Fetching contributors…
Cannot retrieve contributors at this time
671 lines (550 sloc) 19.4 KB
#! /usr/bin/env python
# Last Change: Mon Aug 20 08:00 PM 2007 J
from __future__ import division, print_function, absolute_import
import re
import itertools
import datetime
from functools import partial
import numpy as np
from scipy._lib.six import next
"""A module to read arff files."""
__all__ = ['MetaData', 'loadarff', 'ArffError', 'ParseArffError']
# An Arff file is basically two parts:
# - header
# - data
# A header has each of its components starting by @META where META is one of
# the keyword (attribute of relation, for now).
# - both integer and reals are treated as numeric -> the integer info
# is lost!
# - Replace ValueError by ParseError or something
# We know can handle the following:
# - numeric and nominal attributes
# - missing values for numeric attributes
r_meta = re.compile(r'^\s*@')
# Match a comment
r_comment = re.compile(r'^%')
# Match an empty line
r_empty = re.compile(r'^\s+$')
# Match a header line, that is a line which starts by @ + a word
r_headerline = re.compile(r'^@\S*')
r_datameta = re.compile(r'^@[Dd][Aa][Tt][Aa]')
r_relation = re.compile(r'^@[Rr][Ee][Ll][Aa][Tt][Ii][Oo][Nn]\s*(\S*)')
r_attribute = re.compile(r'^@[Aa][Tt][Tt][Rr][Ii][Bb][Uu][Tt][Ee]\s*(..*$)')
# To get attributes name enclosed with ''
r_comattrval = re.compile(r"'(..+)'\s+(..+$)")
# To get normal attributes
r_wcomattrval = re.compile(r"(\S+)\s+(..+$)")
# Module defined exception
class ArffError(IOError):
class ParseArffError(ArffError):
# Various utilities
# An attribute is defined as @attribute name value
def parse_type(attrtype):
"""Given an arff attribute value (meta data), returns its type.
Expect the value to be a name."""
uattribute = attrtype.lower().strip()
if uattribute[0] == '{':
return 'nominal'
elif uattribute[:len('real')] == 'real':
return 'numeric'
elif uattribute[:len('integer')] == 'integer':
return 'numeric'
elif uattribute[:len('numeric')] == 'numeric':
return 'numeric'
elif uattribute[:len('string')] == 'string':
return 'string'
elif uattribute[:len('relational')] == 'relational':
return 'relational'
elif uattribute[:len('date')] == 'date':
return 'date'
raise ParseArffError("unknown attribute %s" % uattribute)
def get_nominal(attribute):
"""If attribute is nominal, returns a list of the values"""
return attribute.split(',')
def read_data_list(ofile):
"""Read each line of the iterable and put it in a list."""
data = [next(ofile)]
if data[0].strip()[0] == '{':
raise ValueError("This looks like a sparse ARFF: not supported yet")
data.extend([i for i in ofile])
return data
def get_ndata(ofile):
"""Read the whole file to get number of data attributes."""
data = [next(ofile)]
loc = 1
if data[0].strip()[0] == '{':
raise ValueError("This looks like a sparse ARFF: not supported yet")
for i in ofile:
loc += 1
return loc
def maxnomlen(atrv):
"""Given a string containing a nominal type definition, returns the
string len of the biggest component.
A nominal type is defined as seomthing framed between brace ({}).
atrv : str
Nominal type definition
slen : int
length of longest component
maxnomlen("{floup, bouga, fl, ratata}") returns 6 (the size of
ratata, the longest nominal value).
>>> maxnomlen("{floup, bouga, fl, ratata}")
nomtp = get_nom_val(atrv)
return max(len(i) for i in nomtp)
def get_nom_val(atrv):
"""Given a string containing a nominal type, returns a tuple of the
possible values.
