7 Command-Line Tools for Data Science

Blog article by Jeroen Janssens
Sep 19, 2013 • 20 min read

Data science is OSEMN (pronounced as awesome). That is, it involves Obtaining, Scrubbing, Exploring, Modelling, and iNterpreting data. As a data scientist, I spend quite a bit of time on the command-line, especially when there’s data to be obtained, scrubbed, or explored. And I’m not alone in this. Recently, Greg Reda discussed how the classics (e.g., head, cut, grep, sed, and awk) can be used for data science. Prior to that, Seth Brown discussed how to perform basic exploratory data analysis in Unix.

I would like to continue this discussion by sharing seven command-line tools that I have found useful in my day-to-day work. The tools are: jq, json2csv, csvkit, scrape, xml2json, sample, and Rio. (The home-made tools scrape, sample, and Rio can be found in this repository.) Any suggestions, questions, comments, and even pull requests are more than welcome. (Tools suggested by others can be found towards the bottom of the post.) OSEMN, let’s get started with our first tool: jq.

1. jq - sed for JSON

JSON is becoming an increasingly common data format, especially as APIs are appearing everywhere. I remember cooking up the ugliest grep and sed incantations in order to process JSON. Thanks to jq, those days are now in the past.

Imagine we’re interested in the candidate totals of the 2008 presidential election. It so happens that the New York Times has a Campaign Finance API. (You can get your own API keys if you want to access any of their APIs.) Let’s get some JSON using curl:

$ curl -s 'http://api.nytimes.com/svc/elections/us/v3/finances/2008/president/totals.json?api-key=super-secret' > nyt.json

where -s puts curl in silent mode. In its simplest form, i.e., jq '.', the tool transforms the incomprehensible API response we got:

$ cat nyt.json
{"status":"OK","base_uri":"http://api.nytimes.com/svc/elections/us/v3/finances/2008/","cycle":2008,"copyright":"Copyright (c) 2013 The New York Times Company. All Rights Reserved.","results":[{"candidate_name":"Obama, Barack","name":"Barack Obama","party":"D", ...

into nicely indented output:

$ < nyt.json jq '.' | head
{
  "results": [
    {
      "candidate_id": "P80003338",
      "date_coverage_from": "2007-01-01",
      "date_coverage_to": "2008-11-24",
      "candidate_name": "Obama, Barack",
      "name": "Barack Obama",
      "party": "D",

Note that the output isn’t necessarily in the same order as the input. Besides pretty printing, jq can also select, filter, and format JSON data, as illustrated by the following command, which returns the name, cash, and party of each candidate that had at least $1,000,000 in cash:

$ < nyt.json jq -c '.results[] | {name, party, cash: .cash_on_hand} | select(.cash | tonumber > 1000000)' 
{"cash":"29911984.0","party":"D","name":"Barack Obama"}
{"cash":"32812513.75","party":"R","name":"John McCain"}
{"cash":"4428347.5","party":"D","name":"John Edwards"}

Please refer to the jq manual to read about the many other things it can do, but don’t expect it to solve all your data munging problems. Remember, the Unix philosophy favours small programs that do one thing and do it well. And jq’s functionality is more than sufficient I would say! Now that we have the data we need, it’s time to move on to our second tool: json2csv.

2. json2csv - convert JSON to CSV

While JSON is a great format for interchanging data, it’s rather unsuitable for most command-line tools. Not to worry, we can easily convert JSON into CSV using json2csv. Assuming that we stored the data from the last step in million.json, simply invoking json2csv will convert it to some nicely comma-separated values:

$ < million.json json2csv -k name,party,cash | head -n 3
Barack Obama,D,29911984.0
John McCain,R,32812513.75
John Edwards,D,4428347.5

Having the data in CSV format allows us to use the classic tools such as cut -d, and awk -F,. Others like grep and sed don’t really have a notion of fields. Since CSV is the king of tabular file formats, according to the authors of csvkit, they created, well, csvkit.

3. csvkit - suite of utilities for working with CSV

Rather than being one tool, csvkit is a collection of tools that operate on CSV data. Most of these tools expect the CSV data to have a header, so let’s add one. (Since the publication of this post, json2csv has been updated to print the header with the -p option.)

