Bert's blog

Input and output with strings, part 2: input with Python

A short while ago, I wrote a first post in a series about string input and output in various (programming) languages, and I started with the easy part: string output. Since then, I realised that providing a full reference for string input for all languages covered in this first post (Python, C(++) and Fortran) would constitute too much material for a single post. Hence why this second post in the series will be devoted to string input for a single language: Python.

As hinted at before, this is a complicated topic: strings can contain many things in various formats (all the ones covered in the post about string output), and it is impossible to read anything other than strings (e.g. integers or floating point values) from a string without having at least some knowledge about this formatting. Some Python functions are capable of recognising whether a specific part of a string contains a valid number, but these functions will only do this when instructed to do so, and even then, they will still not be able to use this number as an integer/floating point value if you don’t tell them which kind of number you want (or they will use a default that might surprise you).

That being said, Python is Python, so there are a lot of easy hacks that allow you to many powerful things with strings without having to go through the trouble of actually parsing the string, which is the correct and most powerful way to read strings.

Some easy solutions

So let’s focus on some of these easy solutions before embarking upon the difficult journey to understand string parsing. Suppose for example that you want to read a text file that contains a number of data values, arranged in columns. In principle, reading a text file requires you to read the lines from the file as strings, and then reading the values in the columns from those strings. However, since this kind of file input and output is very common, Python modules like numpy actually offer convenient functions that do this for you: numpy.loadtxt reads data from a well-behaved text file (i.e. each row has exactly the same number of columns and all values in a column have the same type and can be parsed as values of that type), while numpy.genfromtxt offers a more flexible (and slower) version that can also handle missing or weirdly formatted data.

Assuming that all values in the file are numbers, reading a text file is as easy as

import numpy as np
data = np.loadtxt("filename.txt")

This will automatically use a 32-bit floating point data type for the values. If your values have another type (e.g. int), you can simply specify this as an additional argument:

data = np.loadtxt("filename.txt", dtype = int)

This function will also automatically ignore lines that start with a # (or any other character passed on to the additional argument comments). If you want to manually skip some rows at the start of the file, or only read specific columns, this is also possible.

In practice, these basic forms of numpy.loadtxt have proven sufficient for about 90% of the cases where I read in data from text files. However, sometimes text files contain data that cannot easily be read in as the same type, e.g. some columns contain floating point values and others (long) integers. Since floating point values cannot be parsed as integers, you can sometimes get around this by simply reading all values as floating point values and then later converting the integer elements to integers from floats. If the integers are really large, this can however lead to round off error which would not occur if you would simply read the integers as integers.

In order to deal with these cases, it is possible to specify a custom numpy data type:

data = np.loadtxt("filename.txt",
                  dtype = {
                    "names": ("longInt", "longFloat"),
                    "formats": ("i8", "f8"),
                  },
)

This will read each line in the file as a tuple with the corresponding listed data types. Individual columns can be easily accessed through the additional names provided in the data type:

print(data["longInt"]) # prints all long integer elements

This even works when one or more columns contain string values:

data = np.loadtxt("filename.txt",
                  dtype = {
                    "names": ("label", "longInt", "longFloat"),
                    "formats": ("s100", "i8", "f8"),
                  },
)

Note that the s100 in this case contains the maximum size (in characters) of a string that can be parsed. This number should be chosen carefully.

What if the data you want to read is not contained in a text file, but already part of a string? Easy: there is a numpy.fromstring method that works in a very similar way to numpy.loadtxt. If the string has a predictable layout (e.g. a fixed number of columns with predictable data types in each column), then this provides a very flexible way of reading it.

What if the string does not really have a column layout? Or you simply want to read a single number from it? If the string has a predictable layout, then a very easy (and dirty) hack is the following:

text = "I want to read the number 42 from this string"
number = int(text.split()[6])

You simply split the string into individual words using split(), and then use int() to convert the 7th (6th when you start counting from 0 like Python) element into an integer. This will work no matter what the value of the number, but will only work as long as the number is always the 7th word in the sentence, and is well-behaved. This has been my go to solution for string parsing in more cases than I would ever want to admit.

The hardcore solution

The solutions above work in almost all cases that I care about, but that is mostly because I usually need to parse strings that I wrote myself, i.e. I can guarantee that the input is well-behaved and has a predictable format. These solutions will not work if

  • the input has no predictable format: e.g. you know that the string contains a single number, but you cannot predict in what position this number will be
  • the number itself is contaminated: e.g. you are trying to read a value that has a unit attached to it (literally; the unit is not separated from the value with a white space), or you are parsing a label that encodes some values in a single word: n10e7s42 (scientists tend to construct file names in this way).

