-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathMVCVAE_train.py
460 lines (385 loc) · 19.6 KB
/
MVCVAE_train.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
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
import os
os.environ['KERAS_BACKEND'] = 'theano'
import numpy as np
from keras.callbacks import Callback
from keras.optimizers import Adam
from keras.layers import Input, Dense, Lambda
from keras.models import Model
from keras import backend as K
from keras import objectives
import scipy.io as scio
import sys
import theano
import theano.tensor as T
import math
from sklearn.cluster import KMeans
from keras.models import model_from_json
from sklearn import preprocessing
from sklearn import metrics as mtr
import metrics
import warnings
import matplotlib.pyplot as plt
from sklearn.manifold import TSNE
import argparse
warnings.filterwarnings("ignore")
theano.config.floatX = 'float32'
class Multiview_VaDE():
def __init__(self, batch_size, num_views, latent_dim, intermediate_dim, config,weights_path, dataset, ispretrain=True,
**kwargs):
self.batch_size = batch_size
self.num_views = num_views
self.latent_dim = latent_dim
self.intermediate_dim = intermediate_dim
self.ispretrain = ispretrain
self.dataset = dataset
self.init = 'variancescaling'
self.weights_path = weights_path
self.original_dim, self.epoch, self.n_centroid, self.lr_nn, self.lr_gmm, self.decay_n, self.decay_nn, self.decay_gmm, self.alpha, self.datatype = config
self.sample_output, self.gamma_output, self.vade_model = self.build()
def encoder(self, original_dim):
means = []
inputs = []
vars = []
for i in range(0, self.num_views):
input = Input(batch_shape=(self.batch_size, original_dim[i]), name='input_%d' %i)
layer1 = Dense(self.intermediate_dim[0], init=self.init, activation='relu', name='encode1_%d' %i)(input)
layer2 = Dense(self.intermediate_dim[1], init=self.init, activation='relu', name='encode2_%d' %i)(layer1)
layer3 = Dense(self.intermediate_dim[2], init=self.init, activation='relu', name='encode3_%d' %i)(layer2)
z_means = Dense(self.latent_dim, init=self.init, activation=None, name='mean_%d' %i)(layer3)
z_vars = Dense(self.latent_dim, init=self.init, activation=None,name='var_%d' %i)(layer3)
means.append(z_means)
vars.append(z_vars)
inputs.append(input)
return means, vars, inputs
def decoder(self, latent, original_dim):
reconst=[]
for i in range(0, self.num_views):
layer4 = Dense(self.intermediate_dim[-1], init=self.init, activation='relu', name='decode1_%d' %i)(latent)
layer5 = Dense(self.intermediate_dim[-2], init=self.init, activation='relu', name='decode2_%d' %i)(layer4)
layer6 = Dense(self.intermediate_dim[-3], init=self.init, activation='relu', name='decode3_%d' %i)(layer5)
decoded= Dense(original_dim[i], init=self.init, activation=self.datatype, name='reconst_%d' %i)(layer6)
reconst.append(decoded)
return reconst
def build(self):
self.gmmpara_init()
self.get_zeta()
means, vars, self.inputs = self.encoder(self.original_dim)
z_mean = Lambda(self.mixture_u, output_shape=(self.latent_dim,))(means)
z_log_var = Lambda(self.mixture_z, output_shape=(self.latent_dim,))(vars)
z = Lambda(self.sampling, output_shape=(self.latent_dim,))([z_mean, z_log_var])
x_decoded = self.decoder(z, self.original_dim)
output=x_decoded
output.append(z)
output.append(z_mean)
output.append(z_log_var)
gamma = Lambda(self.get_gamma, output_shape=(self.n_centroid,))(z)
vade_loss = Lambda(self.vade_loss_function, name='vade_loss', output_shape=([],))(output)
multiview_output = [vade_loss]
return Model(self.inputs, z_mean), Model(self.inputs, gamma), Model(self.inputs, multiview_output)
def gmmpara_init(self):
theta_init = np.ones(self.n_centroid) / self.n_centroid
u_init = np.zeros((self.latent_dim, self.n_centroid))
lambda_init = np.ones((self.latent_dim, self.n_centroid))
self.theta_p = theano.shared(np.asarray(theta_init, dtype=theano.config.floatX), name="pi")
self.u_p = theano.shared(np.asarray(u_init, dtype=theano.config.floatX), name="u")
self.lambda_p = theano.shared(np.asarray(lambda_init, dtype=theano.config.floatX), name="lambda")
def get_zeta(self):
zeta_init = np.ones(self.num_views) / self.num_views
self.zeta = theano.shared(np.asarray(zeta_init, dtype=theano.config.floatX), name="zeta")
def mixture_u(self, args):
u =0
# self.zeta = K.softmax(self.zeta, axis=1)
for i in range(0, self.num_views):
u += self.zeta[i]*args[i]
return u
def mixture_z(self, args):
z =0
for i in range(0, self.num_views):
z += self.zeta[i] * K.exp(args[i])
return K.log(z)
def sampling(self, args):
z_mean, z_log_var = args
epsilon = K.random_normal(shape=(self.batch_size, self.latent_dim), mean=0.)
