forked from rykov8/ssd_keras
-
Notifications
You must be signed in to change notification settings - Fork 84
/
ssd_training.py
187 lines (138 loc) · 7.12 KB
/
ssd_training.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
"""SSD training utils."""
import numpy as np
import matplotlib.pyplot as plt
import keras.backend as K
import tensorflow as tf
from utils.training import smooth_l1_loss, softmax_loss, focal_loss
def compute_metrics(class_true, class_pred, conf, top_k=100):
"""Compute precision, recall, accuracy and f-measure for top_k predictions.
from top_k predictions that are TP FN or FP (TN kept out)
"""
# TODO: does this only work for one class?
top_k = tf.cast(top_k, tf.int32)
eps = K.epsilon()
mask = tf.greater(class_true + class_pred, 0)
#mask = tf.logical_or(tf.greater(class_true, 0), tf.greater(class_pred, 0))
mask_float = tf.cast(mask, tf.float32)
vals, idxs = tf.nn.top_k(conf * mask_float, k=top_k)
top_k_class_true = tf.gather(class_true, idxs)
top_k_class_pred = tf.gather(class_pred, idxs)
true_mask = tf.equal(top_k_class_true, top_k_class_pred)
false_mask = tf.logical_not(true_mask)
pos_mask = tf.greater(top_k_class_pred, 0)
neg_mask = tf.logical_not(pos_mask)
tp = tf.reduce_sum(tf.cast(tf.logical_and(true_mask, pos_mask), tf.float32))
fp = tf.reduce_sum(tf.cast(tf.logical_and(false_mask, pos_mask), tf.float32))
fn = tf.reduce_sum(tf.cast(tf.logical_and(false_mask, neg_mask), tf.float32))
tn = tf.reduce_sum(tf.cast(tf.logical_and(true_mask, neg_mask), tf.float32))
precision = tp / (tp + fp + eps)
recall = tp / (tp + fn + eps)
accuracy = (tp + tn) / (tp + tn + fp + fn + eps)
fmeasure = 2 * (precision * recall) / (precision + recall + eps)
return precision, recall, accuracy, fmeasure
class SSDLoss(object):
"""Multibox loss for SSD.
# Arguments
alpha: Weight of L1-smooth loss.
neg_pos_ratio: Max ratio of negative to positive boxes in loss.
negatives_for_hard: Number of negative boxes to consider
if there is no positive boxes in batch.
# References
https://arxiv.org/abs/1512.02325
"""
def __init__(self, alpha=1.0, neg_pos_ratio=3.0, negatives_for_hard=100.0):
self.alpha = alpha
self.neg_pos_ratio = neg_pos_ratio
self.metric_names = [ 'loss',
'num_pos', 'num_neg',
'pos_conf_loss', 'neg_conf_loss', 'pos_loc_loss',
'precision', 'recall', 'fmeasure', 'accuracy',
]
def compute(self, y_true, y_pred):
