def plot_GPLVM_data_cluster(results_dir, n_clusters=None, VB=False):
# Load relevant datasets
data_array = np.genfromtxt(os.path.join(results_dir, 'summary.csv'), delimiter=',')
X = (np.genfromtxt(os.path.join(results_dir, 'GPLVM-datasets-2.csv'), delimiter=','))
datasets = [line.rstrip('\n') for line in open(os.path.join(results_dir, 'datasets.csv'), 'r').readlines()]
methods = [line.rstrip('\n') for line in open(os.path.join(results_dir, 'methods.csv'), 'r').readlines()]
# Fit a mixture model
if n_clusters is None:
m = DPGMM()
elif VB:
m = VBGMM(alpha = 10, n_components=n_clusters)
else:
m = GMM(n_components=n_clusters, n_init=100)
m.fit(data_array.T)
clusters = m.predict(data_array.T)
# Plot
#clf()
figure(1)
pretty_scatter(X[:,0], X[:,1], clusters, 200*np.ones(X[:,0].shape), datasets)
xlabel('Dimension 1')
ylabel('Dimension 2')
if n_clusters is None:
title('CRP MoG')
elif VB:
title('%d clusters with VB' % n_clusters)
else:
title('%d clusters with EM' % n_clusters)
show()
def _fit_dpgmm(self, x):
# clustering
k = max(self.crange)
for r in xrange(self.repeats):
# info
if self.debug is True:
print '\t[%s][c:%d][r:%d]' % (self.clus_type, k, r + 1),
# fit and evaluate model
model_kwargs = {}
if 'alpha' in self.clus_kwargs:
model_kwargs.update(alpha=self.clus_kwargs['alpha'])
if 'conv_thresh' in self.clus_kwargs:
model_kwargs.update(thresh=self.clus_kwargs['conv_thresh'])
if 'max_iter' in self.clus_kwargs:
model_kwargs.update(n_iter=self.clus_kwargs['max_iter'])
model = DPGMM(n_components=k,
covariance_type=self.cvtype,
**model_kwargs)
model.fit(x)
self._labels[r] = model.predict(x)
self._parameters[r] = model.means_
self._ll[r] = model.score(x).sum()
# evaluate goodness of fit for this run
#self._gof[r] = self.gof(x, self._ll[r], k)
if self.gof_type == 'aic':
self._gof[r] = model.aic(x)
if self.gof_type == 'bic':
self._gof[r] = model.bic(x)
# debug
if self.debug is True:
print self._gof[r], model.n_components, model.weights_.shape[0]
def try_covar(type_str, x_words):
clf = DPGMM(n_components=20, covariance_type=type_str, alpha=30, n_iter=1000)
clf.fit(x_data)
y_ = clf.predict(x_data)
print type_str
print_centers(x_words, y_, clf)
print
def _fit_dpgmm(self, x):
# clustering
k = max(self.crange)
for r in xrange(self.repeats):
# info
if self.debug is True:
print '\t[%s][c:%d][r:%d]' % (self.clus_type, k, r + 1),
# fit and evaluate model
model_kwargs = {}
if 'alpha' in self.clus_kwargs:
model_kwargs.update(alpha=self.clus_kwargs['alpha'])
if 'conv_thresh' in self.clus_kwargs:
model_kwargs.update(thresh=self.clus_kwargs['conv_thresh'])
if 'max_iter' in self.clus_kwargs:
model_kwargs.update(n_iter=self.clus_kwargs['max_iter'])
model = DPGMM(n_components=k, covariance_type=self.cvtype,
**model_kwargs)
model.fit(x)
self._labels[r] = model.predict(x)
self._parameters[r] = model.means_
self._ll[r] = model.score(x).sum()
# evaluate goodness of fit for this run
#self._gof[r] = self.gof(x, self._ll[r], k)
if self.gof_type == 'aic':
self._gof[r] = model.aic(x)
if self.gof_type == 'bic':
self._gof[r] = model.bic(x)
# debug
if self.debug is True:
print self._gof[r], model.n_components, model.weights_.shape[0]
def Dirichlet(cluster_data, identification, iteration_number = 1):
print "In Dirichlet"
for i in range(0,iteration_number):
print "On iteration number ", i
dirichlet = DPGMM(n_components = len(cluster_data)).fit(cluster_data)
#paremeters= dirichlet.get_params #returns parameters of the algorithm as a whole from the fit
predict = dirichlet.predict(cluster_data)
n_clusters_ = len(set(predict)) - (1 if -1 in predict else 0)
print('Estimated number of clusters with Dirichlet: %d' % n_clusters_)
return _make_final_list(identification, predict)
def Dirichlet(cluster_data, identification, iteration_number=1):
print "In Dirichlet"
for i in range(0, iteration_number):