$ echo name,party,cash | cat - million.csv > million-header.csv

We can, for example, sort the candidates by cash with csvsort and display the data using csvlook:

$ < million-header.csv csvsort -rc cash | csvlook
|---------------+-------+--------------|
|  name         | party | cash         |
|---------------+-------+--------------|
|  John McCain  | R     | 32812513.75  |
|  Barack Obama | D     | 29911984.0   |
|  John Edwards | D     | 4428347.5    |
|---------------+-------+--------------|

Looks like the MySQL console doesn’t it? Speaking of databases, you can insert the CSV data into an sqlite database as follows (many other databases are supported as well):

$ csvsql --db sqlite:///myfirst.db --insert million-header.csv
$ sqlite3 myfirst.db
sqlite> .schema million-header

CREATE TABLE "million-header" (
    name VARCHAR(12) NOT NULL, 
    party VARCHAR(1) NOT NULL, 
    cash FLOAT NOT NULL
);

In this case, the database columns have the correct data types because the type is inferred from the CSV data. Other tools within csvkit that might be of interest are: in2csv, csvgrep, and csvjoin. And with csvjson, the data can even be converted back to JSON. All in all, csvkit is worth checking out.

4. scrape - HTML extraction using XPath or CSS selectors

JSON APIs sure are nice, but they aren’t the only source of data; a lot of it is unfortunately still embedded in HTML. scrape is a python script I put together that employs the lxml and cssselect packages to select certain HTML elements by means of an XPath query or CSS selector. Let’s extract the table from this Wikipedia article that lists the border and area ratio of each country.

$ curl -s 'http://en.wikipedia.org/wiki/List_of_countries_and_territories_by_border/area_ratio' |
> scrape -b -e 'table.wikitable > tr:not(:first-child)' | head
<!DOCTYPE html>
<html>
<body>
<tr>
<td>1</td>
<td>Vatican City</td>
<td>3.2</td>
<td>0.44</td>
<td>7.2727273</td>
</tr>

The -b argument lets scrape enclose the output with <html> and <body> tags, which is sometimes required by xml2json to convert correctly the HTML to JSON.

5. xml2json - convert XML to JSON

As its name implies, xml2json takes XML (and HTML) as input and returns JSON as output. Therefore, xml2json is a great liaison between scrape and jq.

$ curl -s 'http://en.wikipedia.org/wiki/List_of_countries_and_territories_by_border/area_ratio' |
> scrape -be 'table.wikitable > tr:not(:first-child)' |
> xml2json |
> jq -c '.html.body.tr[] | {country: .td[1][], border: .td[2][], surface: .td[3][], ratio: .td[4][]}' | head
{"ratio":"7.2727273","surface":"0.44","border":"3.2","country":"Vatican City"}
{"ratio":"2.2000000","surface":"2","border":"4.4","country":"Monaco"}
{"ratio":"0.6393443","surface":"61","border":"39","country":"San Marino"}
{"ratio":"0.4750000","surface":"160","border":"76","country":"Liechtenstein"}
{"ratio":"0.3000000","surface":"34","border":"10.2","country":"Sint Maarten (Netherlands)"}
{"ratio":"0.2570513","surface":"468","border":"120.3","country":"Andorra"}
{"ratio":"0.2000000","surface":"6","border":"1.2","country":"Gibraltar (United Kingdom)"}
{"ratio":"0.1888889","surface":"54","border":"10.2","country":"Saint Martin (France)"}
{"ratio":"0.1388244","surface":"2586","border":"359","country":"Luxembourg"}
{"ratio":"0.0749196","surface":"6220","border":"466","country":"Palestinian territories"}

Of course this JSON data could then be piped into json2csv and so forth.

6. sample - when you’re in debug mode

The second tool I made is sample. (It’s based on two scripts in bitly’s data_hacks, which contains some other tools worth checking out.) When you’re in the process of formulating your data pipeline and you have a lot of data, then debugging your pipeline can be cumbersome. In that case, sample might be useful. The tool serves three purposes (which isn’t very Unix-minded, but since it’s mostly useful when you’re in debug mode, that’s not such a big deal).

The first purpose of sample is to get a subset of the data by outputting only a certain percentage of the input on a line-by-line basis. The second purpose is to add some delay to the output. This comes in handy when the input is a constant stream (e.g., the Twitter firehose), and the data comes in too fast to see what’s going on. The third purpose is to run only for a certain time. The following invocation illustrates all three purposes.

$ seq 10000 | sample -r 20% -d 1000 -s 5 | jq '{number: .}'

This way, every input line has a 20% chance of being forwarded to jq. Moreover, there is a 1000 millisecond delay between each line and after five seconds sample will stop entirely. Please note that each argument is optional. In order to prevent unnecessary computation, try to put sample as early as possible in your pipeline (the same argument holds for head and tail). Once you’re done debugging you can simply take it out of the pipeline.

