-
Notifications
You must be signed in to change notification settings - Fork 10
/
util.py
executable file
·293 lines (243 loc) · 8.84 KB
/
util.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
import h5py
import numpy as np
import os
import tensorflow as tf
import scipy.misc
import time
import logging
fw = tf.contrib.framework
dist = tf.contrib.distributions
def build_vimco_loss(cfg, l, log_q_h):
"""Builds negative VIMCO loss as in the paper.
Reference: Variational Inference for Monte Carlo Objectives, Algorithm 1
https://arxiv.org/abs/1602.06725
"""
k, b = l.get_shape().as_list() # n_samples, batch_size
kf = tf.cast(k, tf.float32)
if cfg['optim/geometric_mean']:
# implicit multi-sample objective (importance-sampled ELBO)
l_logsumexp = tf.reduce_logsumexp(l, [0], keep_dims=True)
L_hat = l_logsumexp - tf.log(kf)
else:
# standard ELBO
L_hat = tf.reduce_mean(l, [0], keep_dims=True)
s = tf.reduce_sum(l, 0, keep_dims=True)
diag_mask = tf.expand_dims(tf.diag(tf.ones([k], dtype=tf.float32)), -1)
off_diag_mask = 1. - diag_mask
diff = tf.expand_dims(s - l, 0) # expand for proper broadcasting
l_i_diag = 1. / (kf - 1.) * diff * diag_mask
l_i_off_diag = off_diag_mask * tf.stack([l] * k)
l_i = l_i_diag + l_i_off_diag
if cfg['optim/geometric_mean']:
L_hat_minus_i = tf.reduce_logsumexp(l_i, [1]) - tf.log(kf)
w = tf.stop_gradient(tf.exp((l - l_logsumexp)))
else:
L_hat_minus_i = tf.reduce_mean(l_i, [1])
w = 1.
local_l = tf.stop_gradient(L_hat - L_hat_minus_i)
if not cfg['optim/geometric_mean']:
# correction factor for multiplying by 1. / (kf - 1.) above
# to verify this, work out 2x2 matrix of samples by hand
local_l = local_l * k
loss = local_l * log_q_h + w * l
return loss / tf.to_float(b)
def relu_zeros(op, tag):
"""Calculate fraction of inputs to a layer that are zero."""
zeros = tf.reduce_mean(
tf.to_float(
tf.less(op.op.inputs[0], tf.cast(0., op.op.inputs[0].dtype))))
zeros = pprint(zeros, tag)
tf.summary.scalar(tag, zeros)
soft_zeros = tf.reduce_sum(tf.minimum(op.op.inputs[0], 0.))
soft_zeros = pprint(soft_zeros, tag)
return soft_zeros
def get_activation(name):
"""Return activation function."""
if name == 'relu':
return tf.nn.relu
elif name == 'tanh':
return tf.tanh
elif name == 'sigmoid':
return tf.sigmoid
elif name == 'elu':
return tf.nn.elu
def provide_data(data_config):
"""Provides batches of MNIST digits.
Args:
config: configuration object
Returns:
data_iterator: an iterator that returns numpy arrays of size [batch_size, 28, 28, 1]
data_mean: mean of the split
data_std: std of the split
"""
cfg = data_config
local_path = os.path.join(cfg['dir'], 'binarized_mnist.hdf5')
if not os.path.exists(local_path):
raise ValueError('need: ', local_path)
f = h5py.File(local_path, 'r')
if cfg['split'] == 'train_and_valid':
train = f['train'][:]
valid = f['valid'][:]
data = np.vstack([train, valid])
else:
data = f[cfg['split']][:]
try:
if cfg['fixed_idx'] is not None:
data = data[cfg['fixed_idx']:cfg['fixed_idx'] + 1]
except:
pass
n_examples = cfg['n_examples']
data = data[0:n_examples]
if 'perturb_data' in cfg:
if cfg['perturb_data']:
np.random.seed(2423232)
flip_indices = np.random.binomial(n=1, p=0.01, size=data.shape)
data = (data + flip_indices) % 2
data_mean = np.mean(data, axis=0)
data_std = np.std(data, axis=0)
def reshape(t): return np.reshape(t, (28, 28, 1))
data_mean = reshape(data_mean)
data_std = reshape(data_std)
# create indexes for the data points.
indexed_data = list(zip(range(len(data)), np.split(data, len(data))))
def data_iterator():
""" A simple data iterator """
batch_idx = 0
batches_per_epoch = int(np.floor(n_examples / cfg['batch_size']))
while True:
# shuffle data
idxs = np.arange(0, len(data))
np.random.shuffle(idxs)
shuf_data = [indexed_data[idx] for idx in idxs]
for n in range(0, batches_per_epoch):
batch_idx = n * cfg['batch_size']
indexed_images_batch = shuf_data[batch_idx:batch_idx +
cfg['batch_size']]
indexes, images_batch = zip(*indexed_images_batch)
images_batch = np.vstack(images_batch)
images_batch = images_batch.reshape(
(cfg['batch_size'], 28, 28, 1))
# yield indexes, images_batch
yield images_batch
return data_iterator(), data_mean, data_std
def remove_dir(config):
"""Delete directory contents if it exists."""
