forked from benrhodes26/tre_code
-
Notifications
You must be signed in to change notification settings - Fork 0
/
losses.py
288 lines (225 loc) · 10.6 KB
/
losses.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
import tensorflow_probability as tfp
tfd = tfp.distributions
from utils.tf_utils import *
class LogisticLoss:
def __init__(self, nu, label_smoothing_alpha=0.0, one_sided_smoothing=True):
self.nu = nu
self.log_nu = tf.log(nu)
self.class1_acc = None
self.class2_acc = None
self.acc = None
self.dawid_statistic_numerator = None
self.dawid_statistic_denominator = None
assert 0 <= label_smoothing_alpha < 0.5, "label smoothing parameter should be between 0 & 0.5"
self.ls_alpha = label_smoothing_alpha
self.one_sided_smoothing = one_sided_smoothing
@tf_var_scope
def loss(self, neg_energy):
"""Returns average over K Logistic losses given negative energies of model
Args:
neg_energy: (2 * n_batch, n_ratios)
"""
neg_energy1, neg_energy2 = tf.split(neg_energy, num_or_size_splits=2, axis=0) # each (n, n_losses)
term1 = tf.log_sigmoid(neg_energy1 - self.log_nu) # (n, n_losses)
term2 = tf.log_sigmoid(self.log_nu - neg_energy2) # (n, n_losses)
if self.ls_alpha > 0:
term1, term2 = self.apply_label_smoothing(neg_energy1, neg_energy2, term1, term2)
self.compute_classification_acc(term1, term2)
self.compute_dawid_statistic(term1, term2)
loss = -tf.reduce_mean(term1, axis=0) - self.nu * tf.reduce_mean(term2, axis=0) # (n_losses, )
return loss, -term1, -self.nu*term2
def apply_label_smoothing(self, neg_energy1, neg_energy2, term1, term2):
term1 *= (1 - self.ls_alpha)
term1 += self.ls_alpha * tf.log_sigmoid(-neg_energy1)
if not self.one_sided_smoothing:
term2 *= (1 - self.ls_alpha)
term2 += self.ls_alpha * tf.log_sigmoid(neg_energy2)
return term1, term2
def compute_classification_acc(self, term1, term2):
class1_scores = tf.where(term1 > -tf.log(2.), tf.ones_like(term1, dtype=tf.float32),
tf.zeros_like(term1, dtype=tf.float32))
class2_scores = tf.where(term2 > -tf.log(2.), tf.ones_like(term2, dtype=tf.float32),
tf.zeros_like(term2, dtype=tf.float32))
self.class1_acc = tf.reduce_mean(class1_scores, axis=0) # (n_losses,)
self.class2_acc = tf.reduce_mean(class2_scores, axis=0) # (n_losses,)
self.acc = 0.5 * (self.class1_acc + self.class2_acc) # (n_losses,)
def compute_dawid_statistic(self, term1, term2):
"""Dawid, A. P. Prequential analysis. Encyclopedia of Statistical Sciences, 1:464–470, 1997"""
num_class1 = tf.cast(shape_list(term1)[0], tf.float32)
class1_p1 = tf.exp(term1)
class2_p1 = 1 - tf.exp(term2)
sum_class1_p1 = tf.reduce_sum(class1_p1, axis=0) # (n_losses)
sum_class2_p1 = tf.reduce_sum(class2_p1, axis=0) # (n_losses)
class1_bernoulli_var = tf.reduce_sum(class1_p1 * (1 - class1_p1), axis=0) # (n_losses)
class2_bernoulli_var = tf.reduce_sum(class2_p1 * (1 - class2_p1), axis=0) # (n_losses)
self.dawid_statistic_numerator = num_class1 - sum_class1_p1 - sum_class2_p1 # (n_losses)
self.dawid_statistic_denominator = class1_bernoulli_var + class2_bernoulli_var # (n_losses)
class DVLoss:
def __init__(self):
