-
Notifications
You must be signed in to change notification settings - Fork 4
/
bayesscdc_unsup.py
357 lines (306 loc) · 13.4 KB
/
bayesscdc_unsup.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
#!/usr/bin/env python
# -*- coding: utf-8 -*-
"""
BayesSCDC model for MNIST without annotations, i.e., unsupervised clustering.
"""
from __future__ import absolute_import
from __future__ import print_function
from __future__ import division
import os
import time
from six.moves import range, zip
import tensorflow as tf
import numpy as np
import zhusuan as zs
from sklearn.metrics import normalized_mutual_info_score, adjusted_mutual_info_score
from utils import dataset, setup_logger, save_image_collections, cluster_acc
from distributions import niw, catgorical, mvn, dirichlet
from distributions import normalize, exp_family_kl
FLAGS = tf.flags.FLAGS
tf.flags.DEFINE_integer("seed", 1234, """Random seed.""")
def get_global_params(scope, d, K, alpha, niw_conc, random_scale=None,
trainable=True):
def init_niw_param():
# nu: scalar, S: (d, d), m: (d,) kappa: scalar
# TODO: nu different to orig code, different init of nu, S?
nu, S, m, kappa = (tf.constant(d + niw_conc, dtype=tf.float32),
(d + niw_conc) * tf.eye(d),
tf.zeros(d),
tf.constant(niw_conc, dtype=tf.float32))
if random_scale:
m = m + random_scale * tf.random_normal(m.shape)
return niw.standard_to_natural(m, kappa, S, nu)
with tf.variable_scope(scope, reuse=tf.AUTO_REUSE):
if random_scale:
dir_nat = tf.random_uniform([K], minval=0, maxval=alpha - 1.)
else:
dir_nat = tf.ones([K]) * (alpha - 1.)
# [K]
dir_params = tf.get_variable(
"dir_params", dtype=tf.float32, initializer=dir_nat,
trainable=trainable)
# [K, d + d^2 + 2]
niw_nat = tf.stack([init_niw_param() for _ in range(K)])
niw_params = tf.get_variable(
"niw_params", dtype=tf.float32, initializer=niw_nat,
trainable=trainable)
return dir_params, niw_params
def global_expected_stats(global_params, d):
dir_params, niw_params = global_params
# [K]
dir_stats = dirichlet.expected_stats(dir_params)
# [K, d + d^2 + 2]
niw_stats = niw.expected_stats(niw_params, d)
return dir_stats, niw_stats
@zs.reuse("encoder")
def encoder(o, d):
h = tf.layers.dense(tf.to_float(o), 500, activation=tf.nn.relu)
h = tf.layers.dense(h, 500, activation=tf.nn.relu)
x_param_1 = tf.layers.dense(h, d)
x_sigma_inv = tf.layers.dense(h, d, activation=tf.nn.softplus)
x_param_2 = -0.5 * x_sigma_inv
return x_param_1, x_param_2
@zs.reuse("decoder")
def decoder(x, o_dim):
h = tf.layers.dense(x, 500, activation=tf.nn.relu)
h = tf.layers.dense(h, 500, activation=tf.nn.relu)
o_logits = tf.layers.dense(h, o_dim)
o_dist = zs.distributions.Bernoulli(o_logits, group_ndims=1)
return o_dist, tf.sigmoid(o_logits)
def x_mean_field(niw_stats, z_stats, x_obs_param, d):
# niw_stats: [K, d + d^2 + 2], z_stats: [M, K]
# x_prior_term: [M, d + d^2]
x_prior_term = tf.matmul(z_stats, niw_stats[:, :-2])
# x_obs_param: [M, d + d^2]
# x_nat_param: [M, d + d^2]
x_nat_param = x_prior_term + x_obs_param
# x_stats: [M, d + d^2]
x_stats = mvn.expected_stats(x_nat_param, d)
return x_nat_param, x_stats
def z_mean_field(global_stats, x_stats):
# dir_stats: [K]
dir_stats, niw_stats = global_stats
# x_stats: [M, d + d^2], niw_stats: [K, d + d^2 + 2]
M = tf.shape(x_stats)[0]
# x_stats_pad: [M, d + d^2 + 2]
x_stats_pad = tf.concat([x_stats, tf.ones([M, 2])], axis=-1)
# z_nat_param: [M, K]
z_nat_param = dir_stats + tf.matmul(x_stats_pad, niw_stats,
transpose_b=True)
# z_stats: [M, K]
z_stats = catgorical.expected_stats(z_nat_param)
return z_nat_param, z_stats
def local_kl_z(z_nat_param, dir_stats, z_stats):
# z_nat_param: [M, K]
# dir_stats: [K]
# z_stats: [M, K]
z_nat_param = z_nat_param - tf.reduce_logsumexp(z_nat_param, axis=-1,
keepdims=True)
nat_param_diff = z_nat_param - dir_stats
# ret: [M]
return exp_family_kl(nat_param_diff, z_stats)