print "On iteration number ", i
dirichlet = DPGMM(n_components=len(cluster_data)).fit(cluster_data)
#paremeters= dirichlet.get_params #returns parameters of the algorithm as a whole from the fit
predict = dirichlet.predict(cluster_data)
n_clusters_ = len(set(predict)) - (1 if -1 in predict else 0)
print('Estimated number of clusters with Dirichlet: %d' % n_clusters_)
return _make_final_list(identification, predict)
def get_best_dpgmm(X, num_c, cv_type, alpha, iters, n_init, rand_state=None):
best_bic = np.inf
bic_dpgmm = None
lbl_vec_dpgmm = np.zeros(X.shape[0])
prob_vec_dpgmm = np.zeros(X.shape[0])
log_prob_dpgmm = None
for i in xrange(n_init):
dpgmm = DPGMM(n_components=num_c, covariance_type=cv_type, \
alpha=alpha, random_state=rand_state)
dpgmm.fit(X)
b = dpgmm.bic(X)
if b < best_bic:
bic_dpgmm = b
lbl_vec = dpgmm.predict(X)
prob_vec = dpgmm.predict_proba(X)
log_prob_dpgmm = np.sum(dpgmm.score(X))
return [lbl_vec, prob_vec, bic_dpgmm, log_prob_dpgmm]
def _Dirichlet(cluster_data, identification):
print "In Dirichlet"
for i in range(0, 3):
print "i is ", i
dirichlet = DPGMM(n_components=len(cluster_data)).fit(cluster_data)
#paremeters= dirichlet.get_params #returns parameters of the algorithm as a whole from the fit
predict = dirichlet.predict(cluster_data)
n_clusters_ = len(set(predict)) - (1 if -1 in predict else 0)
print('Estimated number of clusters with Dirichlet: %d' % n_clusters_)
final = []
for x in range(0, len(identification)):
final.append([identification[x], predict[x]])
print "this is what final sort of looked like"
print final[:3]
return final
def _Dirichlet(cluster_data, identification):
print "In Dirichlet"
for i in range(0, 3):
print "i is ", i
dirichlet = DPGMM(n_components=len(cluster_data)).fit(cluster_data)
# paremeters= dirichlet.get_params #returns parameters of the algorithm as a whole from the fit
predict = dirichlet.predict(cluster_data)
n_clusters_ = len(set(predict)) - (1 if -1 in predict else 0)
print ("Estimated number of clusters with Dirichlet: %d" % n_clusters_)
final = []
for x in range(0, len(identification)):
final.append([identification[x], predict[x]])
print "this is what final sort of looked like"
print final[:3]
return final
def dpgmm_segmenter(factors, width=MEDIAN_WIDTH):
factors = median_filter(factors, size=(MEDIAN_WIDTH, 1), mode='mirror')
factors = pre.scale(factors, axis=1)
best_boundaries = [0, factors.shape[0] - 1]
best_n_types = 1
dpgmm = DPGMM(n_components=10, covariance_type='diag', alpha=10, n_iter=100)
dpgmm.fit(np.tile(factors, (10, 1)))
labels = dpgmm.predict(factors)
boundaries, labels = find_boundaries(labels, width)
if len(np.unique(labels)) > 1:
best_boundaries = boundaries
best_n_types = len(np.unique(labels))
if len(best_boundaries) < best_n_types + 1:
best_n_types = len(best_boundaries) - 1
best_labels = segment_labeling(factors, best_boundaries, c_method='kmeans', k=best_n_types)
best_boundaries = np.array(best_boundaries)
return best_boundaries, best_labels
def main(method,cluster_num=30,alpha=.5):
f ='/Users/davidgreenfield/Downloads/features_csv_tmp.csv'
#f ='/Users/davidgreenfield/Downloads/features_f500.csv'
cols=range(1,4096)
feats =np.loadtxt(open(f,"rb"),delimiter=",",skiprows=1,usecols=(cols))
asins = np.loadtxt(open(f,"rb"),delimiter=",",skiprows=1,usecols=([0]),dtype=str)
if method == 'kmeans':
k_means=cluster.KMeans(n_clusters=cluster_num)
k_means.fit(feats)
y = k_means.labels_
if MAKE_GRAPH==1:
print "hello 1"
create_graph(k_means)
elif method == 'GMM_VB':
gmm_vb = VBGMM.fit(feats,n_components=50,alpha=.5)
y = gmm_vb.predict(feats)
cluster_no = len(np.unique(y))
elif method == 'GMM_DP':
gmm_dp = DPGMM(n_components=50,alpha=alpha)
gmm_dp.fit(feats)
y = gmm_dp.predict(feats)
cluster_no = len(np.unique(y))
clusters=[]
groups={}