# y.shape (batches, priors, 4 x segment_offset + n x class_label)
# TODO: negatives_for_hard?
batch_size = tf.shape(y_true)[0]
num_priors = tf.shape(y_true)[1]
num_classes = tf.shape(y_true)[2] - 4
eps = K.epsilon()
# confidence loss
conf_true = tf.reshape(y_true[:,:,4:], [-1, num_classes])
conf_pred = tf.reshape(y_pred[:,:,4:], [-1, num_classes])
conf_loss = softmax_loss(conf_true, conf_pred)
class_true = tf.argmax(conf_true, axis=1)
class_pred = tf.argmax(conf_pred, axis=1)
conf = tf.reduce_max(conf_pred, axis=1)
neg_mask_float = conf_true[:,0]
neg_mask = tf.cast(neg_mask_float, tf.bool)
pos_mask = tf.logical_not(neg_mask)
pos_mask_float = tf.cast(pos_mask, tf.float32)
num_total = tf.cast(tf.shape(conf_true)[0], tf.float32)
num_pos = tf.reduce_sum(pos_mask_float)
num_neg = num_total - num_pos
pos_conf_loss = tf.reduce_sum(conf_loss * pos_mask_float)
pos_conf_loss = pos_conf_loss / (num_pos + eps)
## take only false positives for hard negative mining
#false_pos_mask = tf.logical_and(neg_mask, tf.not_equal(class_pred, 0))
#num_false_pos = tf.reduce_sum(tf.cast(false_pos_mask, tf.float32))
#num_neg = tf.minimum(self.neg_pos_ratio * num_pos, num_false_pos)
#neg_conf_loss = tf.boolean_mask(conf_loss, false_pos_mask)
num_neg = tf.minimum(self.neg_pos_ratio * num_pos, num_neg)
neg_conf_loss = tf.boolean_mask(conf_loss, neg_mask)
neg_conf_loss = neg_conf_loss / (num_neg + eps)
vals, idxs = tf.nn.top_k(neg_conf_loss, k=tf.cast(num_neg, tf.int32))
#neg_conf_loss = tf.reduce_sum(tf.gather(neg_conf_loss, idxs))
neg_conf_loss = tf.reduce_sum(vals)
conf_loss = pos_conf_loss + neg_conf_loss
# offset loss
loc_true = tf.reshape(y_true[:,:,0:4], [-1, 4])
loc_pred = tf.reshape(y_pred[:,:,0:4], [-1, 4])
loc_loss = smooth_l1_loss(loc_true, loc_pred)
pos_loc_loss = tf.reduce_sum(loc_loss * pos_mask_float) # only for positives
loc_loss = pos_loc_loss / (num_pos + eps)
# total loss
loss = conf_loss + self.alpha * loc_loss
# metrics
precision, recall, accuracy, fmeasure = compute_metrics(class_true, class_pred, conf, top_k=100*batch_size)
return eval('{'+' '.join(['"'+n+'": '+n+',' for n in self.metric_names])+'}')
class SSDFocalLoss(object):
def __init__(self, lambda_conf=100.0, lambda_offsets=1.0, class_weights=1.0):
self.lambda_conf = lambda_conf
self.lambda_offsets = lambda_offsets
self.class_weights = class_weights
self.metric_names = [ 'loss',
'conf_loss', 'loc_loss',
'precision', 'recall', 'fmeasure', 'accuracy',
]
def compute(self, y_true, y_pred):
# y.shape (batches, priors, 4 x bbox_offset + n x class_label)
batch_size = tf.shape(y_true)[0]
num_priors = tf.shape(y_true)[1]
num_classes = tf.shape(y_true)[2] - 4
eps = K.epsilon()
# confidence loss
conf_true = tf.reshape(y_true[:,:,4:], [-1, num_classes])
conf_pred = tf.reshape(y_pred[:,:,4:], [-1, num_classes])
class_true = tf.argmax(conf_true, axis=1)
class_pred = tf.argmax(conf_pred, axis=1)
conf = tf.reduce_max(conf_pred, axis=1)
neg_mask_float = conf_true[:,0]
neg_mask = tf.cast(neg_mask_float, tf.bool)
pos_mask = tf.logical_not(neg_mask)
pos_mask_float = tf.cast(pos_mask, tf.float32)
num_total = tf.cast(tf.shape(conf_true)[0], tf.float32)
num_pos = tf.reduce_sum(pos_mask_float)
num_neg = num_total - num_pos
conf_loss = focal_loss(conf_true, conf_pred, alpha=self.class_weights)
conf_loss = tf.reduce_sum(conf_loss)
conf_loss = conf_loss / (num_total + eps)
# offset loss
loc_true = tf.reshape(y_true[:,:,0:4], [-1, 4])
loc_pred = tf.reshape(y_pred[:,:,0:4], [-1, 4])
loc_loss = smooth_l1_loss(loc_true, loc_pred)
pos_loc_loss = tf.reduce_sum(loc_loss * pos_mask_float) # only for positive ground truth
loc_loss = pos_loc_loss / (num_pos + eps)
# total loss
loss = self.lambda_conf * conf_loss + self.lambda_offsets * loc_loss
# metrics
precision, recall, accuracy, fmeasure = compute_metrics(class_true, class_pred, conf, top_k=100*batch_size)
return eval('{'+' '.join(['"'+n+'": '+n+',' for n in self.metric_names])+'}')