-
Notifications
You must be signed in to change notification settings - Fork 6
/
ASPC.py
198 lines (168 loc) · 8.68 KB
/
ASPC.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
from tensorflow.keras.models import Model
from tensorflow.keras import callbacks
import numpy as np
from time import time
import math
from sklearn.cluster import KMeans
from MyModel import MyImageGenerator, autoencoder, generator
import metrics
class ASPC(object):
def __init__(self, dims, n_clusters):
self.dims = dims
self.n_clusters = n_clusters
self.centers = []
self.y_pred = []
self.datagen = MyImageGenerator(width_shift_range=0.1, height_shift_range=0.1, rotation_range=10)
self.autoencoder, self.encoder = autoencoder(dims=dims)
self.model = self.encoder
self.pretrained = False
def pretrain(self, x, y=None, optimizer='adam', epochs=200, batch_size=256, save_dir='results/temp',
da_s1=False, verbose=1, use_multiprocessing=True):
print('Pretraining......')
self.autoencoder.compile(optimizer=optimizer, loss='mse')
csv_logger = callbacks.CSVLogger(save_dir + '/pretrain_log.csv')
cb = [csv_logger]
if y is not None and verbose > 0:
class PrintACC(callbacks.Callback):
def __init__(self, x, y):
self.x = x
self.y = y
super(PrintACC, self).__init__()
def on_epoch_end(self, epoch, logs=None):
if epochs < 10 or epoch % int(epochs / 10) != 0:
return
feature_model = Model(self.model.input,
self.model.get_layer(index=int(len(self.model.layers) / 2)).output)
features = feature_model.predict(self.x)
km = KMeans(n_clusters=len(np.unique(self.y)), n_init=20, n_jobs=4)
y_pred = km.fit_predict(features)
print(' ' * 8 + '|==> acc: %.4f, nmi: %.4f <==|'
% (metrics.acc(self.y, y_pred), metrics.nmi(self.y, y_pred)))
cb.append(PrintACC(x, y))
# begin pretraining
t0 = time()
if not da_s1:
self.autoencoder.fit(x, x, batch_size=batch_size, epochs=epochs, callbacks=cb, verbose=verbose)
else:
print('-=*' * 20)
print('Using augmentation for pretraining')
print('-=*' * 20)
self.autoencoder.fit_generator(
generator(self.datagen, x, batch_size=batch_size),
steps_per_epoch=math.ceil(x.shape[0] / batch_size), epochs=epochs,
callbacks=cb, verbose=verbose, use_multiprocessing=use_multiprocessing, workers=4)
print('Pretraining time: ', time() - t0)
self.autoencoder.save_weights(save_dir + '/ae_weights.h5')
print('Pretrained weights are saved to %s/ae_weights.h5' % save_dir)
self.pretrained = True
def basic_clustering(self, x):
""" Initialize a clustering result, i.e., labels and cluster centers.
:param x: input data, shape=[n_samples, n_features]
:return: labels and centers
"""
print("Using k-means for initialization by default.")
kmeans = KMeans(n_clusters=self.n_clusters, n_init=20, n_jobs=4)
print(x.shape)
y_pred = kmeans.fit_predict(X=x)
centers = kmeans.cluster_centers_.astype(np.float32)
return y_pred, centers
def update_labels(self, x, centers):
""" Update cluster labels.
:param x: input data, shape=(n_samples, n_features)
:param centers: cluster centers, shape=(n_cluster, n_features)
:return: (labels, loss): labels indicate each sample belongs to which cluster. labels[i]=j means sample i
belongs to cluster j; loss, the average distance between samples and their responding centers
"""
x_norm = np.reshape(np.sum(np.square(x), 1), [-1, 1]) # column vector
center_norm = np.reshape(np.sum(np.square(centers), 1), [1, -1]) # row vector
dists = x_norm - 2 * np.matmul(x, np.transpose(centers)) + center_norm # |x-y|^2 = |x|^2 -2*x*y^T + |y|^2
labels = np.argmin(dists, 1)
losses = np.min(dists, 1)
return labels, losses
def compute_sample_weight(self, losses, t, T):
lam = np.mean(losses) + t*np.std(losses) / T
return np.where(losses < lam, 1., 0.)
