Merge pull request #977 from campagnola/pseudoscatter-performance

Add a faster method for computing pseudoscatter
This commit is contained in:
Ogi Moore 2020-06-13 21:08:33 -07:00 committed by GitHub
commit ecd0642ca9
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23

View File

@ -2322,14 +2322,62 @@ def invertQTransform(tr):
raise Exception("Transform is not invertible.")
return inv[0]
def pseudoScatter(data, spacing=None, shuffle=True, bidir=False, method='exact'):
"""Return an array of position values needed to make beeswarm or column scatter plots.
def pseudoScatter(data, spacing=None, shuffle=True, bidir=False):
"""
Used for examining the distribution of values in a set. Produces scattering as in beeswarm or column scatter plots.
Used for examining the distribution of values in an array.
Given a list of x-values, construct a set of y-values such that an x,y scatter-plot
Given an array of x-values, construct an array of y-values such that an x,y scatter-plot
will not have overlapping points (it will look similar to a histogram).
"""
if method == 'exact':
return _pseudoScatterExact(data, spacing=spacing, shuffle=shuffle, bidir=bidir)
elif method == 'histogram':
return _pseudoScatterHistogram(data, spacing=spacing, shuffle=shuffle, bidir=bidir)
def _pseudoScatterHistogram(data, spacing=None, shuffle=True, bidir=False):
"""Works by binning points into a histogram and spreading them out to fill the bin.
Faster method, but can produce blocky results.
"""
inds = np.arange(len(data))
if shuffle:
np.random.shuffle(inds)
data = data[inds]
if spacing is None:
spacing = 2.*np.std(data)/len(data)**0.5
yvals = np.empty(len(data))
dmin = data.min()
dmax = data.max()
nbins = int((dmax-dmin) / spacing) + 1
bins = np.linspace(dmin, dmax, nbins)
dx = bins[1] - bins[0]
dbins = ((data - bins[0]) / dx).astype(int)
binCounts = {}
for i,j in enumerate(dbins):
c = binCounts.get(j, -1) + 1
binCounts[j] = c
yvals[i] = c
if bidir is True:
for i in range(nbins):
yvals[dbins==i] -= binCounts.get(i, 0) * 0.5
return yvals[np.argsort(inds)] ## un-shuffle values before returning
def _pseudoScatterExact(data, spacing=None, shuffle=True, bidir=False):
"""Works by stacking points up one at a time, searching for the lowest position available at each point.
This method produces nice, smooth results but can be prohibitively slow for large datasets.
"""
inds = np.arange(len(data))
if shuffle:
np.random.shuffle(inds)