data=load_data('./data/boots_aws.csv')
for i in range(0,cluster_num):
groups[i]=np.where(y==i)
ids=asins[groups[i]]
clusters.append(ids)
links=[data[x]['url'] for x in ids]
create_html(links,"templates/groups/group"+str(i)+".html")
output_clusters(clusters,"outputs/clusters.csv")
#X = StandardScaler().fit_transform(fn)
fold = 3
kf = StratifiedKFold(label, n_folds=fold, shuffle=True)
#kf = KFold(len(label), n_folds=fold, shuffle=True)
clf = RFC(n_estimators=100, criterion='entropy')
rounds = 1
acc_sum = [[] for i in range(fold)]
for train, test in kf:
train_fn = fn[train]
#n_class = len(np.unique(label[train]))
d = DPGMM(n_components=50, covariance_type='spherical',alpha=10)
d.fit(train_fn)
#print 'mixture mean', d.means_
preds = d.predict(train_fn)
print '# of M by DP', len(np.unique(preds))
acc_sum[0].append(ARI(label[train], preds))
#acc_sum[0].append(SS(train_fn, preds))
#n_class = len(np.unique(preds))
n_class = 32
g = GMM(n_components=n_class, covariance_type='spherical', init_params='wmc', n_iter=100)
g.fit(train_fn)
#g.means_ = np.array([x_train[y_train == i].mean(axis=0) for i in np.unique(y_train)])
preds = g.predict(train_fn)
#prob = np.sort(g.predict_proba(train_fd))
acc_sum[1].append(ARI(label[train], preds))
#acc_sum[1].append(SS(train_fn, preds))
k = KMeans(init='k-means++', n_clusters=n_class, n_init=10)
class Clustering:
def __init__(self, X, scale=False, features=None):
self.random_seed = 12345 # set to None for random
self.X = X
self.k = None
self.centers = None
self.h = None
self.Y = None
self.scaler = None
self.ids = None
# Visualize().plot( zip(*self.X) ) # FOR DEBUG
# Reduce the number of features
if features is not None:
if isinstance(features, (int, float)):
variances = VarianceThreshold().fit(self.X).variances_
self.ids = sorted(
range(len(variances)), key=lambda i: variances[i]
)[-int(features
):] # indexes of the top n_features values in variances
elif type(features) in [list, tuple]:
self.ids = features
self.X = self.reduceFeatures(self.X)
print("Selected features", self.ids, "on a total of",
len(X[0])) # FOR DEBUG
# Visualize().plot( zip(*self.X) ) # FOR DEBUG
if scale:
self.scaler = StandardScaler(
) # MinMaxScaler() can also be used instead of StandardScaler()
self.X = self.scaler.fit_transform(self.X)
#---------------------------------------
def reduceFeatures(self, X):
if self.ids is None:
return X
else:
return [[v for iv, v in enumerate(x) if iv in self.ids] for x in X]
#---------------------------------------
def affinity(self, k=2): # K is not used here
self.h = AffinityPropagation(damping=0.75,
preference=k,
max_iter=200,
convergence_iter=15,
copy=True,
affinity='euclidean').fit(self.X)
self.Y = self.h.labels_
self.k = k
self.centers = self.getCenters()
return self
#---------------------------------------
def meanshift(self, k=2): # K is not used here
self.h = MeanShift(bandwidth=None,
seeds=None,
bin_seeding=False,
min_bin_freq=1,
cluster_all=True).fit(self.X)
self.Y = self.h.labels_
self.k = k
self.centers = self.getCenters()
return self
#---------------------------------------
def kmeans(self, k=2):
self.h = KMeans(n_clusters=k,
init='k-means++',
n_init=10,
max_iter=1000,
tol=0.00001,
random_state=self.random_seed).fit(self.X)
self.Y = self.h.labels_
self.k = k
self.centers = self.getCenters()
return self
#---------------------------------------
def gmm(self, k=2):
self.h = GMM(n_components=k, random_state=self.random_seed).fit(self.X)
self.Y = self.h.predict(self.X)
self.k = k
self.centers = self.getCenters()
#TODO
# posterior = self.h.predict_proba( self.X[:5] )
# likelihood = self.h.score( self.X[:5] )
return self
#---------------------------------------
''' Dirichlet Process is as likely to start a new cluster for a point as it is to add that point to a cluster with alpha elements (0<alpha<inf).