def load_weights(self, weights): # load weights of DEC model
self.model.load_weights(weights)
def predict(self, x, batch_size=256):
return self.model.predict(x, batch_size)
def predict_labels(self, x): # predict cluster labels using the output of clustering layer
return self.basic_clustering(self.predict(x))[0]
def get_labels(self):
return self.y_pred
def compile(self, optimizer='sgd', loss='mse'):
self.model.compile(optimizer=optimizer, loss=loss)
def fit(self, x, y=None, batch_size=256, epochs=100,
ae_weights=None, save_dir='result/temp', tol=0.001,
use_sp=True, da_s2=False, use_multiprocessing=True):
# prepare folder for saving results
import csv, os
if not os.path.exists(save_dir):
os.makedirs(save_dir)
# pretraining
t0 = time()
if ae_weights is None and not self.pretrained:
print('Pretraining AE...')
self.pretrain(x, save_dir=save_dir)
print('Pretraining time: %.1fs' % (time() - t0))
elif ae_weights is not None:
self.autoencoder.load_weights(ae_weights)
print('Pretrained AE weights are loaded successfully!')
# initialization
t1 = time()
self.y_pred, self.centers = self.basic_clustering(self.predict(x))
t2 = time()
print('Time for initialization: %.1fs' % (t2 - t1))
# logging file
logfile = open(save_dir + '/log.csv', 'w')
logwriter = csv.DictWriter(logfile, fieldnames=['epoch', 'acc', 'nmi', 'Ln', 'Lc'])
logwriter.writeheader()
net_loss = 0
clustering_loss = 0
time_train = 0
sample_weight = np.ones(shape=x.shape[0])
sample_weight[self.y_pred == -1] = 0 # do not use the noisy examples
y_pred_last = np.copy(self.y_pred)
result = None
for epoch in range(epochs+1):
""" Log and check stopping criterion """
if y is not None:
acc = np.round(metrics.acc(y, self.y_pred), 5)
nmi = np.round(metrics.nmi(y, self.y_pred), 5)
print('Epoch-%d: ACC=%.4f, NMI=%.4f, Ln=%.4f, Lc=%.4f; time=%.1f' %
(epoch, acc, nmi, net_loss, clustering_loss, time_train))
logwriter.writerow(dict(epoch=epoch, acc=acc, nmi=nmi, Ln=net_loss, Lc=clustering_loss))
logfile.flush()
# record the initial result
if epoch == 0:
print('ASPC model saved to \'%s/model_init.h5\'' % save_dir)
self.model.save_weights(save_dir + '/model_init.h5')
# check stop criterion
delta_y = np.sum(self.y_pred != y_pred_last).astype(np.float32) / self.y_pred.shape[0]
y_pred_last = np.copy(self.y_pred)
if (epoch > 0 and delta_y < tol) or epoch >= epochs:
result = np.asarray([acc, nmi])
print('Training stopped: epoch=%d, delta_label=%.4f, tol=%.4f' % (epoch, delta_y, tol))
print('ASPC model saved to \'%s/model_final.h5\'' % save_dir)
print('-' * 30 + ' END: time=%.1fs ' % (time()-t0) + '-' * 30)
self.model.save_weights(save_dir + '/model_final.h5')
logfile.close()
break
""" Step 1: train the network """
t0_epoch = time()
if da_s2: # use data augmentation
history = self.model.fit_generator(
generator(self.datagen, x, self.centers[self.y_pred], sample_weight, batch_size),
steps_per_epoch=math.ceil(x.shape[0] / batch_size), epochs=5 if np.any(self.y_pred == -1) and epoch==0 else 1,
use_multiprocessing=use_multiprocessing, workers=4, verbose=0)
else:
history = self.model.fit(x, y=self.centers[self.y_pred], batch_size=batch_size, epochs=1,
sample_weight=sample_weight, verbose=0)
net_loss = history.history['loss'][0]
""" Step 2: update labels """
self.y_pred, losses = self.update_labels(self.predict(x), self.centers)
clustering_loss = np.mean(losses)
""" Step 3: Compute sample weights """
sample_weight = self.compute_sample_weight(losses, epoch, epochs) if use_sp else None
time_train = time() - t0_epoch
return result