return z_mean + K.exp(z_log_var / 2) * epsilon
def get_gamma(self, tempz):
temp_Z = T.transpose(K.repeat(tempz, self.n_centroid), [0, 2, 1])
temp_u_tensor = T.repeat(self.u_p.dimshuffle('x', 0, 1), self.batch_size, axis=0)
temp_lambda_tensor = T.repeat(self.lambda_p.dimshuffle('x', 0, 1), self.batch_size, axis=0)
# version1
temp_theta_tensor = self.theta_p.dimshuffle('x', 'x', 0) * T.ones(
(self.batch_size, self.latent_dim, self.n_centroid))
temp_p_c_z = K.exp(K.sum((K.log(temp_theta_tensor) - 0.5 * K.log(2 * math.pi * temp_lambda_tensor) - \
K.square(temp_Z - temp_u_tensor) / (2 * temp_lambda_tensor)), axis=1)) + 1e-10
return temp_p_c_z / K.sum(temp_p_c_z, axis=-1, keepdims=True)
#version2
# temp_theta_tensor = self.theta_p.dimshuffle('x', 0) * T.ones((self.batch_size, self.n_centroid))
# temp_p_c_z = K.exp(K.log(temp_theta_tensor) - K.sum((0.5 * K.log(2 * math.pi * temp_lambda_tensor) + \
# K.square(temp_Z - temp_u_tensor) / (2 * temp_lambda_tensor)), axis=1)) + 1e-10
# return temp_p_c_z / K.sum(temp_p_c_z, axis=1, keepdims=True)
def vade_loss_function(self, args):
inputs = self.inputs
#reconst = self.x_decoded
reconst, z, z_mean, z_log_var = args[:self.num_views], args[-3], args[-2], args[-1]
Z = T.transpose(K.repeat(z, self.n_centroid), [0, 2, 1])
z_mean_t = T.transpose(K.repeat(z_mean, self.n_centroid), [0, 2, 1])
z_log_var_t = T.transpose(K.repeat(z_log_var, self.n_centroid), [0, 2, 1])
u_tensor3 = T.repeat(self.u_p.dimshuffle('x', 0, 1), self.batch_size, axis=0)
lambda_tensor3 = T.repeat(self.lambda_p.dimshuffle('x', 0, 1), self.batch_size, axis=0)
#version1
theta_tensor3 = self.theta_p.dimshuffle('x', 'x', 0) * T.ones(
(self.batch_size, self.latent_dim, self.n_centroid))
p_c_z = K.exp(K.sum((K.log(theta_tensor3) - 0.5 * K.log(2 * math.pi * lambda_tensor3) - \
K.square(Z - u_tensor3) / (2 * lambda_tensor3)), axis=1)) + 1e-10
gamma = p_c_z / K.sum(p_c_z, axis=-1, keepdims=True)
gamma_t = K.repeat(gamma, self.latent_dim)
#version2
# theta_tensor2 = self.theta_p.dimshuffle('x', 0) * T.ones((self.batch_size, self.n_centroid))
# p_c_z = K.exp(K.log(theta_tensor2) - K.sum((0.5 * K.log(2 * math.pi * lambda_tensor3) + \
# K.square(Z - u_tensor3) / (
# 2 * lambda_tensor3)), axis=1)) + 1e-10
# gamma = p_c_z / K.sum(p_c_z, axis=1, keepdims=True)
reconst_loss=0
for i in range(0, num_views):
#version 1
r_loss = self.original_dim[i] * objectives.mean_squared_error(inputs[i], reconst[i])
#version 2
#r_loss = self.original_dim[i]*objectives.binary_crossentropy(inputs[i], reconst[i])
reconst_loss += r_loss
#version 1
loss = reconst_loss + self.alpha * (K.sum(0.5 * gamma_t * (
self.latent_dim * K.log(math.pi * 2) + K.log(lambda_tensor3) + K.exp(
z_log_var_t) / lambda_tensor3 + K.square(z_mean_t - u_tensor3) / lambda_tensor3), axis=(1, 2)) \
- 0.5 * K.sum(z_log_var + 1, axis=-1) \
- K.sum(
K.log(K.repeat_elements(self.theta_p.dimshuffle('x', 0), self.batch_size, 0)) * gamma, axis=-1) \
+ K.sum(K.log(gamma) * gamma, axis=-1))
#version2