def local_kl_x(x_nat_param, niw_stats, z_stats, x_stats, d):
# x_nat_param: [M, d + d^2]
# niw_stats: [K, d + d^2 + 2]
# z_stats: [M, K]
# x_stats: [M, d + d^2]
# x_prior_term: [M, d + d^2 + 2]
x_prior_term = tf.matmul(z_stats, niw_stats)
# nat_param_diff: [M, d + d^2]
nat_param_diff = x_nat_param - x_prior_term[:, :-2]
# log_partition_diff: [M]
log_z_diff = mvn.log_partition(x_nat_param, d) + tf.reduce_sum(
x_prior_term[:, -2:], axis=-1)
# ret: [M]
return exp_family_kl(nat_param_diff, x_stats, log_z_diff=log_z_diff)
def global_kl(prior_global_params, global_params, global_stats, d):
prior_dir_param, prior_niw_param = prior_global_params
def _kl_helper(log_partition, param, prior_param, stats):
nat_diff = param - prior_param
log_z_diff = log_partition(param) - log_partition(prior_param)
return exp_family_kl(nat_diff, stats, log_z_diff=log_z_diff)
# dir_param: [K], niw_param: [K, d + d^2 + 2]
dir_param, niw_param = global_params
# dir_stats: [K], niw_stats: [K, d + d^2 + 2]
dir_stats, niw_stats = global_stats
# dir_kl: []
dir_kl = _kl_helper(dirichlet.log_partition, dir_param, prior_dir_param,
dir_stats)
# niw_kl: [K]
niw_kl = _kl_helper(lambda x: niw.log_partition(x, d), niw_param,
prior_niw_param, niw_stats)
return dir_kl + tf.reduce_sum(niw_kl, axis=0)
def elbo(log_po_term, local_kl_z, local_kl_x, global_kl, N):
# log_po_term: [M]
# local_kl_z: [M], local_kl_x: [M]
# global_kl: []
obj = tf.reduce_mean(log_po_term - local_kl_z - local_kl_x) - global_kl / N
# ret: []
return obj
def variational_message_passing(prior_global_params, global_params,
o, o_dim, d, K, N, n_iters=100):
global_stats = global_expected_stats(global_params, d)
dir_stats, niw_stats = global_stats
M = tf.shape(o)[0]
z_stats = normalize(tf.random_uniform([M, K], 1e-8, maxval=1))
# h: [M, d], J: [M, d]
h, J = encoder(o, d)
# J: [M, d * d]
J = tf.reshape(tf.matrix_diag(J), [M, d * d])
# x_obs_param: [M, d + d * d]
x_obs_param = tf.concat([h, J], axis=-1)
for t in range(n_iters):
x_nat_param, x_stats = x_mean_field(niw_stats, z_stats, x_obs_param, d)
z_nat_param, z_stats = z_mean_field(global_stats, x_stats)
# x: [M, d]
x = mvn.sample(x_nat_param, d)
o_dist, _ = decoder(x, o_dim)
# log_po_term: [M]
log_po_term = o_dist.log_prob(o)
# log_kl_x_term: [M]
local_kl_x_term = local_kl_x(x_nat_param, niw_stats, z_stats, x_stats, d)
# log_kl_z_term: [M]
local_kl_z_term = local_kl_z(z_nat_param, dir_stats, z_stats)
# global_kl_term: []
global_kl_term = global_kl(
prior_global_params, global_params, global_stats, d)
lower_bound = elbo(log_po_term, local_kl_z_term, local_kl_x_term,
global_kl_term, N)
# Natural gradient for global variational parameters
# z_stats: [M, K], x_stats: [M, d + d^2]
# dir_updates: [K]
dir_updates = tf.reduce_mean(z_stats, axis=0)
# niw_updates: [K, d + d^2 + 2]
niw_updates = tf.matmul(z_stats, tf.concat([x_stats, tf.ones([M, 2])], -1),
transpose_a=True) / tf.to_float(M)
updates = (dir_updates, niw_updates)
nat_grads = [(prior_global_params[i] - global_params[i]) / N + updates[i]
for i in range(2)]
return lower_bound, nat_grads, z_stats, dir_stats, niw_stats
def main():
seed = FLAGS.seed
result_path = "results/mnist_{}_{}".format(time.strftime("%Y%m%d_%H%M%S"), seed)
logger = setup_logger('mnist', __file__, result_path)
np.random.seed(seed)
tf.set_random_seed(seed)
# Load MNIST
data_path = os.path.join('data', 'mnist.pkl.gz')
o_train, t_train, o_valid, t_valid, o_test, t_test = \
dataset.load_mnist_realval(data_path, one_hot=False)
o_train = np.vstack([o_train, o_valid])
t_train = np.hstack([t_train, t_valid])
o_test = np.random.binomial(1, o_test, size=o_test.shape)
n_train, o_dim = o_train.shape
n_test, _ = o_test.shape
# n_class = np.max(t_test) + 1
# Prior parameters
d = 8
K = 50
prior_alpha = 1.05
prior_niw_conc = 0.5
# Variational initialization
alpha = 2.
niw_conc = 1.
random_scale = 3.
