-
Notifications
You must be signed in to change notification settings - Fork 18
/
losses.py
252 lines (212 loc) · 10 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
import tensorflow as tf
from utils.tf_geom import get_dist_mat, rnd_sample, interpolate
def make_detector_loss(pos0, pos1, dense_feat_map0, dense_feat_map1,
score_map0, score_map1, batch_size, num_corr, loss_type, config):
joint_loss = tf.constant(0.)
accuracy = tf.constant(0.)
all_valid_pos0 = []
all_valid_pos1 = []
all_valid_match = []
for i in range(batch_size):
# random sample
valid_pos0, valid_pos1 = rnd_sample([pos0[i], pos1[i]], num_corr)
valid_num = tf.shape(valid_pos0)[0]
valid_feat0 = interpolate(
valid_pos0 / 4, dense_feat_map0[i], batched=False)
valid_feat1 = interpolate(
valid_pos1 / 4, dense_feat_map1[i], batched=False)
valid_feat0 = tf.nn.l2_normalize(valid_feat0, axis=-1)
valid_feat1 = tf.nn.l2_normalize(valid_feat1, axis=-1)
valid_score0 = interpolate(valid_pos0, tf.squeeze(
score_map0[i], axis=-1), nd=False, batched=False)
valid_score1 = interpolate(valid_pos1, tf.squeeze(
score_map1[i], axis=-1), nd=False, batched=False)
if config['det']['corr_weight']:
corr_weight = valid_score0 * valid_score1
else:
corr_weight = None
safe_radius = config['det']['safe_radius']
if safe_radius > 0:
radius_mask_row = get_dist_mat(
valid_pos1, valid_pos1, "euclidean_dist_no_norm")
radius_mask_row = tf.less(radius_mask_row, safe_radius)
radius_mask_col = get_dist_mat(
valid_pos0, valid_pos0, "euclidean_dist_no_norm")
radius_mask_col = tf.less(radius_mask_col, safe_radius)
radius_mask_row = tf.cast(
radius_mask_row, tf.float32) - tf.eye(valid_num)
radius_mask_col = tf.cast(
radius_mask_col, tf.float32) - tf.eye(valid_num)
else:
radius_mask_row = None
radius_mask_col = None
si_loss, si_accuracy, matched_mask = tf.cond(
tf.less(valid_num, 32),
lambda: (tf.constant(0.), tf.constant(1.),
tf.cast(tf.zeros((1, valid_num)), tf.bool)),
lambda: make_structured_loss(
tf.expand_dims(valid_feat0, 0), tf.expand_dims(valid_feat1, 0),
loss_type=loss_type,
radius_mask_row=radius_mask_row, radius_mask_col=radius_mask_col,
corr_weight=tf.expand_dims(corr_weight, 0) if corr_weight is not None else None,
name='si_loss')
)
joint_loss += si_loss / batch_size
accuracy += si_accuracy / batch_size
all_valid_match.append(tf.squeeze(matched_mask, axis=0))
all_valid_pos0.append(valid_pos0)
all_valid_pos1.append(valid_pos1)
return joint_loss, accuracy
def make_quadruple_loss(kpt_m0, kpt_m1, inlier_num):
batch_size = kpt_m0.get_shape()[0].value
num_corr = kpt_m1.get_shape()[1].value
kpt_m_diff0 = tf.linalg.matrix_transpose(
tf.tile(kpt_m0, (1, 1, num_corr))) - kpt_m0
kpt_m_diff1 = tf.linalg.matrix_transpose(
tf.tile(kpt_m1, (1, 1, num_corr))) - kpt_m1
R = kpt_m_diff0 * kpt_m_diff1
quad_loss = 0
accuracy = 0
for i in range(batch_size):
cur_inlier_num = tf.squeeze(inlier_num[i])
inlier_block = R[i, 0:cur_inlier_num, 0:cur_inlier_num]
inlier_block = inlier_block + tf.eye(cur_inlier_num)
inlier_block = tf.maximum(0., 1. - inlier_block)
error = tf.count_nonzero(inlier_block)
cur_inlier_num = tf.cast(cur_inlier_num, tf.float32)
quad_loss += tf.reduce_sum(inlier_block) / \
(cur_inlier_num * (cur_inlier_num - 1))
accuracy += 1. - tf.cast(error, tf.float32) / \
(cur_inlier_num * (cur_inlier_num - 1))
quad_loss /= float(batch_size)
accuracy /= float(batch_size)
return quad_loss, accuracy
def make_structured_loss(feat_anc, feat_pos,
loss_type='RATIO', inlier_mask=None,
radius_mask_row=None, radius_mask_col=None,
corr_weight=None, dist_mat=None, name='loss'):
"""
Structured loss construction.
Args:
feat_anc, feat_pos: Feature matrix.
loss_type: Loss type.