When this is the case, there is really only one solution to still retrieve the values from the string, and that is by using regular expressions. In essence, a regular expression is a combination of characters of various types that tell the parser that reads the string how to recognise a certain part of the string. An unsigned integer for example will always consist of a combination of digits (0123456789) and optional exponential signs (eE), while a signed integer can also contain a sign (+-). Whenever the parser encounters a group of characters that consists solely of these characters, it could decide that these characters make up an integer, and parse that integer, no matter where it is in the string.

While in theory this sounds fairly easy, I have noticed that it usually is not. One of the reasons for this is that there are many regular expression parsers, and that all of them tend to have a slightly different syntax. Furthermore, regular expressions are very basic constructs. This gives them maximum flexibility (which is good), but also means they can be very pedantic. If you want to parse numbers consisting of multiple digits for example, you need to explicitly instruct the parser to do this. This is very useful if you want to parse two different numbers that are not separated by a white space, but very annoying in the more obvious case where you just want it to parse the whole number.

Python has a very powerful regular expression module called re, that is devoted to regular expression parsing. It contains some basic functions that take a string and a regular expression as input and simply output the matching parts of the string, but also has functionality to compile regular expressions that are used multiple times. In this case, a regular expression object is created and optimised and then applied to many strings using the same regular expression.

Let’s start with a simple example:

import re
text = "this string contains the number 43 somewhere in its body"
print(re.findall("[0-9]", text))

This will print the following:

['4', '3']

This is how it works: re.findall will parse the string provided as second argument, and match every occurrence of the regular expression [0-9], i.e. all characters in the range 0-9 (all digits). It will then return a list of all matches as individual strings. This is probably not entirely what you want in this case, since you probably do not want to match individual digits of the number. To match all digits as a single number, we need to change the regular expression to [0-9]+, where the + indicates repeat until the next character does no longer match.

This will work for any integer number containing only digits, and will also work if the string contains multiple numbers (since findall matches all occurrences of the regular expression). It will not work if the integer is written in exponential notation, e.g. 2e7. To match this, we need to add the exponent part: [0-9]+e[0-9]+. However, this will no longer match numbers that do not have an exponent… Note that we also cannot add the e to the square brackets, as this would literally match every character e in the string:

> re.findall("[0-9e]+", text)
['e', 'e', '43', 'e', 'e', 'e']

To allow for both integers with and without an exponent, we need to make the exponent bit optional. To this end, we first need to create a group: [0-9]+(?:e[0-9]+). The group is created by using the round brackets, the additional ?: is required to turn this into a non-capturing group, because re by default only displays the captured part of a match if such a part exists (I will get back to this in a minute). The entire group is made optional by appending a ? to it: [0-9]+(?:e[0-9]+)?. To also allow for capital E exponents, we can change this to [0-9]+(?:[eE][0-9]+)?.

This expression will match all unsigned integers. We can add an optional sign very easily: [-+]?[0-9]+(?:[eE][0-9]+)?. This will now match all integers. For floating point values, we need to go a step further, and add an optional decimal point, followed by additional (optional) digits: [-+]?[0-9]+\.?[0-9]*(?:[eE][-+]?[0-9]+)?. The \ before the . is required to escape this character, since a . is already used by re to denote a regular expression that matches any character except a white space. Note that we also added an optional sign to the exponent in this case. We also introduced the * that, unlike +, also matches 0 repetitions of the preceding expression. This expression now matches almost all possible numbers. However, it will fail in some cases, e.g. when you omit the leading 0 in a purely decimal number like 0.05: .05. This expression will be matched (since the [0-9]+ bit matches the 05) but it will omit the decimal point itself from the match. And this of course leads to a completely different number…

To solve this, we could replace the + for the leading digits with a *. However, this will go horribly wrong, since then the regular expression could match literally anything (since it matches a character containing no digits at all). To have a useful regular expression, we really need to keep at least one +. One possibility is to put the + after the decimal digits, since then the same problem we had above will simply omit the decimal point from expressions like 5., which parses as the correct number. This will still fail to parse numbers like 2.e5, as it will split them in two parts. More correct would be to add an additional \.? after the decimal digits. This works, but will now also match wrongly formatted numbers like 2.4.e5, which might not be desired.