A nominal type is defined as something framed between braces ({}).
atrv : str
Nominal type definition
poss_vals : tuple
possible values
>>> get_nom_val("{floup, bouga, fl, ratata}")
('floup', 'bouga', 'fl', 'ratata')
r_nominal = re.compile('{(.+)}')
m = r_nominal.match(atrv)
if m:
return tuple(i.strip() for i in','))
raise ValueError("This does not look like a nominal string")
def get_date_format(atrv):
r_date = re.compile(r"[Dd][Aa][Tt][Ee]\s+[\"']?(.+?)[\"']?$")
m = r_date.match(atrv)
if m:
pattern =
# convert time pattern from Java's SimpleDateFormat to C's format
datetime_unit = None
if "yyyy" in pattern:
pattern = pattern.replace("yyyy", "%Y")
datetime_unit = "Y"
elif "yy":
pattern = pattern.replace("yy", "%y")
datetime_unit = "Y"
if "MM" in pattern:
pattern = pattern.replace("MM", "%m")
datetime_unit = "M"
if "dd" in pattern:
pattern = pattern.replace("dd", "%d")
datetime_unit = "D"
if "HH" in pattern:
pattern = pattern.replace("HH", "%H")
datetime_unit = "h"
if "mm" in pattern:
pattern = pattern.replace("mm", "%M")
datetime_unit = "m"
if "ss" in pattern:
pattern = pattern.replace("ss", "%S")
datetime_unit = "s"
if "z" in pattern or "Z" in pattern:
raise ValueError("Date type attributes with time zone not "
"supported, yet")
if datetime_unit is None:
raise ValueError("Invalid or unsupported date format")
return pattern, datetime_unit
raise ValueError("Invalid or no date format")
def go_data(ofile):
"""Skip header.
the first next() call of the returned iterator will be the @data line"""
return itertools.dropwhile(lambda x: not r_datameta.match(x), ofile)
# Parsing header
def tokenize_attribute(iterable, attribute):
"""Parse a raw string in header (eg starts by @attribute).
Given a raw string attribute, try to get the name and type of the
attribute. Constraints:
* The first line must start with @attribute (case insensitive, and
space like characters before @attribute are allowed)
* Works also if the attribute is spread on multilines.
* Works if empty lines or comments are in between
attribute : str
the attribute string.
name : str
name of the attribute
value : str
value of the attribute
next : str
next line to be parsed
If attribute is a string defined in python as r"floupi real", will
return floupi as name, and real as value.
>>> iterable = iter([0] * 10) # dummy iterator
>>> tokenize_attribute(iterable, r"@attribute floupi real")
('floupi', 'real', 0)
If attribute is r"'floupi 2' real", will return 'floupi 2' as name,
and real as value.
>>> tokenize_attribute(iterable, r" @attribute 'floupi 2' real ")
('floupi 2', 'real', 0)
sattr = attribute.strip()
mattr = r_attribute.match(sattr)
if mattr:
# atrv is everything after @attribute
atrv =
if r_comattrval.match(atrv):
name, type = tokenize_single_comma(atrv)
next_item = next(iterable)
elif r_wcomattrval.match(atrv):
name, type = tokenize_single_wcomma(atrv)
next_item = next(iterable)
# Not sure we should support this, as it does not seem supported by
# weka.
raise ValueError("multi line not supported yet")
#name, type, next_item = tokenize_multilines(iterable, atrv)
raise ValueError("First line unparsable: %s" % sattr)
if type == 'relational':
raise ValueError("relational attributes not supported yet")
return name, type, next_item
def tokenize_single_comma(val):
# XXX we match twice the same string (here and at the caller level). It is
# stupid, but it is easier for now...
m = r_comattrval.match(val)
if m:
name =
type =
except IndexError:
raise ValueError("Error while tokenizing attribute")
raise ValueError("Error while tokenizing single %s" % val)
return name, type
def tokenize_single_wcomma(val):
# XXX we match twice the same string (here and at the caller level). It is
# stupid, but it is easier for now...
m = r_wcomattrval.match(val)
if m:
name =
type =
except IndexError:
raise ValueError("Error while tokenizing attribute")
raise ValueError("Error while tokenizing single %s" % val)
return name, type
def read_header(ofile):
"""Read the header of the iterable ofile."""