# loss = reconst_loss + self.alpha * (K.sum(0.5 * gamma * K.sum(
# K.log(math.pi * 2) + K.log(lambda_tensor3) + K.exp(
# z_log_var_t) / lambda_tensor3 + K.square(z_mean_t - u_tensor3) / lambda_tensor3, axis=1), axis=1) \
# - 0.5 * K.sum(z_log_var + 1 + K.log(2*math.pi), axis=1) \
# - K.sum(
# K.log(theta_tensor2)* gamma, axis=1) \
# + K.sum(K.log(gamma) * gamma, axis=1))
return loss
def floatX(self, X):
return np.asarray(X, dtype=theano.config.floatX)
def load_pretrain_weights(self, vade, weights_path, dataset, inputs,Y):
vade.load_weights(weights_path + dataset + '.h5')
sample = self.sample_output.predict(inputs, batch_size=self.batch_size)
kmeans = KMeans(n_clusters=self.n_centroid, n_init=20)
kmeans.fit(sample)
self.u_p.set_value(self.floatX(kmeans.cluster_centers_.T))
y_pred = kmeans.predict(sample)
gam = self.gamma_output.predict(inputs, batch_size=batch_size)
gam_acc = metrics.acc(np.argmax(gam, axis=1), Y)
print ('pretrain weights loaded!')
print('Initial_acc:', gam_acc)
return vade
def compile(self, inputs, Y):
if self.ispretrain is True:
self.vade_model = self.load_pretrain_weights(self.vade_model, self.weights_path, self.dataset, inputs, Y)
adam_nn = Adam(lr=self.lr_nn, epsilon=1e-4)
adam_gmm = Adam(lr=self.lr_gmm, epsilon=1e-4)
self.vade_model.compile(optimizer=adam_nn, loss=lambda y_true, y_pred: y_pred,
add_trainable_weights=[self.theta_p, self.u_p, self.lambda_p, self.zeta],
add_optimizer=adam_gmm)
self.epoch_begin = EpochBegin(self.sample_output, self.decay_n, self.gamma_output, self.decay_nn, self.decay_gmm, adam_nn, adam_gmm,
inputs, Y, self.u_p, self.lambda_p, self.theta_p, self.zeta)
def train(self, inputs):
none = np.zeros([np.shape(X1)[0]])
self.vade_model.fit(x=inputs, y=none, shuffle=True, nb_epoch=self.epoch, batch_size=self.batch_size,
callbacks=[self.epoch_begin])
def load_dataset(dataset):
if dataset == 'ORL3':
min_max_scaler = preprocessing.MinMaxScaler(feature_range=(-1,1))
data = scio.loadmat('dataset/ORL_mtv.mat')
X1 = min_max_scaler.fit_transform(np.transpose(data['X'][0][0]))
X2 = min_max_scaler.fit_transform(np.transpose(data['X'][0][1]))
X3 = min_max_scaler.fit_transform(np.transpose(data['X'][0][2]))
# X3 = np.transpose(data['X'][0][2])
Y = data['gt'] - 1
return X1, X2, X3, Y
if dataset == 'UCI6':
min_max_scaler = preprocessing.MinMaxScaler()
path = 'dataset/handwritten.mat'
data = scio.loadmat(path)
x1 = min_max_scaler.fit_transform(data['X'][0][0])
x2 = min_max_scaler.fit_transform(data['X'][0][1])
x3 = min_max_scaler.fit_transform(data['X'][0][2])
x4 = min_max_scaler.fit_transform(data['X'][0][3])
x5 = min_max_scaler.fit_transform(data['X'][0][4])
x6 = min_max_scaler.fit_transform(data['X'][0][5])
y = data['Y']
return x1, x2, x3, x4, x5, x6, y
if dataset == 'UCI2':
min_max_scaler = preprocessing.MinMaxScaler()
path = 'dataset/handwritten.mat'
data = scio.loadmat(path)
x1 = min_max_scaler.fit_transform(data['X'][0][0])
x2 = min_max_scaler.fit_transform(data['X'][0][1])
x3 = min_max_scaler.fit_transform(data['X'][0][2])
x4 = min_max_scaler.fit_transform(data['X'][0][3])
x5 = min_max_scaler.fit_transform(data['X'][0][4])
x6 = min_max_scaler.fit_transform(data['X'][0][5])
y = data['Y']
return x3, x5, y
if dataset == 'NUS5':
