def plot_results_many_gen(infile, outfiles):
df = pickle.load(open(infile, 'r'))
df['ari'] = df.apply(lambda row: rand.compute_adj_rand_index(row['true_assign'],
irm.util.canonicalize_assignment(row['assign'])), axis=1)
df['empirical_class_n'] = df.apply(lambda row : len(np.unique(row['assign'])), axis=1)
df_cc = df[(df['dataset_name'] == 'class_compare_gen') & (df['model'] == 'ld')]
a = df_cc.groupby(['dataset_name', 'jitter', 'nonzero_frac', 'class_n',
'side_n', 'seed', 'truth']).apply(lambda group: group.sort_index(by='score', ascending=False).head(1))
colors = {'distblock' : 'b',
'mixedblock' : 'r',
'bumpblock' : 'g'}
f = pylab.figure(figsize=(4, 3))
ax = f.add_subplot(1, 1, 1)
for g_i, (g_idx, g) in enumerate(a.groupby(['truth'])):
ax.scatter(g.index.get_level_values('class_n') + 0.3*g_i,
g['empirical_class_n'], c=colors[g_idx],
edgecolor='none')
ax.plot([1, 16], [1, 16], c='k')
ax.set_xlabel("true class number")
ax.set_ylabel("estimated class number")
ax.set_xticks([1, 2, 4, 8, 16])
f.tight_layout()
f.savefig(outfiles[0])
colors = {'distblock' : 'b',
'mixedblock' : 'r',
'bumpblock' : 'g'}
offsets = {'distblock' : 0.0,
'mixedblock' : 1.0,
'bumpblock' : 2.0}
f = pylab.figure(figsize=(4, 3))
ax = f.add_subplot(1, 1, 1)
CLASS_SPACE = 3.5
WIDTH = 0.8
N = 0
for g_idx, g in a.groupby(['truth']):
h = g.groupby(['class_n']).mean()
herr = g.groupby(['class_n']).std()
N= len(h)
print "g_idx", g_idx, h['ari']
ax.bar(np.arange(N)*CLASS_SPACE + offsets[g_idx], h['ari'], width=WIDTH,
color=colors[g_idx])
ax.errorbar(np.arange(N)*CLASS_SPACE + offsets[g_idx] + WIDTH/2,
h['ari'], yerr= herr['ari'],
capsize=0,elinewidth=4,ecolor='k', linewidth=0)
ax.set_xlabel("true class number")
ax.set_ylabel("adjusted rand index")
ax.set_ylim(0, 1.0)
ax.set_xticks(np.arange(N)*CLASS_SPACE + 1)
ax.set_xticklabels([1, 2, 4, 8, 16])
f.tight_layout()
f.savefig(outfiles[1])
def plot_results_many_gen(infile, outfiles):
df = pickle.load(open(infile, 'r'))
df['ari'] = df.apply(lambda row: rand.compute_adj_rand_index(
row['true_assign'], irm.util.canonicalize_assignment(row['assign'])),
axis=1)
df['empirical_class_n'] = df.apply(
lambda row: len(np.unique(row['assign'])), axis=1)
df_cc = df[(df['dataset_name'] == 'class_compare_gen')
& (df['model'] == 'ld')]
a = df_cc.groupby([
'dataset_name', 'jitter', 'nonzero_frac', 'class_n', 'side_n', 'seed',
'truth'
]).apply(
lambda group: group.sort_index(by='score', ascending=False).head(1))
colors = {'distblock': 'b', 'mixedblock': 'r', 'bumpblock': 'g'}
f = pylab.figure(figsize=(4, 3))
ax = f.add_subplot(1, 1, 1)
for g_i, (g_idx, g) in enumerate(a.groupby(['truth'])):
ax.scatter(g.index.get_level_values('class_n') + 0.3 * g_i,
g['empirical_class_n'],
c=colors[g_idx],
edgecolor='none')
ax.plot([1, 16], [1, 16], c='k')
ax.set_xlabel("true class number")
ax.set_ylabel("estimated class number")
ax.set_xticks([1, 2, 4, 8, 16])
f.tight_layout()
f.savefig(outfiles[0])
colors = {'distblock': 'b', 'mixedblock': 'r', 'bumpblock': 'g'}
offsets = {'distblock': 0.0, 'mixedblock': 1.0, 'bumpblock': 2.0}
f = pylab.figure(figsize=(4, 3))
ax = f.add_subplot(1, 1, 1)
CLASS_SPACE = 3.5
WIDTH = 0.8
N = 0
for g_idx, g in a.groupby(['truth']):
h = g.groupby(['class_n']).mean()
herr = g.groupby(['class_n']).std()
N = len(h)
print "g_idx", g_idx, h['ari']
ax.bar(np.arange(N) * CLASS_SPACE + offsets[g_idx],
h['ari'],
width=WIDTH,
color=colors[g_idx])
ax.errorbar(np.arange(N) * CLASS_SPACE + offsets[g_idx] + WIDTH / 2,
h['ari'],
yerr=herr['ari'],
capsize=0,
elinewidth=4,
ecolor='k',
linewidth=0)
ax.set_xlabel("true class number")
ax.set_ylabel("adjusted rand index")
ax.set_ylim(0, 1.0)
ax.set_xticks(np.arange(N) * CLASS_SPACE + 1)
