class Clusterer():
# TODO HDBSCAN should be added
def __init__(self,
cluster_model,
n_clusters,
spectral_affinity='nearest_neighbors',
max_try_cnt=5):
self.cluster_model = cluster_model
self.n_clusters = n_clusters
self.max_try_cnt = max_try_cnt
self.trained_model = None
if self.cluster_model in ['KMeans', 'KMEANS', 'km', 'kmeans', 'KM']:
self.cluster_model = "KMeans"
if self.cluster_model in ['GMM', 'GMM_Full', 'GMM_full']:
self.cluster_model = "GMM_full"
if self.cluster_model in ['GMMdiag', 'GMM_Diag', 'GMM_diag']:
self.cluster_model = "GMM_diag"
if self.cluster_model in ['Spectral', 'SC', 'SpectralClustering']:
self.cluster_model = "Spectral"
self.spectral_affinity = spectral_affinity # 'nearest_neighbors', 'rbf'
if self.cluster_model not in [
"KMeans", "GMM_full", "GMM_diag", "Spectral"
]:
os_error(cluster_model + " is not applicable in this Class")
def fit(self,
X,
post_analyze_distribution=False,
verbose=1,
random_state=0):
df = pd_df(np_array(X))
curTol = 0.0001 if self.cluster_model == 'KMeans' else 0.01
max_iter = 300 if self.cluster_model == 'KMeans' else 200
numOf_1_sample_bins = 1
unique_clust_cnt = 1
expCnt = 0
while (unique_clust_cnt == 1 or
numOf_1_sample_bins - expCnt > 0) and expCnt < self.max_try_cnt:
t = time()
if expCnt > 0:
if numOf_1_sample_bins > 0:
print("running ", self.cluster_model, " for the ",
str(expCnt), " time due to numOf_1_sample_bins(",
str(numOf_1_sample_bins), ")")
if unique_clust_cnt == 1:
print("running ", self.cluster_model, " for the ",
str(expCnt), " time due to unique_clust_cnt==1")
if verbose > 0:
print('Clustering the featVec(', X.shape, ') with n_clusters(',
str(self.n_clusters),
') and model = ', self.cluster_model, ", curTol(",
str(curTol), "), max_iter(", str(max_iter), "), at ",
datetime.now().strftime("%H:%M:%S"))
self.kluster_centers = None
self.predictedKlusters = None
if self.cluster_model == 'KMeans':
# default vals for kmeans --> max_iter=300, 1e-4
self.trained_model = KMeans(init='k-means++',
n_clusters=self.n_clusters,
n_init=20,
tol=curTol,
max_iter=max_iter,
random_state=random_state).fit(df)
self.predictedKlusters = self.trained_model.labels_.astype(
float)
self.kluster_centers = self.trained_model.cluster_centers_.astype(
float)
elif self.cluster_model == 'GMM_full':
# default vals for gmm --> max_iter=100, 1e-3
self.trained_model = GaussianMixture(
n_components=self.n_clusters,
covariance_type='full',
tol=curTol,
random_state=random_state,
max_iter=max_iter,
reg_covar=1e-4).fit(df)
_, log_resp = self.trained_model._e_step(X)
self.predictedKlusters = log_resp.argmax(axis=1)
elif self.cluster_model == 'GMM_diag':
self.trained_model = GaussianMixture(
n_components=self.n_clusters,
covariance_type='diag',
tol=curTol,
random_state=random_state,
max_iter=max_iter,
reg_covar=1e-4).fit(df)
_, log_resp = self.trained_model._e_step(X)
self.predictedKlusters = log_resp.argmax(axis=1)
elif self.cluster_model == 'Spectral':
sc = SpectralClustering(n_clusters=self.n_clusters,
affinity=self.spectral_affinity,
random_state=random_state)
self.trained_model = sc.fit(X)
self.predictedKlusters = self.trained_model.labels_
self.kluster_centroids = get_cluster_centroids(
X,
self.predictedKlusters,
kluster_centers=self.kluster_centers,
verbose=0)
if post_analyze_distribution:
numOf_1_sample_bins, histSortedInv = analyzeClusterDistribution(
self.predictedKlusters, self.n_clusters, verbose=verbose)
unique_clust_cnt = len(np_unique(self.predictedKlusters))
curTol = curTol * 10
max_iter = max_iter + 50
expCnt = expCnt + 1
else:
expCnt = self.max_try_cnt
elapsed = time() - t
if verbose > 0:
print('Clustering done in (', getElapsedTimeFormatted(elapsed),
'), ended at ',
