When dealing with numeric matrices and vectors in Python, NumPy makes life a lot easier. For more complex data, however, it leaves a lot to be desired. If you're used to working with data frames in R, doing data analysis directly with NumPy feels like a step back.
Fortunately, some nice folks have written the Python Data Analysis Library (a.k.a. pandas).
Pandas provides an R-like DataFrame, produces high quality plots with matplotlib, and integrates nicely with other libraries that expect NumPy arrays.
In this tutorial, we'll go through the basics of pandas using a year's worth of weather data from Weather Underground. Pandas has a lot of functionality, so we'll only be able to cover a small fraction of what you can do. Check out the (very readable) pandas docs if you want to learn more.
Installing pandas should be an easy process if you use pip:
sudo pip install pandasFor more complex scenarios, please see the installation instructions.
OK, let's get started by importing the pandas library.
In [ ]:
import pandas
In [ ]:
data = pandas.read_csv("data/weather_year.csv")
We can get a summary of the DataFrame by printing the object.
In [ ]:
data
This gives us a lot of information. First, we can see that there are 366 rows (entries) -- a year and a day's worth of weather. Each column is printed along with however many "non-null" values are present.
We'll talk more about null (or missing) values in pandas later, but for now we can note that only the "Max Gust SpeedMPH" and "Events" columns have fewer than 366 non-null values.
Lastly, the data types (dtypes) of the columns are printed at the very bottom. We can see that there are 4 float64, 16 int64, and 3 object columns.
In [ ]:
len(data)
Using len on a DataFrame will give you the number of rows. You can get the column names using the columns property.
In [ ]:
data.columns
Columns can be accessed in two ways. The first is using the DataFrame like a dictionary with string keys:
In [ ]:
data["EDT"]
You can get multiple columns out at the same time by passing in a list of strings.
In [ ]:
data[["EDT", "Mean TemperatureF"]]
The second way to access columns is using the dot syntax. This only works if your column name could also be a Python variable name (i.e., no spaces), and if it doesn't collide with another DataFrame property or function name (e.g., count, sum).
In [ ]:
data.EDT
We'll be mostly using the dot syntax here because you can auto-complete the names in IPython. The first pandas function we'll learn about is head(). This gives us the first 5 items in a column (or the first 5 rows in the DataFrame).
In [ ]:
data.EDT.head()
Passing in a number n gives us the first n items in the column. There is also a corresponding tail() method that gives the last n items or rows.
In [ ]:
data.EDT.head(10)
This also works with the dictionary syntax.
In [ ]:
data["Mean TemperatureF"].head()
In [ ]:
If the data in the column is numeric, you can use describe() to get some stats on it.
In [ ]:
data["Mean TemperatureF"].describe()
The column names in data are a little unweildy, so we're going to rename them. This is as easy as assigning a new list of column names to the columns property of the DataFrame.
In [ ]:
data.columns = ["date", "max_temp", "mean_temp", "min_temp", "max_dew",
"mean_dew", "min_dew", "max_humidity", "mean_humidity",
"min_humidity", "max_pressure", "mean_pressure",
"min_pressure", "max_visibilty", "mean_visibility",
"min_visibility", "max_wind", "mean_wind", "min_wind",
"precipitation", "cloud_cover", "events", "wind_dir"]
These should be in the same order as the original columns. Let's take another look at our DataFrame summary.
In [ ]:
data
Now our columns can all be accessed using the dot syntax!
In [ ]:
data.mean_temp.head()
There are lots useful methods on columns, such as std() to get the standard deviation. Most of pandas' methods will happily ignore missing values like NaN.
In [ ]:
data.mean_temp.std()
Some methods, like plot() and hist() produce plots using matplotlib. We'll go over plotting in more detail later.
In [ ]:
data.mean_temp.hist()
By the way, many of the column-specific methods also work on the entire DataFrame. Instead of a single number, you'll get a result for each column.
In [ ]:
data.std()
In [ ]:
Methods like sum() and std() work on entire columns. We can run our own functions across all values in a column (or row) using apply().
To give you an idea of how this works, let's consider the "date" column in our DataFrame (formally "EDT").
In [ ]:
data.date.head()
We can use the values property of the column to get a list of values for the column. Inspecting the first value reveals that these are strings with a particular format.
In [ ]:
first_date = data.date.values[0]
first_date
The strptime function from the datetime module will make quick work of this date string. There are many more shortcuts available for strptime.
In [ ]:
# Import the datetime class from the datetime module
from datetime import datetime
# Convert date string to datetime object
datetime.strptime(first_date, "%Y-%m-%d")
Using the apply() method, which takes a function (without the parentheses), we can apply strptime to each value in the column. We'll overwrite the string date values with their Python datetime equivalents.
In [ ]:
# Define a function to convert strings to dates
def string_to_date(date_string):
return datetime.strptime(date_string, "%Y-%m-%d")
# Run the function on every date string and overwrite the column
data.date = data.date.apply(string_to_date)
data.date.head()
Let's go one step futher. Each row in our DateFrame represents the weather from a single day. Each row in a DataFrame is associated with an index, which is a label that uniquely identifies a row.