inlier_mask:
Returns:
"""
batch_size = feat_anc.get_shape()[0].value
num_corr = tf.shape(feat_anc)[1]
if inlier_mask is None:
inlier_mask = tf.cast(tf.ones((batch_size, num_corr)), tf.bool)
inlier_num = tf.count_nonzero(tf.cast(inlier_mask, tf.float32), axis=-1)
if loss_type == 'LOG' or loss_type == 'L2NET' or loss_type == 'CIRCLE':
dist_type = 'cosine_dist'
elif loss_type.find('HARD') >= 0:
dist_type = 'euclidean_dist'
else:
raise NotImplementedError()
if dist_mat is None:
dist_mat = get_dist_mat(feat_anc, feat_pos, dist_type)
pos_vec = tf.linalg.diag_part(dist_mat)
if loss_type.find('HARD') >= 0:
neg_margin = 1
dist_mat_without_min_on_diag = dist_mat + \
10 * tf.expand_dims(tf.eye(num_corr), 0)
mask = tf.cast(
tf.less(dist_mat_without_min_on_diag, 0.008), tf.float32)
dist_mat_without_min_on_diag += mask*10
if radius_mask_row is not None:
hard_neg_dist_row = dist_mat_without_min_on_diag + 10 * radius_mask_row
else:
hard_neg_dist_row = dist_mat_without_min_on_diag
if radius_mask_col is not None:
hard_neg_dist_col = dist_mat_without_min_on_diag + 10 * radius_mask_col
else:
hard_neg_dist_col = dist_mat_without_min_on_diag
hard_neg_dist_row = tf.reduce_min(hard_neg_dist_row, axis=-1)
hard_neg_dist_col = tf.reduce_min(hard_neg_dist_col, axis=-2)
if loss_type == 'HARD_TRIPLET':
loss_row = tf.maximum(neg_margin + pos_vec - hard_neg_dist_row, 0)
loss_col = tf.maximum(neg_margin + pos_vec - hard_neg_dist_col, 0)
elif loss_type == 'HARD_CONTRASTIVE':
pos_margin = 0.2
pos_loss = tf.maximum(pos_vec - pos_margin, 0)
loss_row = pos_loss + tf.maximum(neg_margin - hard_neg_dist_row, 0)
loss_col = pos_loss + tf.maximum(neg_margin - hard_neg_dist_col, 0)
else:
raise NotImplementedError()
elif loss_type == 'LOG' or loss_type == 'L2NET':
if loss_type == 'LOG':
with tf.compat.v1.variable_scope(name, reuse=tf.compat.v1.AUTO_REUSE):
log_scale = tf.compat.v1.get_variable('scale_temperature', shape=(), dtype=tf.float32,
initializer=tf.constant_initializer(1))
tf.compat.v1.add_to_collection(
tf.compat.v1.GraphKeys.REGULARIZATION_LOSSES, log_scale)
else:
log_scale = tf.constant(1.)
softmax_row = tf.nn.softmax(log_scale * dist_mat, axis=2)
softmax_col = tf.nn.softmax(log_scale * dist_mat, axis=1)
loss_row = -tf.math.log(tf.linalg.diag_part(softmax_row))
loss_col = -tf.math.log(tf.linalg.diag_part(softmax_col))
elif loss_type == 'CIRCLE':
log_scale = 512
m = 0.1
neg_mask_row = tf.expand_dims(tf.eye(num_corr), 0)
if radius_mask_row is not None:
neg_mask_row += radius_mask_row
neg_mask_col = tf.expand_dims(tf.eye(num_corr), 0)
if radius_mask_col is not None:
neg_mask_col += radius_mask_col
pos_margin = 1 - m
neg_margin = m
pos_optimal = 1 + m
neg_optimal = -m
neg_mat_row = dist_mat - 128 * neg_mask_row
neg_mat_col = dist_mat - 128 * neg_mask_col
lse_positive = tf.math.reduce_logsumexp(-log_scale * (pos_vec[..., None] - pos_margin) * \
tf.stop_gradient(tf.maximum(pos_optimal - pos_vec[..., None], 0)), axis=-1)
lse_negative_row = tf.math.reduce_logsumexp(log_scale * (neg_mat_row - neg_margin) * \
tf.stop_gradient(tf.maximum(neg_mat_row - neg_optimal, 0)), axis=-1)
lse_negative_col = tf.math.reduce_logsumexp(log_scale * (neg_mat_col - neg_margin) * \
tf.stop_gradient(tf.maximum(neg_mat_col - neg_optimal, 0)), axis=-2)
loss_row = tf.math.softplus(lse_positive + lse_negative_row) / log_scale
loss_col = tf.math.softplus(lse_positive + lse_negative_col) / log_scale
else:
raise NotImplementedError()
if dist_type == 'cosine_dist':
err_row = dist_mat - tf.expand_dims(pos_vec, -1)
err_col = dist_mat - tf.expand_dims(pos_vec, -2)
elif dist_type == 'euclidean_dist' or dist_type == 'euclidean_dist_no_norm':
err_row = tf.expand_dims(pos_vec, -1) - dist_mat
err_col = tf.expand_dims(pos_vec, -2) - dist_mat
else:
raise NotImplementedError()
if radius_mask_row is not None:
err_row = err_row - 10 * radius_mask_row
if radius_mask_col is not None:
err_col = err_col - 10 * radius_mask_col
err_row = tf.reduce_sum(tf.maximum(err_row, 0), axis=-1)
err_col = tf.reduce_sum(tf.maximum(err_col, 0), axis=-2)
loss = 0
accuracy = 0
tot_loss = (loss_row + loss_col) / 2
if corr_weight is not None:
tot_loss = tot_loss * corr_weight
for i in range(batch_size):
if corr_weight is not None:
loss += tf.reduce_sum(tot_loss[i][inlier_mask[i]]) / \
(tf.reduce_sum(corr_weight[i][inlier_mask[i]]) + 1e-6)
else:
loss += tf.reduce_mean(tot_loss[i][inlier_mask[i]])
cnt_err_row = tf.count_nonzero(
err_row[i][inlier_mask[i]], dtype=tf.float32)
cnt_err_col = tf.count_nonzero(
err_col[i][inlier_mask[i]], dtype=tf.float32)
tot_err = cnt_err_row + cnt_err_col
accuracy += 1. - \
tf.math.divide_no_nan(tot_err, tf.cast(
inlier_num[i], tf.float32)) / batch_size / 2.
matched_mask = tf.logical_and(tf.equal(err_row, 0), tf.equal(err_col, 0))
matched_mask = tf.logical_and(matched_mask, inlier_mask)
loss /= batch_size
accuracy /= batch_size
return loss, accuracy, matched_mask