datetime.now().strftime("%H:%M:%S"))
removeLastLine()
if verbose > 0:
print('Clustering completed with (',
np_unique(self.predictedKlusters).shape,
') clusters, expCnt(', str(expCnt), ')')
# elif 'OPTICS' in clusterModel:
# N = featVec.shape[0]
# min_cluster_size = int(np.ceil(N / (n_clusters * 4)))
# pars = clusterModel.split('_') # 'OPTICS_hamming_dbscan', 'OPTICS_russellrao_xi'
# # metricsAvail = np.sort(['braycurtis', 'canberra', 'chebyshev', 'correlation', 'dice', 'hamming', 'jaccard', 'kulsinski',
# # 'mahalanobis', 'minkowski', 'rogerstanimoto', 'russellrao', 'seuclidean', 'sokalmichener',
# # 'sokalsneath', 'sqeuclidean', 'yule',
# # 'cityblock', 'cosine', 'euclidean', 'l1', 'l2', 'manhattan'])
# # cluster_methods_avail = ['xi', 'dbscan']
# clust = ClusterOPT(min_samples=50, xi=.05, min_cluster_size=min_cluster_size, metric=pars[1], cluster_method=pars[2])
# clust.fit(featVec)
# predictedKlusters = cluster_optics_dbscan(reachability=clust.reachability_,
# core_distances=clust.core_distances_,
# ordering=clust.ordering_, eps=0.5)
# n1 = np.unique(predictedKlusters)
# print(clusterModel, ' found ', str(n1), ' uniq clusters')
# predictedKlusters = predictedKlusters + 1
return self
def fit_predict(self, X, post_analyze_distribution=False, verbose=1):
self.fit(X,
post_analyze_distribution=post_analyze_distribution,
verbose=verbose)
return np_asarray(self.predictedKlusters,
dtype=int), self.kluster_centroids
def predict(self, X, post_analyze_distribution=False, verbose=1):
df = pd_df(X)
print("started prediction for ", self.cluster_model, " X(", X.shape,
")")
if self.cluster_model == 'KMeans':
# default vals for kmeans --> max_iter=300, 1e-4
self.predictedKlusters = self.trained_model.predict(df).astype(
float)
self.kluster_centers = self.trained_model.cluster_centers_.astype(
float)
elif self.cluster_model == 'GMM_full':
# default vals for gmm --> max_iter=100, 1e-3
_, log_resp = self.trained_model._e_step(df)
self.predictedKlusters = log_resp.argmax(axis=1)
elif self.cluster_model == 'GMM_diag':
_, log_resp = self.trained_model._e_step(df)
self.predictedKlusters = log_resp.argmax(axis=1)
elif self.cluster_model == 'Spectral':
self.predictedKlusters = self.trained_model.predict(X).labels_
self.kluster_centroids = get_cluster_centroids(
X,
self.predictedKlusters,
kluster_centers=self.kluster_centers,
verbose=0)
if post_analyze_distribution:
numOf_1_sample_bins, histSortedInv = analyzeClusterDistribution(
self.predictedKlusters, self.n_clusters, verbose=1)
unique_clust_cnt = len(np_unique(self.predictedKlusters))
print("prediction completed for ",
self.cluster_model, " - unique_clust_cnt(",
str(unique_clust_cnt), "), numOf_1_sample_bins(",
str(numOf_1_sample_bins), ")")
return np_asarray(self.predictedKlusters,
dtype=int), self.kluster_centroids
np.savetxt('../result/arc_roc.txt', np.column_stack((tpr, fpr)), fmt='%f')
plt.figure()
plt.plot(fpr,
tpr,
color='b',
label='ROC (AUC = %0.2f)' % auc,
lw=2,
alpha=.8)
plt.xlabel('False Positive Rate')
plt.ylabel('True Positive Rate')
plt.show()
# online learning
for n_iter in range(100):
log_prob_norm, log_resp = clf._e_step(test_data[label == 0, :-1])
clf._m_step(test_data[label == 0, :-1], log_resp)
scores = clf.score_samples(test_data[:, :-1])
scores = (scores - min(scores)) / (max(scores) - min(scores))
np.savetxt('../result/scores.txt', scores, fmt='%.6f')
exit(0)
thresholds = np.arange(scores.min(), scores.max(), 0.001)
tpr, fpr, F1, accuracy, best_threshold_index = calc_roc(
thresholds, scores, np.logical_not(label))
auc = metrics.auc(fpr, tpr)
print('auc: %s' % auc)
print('accuracy: %s' % np.mean(accuracy))
print('best_threshold: %s' % thresholds[best_threshold_index])
