Quantifying the amount of change in a git diff?

I use git for a slightly unusual purpose–it stores my text as I write fiction. (I know, I know…geeky.)

I am trying to keep track of productivity, and want to measure the degree of difference between subsequent commits. The writer’s proxy for “work” is “words written”, at least during the creation stage. I can’t use straight word count as it ignores editing and compression, both vital parts of writing. I think I want to track:

 (words added)+(words removed)

which will double-count (words changed), but I’m okay with that.

It’d be great to type some magic incantation and have git report this distance metric for any two revisions. However, git diffs are patches, which show entire lines even if you’ve only twiddled one character on the line; I don’t want that, especially since my ‘lines’ are paragraphs. Ideally I’d even be able to specify what I mean by “word” (though \W+ would probably be acceptable).

Is there a flag to git-diff to give diffs on a word-by-word basis? Alternately, is there a solution using standard command-line tools to compute the metric above?

8 Solutions collect form web for “Quantifying the amount of change in a git diff?”

wdiff does word-by-word comparison. Git can be configured to use an external program to do the diffing. Based on those two facts and this blog post, the following should do roughly what you want.

Create a script to ignore most of the unnecessary arguments that git-diff provides and pass them to wdiff. Save the following as ~/wdiff.py or something similar and make it executable.

#!/usr/bin/python

import sys
import os

os.system('wdiff -s3 "%s" "%s"' % (sys.argv[2], sys.argv[5]))

Tell git to use it.

git config --global diff.external ~/wdiff.py
git diff filename

git diff –word-diff works in the latest stable version of git (at git-scm.com)

There are a few options that let you decide what format you want it in, the default is quite readable but you might want –word-diff=porcelain if you’re feeding the output into a script.

I figured out a way to get concrete numbers by building on top of the other answers here. The result is an approximation, but it should be close enough to serve as a useful indicator of the amount characters that were added or removed. Here’s an example with my current branch compared to origin/master:

$ git diff --word-diff=porcelain origin/master | grep -e '^+[^+]' | wc -m
38741
$ git diff --word-diff=porcelain origin/master | grep -e '^-[^-]' | wc -m
46664

The difference between the removed characters (46664) and the added characters (38741) shows that my current branch has removed approximately 7923 characters. Those individual added/removed counts are inflated due to the diff’s +/- and indentation characters, however, the difference should cancel out a significant portion of that inflation in most cases.

Building on James’ and cornmacrelf’s input, I’ve added arithmetic expansion,
and came up with a few reusable alias commands for counting words in a git diff:

alias gitwa='git diff --word-diff=porcelain origin/master | grep -e "^+[^+]" | wc -w | xargs'
alias gitwd='git diff --word-diff=porcelain origin/master | grep -e "^-[^-]" | wc -w | xargs'
alias gitw='echo $(($(gitwa) - $(gitwd)))'

Output from gitwa and gitwd is trimmed using xargs trick.

Git has had (for a long time) a --color-words option for git diff. This doesn’t get you your counting, but it does let you see the diffs.

scompt.com’s suggestion of wdiff is also good; it’s pretty easy to shove in a different differ (see git-difftool). From there you just have to go from the output wdiff can give to the result you really want.

There’s one more exciting thing to share, though, from git’s what’s cooking:

* tr/word-diff (2010-04-14) 1 commit
  (merged to 'next' on 2010-05-04 at d191b25)
 + diff: add --word-diff option that generalizes --color-words

Here’s the commit introducing word-diff. Presumably it will make its way from next into master before long, and then git will be able to do this all internally – either producing its own word diff format or something similar to wdiff. If you’re daring, you could build git from next, or just merge that one commit into your local master to build.

Thanks to Jakub’s comment: you can further customize word diffs if necessary by providing a word regex (config parameter diff.*.wordRegex), documented in gitattributes.

I liked Stoutie’s answer and wanted to make it a bit more configurable to answer some word count questions I had. Ended up with the following solution that works in ZSH and should work in Bash. Each function takes any revision or revision difference, with a default of comparing the current state of the world with origin/master:


# Calculate writing word diff between revisions. Cribbed / modified from:
# https://stackoverflow.com/questions/2874318/quantifying-the-amount-of-change-in-a-git-diff
function git_words_added {
  revision=${1:-origin/master}

  git diff --word-diff=porcelain $revision | \
    grep -e "^+[^+]" | \
    wc -w | \
    xargs
}

function git_words_removed {
  revision=${1:-origin/master}

  git diff --word-diff=porcelain $revision | \
    grep -e "^-[^-]" | \
    wc -w | \
    xargs
}

function git_words_diff {
  revision=${1:-origin/master}

  echo $(($(git_words_added $1) - $(git_words_removed $1)))
}

Then you can use it like so:


$ git_words_added
# => how many words were added since origin/master

$ git_words_removed
# => how many words were removed since origin/master

$ git_words_diff
# => difference of adds and removes since origin/master (net words)

$ git_words_diff HEAD
# => net words since you last committed

$ git_words_diff master@{yesterday}
# => net words written today!