self.term1 = None
self.term2 = None
@tf_var_scope
def loss(self, neg_energy):
"""Returns Donsker-Varadhan loss
Args:
neg_energy: (2 * n_batch,) - a.k.a the log-density ratio
"""
neg_energy1, neg_energy2 = tf.split(neg_energy, num_or_size_splits=2, axis=0) # each (n_ratios,)
n_samples = shape_list(neg_energy2)
self.term1 = -tf.reduce_mean(neg_energy1)
self.term2 = tf.reduce_logsumexp(neg_energy2) - tf.log(n_samples)
return self.term1 + self.term2
class NWJLoss:
def __init__(self):
self.term1 = None
self.term2 = None
self.acc = None
self.class1_acc = None
self.class2_acc = None
self.dawid_statistic_numerator = None
self.dawid_statistic_denominator = None
@tf_var_scope
def loss(self, neg_energy):
"""Returns Donsker-Varadhan loss
Args:
neg_energy: (2 * n_batch,) -
"""
neg_energy1, neg_energy2 = tf.split(neg_energy, num_or_size_splits=2, axis=0) # each (n_ratios,)
class1_scores = tf.where(neg_energy1 > 0.0, tf.ones_like(neg_energy1, dtype=tf.float32),
tf.zeros_like(neg_energy1, dtype=tf.float32))
class2_scores = tf.where(neg_energy2 < 0.0, tf.ones_like(neg_energy2, dtype=tf.float32),
tf.zeros_like(neg_energy2, dtype=tf.float32))
self.class1_acc = tf.reduce_mean(class1_scores, axis=0) # (n_ratios,)
self.class2_acc = tf.reduce_mean(class2_scores, axis=0) # (n_ratios,)
self.dawid_statistic_numerator = tf.zeros_like(self.class1_acc)
self.dawid_statistic_denominator = tf.zeros_like(self.class1_acc)
self.acc = 0.5 * (self.class1_acc + self.class2_acc) # (n_ratios,)
self.term1 = -tf.reduce_mean(neg_energy1, axis=0) - 1
self.term2 = tf.reduce_mean(tf.exp(neg_energy2), axis=0)
return self.term1 + self.term2, -neg_energy1-1, tf.exp(neg_energy2)
class LSQLoss:
def __init__(self):
self.term1 = None
self.term2 = None
self.acc = None
self.class1_acc = None
self.class2_acc = None
self.dawid_statistic_numerator = None
self.dawid_statistic_denominator = None
@tf_var_scope
def loss(self, neg_energy):
"""Returns least-square losss
Args:
neg_energy: (2 * n_batch,) -
"""
neg_energy1, neg_energy2 = tf.split(neg_energy, num_or_size_splits=2, axis=0) # each (n_ratios,)
class1_scores = tf.where(neg_energy1 > 0.0, tf.ones_like(neg_energy1, dtype=tf.float32),
tf.zeros_like(neg_energy1, dtype=tf.float32))
class2_scores = tf.where(neg_energy2 < 0.0, tf.ones_like(neg_energy2, dtype=tf.float32),
tf.zeros_like(neg_energy2, dtype=tf.float32))
self.class1_acc = tf.reduce_mean(class1_scores, axis=0) # (n_ratios,)
self.class2_acc = tf.reduce_mean(class2_scores, axis=0) # (n_ratios,)
self.dawid_statistic_numerator = tf.zeros_like(self.class1_acc)
self.dawid_statistic_denominator = tf.zeros_like(self.class1_acc)
self.acc = 0.5 * (self.class1_acc + self.class2_acc) # (n_ratios,)
term1 = 0.5 * (tf.sigmoid(neg_energy1) - 1)**2
term2 = 0.5 * tf.sigmoid(neg_energy2)**2
loss = tf.reduce_mean(term1, axis=0) + tf.reduce_mean(term2, axis=0)
return loss, term1, term2
class CNCELoss:
def __init__(self, cnce_loss, model_type):
assert model_type == "inverted", "This loss function has currently only be designed" \
"to work for an 'inverted' model."