ax.set_xticklabels([1, 2, 4, 8, 16])
f.tight_layout()
f.savefig(outfiles[1])
def plot_results(infile, outfiles):
df = pickle.load(open(infile, 'r'))
df['ari'] = df.apply(lambda row: rand.compute_adj_rand_index(row['true_assign'],
irm.util.canonicalize_assignment(row['assign'])), axis=1)
df['empirical_class_n'] = df.apply(lambda row : len(np.unique(row['assign'])), axis=1)
for plot_files, dataset_name in zip(outfiles, PLOT_DATASETS):
df_cc = df[df['dataset_name'] == dataset_name]
a = df_cc.groupby(['dataset_name', 'jitter', 'model', 'nonzero_frac', 'class_n',
'side_n', 'seed', 'truth']).apply(lambda group: group.sort_index(by='score', ascending=False).head(1))
colors = {'bb' : 'b',
'ld' : 'r'}
f = pylab.figure(figsize=(4, 3))
ax = f.add_subplot(1, 1, 1)
labels = {'bb' : "conn only",
'ld' : "conn + dist"}
for g_idx, g in a.groupby(['model']):
ax.scatter(g.index.get_level_values('class_n'), g['empirical_class_n'], c=colors[g_idx],
edgecolor='none', label= labels[g_idx])
ax.plot([1, 16], [1, 16], c='k', label="ground truth")
ax.set_xlabel("true type number")
ax.set_ylabel("estimated type number")
ax.set_xticks([1, 2, 4, 8, 16])
ax.legend(loc="upper left", fontsize=10)
ax.set_yticks([0, 70])
ax.set_ylim([-2, 70])
for tic in ax.yaxis.get_major_ticks():
tic.tick1On = tic.tick2On = False
f.tight_layout()
for tic in ax.xaxis.get_major_ticks():
tic.tick1On = tic.tick2On = False
spines_to_remove = ['top', 'right']
for spine in spines_to_remove:
ax.spines[spine].set_visible(False)
f.savefig(plot_files[0])
colors = {'bb' : 'b',
'ld' : 'r'}
offsets = {'bb' : 0.0,
'ld' : 1.0}
f = pylab.figure(figsize=(4, 3))
ax = f.add_subplot(1, 1, 1)
CLASS_SPACE = 2.5
WIDTH = 0.8
N = 0
for g_idx, g in a.groupby(['model']):
h = g.groupby(['class_n']).mean()
herr = g.groupby(['class_n']).std()
N= len(h)
ax.bar(np.arange(N)*CLASS_SPACE + offsets[g_idx], h['ari'], width=WIDTH,
color=colors[g_idx])
ax.errorbar(np.arange(N)*CLASS_SPACE + offsets[g_idx] + WIDTH/2,
h['ari'], yerr= herr['ari'], capsize=0,elinewidth=2, linewidth=0, ecolor='black')
#ax.plot([1, 16], [1, 1], c='k')
ax.set_xlabel("true type number")
ax.set_ylabel("Cluster accuracy (ARI)")
ax.set_ylim(0, 1.0)
ax.set_yticks([0.0, 1.0])
ax.set_xticks(np.arange(N)*CLASS_SPACE + 1)
ax.set_xticklabels([1, 2, 4, 8, 16])
for tic in ax.xaxis.get_major_ticks():
tic.tick1On = tic.tick2On = False
spines_to_remove = ['top', 'right']
for spine in spines_to_remove:
ax.spines[spine].set_visible(False)
f.tight_layout()
f.savefig(plot_files[1])
## The future
def cluster_var(row):
assign = row['assign']
true_assign = row['true_assign']
print type(row['node_pos'])
node_pos = row['node_pos']
def node_to_df(assign, nodes):
return pandas.DataFrame({'cluster' : assign, 'x' : node_pos[:, 0],
'y' : node_pos[:, 1], 'z' : node_pos[:, 2]})
rdf1 = node_to_df(assign, node_pos)
rdf1['truth'] = False
rdf2 = node_to_df(true_assign, node_pos)
rdf2['truth'] = True
rdf = pandas.concat([rdf1, rdf2])
rdf[['x', 'y', 'z']] = rdf[['x', 'y', 'z']].astype(float)
return rdf.groupby(['truth', 'cluster']).var()
#a = df_cc.groupby(['dataset_name', 'jitter', 'model', 'nonzero_frac', 'class_n',
# 'side_n', 'seed', 'truth']).apply(lambda group: group.sort_index(by='score', ascending=False).head(1))
df_vars = []
for rid, r in a.iterrows():
cv = cluster_var(r.to_dict())
cv['model'] = r['model']
cv['seed'] = r['seed']
cv['class_n'] = r['class_n']
df_vars.append(cv)
df_vars = pandas.concat(df_vars)
df_vars['truth'] = df_vars.index.get_level_values('truth')
df_vars['std'] = np.sqrt(df_vars['x'] + df_vars['y'])
f = pylab.figure(figsize=(4.0, 6.5))
bins = np.linspace(0, 3.5, 20)
bin_width = (bins[1] - bins[0])
bar_width = bin_width/4.0
bar_space = bin_width/3.
