def plot_distribution(self):
"Plot histogram of # classes in an image. Return the figure."
table = self.get_cls_ground_truth(with_diff=False, with_trun=True)
fig = plt.figure()
ax = fig.add_subplot(111)
bins = np.arange(1, max(table.sum(1)) + 2)
ax.hist(table.sum(1), bins, align="left", normed=True)
ax.set_xticks(bins)
ax.set_xlabel("Number of classes present in the image")
ax.grid(False)
dirname = config.get_dataset_stats_dir(self)
filename = opjoin(dirname, "num_classes.png")
fig.savefig(filename)
return fig
def plot_distribution(self):
"Plot histogram of # classes in an image. Return the figure."
table = self.get_cls_ground_truth(with_diff=False, with_trun=True)
fig = plt.figure()
ax = fig.add_subplot(111)
bins = np.arange(1, max(table.sum(1)) + 2)
ax.hist(table.sum(1), bins, align='left', normed=True)
ax.set_xticks(bins)
ax.set_xlabel('Number of classes present in the image')
ax.grid(False)
dirname = config.get_dataset_stats_dir(self)
filename = opjoin(dirname, 'num_classes.png')
fig.savefig(filename)
return fig
def plot_coocurrence(
self,
cmap=plt.cm.Reds,
color_anchor=[0, 1],
x_tick_rot=90,
size=None,
title=None,
plot_vals=True,
second_order=False,
):
"""
Plot a heat map of conditional occurence, where cell (i,j) means
P(C_j|C_i). The last column in the K x (K+2) heat map corresponds
to the prior P(C_i).
If second_order, plots (K choose 2) x (K+2) heat map corresponding
to P(C_i|C_j,C_k): second-order correlations.
Return the figure.
"""
table = self.get_cls_ground_truth(with_diff=False, with_trun=True)
# This takes care of most of the difference between normal and second_order
# In the former case, a "combination" is just one class to condition on.
combinations = combination_strs = table.cols
if second_order:
combinations = [x for x in itertools.combinations(table.cols, 2)]
combination_strs = ["%s, %s" % (x[0], x[1]) for x in combinations]
total = table.shape[0]
N = len(table.cols)
K = len(combinations)
data = np.zeros((K, N + 2)) # extra columns are for P("nothing"|C) and P(C)
prior = np.zeros(K)
for i, combination in enumerate(combinations):
if second_order:
cls1 = combination[0]
cls2 = combination[1]
conditioned = table.filter_on_column(cls1).filter_on_column(cls2)
else:
conditioned = table.filter_on_column(combination)
# count all the classes
data[i, :-2] = conditioned.sum()
# count the number of times that cls was the only one present to get
# P("nothing"|C)
if second_order:
data[i, -2] = ((conditioned.sum(1) - 2) == 0).sum()
else:
data[i, -2] = ((conditioned.sum(1) - 1) == 0).sum()
# normalize
max_val = np.max(data[i, :])
data[i, :] /= max_val
data[i, :][data[i, :] == 1] = np.nan
# use the max count to compute the prior
data[i, -1] = max_val / total
m = Table(data, table.cols + ["nothing", "prior"], index=combination_strs)
# If second_order, sort by prior and remove rows with 0 prior
if second_order:
m = m.filter_on_column("prior", 0.001, operator.gt).sort_by_column("prior", descending=True)
# TODO: just take the top K actually, for a side-by-side figure
m.arr = m.arr[: len(self.classes), :]
if size:
fig = plt.figure(figsize=size)
else:
w = max(12, m.shape[1])
h = max(12, m.shape[0])
fig = plt.figure(figsize=(w, h))
ax_im = fig.add_subplot(111)
# make axes for colorbar
divider = make_axes_locatable(ax_im)