A higher alpha means more clusters, as the expected number of clusters is alpha*log(N)'''
def dpgmm(self, k=10, alpha=1.0):
self.h = DPGMM(n_components=k,
alpha=alpha,
random_state=self.random_seed).fit(self.X)
self.Y = self.h.predict(self.X)
self.k = k # this is the max number of components in dpgmm
self.centers = self.getCenters()
#TODO
# posterior = self.h.predict_proba( self.X[:5] )
# likelihood = self.h.score( self.X[:5] )
return self
#---------------------------------------
def done(self):
if self.h is None:
print("Clustering is not yet done !")
return False
else:
return True
#---------------------------------------
def getCenters(self):
if not self.done(): return
try:
return self.h.cluster_centers_
# If the clustering has no centers, compute them based on clusters
except AttributeError:
unique_labels = np.unique(self.Y)
clusters = {ul: [] for ul in unique_labels}
for i in range(len(self.X)):
clusters[self.Y[i]].append(self.X[i])
centers = []
for label in clusters:
centers.append(
[np.mean(col) for col in list(zip(*clusters[label]))])
return centers
#---------------------------------------
def predict(self, x):
if not self.done(): return
x_processed = x
x_processed = self.reduceFeatures([x_processed])[0]
x_processed = x_processed if self.scaler is None else self.scaler.transform(
x_processed)
return self.h.predict(x_processed)[0]
#---------------------------------------
def predictAll(self, X):
if not self.done(): return
X_processed = X
X_processed = self.reduceFeatures(X_processed)
X_processed = X_processed if self.scaler is None else self.scaler.transform(
X_processed)
return list(self.h.predict(X_processed))
#---------------------------------------
def quality(self, X=None):
if not self.done(): return
if X is None: # if X not provided then use the training data and resulting labels
X = self.X
Y = self.Y
else: # if X is provided then use it with the predicted labels (clusters)
Y = self.predictAll(X)
indexs = range(len(X))
shuffle(indexs)
X = np.array([X[i] for i in indexs[:5000]])
Y = np.array([Y[i] for i in indexs[:5000]])
if len(set(Y)) < 2: return 0. # FIXME
return silhouette_score(X, Y, metric='euclidean')
#---------------------------------------
def plot(self, fig=None):
if not self.done(): return
viz = Visualize()
if len(self.X[0]) > 3:
X = viz.PCA_Transform(list(zip(*self.X)))
else:
X = self.X
unique_labels = np.unique(self.Y)
clusters = {ul: [] for ul in unique_labels}
for i in range(len(X)):
clusters[self.Y[i]].append(X[i])
centers_for_plot = [
] # Not the real centers because dimension was reduced using PCA
for label in clusters:
centers_for_plot.append(
[np.mean(col) for col in list(zip(*clusters[label]))])
viz.do_plot(list(zip(*centers_for_plot)), marker='o', color='m')
viz.plot_groups(clusters, fig)
"""if you want clustering on the dissimilarity space uncomment
below and change accordingly"""
# print 'Calculating dissimilarity space for training queries...'
# data_cluster_train_ds = sc.pdist(data_cluster_train, 'euclidean')
# data_cluster_train_ds = sc.squareform(data_cluster_train_ds)
# # plt.figure(1)
# # plt.imshow(data_cluster_train_ds)
# # plt.colorbar()
# # plt.title('Initial dissimilarity')
print 'Training a Dirichlet Process Gaussian Mixture model...'
dpgmm = DPGMM(alpha=1.0, n_iter=100, n_components=50)
dpgmm.fit(data_cluster_train_ds)
prediction = dpgmm.predict(data_cluster_train_ds)
clusters = np.unique(prediction)
print 'Found %i clusters!' % clusters.shape[0]
print clusters
"""create the reordered input data according to the clusters
it is only needed if you want to visuallize the clustering
afterwards"""
#data_cluster = np.zeros((1, data_cluster_train.shape[1]))
# each cluster is a list of lists that contains the indices
# of the queries for each cluster
each_cluster = []
for i in xrange(clusters.shape[0]):
cluster = data_cluster_train[prediction == clusters[i], :]
data_cluster_train_ds = data_cluster_train
"""if you want clustering on the dissimilarity space uncomment
below and change accordingly"""
# print 'Calculating dissimilarity space for training queries...'
# data_cluster_train_ds = sc.pdist(data_cluster_train, 'euclidean')
# data_cluster_train_ds = sc.squareform(data_cluster_train_ds)
# # plt.figure(1)
# # plt.imshow(data_cluster_train_ds)
# # plt.colorbar()
# # plt.title('Initial dissimilarity')
print 'Training a Dirichlet Process Gaussian Mixture model...'