min_max_scaler = preprocessing.MinMaxScaler()
data = scio.loadmat('dataset/NUSWIDEOBJ.mat')
X1 = min_max_scaler.fit_transform(data['X'][0][0])
X2 = min_max_scaler.fit_transform(data['X'][0][1])
X3 = min_max_scaler.fit_transform(data['X'][0][2])
X4 = min_max_scaler.fit_transform(data['X'][0][3])
X5 = min_max_scaler.fit_transform(data['X'][0][4])
Y = data['Y']-1
return X1, X2, X3, X4, X5, Y
if dataset == 'caltech7':
min_max_scaler = preprocessing.MinMaxScaler(feature_range=(-1,1))
data = scio.loadmat('dataset/Caltech101-7.mat')
X1 = min_max_scaler.fit_transform(data['X'][0][0])
X2 = min_max_scaler.fit_transform(data['X'][0][1])
X3 = min_max_scaler.fit_transform(data['X'][0][2])
X4 = min_max_scaler.fit_transform(data['X'][0][3])
X5 = min_max_scaler.fit_transform(data['X'][0][4])
X6 = min_max_scaler.fit_transform(data['X'][0][5])
Y = data['Y'] - 1
return X1, X2, X3, X4, X5, X6, Y
if dataset == 'Cal2':
min_max_scaler = preprocessing.MinMaxScaler(feature_range=(-1,1))
data = scio.loadmat('dataset/Caltech101-7.mat')
#X1 = min_max_scaler.fit_transform(data['X'][0][0])
#X2 = min_max_scaler.fit_transform(data['X'][0][1])
#X3 = min_max_scaler.fit_transform(data['X'][0][2])
#X4 = min_max_scaler.fit_transform(data['X'][0][3])
X5 = min_max_scaler.fit_transform(data['X'][0][4])
X6 = min_max_scaler.fit_transform(data['X'][0][5])
Y = data['Y'] - 1
return X5, X6, Y
if dataset == 'ORL2':
min_max_scaler = preprocessing.MinMaxScaler(feature_range=(-1,1))
data = scio.loadmat('dataset/ORL_mtv.mat')
X1 = min_max_scaler.fit_transform(np.transpose(data['X'][0][0]))
X2 = min_max_scaler.fit_transform(np.transpose(data['X'][0][1]))
X3 = min_max_scaler.fit_transform(np.transpose(data['X'][0][2]))
# X3 = np.transpose(data['X'][0][2])
Y = data['gt'] - 1
return X2, X3, Y
def config_init(dataset):
if dataset == 'UCI6':
return [240, 76, 216, 47, 64, 6], 100, 10, 0.0001, 0.002, 10, 0.9, 0.9, 0.1, 'sigmoid'
if dataset == 'UCI2':
return [216, 64], 100, 10, 0.0001, 0.005, 10, 0.9, 0.9, 0.1, 'sigmoid'
if dataset == 'caltech7':
return [48, 40, 254, 1984, 512, 928], 100, 7, 0.0001, 0.05, 10, 0.9, 0.9, 1, 'linear'
if dataset == 'Cal2':
return [512, 928], 50, 7, 0.0001, 0.05, 5, 0.5, 0.5, 1, 'linear'
if dataset == 'ORL2':
return [3304, 6750], 50, 40, 0.0001, 0.01, 5, 0.9, 0.9, 1, 'linear'
if dataset == 'ORL3':
return [4096, 3304, 6750], 50, 40, 0.0001, 0.05, 10, 1, 1, 5, 'linear'
if dataset == 'NUS5':
return [65, 226, 145, 74, 129], 50, 31, 1e-6, 0.001, 5, 0.9, 0.9, 0.1, 'sigmoid'
class EpochBegin(Callback):
def __init__(self, sample_output, decay_n, gamma_output, decay_nn, decay_gmm, adam_nn, adam_gmm, inputs, Y, u_p, lambda_p, theta_p,
zeta):
self.sample_output = sample_output
self.decay_n = decay_n
self.gamma_output = gamma_output
self.decay_nn = decay_nn
self.decay_gmm = decay_gmm
self.adam_nn = adam_nn
self.adam_gmm = adam_gmm
self.inputs = inputs
self.Y = Y
self.u_p = u_p
self.lambda_p = lambda_p
self.theta_p = theta_p
self.zeta = zeta
def on_epoch_begin(self, epoch, logs={}):
self.epochBegin(epoch)
def plot_embedding(self, data, label, id):
x_min, x_max = np.min(data, 0), np.max(data, 0)
