-
Notifications
You must be signed in to change notification settings - Fork 19
/
Copy pathlosses.py
236 lines (206 loc) · 8.49 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
import torch
LOG_EPSILON = 1e-5
'''
helper functions
'''
def neg_log(x):
return - torch.log(x + LOG_EPSILON)
def log_loss(preds, targs):
return targs * neg_log(preds)
def expected_positive_regularizer(preds, expected_num_pos, norm='2'):
# Assumes predictions in [0,1].
if norm == '1':
reg = torch.abs(preds.sum(1).mean(0) - expected_num_pos)
elif norm == '2':
reg = (preds.sum(1).mean(0) - expected_num_pos)**2
else:
raise NotImplementedError
return reg
'''
loss functions
'''
def loss_bce(batch, P, Z):
# unpack:
preds = batch['preds']
observed_labels = batch['label_vec_obs']
# input validation:
assert not torch.any(observed_labels == -1)
assert P['train_set_variant'] == 'clean'
# compute loss:
loss_mtx = torch.zeros_like(observed_labels)
loss_mtx[observed_labels == 1] = neg_log(preds[observed_labels == 1])
loss_mtx[observed_labels == 0] = neg_log(1.0 - preds[observed_labels == 0])
reg_loss = None
return loss_mtx, reg_loss
def loss_bce_ls(batch, P, Z):
# unpack:
preds = batch['preds']
observed_labels = batch['label_vec_obs']
# input validation:
assert not torch.any(observed_labels == -1)
assert P['train_set_variant'] == 'clean'
# compute loss:
loss_mtx = torch.zeros_like(observed_labels)
loss_mtx[observed_labels == 1] = (1.0 - P['ls_coef']) * neg_log(preds[observed_labels == 1]) + P['ls_coef'] * neg_log(1.0 - preds[observed_labels == 1])
loss_mtx[observed_labels == 0] = (1.0 - P['ls_coef']) * neg_log(1.0 - preds[observed_labels == 0]) + P['ls_coef'] * neg_log(preds[observed_labels == 0])
reg_loss = None
return loss_mtx, reg_loss
def loss_iun(batch, P, Z):
# unpack:
preds = batch['preds']
observed_labels = batch['label_vec_obs']
true_labels = batch['label_vec_true']
# input validation:
assert torch.min(observed_labels) >= 0
# compute loss:
loss_mtx = torch.zeros_like(observed_labels)
loss_mtx[observed_labels == 1] = neg_log(preds[observed_labels == 1])
loss_mtx[true_labels == -1] = neg_log(1.0 - preds[true_labels == -1]) # This loss gets unrealistic access to true negatives.
reg_loss = None
return loss_mtx, reg_loss
def loss_iu(batch, P, Z):
# unpack:
preds = batch['preds']
observed_labels = batch['label_vec_obs']
# input validation:
assert torch.any(observed_labels == 1) # must have at least one observed positive
assert torch.any(observed_labels == -1) # must have at least one observed negative
# compute loss:
loss_mtx = torch.zeros_like(observed_labels)
loss_mtx[observed_labels == 1] = neg_log(preds[observed_labels == 1])
loss_mtx[observed_labels == -1] = neg_log(1.0 - preds[observed_labels == -1])
reg_loss = None
return loss_mtx, reg_loss
def loss_pr(batch, P, Z):
# unpack:
preds = batch['preds']
observed_labels = batch['label_vec_obs']
batch_size = int(batch['label_vec_obs'].size(0))
num_classes = int(batch['label_vec_obs'].size(1))
# input validation:
assert torch.min(observed_labels) >= 0
# compute loss:
loss_mtx = torch.zeros_like(observed_labels)
for n in range(batch_size):
preds_neg = preds[n, :][observed_labels[n, :] == 0]
for i in range(num_classes):
if observed_labels[n, i] == 1:
torch.nonzero(observed_labels[n, :])
loss_mtx[n, i] = torch.sum(torch.clamp(1.0 - preds[n, i] + preds_neg, min=0))
reg_loss = None
return loss_mtx, reg_loss
def loss_an(batch, P, Z):
# unpack:
preds = batch['preds']
observed_labels = batch['label_vec_obs']
# input validation:
assert torch.min(observed_labels) >= 0
# compute loss:
loss_mtx = torch.zeros_like(observed_labels)
loss_mtx[observed_labels == 1] = neg_log(preds[observed_labels == 1])
loss_mtx[observed_labels == 0] = neg_log(1.0 - preds[observed_labels == 0])
reg_loss = None
return loss_mtx, reg_loss
def loss_an_ls(batch, P, Z):
# unpack:
preds = batch['preds']
observed_labels = batch['label_vec_obs']
# input validation:
assert torch.min(observed_labels) >= 0
# compute loss:
loss_mtx = torch.zeros_like(observed_labels)
