Plotting millions of points can be slow. Real slow... π΄
So why not use density maps? β‘
The mpl-scatter-density mini-package provides functionality to make it easy to make your own scatter density maps, both for interactive and non-interactive use. Fast. The following animation shows real-time interactive use with 10 million points, but interactive performance is still good even with 100 million points (and more if you have enough RAM).
When panning, the density map is shown at a lower resolution to keep things responsive (though this is customizable).
To install, simply do:
pip install mpl-scatter-density
This package requires Numpy, Matplotlib, and fast-histogram - these will be installed by pip if they are missing. Both Python 2.7 and Python 3.x are supported, and the package should work correctly on Linux, MacOS X, and Windows.
There are two main ways to use mpl-scatter-density, both of which are explained below.
The easiest way to use this package is to simply import mpl_scatter_density
,
then create Matplotlib axes as usual but adding a
projection='scatter_density'
option (if your reaction is 'wait, what?', see
here).
This will return a ScatterDensityAxes
instance that has a
scatter_density
method in addition to all the usual methods (scatter
,
plot
, etc.).
import numpy as np
import mpl_scatter_density
import matplotlib.pyplot as plt
# Generate fake data
N = 10000000
x = np.random.normal(4, 2, N)
y = np.random.normal(3, 1, N)
# Make the plot - note that for the projection option to work, the
# mpl_scatter_density module has to be imported above.
fig = plt.figure()
ax = fig.add_subplot(1, 1, 1, projection='scatter_density')
ax.scatter_density(x, y)
ax.set_xlim(-5, 10)
ax.set_ylim(-5, 10)
fig.savefig('gaussian.png')
Which gives:
The scatter_density
method takes the same options as imshow
(for example
cmap
, alpha
, norm
, etc.), but also takes the following optional
arguments:
dpi
: this is an integer that is used to determine the resolution of the density map. By default, this is 72, but you can change it as needed, or set it toNone
to use the default for the Matplotlib backend you are using.downres_factor
: this is an integer that is used to determine how much to downsample the density map when panning in interactive mode. Set this to 1 if you don't want any downsampling.color
: this can be set to any valid matplotlib color, and will be used to automatically make a monochromatic colormap based on this color. The colormap will fade to transparent, which means that this mode is ideal when showing multiple density maps together.
Here is an example of using the color
option:
import numpy as np
import matplotlib.pyplot as plt
import mpl_scatter_density # noqa
fig = plt.figure()
ax = fig.add_subplot(1, 1, 1, projection='scatter_density')
n = 10000000
x = np.random.normal(0.5, 0.3, n)
y = np.random.normal(0.5, 0.3, n)
ax.scatter_density(x, y, color='red')
x = np.random.normal(1.0, 0.2, n)
y = np.random.normal(0.6, 0.2, n)
ax.scatter_density(x, y, color='blue')
ax.set_xlim(-0.5, 1.5)
ax.set_ylim(-0.5, 1.5)
fig.savefig('double.png')
Which produces the following output:
If you are a more experienced Matplotlib user, you might want to use the
ScatterDensityArtist
directly (this is used behind the scenes in the
above example). To use this, initialize the ScatterDensityArtist
with
the axes as first argument, followed by any arguments you would have passed
to scatter_density
above (you can also take a look at the docstring for
ScatterDensityArtist
). You should then add the artist to the axes:
from mpl_scatter_density import ScatterDensityArtist
a = ScatterDensityArtist(ax, x, y)
ax.add_artist(a)
In some cases, your density map might have a high dynamic range, and you might
therefore want to show the log of the counts rather than the counts. You can do
this by passing a matplotlib.colors.Normalize
object to the norm
argument
in the same wasy as for imshow
. For example, the astropy package includes a nice framework
for making such a Normalize
object for different functions. The following
example shows how to show the density map on a log scale:
import numpy as np
import mpl_scatter_density
import matplotlib.pyplot as plt
# Make the norm object to define the image stretch
from astropy.visualization import LogStretch
