def run_silhouette_metrics(adata,
batch_key='orig.ident',
cell_label='paper.cell.type',
embed='X_pca'):
# global silhouette coefficient
sil_global = me.silhouette(adata,
group_key='paper.cell.type',
embed=embed,
metric='euclidean')
# silhouette coefficient per batch
_, sil_clus = me.silhouette_batch(adata,
batch_key=batch_key,
group_key=cell_label,
embed=embed,
metric='euclidean',
verbose=False)
sil_clus = sil_clus['silhouette_score'].mean()
il_score_sil = me.isolated_labels(adata,
label_key=cell_label,
batch_key=batch_key,
cluster=False,
n=1,
verbose=False)
il_score_clus = me.isolated_labels(adata,
label_key=cell_label,
batch_key=batch_key,
cluster=True,
n=1,
verbose=False)
return (sil_global, sil_clus, il_score_clus, il_score_sil)
def cost_func(args):
partial_ff = partial(ff, s=args[0])
res = pool.map(partial_ff, cnt)
si = metrics.silhouette(scale(np.array(res)), np.array(Y) - 1)
db = metrics.db(scale(np.array(res)), np.array(Y) - 1)
ch = metrics.ch(scale(np.array(res)), np.array(Y) - 1)
w1, w2, w3 = float(args[1]), float(args[2]), float(args[3])
fit = (np.std(si) + np.mean(1. - si)) * db / (ch + 1e-12)
return (np.mean(si), db, ch, fit)
# casestudy_iris_pca.py
import data_iris
import hierarchical
import matplotlib.pyplot as plt
import metrics
import numpy as np
import pca
import plot_data
# (1) load data
iris = data_iris.iris()
X,class_label = iris.load()
# perform pca and reduce dimension to 2
model_pca = pca.pca()
model_pca.fit(X)
R = model_pca.data_reduced_dimension(reduced_dim=2)
plot_data.plot_scatter_class(R,class_label,"Iris Data Projected to 2 Dimensions using PCA","u0","u1")
# (2) create model
model = hierarchical.hierarchical()
# (3) fit model
model.fit(R)
print("Time fit: {}".format(model.time_fit))
# (4) results
level = -3
print("Purity: {}".format(metrics.purity(model.clustersave[level],class_label)))
print("Davies-Bouldin: {}".format(metrics.davies_bouldin(R,model.clustersave[level])))
print("Silhouette: {}".format(metrics.silhouette(R,model.clustersave[level])))
model.plot_cluster(nlevel=level,title="Hierarchical Clustering for Iris Dataset reduced to 2d",xlabel="u0",ylabel="u1")
metrics.plot_cluster_distribution(model.clustersave[level],class_label)
plt.show()
#!/usr/bin/python
import sys
import descritores
import pylab
import cPickle
import metrics
db = cPickle.load(open(sys.argv[1] + "/classes.txt"))
#names = cPickle.load(open(sys.argv[1]+"/names.pkl"))
names = db.keys()
scales = pylab.loadtxt(sys.argv[2])
X = [
pylab.vstack(([db[f] for f in names],
pylab.array([
pylab.log(descritores.bendenergy(sys.argv[1] + f, s)())
for f in names
]).T)).T for s in scales
]
for x in X:
s = metrics.silhouette(x[:, 1:], x[:, 0].astype(int) - 1)
print pylab.mean(s), pylab.std(s)
for n, x in zip(['nmbe_pso.pkl', 'nmbe_de.pkl', 'nmbe_sa.pkl'], X):
with open(n, "wb") as f:
cPickle.dump(dict(zip(names, x)), f)
mod = kmeans.kmeans(ncluster=2, initialization='kmeans++')
elif model == "GaussianMM":
mod = gaussianmm.gaussianmm(ncluster=2,
initialization='kmeans++')