dpgmm = DPGMM(alpha=1.0, n_iter=100, n_components=50)
dpgmm.fit(data_cluster_train_ds)
prediction = dpgmm.predict(data_cluster_train_ds)
clusters = np.unique(prediction)
print 'Found %i clusters!' % clusters.shape[0]
print clusters
"""create the reordered input data according to the clusters
it is only needed if you want to visuallize the clustering
afterwards"""
#data_cluster = np.zeros((1, data_cluster_train.shape[1]))
# each cluster is a list of lists that contains the indices
# of the queries for each cluster
each_cluster = []
for i in xrange(clusters.shape[0]):
cluster = data_cluster_train[prediction == clusters[i], :]
each_cluster.append(np.where(prediction == clusters[i])[0])
class DPGMMClusterModel(BaseEstimator, TransformerMixin):
def __init__(self, w2v_model=None, n_components=None, no_above=0.9, no_below=8, dataname="", stoplist=None,
dictionary=None, recluster_thresh=1000, alpha=5):
self.w2v_model = w2v_model
self.no_above = no_above
self.no_below = no_below
self.alpha = alpha
self.n_components = n_components
self.n_sub_components = int(n_components / 2)
self.stoplist = stoplist
self.dataname = dataname
self.dictionary = dictionary
self.dpgmm = None
self.scaler = None
self.cluster_info = None
# a list of sub-clusterer
self.feature_crd = {}
self.subdpgmms = []
self.reclustered = []
self.recluster_thresh = recluster_thresh
def should_cluster_word(self, word):
return (word in self.dictionary.token2id) and (len(word) > 1) and \
(self.w2v_model is None or word in self.w2v_model) and \
(self.stoplist is None or word not in self.stoplist)
# constructs a dictionary and a DPGMM model on 9000 middle frequency words from X
# X is a sequence of texts
def fit(self, X, y=None):
# either consturct a dictionary from X, trim it
if self.dictionary is None:
self.dictionary = corpora.Dictionary(X)
# or use an existing dictionary and trim the given set of words
self.dictionary.filter_extremes(no_below=self.no_below, no_above=self.no_above, keep_n=9000)
if self.w2v_model is None:
w2v_corpus = [[word for word in text if self.should_cluster_word(word)] for text in X]
self.w2v_model = w2v_models.build_word2vec(w2v_corpus, size=100, window=10, min_count=self.no_below,
dataname=self.dataname+"_dpgmm")
word_list = np.array([word for word in self.dictionary.token2id.iterkeys() if self.should_cluster_word(word)])
# This was reclustering clause - I need to re-write this
# else:
# # note the double loop here!!
# word_list = np.array([word for text in X for word in text if self.should_cluster_word(word)])
# construct a list of words to cluster
# remove rare and frequent words
# remove words of length 1
# remove stopwords
vec_list = [self.w2v_model[word] for word in word_list]
logging.info("DPGMM received %i words" % len(vec_list))
# save word representations
filename = "w2v_vocab_%s_%.1f_%.0f.lcsv" % (self.dataname, self.no_above, self.no_below)
io.save_words_representations(filename, word_list, vec_list)
self.scaler = StandardScaler()
vecs = self.scaler.fit_transform(np.array(vec_list))
self.dpgmm = DPGMM(n_components=self.n_components, covariance_type='diag', alpha=self.alpha,
n_iter=1000, tol=0.0001)
self.dpgmm.fit(vecs)
logging.info("DPGMM converged: %s" % self.dpgmm.converged_)
# save information about found clusters
self.cluster_info = []
y_ = self.dpgmm.predict(vecs)
for i, cluster_center in enumerate(self.dpgmm.means_):
cluster_words = word_list[y_ == i]
cluster_size = len(cluster_words)
if cluster_size > self.recluster_thresh and self.recluster_thresh > 0:
logging.info("DPGMM: reclustering %i words for cluster %i" % (len(cluster_words), i))
sub_dpgmm = DPGMMClusterModel(w2v_model=self.w2v_model,
n_components=self.n_sub_components,
dictionary=self.dictionary,
dataname="%s-%i" % (self.dataname, i), stoplist=self.stoplist)
# recluster words. Note the double array
sub_dpgmm.fit([cluster_words])
self.subdpgmms.append(sub_dpgmm)
self.reclustered.append(i)
if cluster_size > 0:
#cluster_center_original = self.scaler.inverse_transform(cluster_center)
#similar_words = self.w2v_model.most_similar_cosmul(positive=[cluster_center_original], topn=cluster_size)
#central_words = [word for word, _ in similar_words if word in cluster_words]
central_words = cluster_words[0:10]
else:
central_words = []
self.cluster_info.append({'cnt': i, 'size': cluster_size, 'words': central_words})
filename = "clusters_%s_%i_%.1f_%.0f.txt" % (self.dataname, self.n_components, self.no_above, self.no_below)
io.save_cluster_info(filename, self.cluster_info)
# setting up the coordinates for the features
self.feature_crd = {'global': range(0, self.n_components),
'reclustered': [i for i in range(0, self.n_components + self.n_sub_components*len(self.reclustered))
if i not in self.reclustered]}
return self
# calculate cluster counts for one text
def clusterize(self, text):
word_list = [word for word in text if self.should_cluster_word(word)]
vec_list = np.array([self.w2v_model[word] for word in word_list])
bincounts = np.zeros((self.n_components+self.n_sub_components*len(self.reclustered),))
if len(vec_list) > 0:
# assign words to clusters
predictions = self.dpgmm.predict(self.scaler.transform(np.array(vec_list)))
global_bincount = np.bincount(predictions, minlength=self.n_components)
# re-assign words in large clusters
bincounts[0:self.n_components] = global_bincount #reshape((1,len(global_bincount)))
start = self.n_components
for i, subdpgmm in zip(self.reclustered, self.subdpgmms):
# if words in respective clusters exists - recluster them
vecs_torecluster = vec_list[predictions == i]
if len(vecs_torecluster) > 0:
predictions = subdpgmm.dpgmm.predict(subdpgmm.scaler.transform(np.array(vecs_torecluster)))
bincounts[start:start+subdpgmm.dpgmm.n_components] = \
np.bincount(predictions, minlength=subdpgmm.dpgmm.n_components) #.reshape((1, subdpgmm.n_components))