for i, class_n in enumerate([4, 8, 16]):
ax = f.add_subplot(3, 1, i + 1)
for model_i, (model, color) in enumerate([('bb', 'b'),
('ld', 'r'),]):
df2 = df_vars[(df_vars['model'] == model) & (df_vars['class_n']==class_n) & (df_vars['truth']==False)]
hist, _ = np.histogram(df2.dropna()['std'], bins=bins, density=True)
ax.bar(bins[:-1] + model_i * bar_space, hist*bin_width, width=bar_width, color=color,
label=model, linewidth=0.0)
df2 = df_vars[(df_vars['model'] == model) & (df_vars['class_n']==class_n) & (df_vars['truth']==True)]
hist, _ = np.histogram(df2.dropna()['std'], bins=bins, density=True)
print "Histogram=", hist
ax.bar(bins[:-1] + 2*bar_space, hist*bin_width,
width=bar_width, color='k',
label='truth', linewidth=0.0)
ax.set_yticks([0.0, 1.0])
ax.set_ylim(0.0, 1.05)
ax.set_ylabel("frac (class=%d)" % class_n)
ax.set_xticks([0.0, 3.5])
if i == 0:
handles, labels = ax.get_legend_handles_labels()
ax.legend(handles, [
'conn only',
'conn + dist',
'Ground Truth',
],
loc='upper left',
fontsize=12)
if i < 2:
ax.set_xticklabels([])
for tic in ax.xaxis.get_major_ticks():
tic.tick1On = tic.tick2On = False
for tic in ax.yaxis.get_major_ticks():
tic.tick1On = tic.tick2On = False
spines_to_remove = ['top', 'right']
for spine in spines_to_remove:
ax.spines[spine].set_visible(False)
ax.set_xlabel("size of clusters (2D std dev)")
f.tight_layout()
f.savefig(plot_files[2])
def plot_results(infile, outfiles):
df = pickle.load(open(infile, 'r'))
df['ari'] = df.apply(lambda row: rand.compute_adj_rand_index(
row['true_assign'], irm.util.canonicalize_assignment(row['assign'])),
axis=1)
df['empirical_class_n'] = df.apply(
lambda row: len(np.unique(row['assign'])), axis=1)
for plot_files, dataset_name in zip(outfiles, PLOT_DATASETS):
df_cc = df[df['dataset_name'] == dataset_name]
a = df_cc.groupby([
'dataset_name', 'jitter', 'model', 'nonzero_frac', 'class_n',
'side_n', 'seed', 'truth'
]).apply(lambda group: group.sort_index(by='score', ascending=False).
