The right* way to plot: gridspec¶

* It really is ...

PyCoffee @ Vitacura - 7 April 2016 - fvogt@eso.org¶

premise:

matplotlib is the one-stop shop for all 2-D plotting in Python. Easy to make rapid plots & tons of examples.
But at times frustrating to get beautiful & complex plots:
- subplots locations, sizes, ratios
- colorbar location, size, width, alignment
- labels outside plotting window
- layout updates
- ...

gridspec:

a simple, elegant, quick and essential way to handle multiple plotting areas in Python/matplotlib
http://matplotlib.org/users/gridspec.html

utility:

essential for complex subplots
extremely useful for plots with colorbars
great (and good practice) for ANY plot

cost:

2 lines +
1 sub-module import

gain:

total control over the plot
accuracy
elegance
infinite powers
...

take-home message :

$\texttt{plt.subplot(2,1,1)}$ $\leftarrow$ **NEVER EVER AGAIN !**
$\texttt{plt.colorbar()}$ $\leftarrow$ **NEVER EVER AGAIN !**

A-1) The goal¶

demo-image

A-2) The concept¶

demo-image

B) Set the stage: import the required modules¶

import numpy as np # For creating some fake data

# Use some notebook magic - forget the next line in normal scripted code
%pylab 

import matplotlib.pyplot as plt # Gives access to basic plotting functions

import matplotlib.gridspec as gridspec # GRIDSPEC !

from matplotlib.colorbar import Colorbar # For dealing with Colorbars the proper way - TBD in a separate PyCoffee ?

Using matplotlib backend: MacOSX
Populating the interactive namespace from numpy and matplotlib

C) Create some fake datasets to play with ...¶

# Ok, let us start by creating some fake data
x = np.random.randn(1000)
y = np.random.randn(1000)
z = np.sqrt(x**2+y**2)

# `randn` generates an array of shape ``(d0, d1, ..., dn)``, filled
# with random floats sampled from a univariate "normal" (Gaussian)
# distribution of mean 0 and variance 1.

D-1) Get started with the plotting¶

# First, create the figure
plt.close(1)
fig = plt.figure(1, figsize=(15,8))

fig.show()

# Now, create the gridspec structure, as required
gs = gridspec.GridSpec(3,4, height_ratios=[0.05,1,0.2], width_ratios=[1,0.2,0.2,1])

# 3 rows, 4 columns, each with the required size ratios. 
# Also make sure the margins and spacing are apropriate

gs.update(left=0.05, right=0.95, bottom=0.08, top=0.93, wspace=0.02, hspace=0.03)

# Note: I set the margins to make it look good on my screen ...
# BUT: this is irrelevant for the saved image, if using bbox_inches='tight'in savefig !

# Note: Here, I use a little trick. I only have three vertical layers of plots : 
# a scatter plot, a histogram, and a line plot. So, in principle, I could use a 3x3 structure. 
# However, I want to have the histogram 'closer' from the scatter plot than the line plot.
# So, I insert a 4th layer between the histogram and line plot, 
# keep it empty, and use its thickness (the 0.2 above) to adjust the space as required.

fig.show()

D-2) The scatter plot & colorbar¶

# First, the scatter plot
# Use the gridspec magic to place it
# --------------------------------------------------------
ax1 = plt.subplot(gs[1,0]) # place it where it should be.
# --------------------------------------------------------

# The plot itself
plt1 = ax1.scatter(x, y, c = z, 
                   marker = 's', s=20, edgecolor = 'none',alpha =1,
                   cmap = 'magma_r', vmin =0 , vmax = 4)

# Define the limits, labels, ticks as required
ax1.grid(True)
ax1.set_xlim([-4,4])
ax1.set_ylim([-4,4])
ax1.set_xlabel(r' ') # Force this empty !
ax1.set_xticks(np.linspace(-4,4,9)) # Force this to what I want - for consistency with histogram below !
ax1.set_xticklabels([]) # Force this empty !
ax1.set_ylabel(r'My y label')

# and let us not forget the colorbar  above !
# --------------------------------------------------------
cbax = plt.subplot(gs[0,0]) # Place it where it should be.
# --------------------------------------------------------

cb = Colorbar(ax = cbax, mappable = plt1, orientation = 'horizontal', ticklocation = 'top')
cb.set_label(r'Colorbar !', labelpad=10)

fig.show()

D-3) The first histogram (vertical)¶

# NOTE: I guess that a kernel density plot on top of the histogram would be better from a scientific standpoint. 
# But this is only meant as an illustration of a side-plot, so who cares ?