start += subdpgmm.dpgmm.n_components
# erase the count inthe global counts
# returns a vector of cluster bin counts: [ global, reclustered1, reclustered2, ...]
return bincounts.reshape((1, len(bincounts)))
# for a text, constructs a bincount of clusters present in the sentence
# X is a list of texts. One text is one string! Not tokenized
def transform(self, X):
# Text pre-processing
x_clean = [tu.normalize_punctuation(text).split() for text in X]
logging.info("DPGGM: Text prepocessed")
# Vectorize using W2V model
if self.dpgmm is not None:
logging.info("Vectorizing a corpus")
size = self.w2v_model.layer1_size
if len(X) > 0:
vecs = np.concatenate([self.clusterize(z) for z in x_clean], axis=0)
else:
vecs = np.zeros(size).reshape((1, size))
logging.info("DPGMM: returning pre-processed data of shape %s" % (vecs.shape, ))
else:
logging.info("W2V Averaged: no model was provided.")
vecs = np.zeros((len(X), 1))
return vecs
return X, Y
def test1():
print 'test1'
model = VDPGMM(T = 10, alpha = 1, max_iter = 50)
X, Y = getXY('iris')
model.fit(X)
y = model.predict(X)
print 'VDPGMM'
print len(np.unique(y)), np.unique(y)
print [np.sum(y == label) for label in np.unique(y)]
from sklearn.mixture import DPGMM
model = DPGMM(n_components = 10, alpha = 1, n_iter = 50)
model.fit(X)
y = model.predict(X)
print 'DPGMM'
print len(np.unique(y)), np.unique(y)
print [np.sum(y == label) for label in np.unique(y)]
def test2():
print 'test2'
np.random.seed(1)
X = np.concatenate((2 + np.random.randn(100, 2), 5 + np.random.randn(100, 2), 10 + np.random.randn(100, 2)))
T = 10
model = VDPGMM(T=T, alpha=.5, max_iter=100, thresh=1e-5)
model.fit(X)
plt.clf()
h = plt.subplot()
color = 'rgbcmykw'
img.reshape((1, img.shape[0], img.shape[1], img.shape[2])), -1, 1)
img = vgg16.preprocess_input(img.astype('float32'))
""" Scaling activations to fit random initialization scheme"""
actvs = get_activations(model, layer, img).squeeze()
actvs /= np.max(actvs) * 0.1
""" Clustering with dirichlet process Gaussian Mixture Model"""
dpgmm = DPGMM(n_components=50,
alpha=1,
verbose=2,
tol=0.01,
n_iter=250,
min_covar=1e-6)
#dpgmm = BayesianGaussianMixture(n_components=50, covariance_type="diag", reg_covar = 1e-6,
# weight_concentration_prior_type="dirichlet_process",
# weight_concentration_prior=1, verbose=2,
# tol=0.01, max_iter=250, init_params='random',
# mean_precision_prior=actvs.std(),
# mean_prior=np.repeat(actvs.max()/5,actvs.shape[0]))
dpgmm.fit(
np.transpose(actvs.reshape(actvs.shape[0],
actvs.shape[1] * actvs.shape[2])))
labels = dpgmm.predict(
np.transpose(actvs.reshape(actvs.shape[0],
actvs.shape[1] * actvs.shape[2])))
labels = labels.reshape((actvs.shape[1], actvs.shape[2]))
plt.subplot(1, 2, 2)
plt.imshow(labels, interpolation="nearest")
plt.title('Labelmap from layer ' + str(layer))
ax1_min, ax1_max, ax2_min, ax2_max = plt.axis()
plt.xlim((x1_min, x1_max))
plt.ylim((x2_min, x2_max))
plt.title(u'GMM', fontsize=20)
plt.grid(True)
# DPGMM
n_components = 3
dpgmm = DPGMM(n_components=n_components, alpha=1, covariance_type='full', random_state=0)
dpgmm.fit(x)
centers = dpgmm.means_
covs = dpgmm._get_covars()
print 'DPGMM均值 = \n', centers
print 'DPGMM方差 = \n', covs
y_hat = dpgmm.predict(x)
# print y_hat
ax = plt.subplot(212)
grid_hat = dpgmm.predict(grid_test)
grid_hat = grid_hat.reshape(x1.shape)
plt.pcolormesh(x1, x2, grid_hat, cmap=cm)
plt.scatter(x[:, 0], x[:, 1], s=30, c=y, cmap=cm, marker='o')
for i, cc in enumerate(zip(centers, covs)):
if i not in y_hat:
continue
center, cov = cc
value, vector = sp.linalg.eigh(cov)
width, height = value[0], value[1]
v = vector[0] / sp.linalg.norm(vector[0])
def dpgmm_simple(X, init_numC, random_state):
model = DPGMM(n_components = init_numC, n_iter=100, tol=0.000001, random_state=random_state)