head(1))
colors = {'bb': 'b', 'ld': 'r'}
f = pylab.figure(figsize=(4, 3))
ax = f.add_subplot(1, 1, 1)
labels = {'bb': "conn only", 'ld': "conn + dist"}
for g_idx, g in a.groupby(['model']):
ax.scatter(g.index.get_level_values('class_n'),
g['empirical_class_n'],
c=colors[g_idx],
edgecolor='none',
label=labels[g_idx])
ax.plot([1, 16], [1, 16], c='k', label="ground truth")
ax.set_xlabel("true type number")
ax.set_ylabel("estimated type number")
ax.set_xticks([1, 2, 4, 8, 16])
ax.legend(loc="upper left", fontsize=10)
ax.set_yticks([0, 70])
ax.set_ylim([-2, 70])
for tic in ax.yaxis.get_major_ticks():
tic.tick1On = tic.tick2On = False
f.tight_layout()
for tic in ax.xaxis.get_major_ticks():
tic.tick1On = tic.tick2On = False
spines_to_remove = ['top', 'right']
for spine in spines_to_remove:
ax.spines[spine].set_visible(False)
f.savefig(plot_files[0])
colors = {'bb': 'b', 'ld': 'r'}
offsets = {'bb': 0.0, 'ld': 1.0}
f = pylab.figure(figsize=(4, 3))
ax = f.add_subplot(1, 1, 1)
CLASS_SPACE = 2.5
WIDTH = 0.8
N = 0
for g_idx, g in a.groupby(['model']):
h = g.groupby(['class_n']).mean()
herr = g.groupby(['class_n']).std()
N = len(h)
ax.bar(np.arange(N) * CLASS_SPACE + offsets[g_idx],
h['ari'],
width=WIDTH,
color=colors[g_idx])
ax.errorbar(np.arange(N) * CLASS_SPACE + offsets[g_idx] +
WIDTH / 2,
h['ari'],
yerr=herr['ari'],
capsize=0,
elinewidth=2,
linewidth=0,
ecolor='black')
#ax.plot([1, 16], [1, 1], c='k')
ax.set_xlabel("true type number")
ax.set_ylabel("Cluster accuracy (ARI)")
ax.set_ylim(0, 1.0)
ax.set_yticks([0.0, 1.0])
ax.set_xticks(np.arange(N) * CLASS_SPACE + 1)
ax.set_xticklabels([1, 2, 4, 8, 16])
for tic in ax.xaxis.get_major_ticks():
tic.tick1On = tic.tick2On = False
spines_to_remove = ['top', 'right']
for spine in spines_to_remove:
ax.spines[spine].set_visible(False)
f.tight_layout()
f.savefig(plot_files[1])
## The future
def cluster_var(row):
assign = row['assign']
true_assign = row['true_assign']
print type(row['node_pos'])
node_pos = row['node_pos']
def node_to_df(assign, nodes):
return pandas.DataFrame({
'cluster': assign,
'x': node_pos[:, 0],
'y': node_pos[:, 1],
'z': node_pos[:, 2]
})
rdf1 = node_to_df(assign, node_pos)
rdf1['truth'] = False
rdf2 = node_to_df(true_assign, node_pos)
rdf2['truth'] = True
rdf = pandas.concat([rdf1, rdf2])
rdf[['x', 'y', 'z']] = rdf[['x', 'y', 'z']].astype(float)
return rdf.groupby(['truth', 'cluster']).var()
#a = df_cc.groupby(['dataset_name', 'jitter', 'model', 'nonzero_frac', 'class_n',
# 'side_n', 'seed', 'truth']).apply(lambda group: group.sort_index(by='score', ascending=False).head(1))
df_vars = []
for rid, r in a.iterrows():
cv = cluster_var(r.to_dict())
cv['model'] = r['model']
cv['seed'] = r['seed']
cv['class_n'] = r['class_n']
df_vars.append(cv)
df_vars = pandas.concat(df_vars)
df_vars['truth'] = df_vars.index.get_level_values('truth')
df_vars['std'] = np.sqrt(df_vars['x'] + df_vars['y'])
f = pylab.figure(figsize=(4.0, 6.5))
bins = np.linspace(0, 3.5, 20)
bin_width = (bins[1] - bins[0])
bar_width = bin_width / 4.0
bar_space = bin_width / 3.
for i, class_n in enumerate([4, 8, 16]):
ax = f.add_subplot(3, 1, i + 1)
for model_i, (model, color) in enumerate([
('bb', 'b'),
('ld', 'r'),
]):
df2 = df_vars[(df_vars['model'] == model)
& (df_vars['class_n'] == class_n) &
(df_vars['truth'] == False)]
hist, _ = np.histogram(df2.dropna()['std'],
bins=bins,
density=True)
ax.bar(bins[:-1] + model_i * bar_space,
hist * bin_width,
width=bar_width,
color=color,
label=model,
linewidth=0.0)
df2 = df_vars[(df_vars['model'] == model)
& (df_vars['class_n'] == class_n) &
(df_vars['truth'] == True)]
hist, _ = np.histogram(df2.dropna()['std'],
bins=bins,
density=True)
print "Histogram=", hist
ax.bar(bins[:-1] + 2 * bar_space,
hist * bin_width,
width=bar_width,
color='k',
label='truth',
linewidth=0.0)
ax.set_yticks([0.0, 1.0])
ax.set_ylim(0.0, 1.05)
ax.set_ylabel("frac (class=%d)" % class_n)
ax.set_xticks([0.0, 3.5])
if i == 0:
handles, labels = ax.get_legend_handles_labels()
ax.legend(handles, [
'conn only',
'conn + dist',
'Ground Truth',
],
loc='upper left',
fontsize=12)
if i < 2:
ax.set_xticklabels([])
for tic in ax.xaxis.get_major_ticks():
tic.tick1On = tic.tick2On = False
for tic in ax.yaxis.get_major_ticks():
tic.tick1On = tic.tick2On = False
spines_to_remove = ['top', 'right']
for spine in spines_to_remove:
ax.spines[spine].set_visible(False)
ax.set_xlabel("size of clusters (2D std dev)")
f.tight_layout()
f.savefig(plot_files[2])