# And now the histogram
# Use the gridspec magic to place it
# --------------------------------------------------------
ax1v = plt.subplot(gs[1,1])
# --------------------------------------------------------


# Plot the data
bins = np.arange(-4,4,0.1)
ax1v.hist(y,bins=bins, orientation='horizontal', color='k', edgecolor='w')

# Define the limits, labels, ticks as required
ax1v.set_yticks(np.linspace(-4,4,9)) # Ensures we have the same ticks as the scatter plot !
ax1v.set_xticklabels([])
ax1v.set_yticklabels([])
ax1v.set_ylim([-4,4])
ax1v.grid(True)

fig.show()

D-4) The second histogram (horizontal)¶

# And now another histogram
# Use the gridspec magic to place it
# --------------------------------------------------------
ax1h = plt.subplot(gs[2,0])
# --------------------------------------------------------


# Plot the data
bins = np.arange(-4,4,0.1)
ax1h.hist(x, bins=bins, orientation='vertical', color='k', edgecolor='w')

# Define the limits, labels, ticks as required
ax1h.set_xticks(np.linspace(-4,4,9)) # Ensures we have the same ticks as the scatter plot !
ax1h.set_yticklabels([])
ax1h.set_xlim([-4,4])
ax1h.set_xlabel(r'My x label')
ax1h.grid(True)

fig.show()

D-5) The line plot¶

# Finally, show some 'spectra' in the right panel
# Use the gridspec magic to place it
# --------------------------------------------------------
ax2 = plt.subplot(gs[0:2,3]) # Make it span the entire height of the figure (3 rows)
# --------------------------------------------------------

# Plot the data
plt.plot(x[::20], ls = '-', color='darkviolet', lw=2)
plt.plot(y[::20], ls = '-', color ='tomato', lw=2)

# Define the limits, labels, ticks as required
ax2.set_xlabel('My other x label')
ax2.set_ylabel('My other y label')
ax2.set_ylim([-4,4])
ax2.grid(True)

fig.show()

D-6) Plot and save it all¶

# Save it and display it
fig.savefig('gridspec_demo.pdf', bbox_inches='tight') 
# bbox_inches -> crops the extra space if any !

fig.show()

E) Final comments¶

gridspec does not do "insets" of plots.
imshow always a bit tricky if axis ratio is set.

F) rcparams anyone ?¶

G) Python 2.7 vs Python 3.x. Thoughts ?¶

H) Gridspec & Jupyter notebooks (inline)¶

gridspec also works fine inside a jupyter notebook. The examples before used an external plot window for clarity during the setp-by-step construction process of the final plot. Below, you can get the same figure inside the notebook. Note that this will require to restart the kernel before running it.

import numpy as np # For creating some fake data

# Use some notebook magic - forget the next line in normal scripted code
%matplotlib inline

import matplotlib.pyplot as plt # Gives access to basic plotting functions
import matplotlib.gridspec as gridspec # GRIDSPEC !
from matplotlib.colorbar import Colorbar # For dealing with Colorbars the proper way - TBD in a separate PyCoffee ?

# Ok, let us start by creating some fake data
x = np.random.randn(1000)
y = np.random.randn(1000)
z = np.sqrt(x**2+y**2)

# `randn` generates an array of shape ``(d0, d1, ..., dn)``, filled
# with random floats sampled from a univariate "normal" (Gaussian)
# distribution of mean 0 and variance 1.


fig = plt.figure(1, figsize=(15,8))


# Now, create the gridspec structure, as required
gs = gridspec.GridSpec(3,4, height_ratios=[0.05,1,0.2], width_ratios=[1,0.2,0.2,1])

# 3 rows, 4 columns, each with the required size ratios. 
# Also make sure the margins and spacing are apropriate

gs.update(left=0.05, right=0.95, bottom=0.08, top=0.93, wspace=0.02, hspace=0.03)

# Note: I set the margins to make it look good on my screen ...
# BUT: this is irrelevant for the saved image, if using bbox_inches='tight'in savefig !