loss_mtx[observed_labels == 1] = (1.0 - P['ls_coef']) * neg_log(preds[observed_labels == 1]) + P['ls_coef'] * neg_log(1.0 - preds[observed_labels == 1])
loss_mtx[observed_labels == 0] = (1.0 - P['ls_coef']) * neg_log(1.0 - preds[observed_labels == 0]) + P['ls_coef'] * neg_log(preds[observed_labels == 0])
reg_loss = None
return loss_mtx, reg_loss
def loss_wan(batch, P, Z):
# unpack:
preds = batch['preds']
observed_labels = batch['label_vec_obs']
# input validation:
assert torch.min(observed_labels) >= 0
# compute loss:
loss_mtx = torch.zeros_like(observed_labels)
loss_mtx[observed_labels == 1] = neg_log(preds[observed_labels == 1])
loss_mtx[observed_labels == 0] = neg_log(1.0 - preds[observed_labels == 0]) / float(P['num_classes'] - 1)
reg_loss = None
return loss_mtx, reg_loss
def loss_epr(batch, P, Z):
# unpack:
preds = batch['preds']
observed_labels = batch['label_vec_obs']
# input validation:
assert torch.min(observed_labels) >= 0
# compute loss w.r.t. observed positives:
loss_mtx = torch.zeros_like(observed_labels)
loss_mtx[observed_labels == 1] = neg_log(preds[observed_labels == 1])
# compute regularizer:
reg_loss = expected_positive_regularizer(preds, P['expected_num_pos'], norm='2') / (P['num_classes'] ** 2)
return loss_mtx, reg_loss
def loss_role(batch, P, Z):
# unpack:
preds = batch['preds']
observed_labels = batch['label_vec_obs']
estimated_labels = batch['label_vec_est']
# input validation:
assert torch.min(observed_labels) >= 0
# (image classifier) compute loss w.r.t. observed positives:
loss_mtx_pos_1 = torch.zeros_like(observed_labels)
loss_mtx_pos_1[observed_labels == 1] = neg_log(preds[observed_labels == 1])
# (image classifier) compute loss w.r.t. label estimator outputs:
estimated_labels_detached = estimated_labels.detach()
loss_mtx_cross_1 = estimated_labels_detached * neg_log(preds) + (1.0 - estimated_labels_detached) * neg_log(1.0 - preds)
# (image classifier) compute regularizer:
reg_1 = expected_positive_regularizer(preds, P['expected_num_pos'], norm='2') / (P['num_classes'] ** 2)
# (label estimator) compute loss w.r.t. observed positives:
loss_mtx_pos_2 = torch.zeros_like(observed_labels)
loss_mtx_pos_2[observed_labels == 1] = neg_log(estimated_labels[observed_labels == 1])
# (label estimator) compute loss w.r.t. image classifier outputs:
preds_detached = preds.detach()
loss_mtx_cross_2 = preds_detached * neg_log(estimated_labels) + (1.0 - preds_detached) * neg_log(1.0 - estimated_labels)
# (label estimator) compute regularizer:
reg_2 = expected_positive_regularizer(estimated_labels, P['expected_num_pos'], norm='2') / (P['num_classes'] ** 2)
# compute final loss matrix:
reg_loss = 0.5 * (reg_1 + reg_2)
loss_mtx = 0.5 * (loss_mtx_pos_1 + loss_mtx_pos_2)
loss_mtx += 0.5 * (loss_mtx_cross_1 + loss_mtx_cross_2)
return loss_mtx, reg_loss
loss_functions = {
'bce': loss_bce,
'bce_ls': loss_bce_ls,
'iun': loss_iun,
'iu': loss_iu,
'pr': loss_pr,
'an': loss_an,
'an_ls': loss_an_ls,
'wan': loss_wan,
'epr': loss_epr,
'role': loss_role,
}
'''
top-level wrapper
'''
def compute_batch_loss(batch, P, Z):
assert batch['preds'].dim() == 2
batch_size = int(batch['preds'].size(0))
num_classes = int(batch['preds'].size(1))
loss_denom_mtx = (num_classes * batch_size) * torch.ones_like(batch['preds'])
# input validation:
assert torch.max(batch['label_vec_obs']) <= 1
assert torch.min(batch['label_vec_obs']) >= -1
assert batch['preds'].size() == batch['label_vec_obs'].size()
assert P['loss'] in loss_functions
# validate predictions:
assert torch.max(batch['preds']) <= 1
assert torch.min(batch['preds']) >= 0
# compute loss for each image and class:
loss_mtx, reg_loss = loss_functions[P['loss']](batch, P, Z)
main_loss = (loss_mtx / loss_denom_mtx).sum()
if reg_loss is not None:
batch['loss_tensor'] = main_loss + reg_loss
batch['reg_loss_np'] = reg_loss.clone().detach().cpu().numpy()
else:
batch['loss_tensor'] = main_loss
batch['reg_loss_np'] = 0.0
batch['loss_np'] = batch['loss_tensor'].clone().detach().cpu().numpy()
return batch