Seaborn tsplot does not show datetimes on x axis well

Question

Below I have the following script which creates a simple time series plot:

%matplotlib inline
import datetime
import pandas as pd
import seaborn as sns
import matplotlib.pyplot as plt

fig, ax = plt.subplots()

df = []
start_date = datetime.datetime(2015, 7, 1)
for i in range(10):
    for j in [1,2]:
        unit = 'Ones' if j == 1 else 'Twos'
        date = start_date + datetime.timedelta(days=i)
        
        df.append({
                'Date': date.strftime('%Y%m%d'),
                'Value': i * j,
                'Unit': unit
            })
    
df = pd.DataFrame(df)

sns.tsplot(df, time='Date', value='Value', unit='Unit', ax=ax)
fig.autofmt_xdate()

And the result of this is the following:

As you can see the x-axis has strange numbers for the datetimes, and not the usual "nice" representations that come with matplotlib and other plotting utilities. I've tried many things, re-formatting the data but it never comes out clean. Anyone know a way around?

Solution 1

Matplotlib represents dates as floating point numbers (in days), thus unless you (or pandas or seaborn), tell it that your values are representing dates, it will not format the ticks as dates. I'm not a seaborn expert, but it looks like it (or pandas) does convert the datetime objects to matplotlib dates, but then does not assign proper locators and formatters to the axes. This is why you get these strange numbers, which are in fact just the days since 0001.01.01. So you'll have to take care of the ticks manually (which, in most cases, is better anyways as it gives you more control).

So you'll have to assign a date locator, which decides where to put ticks, and a date formatter, which will then format the strings for the tick labels.

import datetime
import pandas as pd
import seaborn as sns
import matplotlib.pyplot as plt
import matplotlib.dates as mdates

# build up the data
df = []
start_date = datetime.datetime(2015, 7, 1)
for i in range(10):
    for j in [1,2]:
        unit = 'Ones' if j == 1 else 'Twos'
        date = start_date + datetime.timedelta(days=i)

        # I believe it makes more sense to directly convert the datetime to a
        # "matplotlib"-date (float), instead of creating strings and then let
        # pandas parse the string again
        df.append({
                'Date': mdates.date2num(date),
                'Value': i * j,
                'Unit': unit
            })
df = pd.DataFrame(df)

# build the figure
fig, ax = plt.subplots()
sns.tsplot(df, time='Date', value='Value', unit='Unit', ax=ax)

# assign locator and formatter for the xaxis ticks.
ax.xaxis.set_major_locator(mdates.AutoDateLocator())
ax.xaxis.set_major_formatter(mdates.DateFormatter('%Y.%m.%d'))

# put the labels at 45deg since they tend to be too long
fig.autofmt_xdate()
plt.show()

Result:

Solution 2

For me, @hitzg's answer results in "OverflowError: signed integer is greater than maximum" in the depths of DateFormatter.

Looking at my dataframe, my indices are datetime64, not datetime. Pandas converts these nicely though. The following works great for me:

import matplotlib as mpl

def myFormatter(x, pos):
    return pd.to_datetime(x)

[ . . . ]

ax.xaxis.set_major_formatter(mpl.ticker.FuncFormatter(myFormatter))

Solution 3

Here is a potentially inelegant solution, but it's the only one I have ... Hope it helps!

    g = sns.pointplot(x, y, data=df, ci=False);

    unique_dates = sorted(list(df['Date'].drop_duplicates()))
    date_ticks = range(0, len(unique_dates), 5)

    g.set_xticks(date_ticks);
    g.set_xticklabels([unique_dates[i].strftime('%d %b') for i in date_ticks], rotation='vertical');
    g.set_xlabel('Date');

Let me know if you see any issues!

Comments

[kurious]

I get this error: OverflowError: Python int too large to convert to C long

[hitzg]

Works on my side. Do you use my exact example? Maybe you need to update your distro.

[Bryan P]

@kurious: I'm getting something similar: OverflowError: signed integer is greater than maximum from matplotlib/dates.pyc (line 214) in _from_ordinalf(x, tz). Just pip installed on Linux with Python 2.7.

[MERose]

And then how can we format it? ax.xaxis.set_major_formatter(mdates.DateFormatter('%Y.%m.%d'‌​)) won't work.

[CertainPerformance]

I'm curious, is there a reason for for i in <tons of spaces> date_ticks?

[ltjds]

This was something that got weird for me ... As it turns out unique_dates is a list of datetime objects. That is, the list doesn't have the method strftime, so we need to take the ith item from the list and apply the method.