ax_cb = divider.new_vertical(size="5%", pad=0.1, pack_start=True)
fig.add_axes(ax_cb)
# The call to imshow produces the matrix plot:
im = ax_im.imshow(
m.arr, origin="upper", interpolation="nearest", vmin=color_anchor[0], vmax=color_anchor[1], cmap=cmap
)
# Formatting:
ax = ax_im
ax.set_xticks(np.arange(m.shape[1]))
ax.set_xticklabels(m.cols)
for tick in ax.xaxis.iter_ticks():
tick[0].label2On = True
tick[0].label1On = False
tick[0].label2.set_rotation(x_tick_rot)
tick[0].label2.set_fontsize("x-large")
ax.set_yticks(np.arange(m.shape[0]))
ax.set_yticklabels(m.index, size="x-large")
ax.yaxis.set_minor_locator(matplotlib.ticker.FixedLocator(np.arange(-0.5, m.shape[0] + 0.5)))
ax.xaxis.set_minor_locator(matplotlib.ticker.FixedLocator(np.arange(-0.5, m.shape[1] - 0.5)))
ax.grid(False, which="major")
ax.grid(True, which="minor", ls="-", lw=7, c="w")
# Make the major and minor tick marks invisible
for line in ax.xaxis.get_ticklines() + ax.yaxis.get_ticklines():
line.set_markeredgewidth(0)
for line in ax.xaxis.get_minorticklines() + ax.yaxis.get_minorticklines():
line.set_markeredgewidth(0)
# Limit the area of the plot
ax.set_ybound([-0.5, m.shape[0] - 0.5])
ax.set_xbound([-0.5, m.shape[1] - 0.5])
# The following produces the colorbar and sets the ticks
# Set the ticks - if 0 is in the interval of values, set that, as well
# as the maximal and minimal values:
# Extract the minimum and maximum values for scaling
max_val = np.nanmax(m.arr)
min_val = np.nanmin(m.arr)
if min_val < 0:
ticks = [color_anchor[0], min_val, 0, max_val, color_anchor[1]]
# Otherwise - only set the maximal value:
else:
ticks = [color_anchor[0], max_val, color_anchor[1]]
# Plot line separating 'nothing' and 'prior' from rest of plot
l = ax.add_line(
mpl.lines.Line2D([m.shape[1] - 2.5, m.shape[1] - 2.5], [-0.5, m.shape[0] - 0.5], ls="--", c="gray", lw=2)
)
l.set_zorder(3)
# Display the actual values in the cells
if plot_vals:
for i in xrange(0, m.shape[0]):
for j in xrange(0, m.shape[1]):
val = m.arr[i, j]
if np.isnan(val):
continue
if val > 0.5:
ax.text(j - 0.2, i + 0.1, "%.2f" % val, color="w")
else:
ax.text(j - 0.2, i + 0.1, "%.2f" % val, color="k")
# Hide the black frame around the plot
# Doing ax.set_frame_on(False) results in weird thin lines
# from imshow() at the edges. Instead, we set the frame to white.
for spine in ax.spines.values():
spine.set_edgecolor("w")
# Set title
if title is not None:
ax.set_title(title)
# Plot the colorbar and remove its frame as well.
cb = fig.colorbar(im, cax=ax_cb, orientation="horizontal", cmap=cmap, ticks=ticks, format="%.2f")
cb.ax.artists.remove(cb.outline)
# Save figure
dirname = config.get_dataset_stats_dir(self)
suffix = "_second_order" if second_order else ""
filename = opjoin(dirname, "cooccur%s.png" % suffix)
fig.savefig(filename)
return fig
def plot_coocurrence(self,
cmap=plt.cm.Reds,
color_anchor=[0, 1],
x_tick_rot=90,
size=None,
title=None,
plot_vals=True,
second_order=False):
"""
Plot a heat map of conditional occurence, where cell (i,j) means
P(C_j|C_i). The last column in the K x (K+2) heat map corresponds
to the prior P(C_i).
If second_order, plots (K choose 2) x (K+2) heat map corresponding
to P(C_i|C_j,C_k): second-order correlations.
Return the figure.
"""
table = self.get_cls_ground_truth(with_diff=False, with_trun=True)