# Note: Here, I use a little trick. I only have three vertical layers of plots : 
# a scatter plot, a histogram, and a line plot. So, in principle, I could use a 3x3 structure. 
# However, I want to have the histogram 'closer' from the scatter plot than the line plot.
# So, I insert a 4th layer between the histogram and line plot, 
# keep it empty, and use its thickness (the 0.2 above) to adjust the space as required.

# First, the scatter plot
# Use the gridspec magic to place it
# --------------------------------------------------------
ax1 = plt.subplot(gs[1,0]) # place it where it should be.
# --------------------------------------------------------

# The plot itself
plt1 = ax1.scatter(x, y, c = z, 
                   marker = 's', s=20, edgecolor = 'none',alpha =1,
                   cmap = 'magma_r', vmin =0 , vmax = 4)

# Define the limits, labels, ticks as required
ax1.grid(True)
ax1.set_xlim([-4,4])
ax1.set_ylim([-4,4])
ax1.set_xlabel(r' ') # Force this empty !
ax1.set_xticks(np.linspace(-4,4,9)) # Force this to what I want - for consistency with histogram below !
ax1.set_xticklabels([]) # Force this empty !
ax1.set_ylabel(r'My y label')

# and let us not forget the colorbar  above !
# --------------------------------------------------------
cbax = plt.subplot(gs[0,0]) # Place it where it should be.
# --------------------------------------------------------

cb = Colorbar(ax = cbax, mappable = plt1, orientation = 'horizontal', ticklocation = 'top')
cb.set_label(r'Colorbar !', labelpad=10)

# NOTE: I guess that a kernel density plot on top of the histogram would be better from a scientific standpoint. 
# But this is only meant as an illustration of a side-plot, so who cares ?

# And now the histogram
# Use the gridspec magic to place it
# --------------------------------------------------------
ax1v = plt.subplot(gs[1,1])
# --------------------------------------------------------


# Plot the data
bins = np.arange(-4,4,0.1)
ax1v.hist(y,bins=bins, orientation='horizontal', color='k', edgecolor='w')

# Define the limits, labels, ticks as required
ax1v.set_yticks(np.linspace(-4,4,9)) # Ensures we have the same ticks as the scatter plot !
ax1v.set_xticklabels([])
ax1v.set_yticklabels([])
ax1v.set_ylim([-4,4])
ax1v.grid(True)

# And now another histogram
# Use the gridspec magic to place it
# --------------------------------------------------------
ax1h = plt.subplot(gs[2,0])
# --------------------------------------------------------


# Plot the data
bins = np.arange(-4,4,0.1)
ax1h.hist(x, bins=bins, orientation='vertical', color='k', edgecolor='w')

# Define the limits, labels, ticks as required
ax1h.set_xticks(np.linspace(-4,4,9)) # Ensures we have the same ticks as the scatter plot !
ax1h.set_yticklabels([])
ax1h.set_xlim([-4,4])
ax1h.set_xlabel(r'My x label')
ax1h.grid(True)

# Finally, show some 'spectra' in the right panel
# Use the gridspec magic to place it
# --------------------------------------------------------
ax2 = plt.subplot(gs[0:2,3]) # Make it span the entire height of the figure (3 rows)
# --------------------------------------------------------

# Plot the data
plt.plot(x[::20], ls = '-', color='darkviolet', lw=2)
plt.plot(y[::20], ls = '-', color ='tomato', lw=2)

# Define the limits, labels, ticks as required
ax2.set_xlabel('My other x label')
ax2.set_ylabel('My other y label')
ax2.set_ylim([-4,4])
ax2.grid(True)