,7,3, 6 ,11)
Slng = 800
rateS = Slng // 100

tStr = "".join(upDouble(tbStr , rateS))
tBytes = tStr.encode("GBK")
spltArgs = upDouble( a20 , rateS)
spltArgsU = upDouble( a20U , rateS)

testList = []
timesI = 100000
roundI = 5
print(f"test round = {roundI} timesI = {timesI} sourceLng = {len(tStr)} argFieldCount = {len(spltArgs)}")


print(f"pure str n{''.ljust(60,'-')}")
# ==========================================
def str_parser(sArgs):
    def prsr(oStr):
        r = []
        r_ap = r.append
        stt=0
        for lng in sArgs:
            end = stt + lng 
            r_ap(oStr[stt:end])
            stt = end 
        return tuple(r)
    return prsr

Str_P = str_parser(spltArgsU)
# eff2("Str_P(tStr)")
testList.append("Str_P(tStr)")

print(f"pure bytes n{''.ljust(60,'-')}")
# ==========================================
def byte_parser(sArgs):
    def prsr(oBytes):
        r, stt = [], 0
        r_ap = r.append
        for lng in sArgs:
            end = stt + lng
            r_ap(oBytes[stt:end])
            stt = end
        return r
    return prsr
Byte_P = byte_parser(spltArgs)
# eff2("Byte_P(tBytes)")
testList.append("Byte_P(tBytes)")

# re,bytes
print(f"re compile object n{''.ljust(60,'-')}")
# ==========================================


def rebc_parser(sArgs,otype="b"):
    re_Args = "".join([f"(.{{{n}}})" for n in sArgs])
    if otype == "b":
        rebc_Args = re.compile(re_Args.encode("GBK"))
    else:
        rebc_Args = re.compile(re_Args)
    def prsr(oBS):
        return rebc_Args.match(oBS).groups()
    return prsr
Rebc_P = rebc_parser(spltArgs)
# eff2("Rebc_P(tBytes)")
testList.append("Rebc_P(tBytes)")

Rebc_Ps = rebc_parser(spltArgsU,"s")
# eff2("Rebc_Ps(tStr)")
testList.append("Rebc_Ps(tStr)")


print(f"struct n{''.ljust(60,'-')}")
# ==========================================

import struct
def struct_parser(sArgs):
    struct_Args = " ".join(map(lambda x: str(x) + "s", sArgs))
    def prsr(oBytes):
        return struct.unpack(struct_Args, oBytes)
    return prsr
Struct_P = struct_parser(spltArgs)
# eff2("Struct_P(tBytes)")
testList.append("Struct_P(tBytes)")

print(f"List Comprehensions + slice n{''.ljust(60,'-')}")
# ==========================================
import itertools
def slice_parser(sArgs):
    tl = tuple(itertools.accumulate(sArgs))
    slice_Args = tuple(zip((0,)+tl,tl))
    def prsr(oBytes):
        return [oBytes[s:e] for s, e in slice_Args]
    return prsr
Slice_P = slice_parser(spltArgs)
# eff2("Slice_P(tBytes)")
testList.append("Slice_P(tBytes)")

def sliceObj_parser(sArgs):
    tl = tuple(itertools.accumulate(sArgs))
    tl2 = tuple(zip((0,)+tl,tl))
    sliceObj_Args = tuple(slice(s,e) for s,e in tl2)
    def prsr(oBytes):
        return [oBytes[so] for so in sliceObj_Args]
    return prsr
SliceObj_P = sliceObj_parser(spltArgs)
# eff2("SliceObj_P(tBytes)")
testList.append("SliceObj_P(tBytes)")

SliceObj_Ps = sliceObj_parser(spltArgsU)
# eff2("SliceObj_Ps(tStr)")
testList.append("SliceObj_Ps(tStr)")


print(f"operator.itemgetter + slice object n{''.ljust(60,'-')}")
# ==========================================

def oprt_parser(sArgs):
    sum_arg = tuple(accumulate(abs(i) for i in sArgs))
    cuts = tuple(i for i,num in enumerate(sArgs) if num < 0)
    ig_Args = tuple(item for i,item in enumerate(zip((0,)+sum_arg,sum_arg)) if i not in cuts)
    oprtObj =itemgetter(*[slice(s,e) for s,e in ig_Args])
    return oprtObj

Oprt_P = oprt_parser(spltArgs)
# eff2("Oprt_P(tBytes)")
testList.append("Oprt_P(tBytes)")

Oprt_Ps = oprt_parser(spltArgsU)
# eff2("Oprt_Ps(tStr)")
testList.append("Oprt_Ps(tStr)")

print("|".join([s.split("(")[0].center(11," ") for s in testList]))
print("|".join(["".center(11,"-") for s in testList]))
print("|".join([eff2(s,True).rjust(11," ") for s in testList]))

Output:

Test round = 5 timesI = 100000 sourceLng = 744 argFieldCount = 20
...
...
   Str_P | Byte_P | Rebc_P | Rebc_Ps | Struct_P | Slice_P | SliceObj_P|SliceObj_Ps| Oprt_P | Oprt_Ps
-----------|-----------|-----------|-----------|-- ---------|-----------|-----------|-----------|---- -------|-----------
     9.6315| 7.5952| 4.4187| 5.6867| 1.5123| 5.2915| 4.2673| 5.7121| 2.4713| 3.9051

This is how I solved with a dictionary that contains where fields start and end. Giving start and end points helped me to manage changes at the length of the column also.

# fixed length
#      '---------- ------- ----------- -----------'
line = '20.06.2019 myname  active      mydevice   '
SLICES = {'date_start': 0,
         'date_end': 10,
         'name_start': 11,
         'name_end': 18,
         'status_start': 19,
         'status_end': 30,
         'device_start': 31,
         'device_end': 42}

def get_values_as_dict(line, SLICES):
    values = {}
    key_list = {key.split("_")[0] for key in SLICES.keys()}
    for key in key_list:
       values[key] = line[SLICES[key+"_start"]:SLICES[key+"_end"]].strip()
    return values

>>> print (get_values_as_dict(line,SLICES))
{'status': 'active', 'name': 'myname', 'date': '20.06.2019', 'device': 'mydevice'}

I like to process text files containing fixed width fields using regular expressions. More specifically, using named capture groups. It’s fast, does not require importing large libraries and is quite descriptive and convenient (in my opinion).

I also like the fact that the named capture groups are basically auto-documenting the data format, acting as a sort of data specification, since each capture group can be written to define each fields’ name, data type and length.

Here’s simple example…

import re

data = [
    "1234ABCDEFGHIJ5", 
    "6789KLMNOPQRST0"
]

record_regex = (
    r"^"
    r"(?P<firstnumbers>[0-9]{4})"
    r"(?P<middletext>[a-zA-Z0-9_-s]{10})"
    r"(?P<lastnumber>[0-9]{1})"
    r"$"
)

records = []

for line in data:
    match = re.match(record_regex, line)
    if match:
        records.append(match.groupdict())

print(records)

…that yields a convenient dictionary of each record:

[
    {'firstnumbers': '1234', 'lastnumber': '5', 'middletext': 'ABCDEFGHIJ'},
    {'firstnumbers': '6789', 'lastnumber': '0', 'middletext': 'KLMNOPQRST'}
]

Helpful tools, like the online regex tester and debugger, are available if you are not familiar (or comfortable) with Python regular expressions or named capture groups.