# This takes care of most of the difference between normal and second_order
# In the former case, a "combination" is just one class to condition on.
combinations = combination_strs = table.cols
if second_order:
combinations = [x for x in itertools.combinations(table.cols, 2)]
combination_strs = ['%s, %s' % (x[0], x[1]) for x in combinations]
total = table.shape[0]
N = len(table.cols)
K = len(combinations)
data = np.zeros(
(K, N + 2)) # extra columns are for P("nothing"|C) and P(C)
prior = np.zeros(K)
for i, combination in enumerate(combinations):
if second_order:
cls1 = combination[0]
cls2 = combination[1]
conditioned = table.filter_on_column(cls1).filter_on_column(
cls2)
else:
conditioned = table.filter_on_column(combination)
# count all the classes
data[i, :-2] = conditioned.sum()
# count the number of times that cls was the only one present to get
# P("nothing"|C)
if second_order:
data[i, -2] = ((conditioned.sum(1) - 2) == 0).sum()
else:
data[i, -2] = ((conditioned.sum(1) - 1) == 0).sum()
# normalize
max_val = np.max(data[i, :])
data[i, :] /= max_val
data[i, :][data[i, :] == 1] = np.nan
# use the max count to compute the prior
data[i, -1] = max_val / total
m = Table(data,
table.cols + ['nothing', 'prior'],
index=combination_strs)
# If second_order, sort by prior and remove rows with 0 prior
if second_order:
m = m.filter_on_column('prior',0.001,operator.gt).\
sort_by_column('prior',descending=True)
# TODO: just take the top K actually, for a side-by-side figure
m.arr = m.arr[:len(self.classes), :]
if size:
fig = plt.figure(figsize=size)
else:
w = max(12, m.shape[1])
h = max(12, m.shape[0])
fig = plt.figure(figsize=(w, h))
ax_im = fig.add_subplot(111)
# make axes for colorbar
divider = make_axes_locatable(ax_im)
ax_cb = divider.new_vertical(size="5%", pad=0.1, pack_start=True)
fig.add_axes(ax_cb)
#The call to imshow produces the matrix plot:
im = ax_im.imshow(m.arr,
origin='upper',
interpolation='nearest',
vmin=color_anchor[0],
vmax=color_anchor[1],
cmap=cmap)
#Formatting:
ax = ax_im
ax.set_xticks(np.arange(m.shape[1]))
ax.set_xticklabels(m.cols)
for tick in ax.xaxis.iter_ticks():
tick[0].label2On = True
tick[0].label1On = False
tick[0].label2.set_rotation(x_tick_rot)
tick[0].label2.set_fontsize('x-large')
ax.set_yticks(np.arange(m.shape[0]))
ax.set_yticklabels(m.index, size='x-large')
ax.yaxis.set_minor_locator(
matplotlib.ticker.FixedLocator(np.arange(-.5, m.shape[0] + 0.5)))
ax.xaxis.set_minor_locator(
matplotlib.ticker.FixedLocator(np.arange(-.5, m.shape[1] - 0.5)))
ax.grid(False, which='major')
ax.grid(True, which='minor', ls='-', lw=7, c='w')
# Make the major and minor tick marks invisible
for line in ax.xaxis.get_ticklines() + ax.yaxis.get_ticklines():
line.set_markeredgewidth(0)
for line in ax.xaxis.get_minorticklines(
) + ax.yaxis.get_minorticklines():
line.set_markeredgewidth(0)
# Limit the area of the plot
ax.set_ybound([-0.5, m.shape[0] - 0.5])
ax.set_xbound([-0.5, m.shape[1] - 0.5])
#The following produces the colorbar and sets the ticks
#Set the ticks - if 0 is in the interval of values, set that, as well
#as the maximal and minimal values:
#Extract the minimum and maximum values for scaling
max_val = np.nanmax(m.arr)
min_val = np.nanmin(m.arr)
if min_val < 0:
ticks = [color_anchor[0], min_val, 0, max_val, color_anchor[1]]
#Otherwise - only set the maximal value:
else:
ticks = [color_anchor[0], max_val, color_anchor[1]]
# Plot line separating 'nothing' and 'prior' from rest of plot
l = ax.add_line(
mpl.lines.Line2D([m.shape[1] - 2.5, m.shape[1] - 2.5],
[-.5, m.shape[0] - 0.5],
ls='--',
c='gray',
lw=2))
l.set_zorder(3)
# Display the actual values in the cells
if plot_vals:
for i in xrange(0, m.shape[0]):
for j in xrange(0, m.shape[1]):
val = m.arr[i, j]
if np.isnan(val):
continue
if val > 0.5:
ax.text(j - 0.2, i + 0.1, '%.2f' % val, color='w')
else:
ax.text(j - 0.2, i + 0.1, '%.2f' % val, color='k')
# Hide the black frame around the plot
# Doing ax.set_frame_on(False) results in weird thin lines
# from imshow() at the edges. Instead, we set the frame to white.
for spine in ax.spines.values():
spine.set_edgecolor('w')
# Set title
if title is not None:
ax.set_title(title)
# Plot the colorbar and remove its frame as well.
cb = fig.colorbar(im,
cax=ax_cb,
orientation='horizontal',
cmap=cmap,
ticks=ticks,
format='%.2f')
cb.ax.artists.remove(cb.outline)
# Save figure
dirname = config.get_dataset_stats_dir(self)
suffix = '_second_order' if second_order else ''
filename = opjoin(dirname, 'cooccur%s.png' % suffix)
fig.savefig(filename)
return fig
