Plotting with Seaborn

# Some library functionality we shall use.
import seaborn as sns
import pandas  as pd
import numpy   as np
import matplotlib.pyplot as plt

# see sns.set() for selecting a few basic parameters: style,
# colorpalette, font family, etc.
sns.set()

# load the data
df = pd.read_csv('https://raw.githubusercontent.com/OpenNeuroLab/metasearch/master/docs/data/phenotype_mri.csv')

# get types of data: data frame method: ``select_dtypes``

# To select all *numeric* types use the numpy dtype ``numpy.number``
var_qnt = list(df.select_dtypes(include=[np.number]))

# To get some quick summary info on each quant variable across groups,
# use the following (returns a DataFrame):
df.describe()

# To select Pandas categorical dtypes, use 'category'
var_cat = list(df.select_dtypes(include=['category']))

# However, in this case, we haven't made any specific 'category'
# variables. See the types we have for each col (returns a Series):
df.dtypes

# So there are type "object" for non-quant ones, which we can select
# as non-numerical
var_cat = list(df.select_dtypes(exclude=[np.number]))
# or could include 'object' type:
# var_cat = list(df.select_dtypes(include=['object']))

# some data will be missing, possibly.  Too see how many non-Nans are
# there for each variable, use (returns a Series):
df.count()

# If one has missing data in columns, some seaborn plotting functions
# can remove them; some have a ``dropna=True`` ability to remove NaNs
# from a column.  For many that *don't*, we can make a mini data
# frame.  Note this function has two outputs: a (potentially) smaller
# data frame, and the reduction in the number of rows.  There could be
# much better ways of writing a related function, this is just a
# quick example.
def NoNoNaNette( df, cols=[] ):

    if not(cols):
        df2 = df[:].dropna()
    else:
        df2 = df[cols].dropna()

    shin  = df.shape
    shout = df2.shape

    print("Input  shape:", shin)
    print("Output shape:", shout)

    # might be useful to return how many variables were left *out*
    return df2, shin[0] - shout[0]

# -------------------  1a) Basic scatterplot -------------------------
# quant vs quant, plots a confidence interval of fit by default.  This
# function does *not* deal well with NaNs, so use the remover:
df3, diff    = NoNoNaNette( df, ['age', 'full_iq'])
my_lmplot    = sns.lmplot( 'age', 'full_iq', data=df3 )
plt.show()

# -------------------  1b) Basic scatterplot, extra cat ----------------

# Can also divvy up with a category; here and below, generally given
# as 'hue' because it adds a color variation to the plot
df3, diff    = NoNoNaNette( df, ['age', 'full_iq', 'diagnosis'] )
my_lmplot    = sns.lmplot( 'age', 'full_iq', data=df3, hue='diagnosis' )
plt.show()

# -------------------  1c) Basic scatterplot, extra opts ----------------

df3, diff    = NoNoNaNette( df, ['age', 'full_iq', 'diagnosis'] )
my_lmplot    = sns.lmplot( 'age', 'full_iq', data=df3, hue='diagnosis' )
sns.plt.suptitle('Important data')   # useful plot title
sns.plt.xlabel('age'+' (years)')     # change labels, e.g., with units
#plt.tight_layout()  # see about this...
plt.show()

# ------------------ 2a) all quantitative pairs -------------------

# Plot all quant pairs vs each other, and histograms along diagonal;
# leave off first 'quant' var because it is just subj ID integer.

# This plotting function *does* have a ``dropna=True`` in it.
# However, the diagonal-plotter STILL complains about NaNs, so we can
# either use the above NoNoNaNette(), or set the diagonal to be the
# "kernel density estimate" instead of the bog standard histogram
# along the diagonals.  NB: in this case, the NoNoNaNette-d version
# would produce something different, because it gets rid of *any* row
# with a nonzero, then plots the pairs; the other way, two variables
# are selected and *then* the NaNy-rows between that pair are
# deleted-- that can generally have different behavior!

my_pairplot = sns.pairplot( df, diag_kind='kde' )
plt.show()

# What are default sundry values for various plot settings?  Check 'em
# out: 
sns.plotting_context()

# these can be changed in one fell swoop with some batched presets:
sns.set_context("talk")
# also have: "notebook", "paper" and "poster"

# ... and check the new normal:
sns.plotting_context()

# To change some of these, one can do, for example:
sns.set_context( rc={'lines.markeredgewidth': 0.0,
                     'xtick.labelsize': 8.0})
my_pairplot = sns.pairplot( df, diag_kind='kde')
plt.show()

# -------------- 2b) all quantitative pairs, no NaNs ----------------

# might be fun:
sns.set_context("poster")
df4, diff = NoNoNaNette(df, var_qnt[1:]) 
my_pairplot = sns.pairplot( df4 )
#plt.show()

# Either plot or save fig (or both); one can control, for example, the
# DPI, etc.
fname = "plot.png" 
plt.savefig(fname, dpi=300)
#plt.show()

# ------------- 2c) all quantitative pairs, corr values --------------

# trivially modified from above (-> names changed to protect the
# innocent):
import scipy.stats as spst
def corrfunc(x, y, **kws):
    r, _ = spst.pearsonr(x, y)
    ax = plt.gca()
    ax.annotate("r = {:.2f}".format(r),
                xy=(.1, .9), xycoords=ax.transAxes)

my_pairplot = sns.pairplot( df4,diag_kind='kde' )
# this just puts the corr values into the LHT
my_pairplot.map_lower(corrfunc)
plt.show()