from astropy.visualization.mpl_normalize import ImageNormalize
norm = ImageNormalize(vmin=0., vmax=1000, stretch=LogStretch())
N = 10000000
x = np.random.normal(4, 2, N)
y = np.random.normal(3, 1, N)
fig = plt.figure()
ax = fig.add_subplot(1, 1, 1, projection='scatter_density')
ax.scatter_density(x, y, norm=norm)
ax.set_xlim(-5, 10)
ax.set_ylim(-5, 10)
fig.savefig('gaussian_log.png')
Which produces the following output:
You can show a colorbar in the same way as you would for an image - the following example shows how to do it:
import numpy as np
import mpl_scatter_density
import matplotlib.pyplot as plt
N = 10000000
x = np.random.normal(4, 2, N)
y = np.random.normal(3, 1, N)
fig = plt.figure()
ax = fig.add_subplot(1, 1, 1, projection='scatter_density')
density = ax.scatter_density(x, y)
ax.set_xlim(-5, 10)
ax.set_ylim(-5, 10)
fig.colorbar(density, label='Number of points per pixel')
fig.savefig('gaussian_colorbar.png')
Which produces the following output:
In the same way that a 1-D array of values can be passed to Matplotlib's
scatter
function/method, a 1-D array of values can be passed to
scatter_density
using the c=
argument:
import numpy as np
import mpl_scatter_density
import matplotlib.pyplot as plt
N = 10000000
x = np.random.normal(4, 2, N)
y = np.random.normal(3, 1, N)
c = x - y + np.random.normal(0, 5, N)
fig = plt.figure()
ax = fig.add_subplot(1, 1, 1, projection='scatter_density')
ax.scatter_density(x, y, c=c, vmin=-10, vmax=+10, cmap=plt.cm.RdYlBu)
ax.set_xlim(-5, 13)
ax.set_ylim(-5, 11)
fig.savefig('gaussian_color_coded.png')
Which produces the following output:
Note that to keep performance as good as possible, the values from the c
attribute are averaged inside each pixel of the density map, then the colormap
is applied. This is a little different to what scatter
would converge to in
the limit of many points (since in that case it would apply the color to all the
markers than average the colors).
This follows the same ideas as datashader, but the aim of mpl-scatter-density is specifically to bring datashader-like functionality to Matplotlib users. Furthermore, mpl-scatter-density is intended to be very easy to install - for example it can be installed with pip. But if you have datashader installed and regularly use bokeh, mpl-scatter-density won't do much for you. Note that if you are interested in datashader and Matplotlib together, there is a work in progress (pull request) by @tacaswell to create a Matplotlib artist similar to that in this package but powered by datashader.
Vaex is a powerful package to visualize large datasets on N-dimensional grids, and therefore has some functionality that overlaps with what is here. However, the aim of mpl-scatter-density is just to provide a lightweight solution to make it easy for users already using Matplotlib to add scatter density maps to their plots rather than provide a complete environment for data visualization. I highly recommend that you take a look at Vaex and determine which approach is right for you!
If you are a Matplotlib developer: I truly am sorry for distorting the intended
purpose of projection
π. But you have to admit that it's a pretty
convenient way to have users get a custom Axes sub-class even if it has nothing
to do with actual projection!
There are a number of things we could add to this package, for example a way to plot density maps as contours, or a way to color code each point by a third quantity and have that reflected in the density map. If you have ideas, please open issues, and even better contribute a pull request! π
I'm glad you asked - of course you are very welcome to contribute! If you have some ideas, you can open issues or create a pull request directly. Even if you don't have time to contribute actual code changes, I would love to hear from you if you are having issues using this package.
[![Build Status](https://dev.azure.com/thomasrobitaille/mpl-scatter-density/_apis/build/status/astrofrog.mpl-scatter-density?branchName=master)](https://dev.azure.com/thomasrobitaille/mpl-scatter-density/_build/latest?definitionId=17&branchName=master)
To run the tests, you will need pytest and the pytest-mpl plugin. You can then run the tests with:
pytest mpl_scatter_density --mpl