elif model == "DBSCAN":
mod = dbscan.dbscan(minpts=5, epsilon=0.18)
# fit model
print("Model: {}".format(model))
if model == "DBSCAN":
mod.fit(X[dataset])
else:
mod.fit(X[dataset], 100, 1e-5, False)
print("Time fit: {}".format(mod.time_fit))
# davies-bouldin and silhouette
db = metrics.davies_bouldin(X[dataset], mod.clustersave[-1])
s = metrics.silhouette(X[dataset], mod.clustersave[-1])
print("Davies-Bouldin: {}".format(db))
print("Silhouette: {}".format(s))
colors = (mod.clustersave[-1] + 1) / mod.ncluster
axes[i, j].scatter(X[dataset][0, :],
X[dataset][1, :],
color=cm.jet(colors),
s=15)
axes[i, j].set_xticklabels([])
axes[i, j].set_yticklabels([])
if i == 0:
title = model + "\ndb:{:.2f} s:{:.2f} t:{:.3f}".format(
db, s, mod.time_fit)
else:
title = "db: {:.2f} s:{:.2f} t:{:.3f}".format(db, s, mod.time_fit)
axes[i, j].set_title(title)
display_clusters(y, "Настоящие метки")
def display_metrics(n_clusters, metrics, title):
plt.figure(figsize=(8, 6))
plt.grid(linestyle='--')
plt.plot(n_clusters, metrics, linestyle='-', marker='.', color='r')
plt.title(title)
plt.xlabel("Количество кластеров")
plt.ylabel("Значение метрики")
plt.show()
external_metrics = []
internal_metrics = []
for i in range(1, 11):
kMean = KMeans(k=i)
centroids = kMean.fit(X_norm)
y_pred = kMean.predict(X_norm)
if i == 1:
internal_metrics.append(0.0)
else:
internal_metrics.append(silhouette(X_norm, y_pred, centroids))
external_metrics.append(adjusted_rand_index(y, y_pred))
display_clusters(y_pred, str(i) + ' кластеров')
display_metrics(range(1, 11), external_metrics, 'Внешняя метрика')
display_metrics(range(1, 11), internal_metrics, 'Внутренняя метрика')
def cost_func(args): partial_ff = partial(ff, s = args) res = pool.map(partial_ff,cnt) s = metrics.silhouette(scale(np.array(res)),np.array(Y)-1) return np.median(np.abs(1.-s))
#!/usr/bin/python import sys import descritores import pylab import cPickle import metrics db = cPickle.load(open(sys.argv[1]+"/classes.txt")) #names = cPickle.load(open(sys.argv[1]+"/names.pkl")) names = db.keys() scales = pylab.loadtxt(sys.argv[2]) X = [pylab.vstack(([db[f] for f in names],pylab.array([pylab.log(descritores.bendenergy(sys.argv[1]+f,s)()) for f in names]).T)).T for s in scales] for x in X: s = metrics.silhouette(x[:,1:],x[:,0].astype(int)-1) print pylab.mean(s),pylab.std(s) for n,x in zip(['nmbe_pso.pkl','nmbe_de.pkl','nmbe_sa.pkl'],X): with open(n,"wb") as f: cPickle.dump(dict(zip(names,x)),f)
#!/usr/bin/python
import cPickle
import sys
import metrics
import pylab
color = ["b","r","g"]
db = cPickle.load(open(sys.argv[1]))
X,Y = db[:,1:],db[:,0].astype(int)
print "CE = {0}".format(metrics.CE(X,Y-1))
print "PC = {0}".format(metrics.PC(X,Y-1))
print "CS = {0}".format(metrics.CS(X,Y-1))
print "DB = {0}".format(metrics.db(X,Y-1))
print "DI = {0}".format(metrics.di(X,Y))
print "Silhouette = {0}".format(metrics.silhouette(X,Y-1).mean())
for c,i in zip(color,range(1,4)):
pylab.plot(X[pylab.where(Y == i)][:,0],X[pylab.where(Y == i)][:,1],"o"+c)
pylab.show()