$ git_words_diff HEAD^..HEAD
# => net words in the last commit

$ git_words_diff ABC123..DEF456
# => net words between two arbitrary commits

Hope this helps someone!

Since Git 1.6.3 there is also git difftool, which can be configured to run nearly any external diff tool. This is a lot easier than some of the solutions which require creating scripts etc. If you like the output of wdiff -s you can configure something like:

git config --global difftool.wdiffs.cmd 'wdiff -s "$LOCAL" "$REMOTE"'
git config --global alias.wdiffs 'difftool -t wdiffs'

Now you can just run git difftool -t wdiffs or its alias git wdiffs.

If you prefer to get statistics for all modified files together, instead do something like:

git config --global difftool.wdiffs.cmd 'diff -pdrU3 "$LOCAL" "$REMOTE" | wdiff -sd'
git config --global alias.wdiffs 'difftool -d -t wdiffs'

This takes the output of a typical unified diff and pipes it into wdiff with its -d option set to just interpret the input. In contrast, the extra -d argument to difftool in the alias tells git to copy all modified files to a temporary directory before doing the diff.

The above answers fail for some use cases where you need to exclude moved text (e.g., if I move a function in code or paragraph in latex further down the document, I don’t want to count all of those as changes!)

For that, you can also calculate the number of duplicate lines, and exclude those from your query if there are too many duplicates.

For example, building on the other answers, I can do:

git diff $sha~1..$sha|grep -e"^+[^+]" -e"^-[^-]"|sed -e's/.//'|sort|uniq -d|wc -w|xargs

calculates the number of duplicate words in the diff, where sha is your commit.

You can do this for all the commits within the last day (since 6 am) by:

for sha in $(git rev-list --since="6am" master | sed -e '$ d'); do
     echo $(git diff --word-diff=porcelain $sha~1..$sha|grep -e"^+[^+]"|wc -w|xargs),\
     $(git diff --word-diff=porcelain $sha~1..$sha|grep -e"^-[^-]"|wc -w|xargs),\
     $(git diff $sha~1..$sha|grep -e"^+[^+]" -e"^-[^-]"|sed -e's/.//'|sort|uniq -d|wc -w|xargs)
done

Prints: added, deleted, duplicates

(I take the line diff for duplicates, as it excludes the times where git diff tries to be too clever, and assumes you have actually just changed text rather than moved it. It also discounts instances where a single word is counted as a duplicate.)

Or, if you want to be sophisticated about it, you can exclude commits entirely if there is more than 80% duplication, and sum up the rest:

total=0
for sha in $(git rev-list --since="6am" master | sed -e '$ d'); do
    added=$(git diff --word-diff=porcelain $sha~1..$sha|grep -e"^+[^+]"|wc -w|xargs)
    deleted=$(git diff --word-diff=porcelain $sha~1..$sha|grep -e"^-[^-]"|wc -w|xargs)
    duplicated=$(git diff $sha~1..$sha|grep -e"^+[^+]" -e"^-[^-]"|sed -e's/.//'|sort|uniq -d|wc -w|xargs)
    if [ "$added" -eq "0" ]; then
        changed=$deleted
        total=$((total+deleted))
        echo "added:" $added, "deleted:" $deleted, "duplicated:"\
             $duplicated, "changed:" $changed
    elif [ "$(echo "$duplicated/$added > 0.8" | bc -l)" -eq "1" ]; then
        echo "added:" $added, "deleted:" $deleted, "duplicated:"\
             $duplicated, "changes counted:" 0
    else
        changed=$((added+deleted))
        total=$((total+changed))
        echo "added:" $added, "deleted:" $deleted, "duplicated:"\
             $duplicated, "changes counted:" $changed
    fi
done
echo "Total changed:" $total

I have this script to do it here: https://github.com/MilesCranmer/git-stats.

This prints out:

➜  bifrost_paper git:(master) ✗ count_changed_words "6am" 

added: 38, deleted: 76, duplicated: 3, changes counted: 114
added: 14, deleted: 19, duplicated: 0, changes counted: 33
added: 1113, deleted: 1112, duplicated: 1106, changes counted: 0
added: 1265, deleted: 1275, duplicated: 1225, changes counted: 0
added: 4207, deleted: 4208, duplicated: 4391, changes counted: 0
Total changed: 147

The commits where I am just moving around things are obvious, so I don’t count those changes. It counts up everything else and tells me the total number of changed words.

Git Baby is a git and github fan, let's start git clone.