# learning rate
learning_rate = 1e-3
nat_grad_scale = 1e4
prior_global_params = get_global_params("prior", d, K, prior_alpha,
prior_niw_conc, trainable=False)
global_params = get_global_params("variational", d, K, alpha, niw_conc,
random_scale=random_scale, trainable=True)
# n_particles = tf.placeholder(tf.int32, shape=[], name='n_particles')
o_input = tf.placeholder(tf.float32, shape=[None, o_dim], name='o')
o = tf.to_int32(tf.random_uniform(tf.shape(o_input)) <= o_input)
lower_bound, global_nat_grads, z_stats, dir_stats, niw_stats = \
variational_message_passing(prior_global_params, global_params,
o, o_dim, d, K, n_train, n_iters=4)
z_pred = tf.argmax(z_stats, axis=-1)
optimizer = tf.train.MomentumOptimizer(learning_rate=learning_rate,
momentum=0.9)
net_vars = (tf.trainable_variables(scope="encoder") +
tf.trainable_variables(scope="decoder"))
net_grads_and_vars = optimizer.compute_gradients(
-lower_bound, var_list=net_vars)
global_nat_grads = [-nat_grad_scale * g for g in global_nat_grads]
global_grads_and_vars = list(zip(global_nat_grads, global_params))
infer_op = optimizer.apply_gradients(net_grads_and_vars +
global_grads_and_vars)
# Generation
# niw_stats: [K, d + d^2 + 2]
gen_mvn_params = niw_stats[:, :-2]
# transparency: [K]
transp = tf.exp(dir_stats) / tf.reduce_max(tf.exp(dir_stats))
# x_samples: [K, d, 10]
x_samples = mvn.sample(gen_mvn_params, d, n_samples=10)
# o_mean: [10, K, o_dim]
_, o_mean = decoder(tf.transpose(x_samples, [2, 0, 1]), o_dim)
# o_gen: [10 * K, 28, 28, 1]
o_gen = tf.reshape(o_mean * transp[:, None], [-1, 28, 28, 1])
# Define training parameters
epochs = 200
batch_size = 128
iters = o_train.shape[0] // batch_size
save_freq = 1
test_freq = 10
test_batch_size = 400
test_iters = o_test.shape[0] // test_batch_size
def _evaluate(pred_batches, labels):
preds = np.hstack(pred_batches)
truths = labels[:preds.size]
acc, _ = cluster_acc(preds, truths)
nmi = adjusted_mutual_info_score(truths, labels_pred=preds)
return acc, nmi
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
for epoch in range(1, epochs + 1):
time_epoch = -time.time()
indices = np.random.permutation(n_train)
o_train = o_train[indices]
t_train = t_train[indices]
lbs = []
t_preds = []
for t in range(iters):
o_batch = o_train[t * batch_size:(t + 1) * batch_size]
_, lb, t_pred = sess.run(
[infer_op, lower_bound, z_pred],
feed_dict={o_input: o_batch})
# print("lb: {}".format(lb))
lbs.append(lb)
t_preds.append(t_pred)
time_epoch += time.time()
train_acc, train_nmi = _evaluate(t_preds, t_train)
logger.info(
'Epoch {} ({:.1f}s): Lower bound = {}, acc = {}, nmi = {}'
.format(epoch, time_epoch, np.mean(lbs), train_acc, train_nmi))
if epoch % test_freq == 0:
time_test = -time.time()
test_lbs = []
test_t_preds = []
for t in range(test_iters):
test_o_batch = o_test[t * test_batch_size:
(t + 1) * test_batch_size]
test_lb, test_t_pred = sess.run([lower_bound, z_pred],
feed_dict={o: test_o_batch})
test_lbs.append(test_lb)
test_t_preds.append(test_t_pred)
time_test += time.time()
test_acc, test_nmi = _evaluate(test_t_preds, t_test)
logger.info('>>> TEST ({:.1f}s)'.format(time_test))
logger.info('>> Test lower bound = {}, acc = {}, nmi = {}'
.format(np.mean(test_lbs), test_acc, test_nmi))
if epoch == epochs:
with open('results/mnist_bayesSCDC_unsup.txt', "a") as myfile:
myfile.write("seed: %d train_acc: %f train_nmi: %f "
"test_acc: %f test_nmi: %f" % (
seed, train_acc, train_nmi, test_acc, test_nmi))
myfile.write('\n')
myfile.close()
if epoch % save_freq == 0:
logger.info('Saving images...')
images = sess.run(o_gen)
name = os.path.join(result_path,
"vae.epoch.{}.png".format(epoch))
save_image_collections(images, name, shape=(10, K))
if __name__ == "__main__":
main()