Our row indices up to now have been auto-generated by pandas, and are simply integers from 0 to 365. If we use dates instead of integers for our index, we will get some extra benefits from pandas when plotting later on. Overwriting the index is as easy as assigning to the index property of the DataFrame.
In [ ]:
data.index = data.date
data
Now we can quickly look up a row by its date with the ix[] property.
In [ ]:
data.ix[datetime(2012, 8, 19)]
With all of the dates in the index now, we no longer need the "date" column. Let's drop it.
In [ ]:
data = data.drop("date", axis=1)
data.columns
Note that we need to pass in axis=1 in order to drop a column. For more details, check out the documentation for drop.
In [ ]:
datetime(2012, 5, 1) # May 1st of 2012
In [ ]:
Pandas considers values like NaN and None to represent missing data. The pandas.isnull function can be used to tell whether or not a value is missing.
Let's use apply() across all of the columns in our DataFrame to figure out which values are missing.
In [ ]:
empty = data.apply(lambda col: pandas.isnull(col))
empty
We got back a dataframe (empty) with boolean values for all 22 columns and 366 rows. Inspecting the first 10 values of the "events", column we can see that there are some missing values because a True was returned from pandas.isnull.
In [ ]:
empty.events.head(10)
Looking at the corresponding rows in the original DataFrame reveals that pandas has filled in NaN for empty values in the "events" column.
In [ ]:
data.events.head(10)
This isn't exactly what we want. One option is to drop all rows in the DataFrame with missing "events" values.
In [ ]:
data.dropna(subset=["events"])
The DataFrame we get back has only 162 rows, so we can infer that there were 366 - 162 = 204 missing values in the "events" column. Note that this didn't affect data; we're just looking at a copy.
Instead of dropping the rows with missing values, let's fill them with empty strings (you'll see why in a moment). This is easily done with the fillna() function. We'll go ahead and overwrite the "events" column with empty string missing values instead of NaN.
In [ ]:
data.events = data.events.fillna("")
data.events.head(10)
Sometimes you need to access individual rows in your DataFrame. The irow() function lets you grab the ith row from a DataFrame (starting from 0).
In [ ]:
data.irow(0)
Of course, another option is to use the index.
In [ ]:
data.ix[datetime(2013, 1, 1)]
You can iterate over each row in the DataFrame with iterrows(). Note that this function returns both the index and the row. Also, you must access columns in the row you get back from iterrows() with the dictionary syntax.
In [ ]:
num_rain = 0
for idx, row in data.iterrows():
if "Rain" in row["events"]:
num_rain += 1
"Days with rain: {0}".format(num_rain)
Was there any November rain?
Hint: check out the strftime() function on datetime objects and the documentation
In [ ]:
d = datetime(2012, 1, 1)
d.strftime("%B")
In [ ]:
Most of your time using pandas will likely be devoted to selecting rows of interest from a DataFrame. In addition to strings, the dictionary syntax accepts things like this:
In [ ]:
freezing_days = data[data.max_temp <= 32]
freezing_days
We get back another DataFrame with fewer rows (21 in this case). This DataFrame can be filtered down even more.
In [ ]:
freezing_days[freezing_days.min_temp >= 20]
Or, using boolean operations, we could apply both filters to the original DataFrame at the same time.
In [ ]:
data[(data.max_temp <= 32) & (data.min_temp >= 20)]
It's important to understand what's really going on underneath with filtering. Let's look at what kind of object we actually get back when creating a filter.
In [ ]:
temp_max = data.max_temp <= 32
type(temp_max)
This is a pandas Series object, which is the one-dimensional equivalent of a DataFrame. Because our DataFrame uses datetime objects for the index, we have a specialized TimeSeries object.
What's inside the filter?
In [ ]:
temp_max
Our filter is nothing more than a Series with a boolean value for every item in the index. When we "run the filter" as so:
In [ ]:
data[temp_max]
pandas lines up the rows of the DataFrame and the filter using the index, and then keeps the rows with a True filter value. That's it.
Let's create another filter.
In [ ]:
temp_min = data.min_temp >= 20
temp_min
Now we can see what the boolean operations are doing. Something like & (not and)...
In [ ]:
temp_min & temp_max
...is just lining up the two filters using the index, performing a boolean AND operation, and returning the result as another Series.
We can do other boolean operations too, like OR:
In [ ]:
temp_min | temp_max
Because the result is just another Series, we have all of the regular pandas functions at our disposal. The any() function returns True if any value in the Series is True.
In [ ]:
temp_both = temp_min & temp_max
temp_both.any()
Sometimes filters aren't so intuitive. This (sadly) doesn't work:
In [ ]:
try:
data["Rain" in data.events]
except:
pass # "KeyError: no item named False"
We can wrap it up in an apply() call fairly easily, though:
In [ ]:
data[data.events.apply(lambda e: "Rain" in e)]
In [ ]:
data.precipitation.head()
We'll replace "T" with a very small number, and convert the rest of the strings to floats:
In [ ]:
# Convert precipitation to floating point number
# "T" means "trace of precipitation"
def precipitation_to_float(precip_str):
if precip_str == "T":
return 1e-10 # Very small value
return float(precip_str)
data.precipitation = data.precipitation.apply(precipitation_to_float)
data.precipitation.head()
Now on to the exercise.