model.fit(X)
y = model.predict(X)
cluster_num = len(np.unique(y))
return cluster_num, y
# Sample the specified number of points from X_unlabeled
size = np.cumsum(chunk_sizes[:chunks])[-1]
# Fit a Dirichlet process mixture of Gaussians using up to ten components
dpgmm = DPGMM(n_components=10, alpha=10.0, covariance_type='full')
indices = np.arange(X_unlabeled.shape[0])
np.random.shuffle(indices)
X = X_unlabeled[indices[:size],]
print("fitting a model with", size, "data points")
with timeit():
dpgmm.fit(X)
print("Done!")
print("AIC for this model & data: ", dpgmm.aic(X))
print("BIC for this model & data: ", dpgmm.bic(X))
Y_hat = dpgmm.predict(X)
print ("Model assigned points to", np.max(Y_hat), "components")
# How can I best check this out?
#color_iter = itertools.cycle(['r', 'g', 'b', 'c', 'm'])
#for i, (clf, title) in enumerate([(gmm, 'GMM'),
#(dpgmm, 'Dirichlet Process GMM')]):
#splot = plt.subplot(2, 1, 1 + i)
#Y_ = clf.predict(X)
#for i, (mean, covar, color) in enumerate(zip(
#clf.means_, clf._get_covars(), color_iter)):
#v, w = linalg.eigh(covar)
#u = w[0] / linalg.norm(w[0])
## as the DP will not use every component it has access to
## unless it needs it, we shouldn't plot the redundant
class Clustering:
def __init__(self, X, scale=False, features=None):
self.random_seed = 12345 # set to None for random
self.X = X
self.k = None
self.centers = None
self.h = None
self.Y = None
self.scaler = None
self.ids = None
# Visualize().plot( zip(*self.X) ) # FOR DEBUG
# Reduce the number of features
if features is not None:
if isinstance( features, (int, float) ):
variances = VarianceThreshold().fit(self.X).variances_
self.ids = sorted(range(len(variances)), key=lambda i: variances[i])[-int(features):] # indexes of the top n_features values in variances
elif type(features) in [list,tuple]:
self.ids = features
self.X = self.reduceFeatures(self.X)
print("Selected features", self.ids, "on a total of", len(X[0])) # FOR DEBUG
# Visualize().plot( zip(*self.X) ) # FOR DEBUG
if scale:
self.scaler = StandardScaler() # MinMaxScaler() can also be used instead of StandardScaler()
self.X = self.scaler.fit_transform(self.X)
#---------------------------------------
def reduceFeatures(self, X):
if self.ids is None:
return X
else:
return [ [v for iv,v in enumerate(x) if iv in self.ids] for x in X ]
#---------------------------------------
def affinity(self, k=2): # K is not used here
self.h = AffinityPropagation(damping=0.75, preference=k, max_iter=200, convergence_iter=15, copy=True, affinity='euclidean').fit( self.X )
self.Y = self.h.labels_
self.k = k
self.centers = self.getCenters()
return self
#---------------------------------------
def meanshift(self, k=2): # K is not used here
self.h = MeanShift(bandwidth=None, seeds=None, bin_seeding=False, min_bin_freq=1, cluster_all=True).fit( self.X )
self.Y = self.h.labels_
self.k = k
self.centers = self.getCenters()
return self
#---------------------------------------
def kmeans(self, k=2):
self.h = KMeans(n_clusters = k, init = 'k-means++', n_init = 10, max_iter = 1000, tol = 0.00001, random_state = self.random_seed).fit( self.X )
self.Y = self.h.labels_
self.k = k
self.centers = self.getCenters()
return self
#---------------------------------------
def gmm(self, k=2):
self.h = GMM(n_components=k, random_state = self.random_seed).fit( self.X )
self.Y = self.h.predict( self.X )
self.k = k
self.centers = self.getCenters()
#TODO
# posterior = self.h.predict_proba( self.X[:5] )
# likelihood = self.h.score( self.X[:5] )
return self
#---------------------------------------
''' Dirichlet Process is as likely to start a new cluster for a point as it is to add that point to a cluster with alpha elements (0<alpha<inf).