Another option is to allow for two possible floating point signatures, using the regular expression or, |: [-+]?(?:(?:[0-9]*\.?[0-9]+)|(?:[0-9]+\.?[0-9]*))(?:[eE][-+]?[0-9]+)?. Unfortunately, this expression will still split values like 2.e5 in half… The reason for that is that | first tries to match the first expression, and only parses the second if that fails. Since the first expression provides a correct match for 2., it will not even try to match the second expression, even though that expression provides a more complete match. The solution is to swap the order of the two expressions:

> text = "5 5. .4 1e5 1.435e-9 .2E3 2.e+5"
> re.findall(
    "[-+]?(?:(?:[0-9]+\.?[0-9]*)|(?:[0-9]*\.?[0-9]+))(?:[eE][-+]?[0-9]+)?",
    text)
['5', '5.', '.4', '1e5', '1.435e-9', '.2E3', '2.e+5']

As you can see, this now matches pretty much anything that we would like. Unfortunately, the regular expression has become quite large by now…

If all of this was not complicated enough, there are an extensive number of additional options to add more constraints to regular expressions, as well as aliases for common groups of characters, like \d for [0-9]. A full list can be found in the documentation for re.

As mentioned before, re can be instructed to only capture specific parts of the matching string, e.g. the contents of groups within the regular expression. If we were to omit the ?: in our above expression for example, then the expression would still match all the numbers, but its output would be more interesting:

> re.findall(
    "[-+]?(([0-9]+\.?[0-9]*)|([0-9]*\.?[0-9]+))([eE][-+]?[0-9]+)?",
    text)
[('5', '5', '', ''),
 ('5.', '5.', '', ''),
 ('.4', '', '.4', ''),
 ('1', '1', '', 'e5'),
 ('1.435', '1.435', '', 'e-9'),
 ('.2', '', '.2', 'E3'),
 ('2.', '2.', '', 'e+5')]

In this case, for each of the 7 matches, the contents of the individual groups that match is displayed. Great for debugging, but not very useful, unless you are interested in matching specific parts of the floating point number, of course. In that case, it could also be useful to provide name labels for specific parts of the match. This can be achieved by naming groups: ?P<name>:

> re.findall(
    "[-+]?(?P<base>(?:[0-9]+\.?[0-9]*)|(?:[0-9]*\.?[0-9]+))"
    "(?P<exponent>[eE][-+]?[0-9]+)?",
    text)
[('5', ''), ('5.', ''), ('.4', ''), ('1', 'e5'), ('1.435', 'e-9'),
 ('.2', 'E3'), ('2.', 'e+5')]

re.findall will ignore the name labels, but other re functions do not. Most other functions do not return a string (or list of strings), but instead return a Match object (or list of these objects). This object contains functionality to retrieve the original string and regular expression that were used to create it, and also offers lots of options to retrieve the different groups contained in the match, e.g. the labels created above. For the example above, we could for example use

> for match in re.finditer(
>    "[-+]?(?P<base>(?:[0-9]+\.?[0-9]*)|(?:[0-9]*\.?[0-9]+))"
>    "(?P<exponent>[eE][-+]?[0-9]+)?", text):
>   print(match["base"], match["exponent"])
5 None
5. None
.4 None
1 e5
1.435 e-9
.2 E3
2. e+5

As already mentioned before, regular expressions can be compiled into a regular expression object, so that you can reuse them more efficiently. An example:

import re
floatregexp = re.compile(
  "[-+]?(?:(?:[0-9]+\.?[0-9]*)|(?:[0-9]*\.?[0-9]+))(?:[eE][-+]?[0-9]+)?"
)
text = "5 5. .4 1e5 1.435e-9 .2E3 2.e+5"
floatregexp.findall(text)

The compiled regular expression objects have the same functions as the re module itself, but always use the same regular expression.

Note that apart from searching specific expressions in a string, you can also use regular expressions to check if a string (as a whole) matches a regular expression, using re.fullmatch. This will only produce a match if the entire string matches the expression.

The regular expressions above allow you to very flexibly locate and retrieve numbers (and other parts) of a string. However, they still only return strings themselves, i.e. you will still need to convert the resulting strings into the actual data types you want (using int() and float() for example). This is usually easy enough, but can get quite cumbersome (and slow) if you are reading a large number of values e.g. from a file. That is why numpy also provides an alternative for loadtxt that use regular expressions: numpy.fromregex. I have never used this function myself (I only discovered it while doing my research for this post), but here is an example. Suppose you have a text file, test.txt, with the following contents:

This line contains the number 5.
And a 672 is hidden somewhere in this line.

Then you can retrieve the numbers in an array as follows:

> import numpy as np
> test = np.fromregex("test.txt", "[0-9]+", dtype = [("number", "u4")])
> print(test)
[(  5,) (672,)]

Note that the dtype argument is required in this case and needs to be a custom data type like the one we already encountered before.


Professional astronomer.