What was the coldest it ever got when there was no cloud cover and no precipitation?
In [ ]:
Besides apply(), another great DataFrame function is groupby().
It will group a DataFrame by one or more columns, and let you iterate through each group.
As an example, let's group our DataFrame by the "cloud_cover" column (a value ranging from 0 to 8).
In [ ]:
cover_temps = {}
for cover, cover_data in data.groupby("cloud_cover"):
cover_temps[cover] = cover_data.mean_temp.mean() # The mean mean temp!
cover_temps
When you iterate through the result of groupby(), you will get a tuple.
The first item is the column value, and the second item is a filtered DataFrame (where the column equals the first tuple value).
You can group by more than one column as well.
In this case, the first tuple item returned by groupby() will itself be a tuple with the value of each column.
In [ ]:
for (cover, events), group_data in data.groupby(["cloud_cover", "events"]):
print "Cover: {0}, Events: {1}, Count: {2}".format(cover, events, len(group_data))
Weather events in our DataFrame are stored in strings like "Rain-Thunderstorm" to represent that it rained and there was a thunderstorm that day. Let's split them out into boolean "rain", "thunderstorm", etc. columns.
First, let's discover the different kinds of weather events we have with unique().
In [ ]:
data.events.unique()
Looks like we have "Rain", "Thunderstorm", "Fog", and "Snow" events. Creating a new column for each of these event kinds is a piece of cake with the dictionary syntax.
In [ ]:
for event_kind in ["Rain", "Thunderstorm", "Fog", "Snow"]:
col_name = event_kind.lower() # Turn "Rain" into "rain", etc.
data[col_name] = data.events.apply(lambda e: event_kind in e)
data
Our new columns show up at the bottom. We can access them now with the dot syntax.
In [ ]:
data.rain
We can also do cool things like find out how many True values there are (i.e., how many days had rain)...
In [ ]:
data.rain.sum()
...and get all the days that had both rain and snow!
In [ ]:
data[data.rain & data.snow]
In [ ]:
We've already seen how the hist() function makes generating histograms a snap. Let's look at the plot() function now.
In [ ]:
data.max_temp.plot()
That one line of code did a lot for us. First, it created a nice looking line plot using the maximum temperature column from our DataFrame. Second, because we used datetime objects in our index, pandas labeled the x-axis appropriately.
Pandas is smart too. If we're only looking at a couple of days, the x-axis looks different:
In [ ]:
data.max_temp.tail().plot()
Prefer a bar plot? Pandas has got your covered.
In [ ]:
data.max_temp.tail().plot(kind="bar", rot=10)
The plot() function returns a matplotlib AxesSubPlot object. You can pass this object into subsequent calls to plot() in order to compose plots.
Although plot() takes a variety of parameters to customize your plot, users familiar with matplotlib will feel right at home with the AxesSubPlot object.
In [ ]:
ax = data.max_temp.plot(title="Min and Max Temperatures")
data.min_temp.plot(style="red", ax=ax)
ax.set_ylabel("Temperature (F)")
Writing data out in pandas is as easy as getting data in. To save our DataFrame out to a new csv file, we can just do this:
In [ ]:
data.to_csv("data/weather-mod.csv")
Want to make that tab separated instead? No problem.
In [ ]:
data.to_csv("data/weather-mod.tsv", sep="\t")
There's also support for reading and writing Excel files, if you need it.
We've only covered a small fraction of the pandas library here. Before I wrap up, however, there are a few miscellaneous tips I'd like to go over.
First, it can be confusing to know when an operation will modify a DataFrame and when it will return a copy to you. Pandas behavior here is entirely dictated by NumPy, and some situations are unintuitive.
For example, what do you think will happen here?
In [ ]:
for idx, row in data.iterrows():
row["max_temp"] = 0
data.max_temp.head()
Contrary to what you might expect, modifying row did not modify data!
This is because row is a copy, and does not point back to the original DataFrame.
Here's the right way to do it:
In [ ]:
for idx, row in data.iterrows():
data.ix[idx, "max_temp"] = 0
any(data.max_temp != 0) # Any rows with max_temp not equal to zero?
Just to make you even more confused, this also doesn't work:
In [ ]:
for idx, row in data.iterrows():
data.ix[idx]["max_temp"] = 100
data.max_temp.head()
When using apply(), the default behavior is to go over columns.
In [ ]:
data.apply(lambda c: c.name)
You can make apply() go over rows by passing axis=1
In [ ]:
data.apply(lambda r: r["max_pressure"] - r["min_pressure"], axis=1)
When you call drop(), though, it's flipped. To drop a column, you need to pass axis=1
In [ ]:
data.drop("events", axis=1).columns