self.cnce_loss = cnce_loss
self.c = 0
self.data_acc = -1
self.noise_acc = -1 # ill-defined for CNCE
self.acc = -1 # ill-defined for CNCE
@tf_var_scope
def loss(self, condprob_data_given_noise, condprob_noise_given_data, ebm_log_p_data, ebm_log_p_noise):
if self.cnce_loss == "symmetric":
log_class_prob = tf.nn.softplus(ebm_log_p_noise - ebm_log_p_data)
else:
log_class_prob = tf.nn.softplus(ebm_log_p_noise + condprob_data_given_noise
- ebm_log_p_data - condprob_noise_given_data)
self.acc = tf.reduce_mean(tf.where(log_class_prob < tf.log(2.),
tf.ones_like(log_class_prob, dtype=tf.float32),
tf.zeros_like(log_class_prob, dtype=tf.float32)))
return 2 * tf.reduce_mean(log_class_prob)
class ScoreMatchingLoss:
def __init__(self):
pass
@tf_var_scope
def loss(self, log_prob, unstacked_data):
"""Compute the SM loss, looping over each dimension"""
n, n_dims = tf.shape(unstacked_data[0])[0], len(unstacked_data)
log_prob = tf.squeeze(log_prob)
ssq = tf.constant(0., dtype=tf.float32) # score function squared
lap = tf.constant(0., dtype=tf.float32) # laplacian
for i in range(n_dims):
feat = unstacked_data[i]
score = tf.gradients(log_prob, feat)[0]
ssq += tf.reduce_sum(score ** 2)
lap += tf.reduce_sum(tf.gradients(tf.reduce_sum(score), feat)[0])
return tf.cast(1 / n, tf.float32) * (lap + 0.5 * ssq)
# class MMDLoss:
#
# def __init__(self, sigma):
# self.sigma = sigma
#
# params = {
# "batch_size": 64,
# "image_dim": 32 * 32 * 3,
# "c": 3,
# "h": 32,
# "w": 32
# }
#
# def makeScaleMatrix(self, num_gen, num_orig):
#
# # first 'N' entries have '1/N', next 'M' entries have '-1/M'
# s1 = tf.constant(1.0 / num_gen, shape=[num_gen, 1])
# s2 = -tf.constant(1.0 / num_orig, shape=[num_orig, 1])
#
# return tf.concat([s1, s2], 0)
#
# def computeMMD(self, x, gen_x, sigma = [1, 2, 4, 8, 16]):
#
# x = slim.flatten(x)
# gen_x = slim.flatten(gen_x)
#
# # concatenation of the generated images and images from the dataset
# # first 'N' rows are the generated ones, next 'M' are from the data
# X = tf.concat([gen_x, x],0)
#
# # dot product between all combinations of rows in 'X'
# XX = tf.matmul(X, tf.transpose(X))
#
# # dot product of rows with themselves
# X2 = tf.reduce_sum(X * X, 1, keep_dims = True)
#
# # exponent entries of the RBF kernel (without the sigma) for each
# # combination of the rows in 'X'
# # -0.5 * (x^Tx - 2 * x^Ty + y^Ty)
# exponent = XX - 0.5 * X2 - 0.5 * tf.transpose(X2)
#
# # scaling constants for each of the rows in 'X'
# s = makeScaleMatrix(params['batch_size'], params['batch_size'])
#
# # scaling factors of each of the kernel values, corresponding to the
# # exponent values
# S = tf.matmul(s, tf.transpose(s))
#
# mmd_sq = 0
#
# # for each bandwidth parameter, compute the MMD value and add them all
# n = params['batch_size']
# n_sq = float(n * n)
# for i in range(len(sigma)):
#
# # kernel values for each combination of the rows in 'X'
# kernel_val = tf.exp(1.0 / sigma[i] * exponent)
#
# mmd_sq += tf.reduce_sum(S * kernel_val)
#
# return tf.sqrt(mmd_sq)