cfg = config
for f in tf.gfile.ListDirectory(cfg['log/dir']):
path = os.path.join(cfg['log/dir'], f)
if os.path.isdir(path):
tf.gfile.DeleteRecursively(path)
else:
tf.gfile.Remove(path)
def make_logdir(cfg, logdir_name):
"""Create a date directory and experiment directory."""
date = time.strftime("%Y-%m-%d")
project_dir = os.path.join(cfg['log/dir'], 'proximity_vi')
date_dir = os.path.join(project_dir, date)
experiment_dir = os.path.join(date_dir, cfg['log/experiment'])
if not os.path.exists(experiment_dir):
os.makedirs(experiment_dir)
train_dir = os.path.join(experiment_dir, logdir_name)
if tf.gfile.Exists(train_dir) and cfg['log/clear_dir']:
for f in tf.gfile.ListDirectory(train_dir):
tf.gfile.Remove(os.path.join(train_dir, f))
else:
tf.gfile.MakeDirs(train_dir)
with open(os.path.join(train_dir, 'config.yml'), 'w') as f:
f.write(str(cfg))
return train_dir + '/'
def save_prior_posterior_predictives(
cfg, sess, inference, prior_predictive, posterior_predictive, feed_dict,
images):
"""Save prior and posterior samples."""
np_step = sess.run(inference.global_step)
np_prior_predictive = sess.run(prior_predictive, feed_dict)
np_posterior_predictive = sess.run(posterior_predictive, feed_dict)
for k in range(cfg['batch_size']):
im_name = 'i_%d_k_%d_' % (np_step, k)
orig_name = im_name + 'original.jpg'
scipy.misc.imsave(os.path.join(cfg['log/dir'], orig_name),
images[k, :, :, 0])
n = 0
if len(np_prior_predictive.shape) == 5:
def get_sample(x): return x[n, k, :, :, 0]
elif len(np_prior_predictive.shape) == 4:
def get_sample(x): return x[k, :, :, 0]
prior_name = im_name + 'prior_predictive_%d.jpg' % n
posterior_name = im_name + 'posterior_predictive_%d.jpg' % n
scipy.misc.imsave(os.path.join(cfg['log/dir'], prior_name),
get_sample(np_prior_predictive))
scipy.misc.imsave(os.path.join(cfg['log/dir'], posterior_name),
get_sample(np_posterior_predictive))
class empty_scope():
"""Empty scope helper."""
def __init__(self):
pass
def __enter__(self):
pass
def __exit__(self, type, value, traceback):
pass
def get_or_create_scope(name, reuse=False):
"""Create a scope or, if it exists, stay in a scope (return empty scope)."""
scope = tf.get_variable_scope()
if scope.name == name:
return empty_scope()
else:
return tf.variable_scope(name, reuse=reuse)
def identity_initializer():
"""Identity initialization for unitary eigenvalues."""
def _initializer(shape, dtype=tf.float32):
if len(shape) == 1:
return tf.constant_op.constant(0., dtype=dtype, shape=shape)
elif len(shape) == 2 and shape[0] == shape[1]:
return tf.constant_op.constant(np.identity(shape[0], dtype))
elif len(shape) == 4 and shape[2] == shape[3]:
array = np.zeros(shape, dtype=float)
cx, cy = shape[0] / 2, shape[1] / 2
for i in range(shape[2]):
array[cx, cy, i, i] = 1
return tf.constant_op.constant(array, dtype=dtype)
else:
raise
return _initializer
def get_initializer(name):
if name == 'identity':
return identity_initializer()
elif name == 'orthogonal':
return tf.orthogonal_initializer(gain=1.0)
elif name == 'truncated_normal':
return tf.truncated_normal_initializer()
def softplus(x):
return np.log(np.exp(x) + 1.)
def inv_softplus(x):
return np.log(np.exp(x) - 1.)
def tf_inv_softplus(x):
return tf.log(tf.exp(x) - 1.)
def latest_checkpoint(cfg):
if cfg['ckpt_to_restore'] is not None:
ckpt = cfg['ckpt_to_restore']
else:
ckpt = tf.train.latest_checkpoint(cfg['log/dir'])
if ckpt is not None:
print('restoring from: ', ckpt)
print(fw.list_variables(ckpt))
return ckpt
def list_to_str(lst):
return '_'.join([str(x) for x in lst])
def tensor_name(tensor):
return tensor.name.split(':')[0]
def get_scalar_var(name):
return tf.get_variable(
name, shape=[], dtype=tf.float32,
initializer=tf.zeros_initializer(), trainable=False)
def norm(t):
return tf.sqrt(tf.nn.l2_loss(t) * 2)
def log_to_file(filename):
logging.basicConfig(level=logging.INFO,
format='%(asctime)s %(name)-4s %(levelname)-4s %(message)s',
datefmt='%m-%d %H:%M',
filename=filename,
filemode='a')
console = logging.StreamHandler()
console.setLevel(logging.INFO)
logging.getLogger('').addHandler(console)