A higher alpha means more clusters, as the expected number of clusters is alpha*log(N)'''
def dpgmm(self, k=10, alpha=1.0):
self.h = DPGMM(n_components=k, alpha=alpha, random_state = self.random_seed).fit( self.X )
self.Y = self.h.predict( self.X )
self.k = k # this is the max number of components in dpgmm
self.centers = self.getCenters()
#TODO
# posterior = self.h.predict_proba( self.X[:5] )
# likelihood = self.h.score( self.X[:5] )
return self
#---------------------------------------
def done(self):
if self.h is None:
print("Clustering is not yet done !")
return False
else:
return True
#---------------------------------------
def getCenters(self):
if not self.done(): return
try:
return self.h.cluster_centers_
# If the clustering has no centers, compute them based on clusters
except AttributeError:
unique_labels = np.unique(self.Y)
clusters = { ul:[] for ul in unique_labels }
for i in range( len(self.X) ):
clusters[ self.Y[i] ].append( self.X[i] )
centers = []
for label in clusters:
centers.append( [np.mean(col) for col in list(zip(* clusters[label] )) ] )
return centers
#---------------------------------------
def predict(self, x):
if not self.done(): return
x_processed = x
x_processed = self.reduceFeatures([x_processed])[0]
x_processed = x_processed if self.scaler is None else self.scaler.transform(x_processed)
return self.h.predict(x_processed)[0]
#---------------------------------------
def predictAll(self, X):
if not self.done(): return
X_processed = X
X_processed = self.reduceFeatures(X_processed)
X_processed = X_processed if self.scaler is None else self.scaler.transform(X_processed)
return list(self.h.predict(X_processed))
#---------------------------------------
def quality(self, X=None):
if not self.done(): return
if X is None: # if X not provided then use the training data and resulting labels
X = self.X
Y = self.Y
else: # if X is provided then use it with the predicted labels (clusters)
Y = self.predictAll(X)
indexs = range(len(X)); shuffle(indexs)
X = np.array([ X[i] for i in indexs[:5000] ])
Y = np.array([ Y[i] for i in indexs[:5000] ])
if len(set(Y)) < 2: return 0. # FIXME
return silhouette_score(X, Y, metric='euclidean')
#---------------------------------------
def plot(self, fig=None):
if not self.done(): return
viz = Visualize()
if len(self.X[0]) > 3:
X = viz.PCA_Transform( list(zip(*self.X)) )
else:
X = self.X
unique_labels = np.unique(self.Y)
clusters = { ul:[] for ul in unique_labels }
for i in range( len(X) ):
clusters[ self.Y[i] ].append( X[i] )
centers_for_plot = [] # Not the real centers because dimension was reduced using PCA
for label in clusters:
centers_for_plot.append( [np.mean(col) for col in list(zip(* clusters[label] )) ] )
viz.do_plot(list(zip(*centers_for_plot)), marker='o', color='m')
viz.plot_groups(clusters, fig)
plt.ylim((x2_min, x2_max))
plt.title(u'GMM', fontsize=20)
plt.grid(True)
# DPGMM
n_components = 3
dpgmm = DPGMM(n_components=n_components,
alpha=1,
covariance_type='full',
random_state=0)
dpgmm.fit(x)
centers = dpgmm.means_
covs = dpgmm._get_covars()
print 'DPGMM均值 = \n', centers
print 'DPGMM方差 = \n', covs
y_hat = dpgmm.predict(x)
# print y_hat
ax = plt.subplot(212)
grid_hat = dpgmm.predict(grid_test)
grid_hat = grid_hat.reshape(x1.shape)
plt.pcolormesh(x1, x2, grid_hat, cmap=cm)
plt.scatter(x[:, 0], x[:, 1], s=30, c=y, cmap=cm, marker='o')
for i, cc in enumerate(zip(centers, covs)):
if i not in y_hat:
continue
center, cov = cc
value, vector = sp.linalg.eigh(cov)
width, height = value[0], value[1]
v = vector[0] / sp.linalg.norm(vector[0])
