Future-proof your skills with Linux, Python, vim & git as I share with you the most timeless and love-worthy tools in tech through my two great projects that work great together.

The Python Pandas .iloc Interface Explained For The Befuddled

I'm sharing my knowledge of how to use Pandas, a data manipulation tool, to select columns from a DataFrame by using its name or the .loc or .iloc interfaces. I'll explain how to use slices to specify row and column ranges, how to use a colon to select all rows, and why the Pandas API has chosen to drop the use of square-brackets for slices.

Master the .iloc Interface of Pandas for Data Manipulation

By Michael Levin

Friday, December 16, 2022

category: pandas

If your data is in rows and columns, chances are, Pandas is your answer. You don’t even need to use a database. Rather, Pandas is based on loading data into memory, doing manipulations as you would in Excel or SQL, and then sending it along. Loading and saving .csv files is very common.

Selecting Column From Pandas DataFrame By a Numeric Index

If you want to grab a column out of data, you just name the column directly on your DataFrame object. So with the dataframe created like this:

cols = ["col1", "col2", "col3"]
lol = [("foo", "bar", "baz"), ("fee", "fi", "fo"), ("bippity", "boppity", "boo")]
df = pd.DataFrame(lol, columns=cols)

So your Pandas DataFrame (df) looks like this:

      col1     col2 col3
0     foo,      bar  baz
1      fee       fi   fo
2  bippity  boppity  boo

You can easily ask for the first column of data by simply using its name (a.k.a. column-label) like this and it is by far the most common use-case.


…and it would return column 1 as a Pandas Series:

0        foo
1        fee
2    bippity
Name: col1, dtype: object

Sometimes you want your column-selection to return a Pandas DataFrame instead of a series. For that, you simply use square-brackets around your requested column, implying you might ask for more columns, and thus a DataFrame is more appropriate to return than a Series:



0     foo,
1      fee
2  bippity

…which is a DataFrame instead of a series. Get it? Ask for the column without the request being in a list and you get a Pandas Series. Ask for a column (or columns) inside a list, and you get a Pandas DataFrame back. The optional use of square-brackets in Pandas is very important.

The movement towards .loc and .iloc

There’s been a movement in Pandas to go through the more formal interfaces .loc and .iloc to do such manipulations. You have less chance of altering the original DataFrame or blowing up memory by making accidental copies. So the above is also accomplished with:

df.loc[:, ["col1"]]

…which also returns:

0     foo,
1      fee
2  bippity

…but now you’re starting to be exposed to the strange colon comma colon interface of .loc and .iloc. They both take row and column descriptors which could be either lists or slices.

If you didn’t know the column label, you can simply switch from the .loc interface to the .iloc interface. Again we see the mysterious colon-comma:

df.iloc[:, [0]]

…which also returns:

0     foo,
1      fee
2  bippity

At first such an API arrangement can be confusing. What the heck is the colon-comma before the list? This is about the most confusing thing in the world, because why wouldn’t it just be df[0] for the zero’th column or df[1] for the 2nd column like everything else in Python?

Why the :, and why the list to specify the 1st column? Well, the answer is that we naturally speak rows and columns, so the row indexing information comes first, and there’s a bit of Pandas API magic going on here. First, understand that you can get the whole dataframe with this:

df.iloc[:, :]

In the case of this particular DataFrame, another way to get back the identical results would be to say:

df.iloc[0:3, 0:3]

This is using the slice-API where the slices are row and column ranges, respectively. How do we do the same thing with the list API? Easy, we address each row and each column by numeric index:

df.iloc[[0, 1, 2], [0, 1, 2]]

With positional arguments, you can’t skip the first and set the second without a comma. The mysterious naked colons go away because there are lists in their place. The only way to get rid of the comma is to only ask for rows:

df.iloc[[0, 1, 2]]

…but if we ask for only columns, we have to say “all rows” like this:

df.iloc[:, [0, 1, 2]]

Crazy, right? But in regular Python, a colon with no numbers has exactly the same meaning, and Pythonistas worldwide use variable[:] in place of variable.copy() all the time. Same thing. A refresher on slicers. Say you wanted Hello World to say ello orl. You’d do this:

x = "Hello"
y = "World"
print((x[1:4], y[1:4]))

Which outputs:

('ell', 'orl')

So you see there are 2 slices defined and they’re wrapped in the parenthesis of the tuple so they can be fed together into the print statement. Leaving off the value after the colon means “to the end”:

print((x[1:], y[1:]))


('ello', 'orld')

And likewise leaving the number off before the colon implies from the beginning:

print((x[:4], y[:4]))


('Hell', 'Worl')

So if you leave off both, you get the whole slice of each:

print((x[:], y[:]))


('Hello', 'World')

Theefore if you have a required positional argument that supports either a slice or a list and you want to say “everything” (all rows in this case), you can simply put a naked colon in the required positional argument:

df.iloc[:, [0]]

The way to remember this is that .loc and .iloc support 2 arguments, one on either side of the comma, but it’s always in the order: rows, columns.

You can’t skip saying something for rows because it’s in the required first positional argument, and since it supports the slice API, you can just throw in a colon comma (saying all rows) and get onto specifying columns.

It’s also important to mention that the Pandas API has decided to drop the use of square-brackets for slices in their API and use them only instead for lists. It’s part of the power of Python that package developers can do whatever the heck they want with APIs, overwriting traditional behavior or the meaning of symbols. For example, the pathlib changes the meaning of forward-slash from the mathematical division operator to appending Path(directory / locations), and so too is Pandas changing how slices are defined inside its own objects.

And I guess it has to be mentioned that things get convoluted because rows comes first. Nobody says “columns and rows”. We all say “rows and columns”. If we lived in a bizarro world where we said things backwards, using the .iloc interface to select columns would be much easier:


…but that actually chooses row 1! It still works, but it’s choosing row 1 by its numeric index location. Things are much simpler when you’re using only the first position in location-sensitive APIs. But because we don’t live in that bizarro world where we say “columns and rows”, we have to put the syntax for “all rows” in position 1 so we can get to specifying columns. And those who think Pythonically know that the way to specify a slice from the beginning to the end is this:


…and since Pandas leaves off the extra nested square-brackets as implied, and Pandas insists on a slice for even a single column (least we get a Series back), the counter-intuitive but extremely common use-case for grabbing columns by list-index:

df.iloc[:, [0]]

or by slice-index:

df.iloc[:, 0:1]

Pshwew! Hope that helps. If so, follow me on YouTube and the social medias for more.