i = next(ofile)
# Pass first comments
while r_comment.match(i):
i = next(ofile)
# Header is everything up to DATA attribute ?
relation = None
attributes = []
while not r_datameta.match(i):
m = r_headerline.match(i)
if m:
isattr = r_attribute.match(i)
if isattr:
name, type, i = tokenize_attribute(ofile, i)
attributes.append((name, type))
isrel = r_relation.match(i)
if isrel:
relation =
raise ValueError("Error parsing line %s" % i)
i = next(ofile)
i = next(ofile)
return relation, attributes
# Parsing actual data
def safe_float(x):
"""given a string x, convert it to a float. If the stripped string is a ?,
return a Nan (missing value).
x : str
string to convert
f : float
where float can be nan
>>> safe_float('1')
>>> safe_float('1\\n')
>>> safe_float('?\\n')
if '?' in x:
return np.nan
return float(x)
def safe_nominal(value, pvalue):
svalue = value.strip()
if svalue in pvalue:
return svalue
elif svalue == '?':
return svalue
raise ValueError("%s value not in %s" % (str(svalue), str(pvalue)))
def safe_date(value, date_format, datetime_unit):
date_str = value.strip().strip("'").strip('"')
if date_str == '?':
return np.datetime64('NaT', datetime_unit)
dt = datetime.datetime.strptime(date_str, date_format)
return np.datetime64(dt).astype("datetime64[%s]" % datetime_unit)
class MetaData(object):
"""Small container to keep useful informations on a ARFF dataset.
Knows about attributes names and types.
data, meta = loadarff('iris.arff')
# This will print the attributes names of the iris.arff dataset
for i in meta:
print i
# This works too
# Getting attribute type
types = meta.types()
Also maintains the list of attributes in order, i.e. doing for i in
meta, where meta is an instance of MetaData, will return the
different attribute names in the order they were defined.
def __init__(self, rel, attr): = rel
# We need the dictionary to be ordered
# XXX: may be better to implement an ordered dictionary
self._attributes = {}
self._attrnames = []
for name, value in attr:
tp = parse_type(value)
if tp == 'nominal':
self._attributes[name] = (tp, get_nom_val(value))
elif tp == 'date':
self._attributes[name] = (tp, get_date_format(value)[0])
self._attributes[name] = (tp, None)
def __repr__(self):
msg = ""
msg += "Dataset: %s\n" %
for i in self._attrnames:
msg += "\t%s's type is %s" % (i, self._attributes[i][0])
if self._attributes[i][1]:
msg += ", range is %s" % str(self._attributes[i][1])
msg += '\n'
return msg
def __iter__(self):
return iter(self._attrnames)
def __getitem__(self, key):
return self._attributes[key]
def names(self):
"""Return the list of attribute names."""
return self._attrnames
def types(self):
"""Return the list of attribute types."""
attr_types = [self._attributes[name][0] for name in self._attrnames]
return attr_types
def loadarff(f):
Read an arff file.
The data is returned as a record array, which can be accessed much like
a dictionary of numpy arrays. For example, if one of the attributes is
called 'pressure', then its first 10 data points can be accessed from the
``data`` record array like so: ``data['pressure'][0:10]``
f : file-like or str
File-like object to read from, or filename to open.
data : record array
The data of the arff file, accessible by attribute names.
meta : `MetaData`
Contains information about the arff file such as name and
type of attributes, the relation (name of the dataset), etc...
This is raised if the given file is not ARFF-formatted.
The ARFF file has an attribute which is not supported yet.
This function should be able to read most arff files. Not
implemented functionality include:
* date type attributes
* string type attributes
It can read files with numeric and nominal attributes. It cannot read
files with sparse data ({} in the file). However, this function can
read files with missing data (? in the file), representing the data
points as NaNs.