data = (data - x_min) / (x_max - x_min)
fig = plt.figure()
ax = plt.subplot(111)
for i in range(data.shape[0]):
plt.text(data[i, 0], data[i, 1], str(label[i]),
color=plt.cm.Set1(label[i] / 10.),
fontdict={'weight': 'bold', 'size': 9})
plt.axis('off')
plt.savefig('vis_%d' %id)
return fig
def epochBegin(self, epoch):
if epoch % self.decay_n == 0 and epoch != 0:
self.lr_decay()
gamma = self.gamma_output.predict(self.inputs, batch_size=batch_size)
pred = np.argmax(gamma, axis=1)
acc = self.cluster_acc(pred, self.Y)
Y = np.reshape(self.Y, [self.Y.shape[0]])
nmi = metrics.nmi(Y, pred)
ari = metrics.ari(Y, pred)
purity = self.purity_score(Y, pred)
global accuracy
accuracy = []
accuracy += [acc[0]]
if epoch > 0:
print ('ACC:%0.8f' % acc[0])
print ('NMI:', nmi)
print ('ARI:', ari)
print ('Purity', purity)
if epoch == 1 and dataset == 'har' and acc[0] < 0.77:
print ('=========== HAR dataset:bad init!Please run again! ============')
sys.exit(0)
def purity_score(self, y_true, y_pred):
# compute contingency matrix (also called confusion matrix)
contingency_matrix = mtr.cluster.contingency_matrix(y_true, y_pred)
# return purity
return np.sum(np.amax(contingency_matrix, axis=0)) / np.sum(contingency_matrix)
def cluster_acc(self, Y_pred, Y):
from sklearn.utils.linear_assignment_ import linear_assignment
assert Y_pred.size == Y.size
D = max(Y_pred.max(), Y.max()) + 1
w = np.zeros((D, D), dtype=np.int64)
for i in range(Y_pred.size):
w[Y_pred[i], Y[i]] += 1
ind = linear_assignment(w.max() - w)
return sum([w[i, j] for i, j in ind]) * 1.0 / Y_pred.size, w
def lr_decay(self):
if dataset == 'mnist' or dataset == 'bbcsport':
self.adam_nn.lr.set_value(self.floatX(max(self.adam_nn.lr.get_value() * self.decay_nn, 0.0002)))
self.adam_gmm.lr.set_value(self.floatX(max(self.adam_gmm.lr.get_value() * self.decay_gmm, 0.0002)))
else:
self.adam_nn.lr.set_value(self.floatX(self.adam_nn.lr.get_value() * self.decay_nn))
self.adam_gmm.lr.set_value(self.floatX(self.adam_gmm.lr.get_value() * self.decay_gmm))
print ('lr_nn:%f' % self.adam_nn.lr.get_value())
print ('lr_gmm:%f' % self.adam_gmm.lr.get_value())
def floatX(self, X):
return np.asarray(X, dtype=theano.config.floatX)
parser = argparse.ArgumentParser()
parser.add_argument('--dataset', dest='dataset', default='UCI2', choices=['UCI2', 'NUS5', 'caltech7', 'UCI6', 'Cal2', 'ORL3', 'ORL2'])
args = parser.parse_args()
dataset = args.dataset
ispretrain = True
batch_size = 200
latent_dim = 10
intermediate_dim = [500, 500, 2000]
# theano.config.floatX='float32'
if dataset in ['UCI2', 'Cal2', 'ORL2']:
X1, X2, Y = load_dataset(dataset)
X=[X1, X2]
elif dataset in ['ORL3']:
X1, X2, X3,Y = load_dataset(dataset)
X=[X1, X2, X3]
elif dataset in ['UCI6', 'caltech7']:
X1, X2, X3, X4, X5, X6, Y = load_dataset(dataset)
X=[X1, X2, X3, X4, X5, X6]
elif dataset in ['NUS5']:
X1, X2, X3, X4, X5, Y = load_dataset(dataset)
X=[X1, X2, X3, X4, X5]
num_views = len(X)
weights_path = 'MVCVAE_pretrain/'
vade = Multiview_VaDE(batch_size, num_views, latent_dim, intermediate_dim, config_init(dataset), weights_path, dataset,
ispretrain=True)
vade.compile(X,Y)
vade.train(X)