7. Rio - making R part of the pipeline

This post wouldn’t be complete without some R. It’s not straightforward to make R/Rscript part of the pipeline since they don’t work with stdin and stdout out of the box. Therefore, as a proof of concept, I put together a bash script called Rio.

Rio works as follows. First, the CSV provided to stdin is redirected to a temporary file and lets R read that into a data frame df. Second, the specified commands in the -e option are executed. Third, the output of the last command is redirected to stdout. Allow me to demonstrate three one-liners that use the Iris dataset (don’t mind the url).

Display the five-number-summary of each field.

$ curl -s 'https://raw.github.com/pydata/pandas/master/pandas/tests/data/iris.csv' > iris.csv
$ < iris.csv Rio -e 'summary(df)'
  SepalLength      SepalWidth     PetalLength      PetalWidth   
 Min.   :4.300   Min.   :2.000   Min.   :1.000   Min.   :0.100  
 1st Qu.:5.100   1st Qu.:2.800   1st Qu.:1.600   1st Qu.:0.300  
 Median :5.800   Median :3.000   Median :4.350   Median :1.300  
 Mean   :5.843   Mean   :3.054   Mean   :3.759   Mean   :1.199  
 3rd Qu.:6.400   3rd Qu.:3.300   3rd Qu.:5.100   3rd Qu.:1.800  
 Max.   :7.900   Max.   :4.400   Max.   :6.900   Max.   :2.500  
     Name          
 Length:150        
 Class :character  
 Mode  :character

If you specify the -s option, the sqldf package will be imported. In case the output is a data frame, CSV will be written to stdout. This enables you to further process that data using other tools.

$ < iris.csv Rio -se 'sqldf("select * from df where df.SepalLength > 7.5")' | csvlook
|--------------+------------+-------------+------------+-----------------|
|  SepalLength | SepalWidth | PetalLength | PetalWidth | Name            |
|--------------+------------+-------------+------------+-----------------|
|  7.6         | 3          | 6.6         | 2.1        | Iris-virginica  |
|  7.7         | 3.8        | 6.7         | 2.2        | Iris-virginica  |
|  7.7         | 2.6        | 6.9         | 2.3        | Iris-virginica  |
|  7.7         | 2.8        | 6.7         | 2          | Iris-virginica  |
|  7.9         | 3.8        | 6.4         | 2          | Iris-virginica  |
|  7.7         | 3          | 6.1         | 2.3        | Iris-virginica  |
|--------------+------------+-------------+------------+-----------------|

If you specify the -g option, ggplot2 gets imported and a ggplot object called g with df as the data is initialised. If the final output is a ggplot object, a PNG will be written to stdout.

$ < iris.csv Rio -ge 'g + geom_point(aes(x = SepalLength, y = SepalWidth, colour = Name))' > iris.png

I made this tool so that I could take advantage of the power of R on the command-line. Of course it has its limits, but at least there’s no need to learn gnuplot any more.

Command-line tools suggested by others

Below is an uncurated list of tools and repositories that others have suggested via Twitter or Hacker News (last updated on 23-09-2013 07:15 EST). Thanks everybody.

Conclusion

I have shown you seven command-line tools that I use in my daily work as a data scientist. While each tool is useful in its own way, I often find myself combining them with, or just resorting to, the classics such as grep, sed, and awk. Combining such small tools into a larger pipeline is what makes them really powerful.

I’m curious to hear what you think about this list and what command-line tools you like to use. Also, if you’ve made any tools yourself, you’re more than welcome to add them to this data science toolbox.

Don’t worry if you don’t regard yourself as a toolmaker. The next time you’re cooking up that exotic pipeline, consider to put it in a file, add a shebang, parametrise it with some $1s and $2s, and chmod +x it. That’s all there is to it. Who knows, you might even become interested in applying the Unix philosophy.

While the power of the command-line should not be underestimated when it comes to Obtaining, Scrubbing, and Exploring data, it can only get you so far. When you’re ready to do some more serious Exploring, Modelling, and iNterpretation of your data, you’re probably better off continuing your work in a statistical computing environment, such as R or Jupyter Notebook+pandas.

— Jeroen

Would you like to receive an email whenever I have a new blog post, organize an event, or have an important announcement to make? Sign up to my newsletter:

&copy 2013–2024 Jeroen Janssens