>>> from import arff
>>> from cStringIO import StringIO
>>> content = \"\"\"
... @relation foo
... @attribute width numeric
... @attribute height numeric
... @attribute color {red,green,blue,yellow,black}
... @data
... 5.0,3.25,blue
... 4.5,3.75,green
... 3.0,4.00,red
... \"\"\"
>>> f = StringIO(content)
>>> data, meta = arff.loadarff(f)
>>> data
array([(5.0, 3.25, 'blue'), (4.5, 3.75, 'green'), (3.0, 4.0, 'red')],
dtype=[('width', '<f8'), ('height', '<f8'), ('color', '|S6')])
>>> meta
Dataset: foo
\twidth's type is numeric
\theight's type is numeric
\tcolor's type is nominal, range is ('red', 'green', 'blue', 'yellow', 'black')
if hasattr(f, 'read'):
ofile = f
ofile = open(f, 'rt')
return _loadarff(ofile)
if ofile is not f: # only close what we opened
def _loadarff(ofile):
# Parse the header file
rel, attr = read_header(ofile)
except ValueError as e:
msg = "Error while parsing header, error was: " + str(e)
raise ParseArffError(msg)
# Check whether we have a string attribute (not supported yet)
hasstr = False
for name, value in attr:
type = parse_type(value)
if type == 'string':
hasstr = True
meta = MetaData(rel, attr)
# XXX The following code is not great
# Build the type descriptor descr and the list of convertors to convert
# each attribute to the suitable type (which should match the one in
# descr).
# This can be used once we want to support integer as integer values and
# not as numeric anymore (using masked arrays ?).
acls2dtype = {'real': float, 'integer': float, 'numeric': float}
acls2conv = {'real': safe_float,
'integer': safe_float,
'numeric': safe_float}
descr = []
convertors = []
if not hasstr:
for name, value in attr:
type = parse_type(value)
if type == 'date':
date_format, datetime_unit = get_date_format(value)
descr.append((name, "datetime64[%s]" % datetime_unit))
convertors.append(partial(safe_date, date_format=date_format,
elif type == 'nominal':
n = maxnomlen(value)
descr.append((name, 'S%d' % n))
pvalue = get_nom_val(value)
convertors.append(partial(safe_nominal, pvalue=pvalue))
descr.append((name, acls2dtype[type]))
#sdescr.append((name, acls2sdtype[type]))
# How to support string efficiently ? Ideally, we should know the max
# size of the string before allocating the numpy array.
raise NotImplementedError("String attributes not supported yet, sorry")
ni = len(convertors)
def generator(row_iter, delim=','):
# TODO: this is where we are spending times (~80%). I think things
# could be made more efficiently:
# - We could for example "compile" the function, because some values
# do not change here.
# - The function to convert a line to dtyped values could also be
# generated on the fly from a string and be executed instead of
# looping.
# - The regex are overkill: for comments, checking that a line starts
# by % should be enough and faster, and for empty lines, same thing
# --> this does not seem to change anything.
# 'compiling' the range since it does not change
# Note, I have already tried zipping the converters and
# row elements and got slightly worse performance.
elems = list(range(ni))
for raw in row_iter:
# We do not abstract skipping comments and empty lines for
# performance reasons.
if r_comment.match(raw) or r_empty.match(raw):
row = raw.split(delim)
yield tuple([convertors[i](row[i]) for i in elems])
a = generator(ofile)
# No error should happen here: it is a bug otherwise
data = np.fromiter(a, descr)
return data, meta
# Misc
def basic_stats(data):
nbfac = data.size * 1. / (data.size - 1)
return np.nanmin(data), np.nanmax(data), np.mean(data), np.std(data) * nbfac
def print_attribute(name, tp, data):
type = tp[0]
if type == 'numeric' or type == 'real' or type == 'integer':
min, max, mean, std = basic_stats(data)
print("%s,%s,%f,%f,%f,%f" % (name, type, min, max, mean, std))
msg = name + ",{"
for i in range(len(tp[1])-1):
msg += tp[1][i] + ","
msg += tp[1][-1]
msg += "}"
def test_weka(filename):
data, meta = loadarff(filename)
for i in meta:
print_attribute(i, meta[i], data[i])
# make sure nose does not find this as a test
test_weka.__test__ = False
if __name__ == '__main__':
import sys
filename = sys.argv[1]