forked from 4kssoft/CUTIE
-
Notifications
You must be signed in to change notification settings - Fork 1
/
model_cutie2.py
147 lines (122 loc) · 7.06 KB
/
model_cutie2.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
# written by Xiaohui Zhao
# 2019-04
import tensorflow as tf
from model_framework import Model
class CUTIE2(Model):
def __init__(self, num_vocabs, num_classes, params, trainable=True):
self.name = "CUTTIE_benchmark" #
self.data_grid = tf.placeholder(tf.int32, shape=[None, None, None, 1], name='grid')
self.data_image = tf.placeholder(tf.uint8, shape=[None, None, None, 3], name='image')
self.gt_classes = tf.placeholder(tf.int32, shape=[None, None, None], name='gt_classes')
self.use_ghm = tf.equal(1, params.use_ghm) if hasattr(params, 'use_ghm') else tf.equal(1, 0) #params.use_ghm
self.activation = 'sigmoid' if (hasattr(params, 'use_ghm') and params.use_ghm) else 'relu'
self.dropout = params.data_augmentation_dropout if hasattr(params, 'data_augmentation_dropout') else 1
self.ghm_weights = tf.placeholder(tf.float32, shape=[None, None, None, num_classes], name='ghm_weights')
self.layers = dict({'data_grid': self.data_grid, 'data_image': self.data_image,
'gt_classes': self.gt_classes, 'ghm_weights':self.ghm_weights})
self.num_vocabs = num_vocabs
self.num_classes = num_classes
self.trainable = trainable
self.embedding_size = params.embedding_size
self.weight_decay = params.weight_decay if hasattr(params, 'weight_decay') else 0.0
self.hard_negative_ratio = params.hard_negative_ratio if hasattr(params, 'hard_negative_ratio') else 0.0
self.batch_size = params.batch_size if hasattr(params, 'batch_size') else 0
self.layer_inputs = []
self.setup()
def setup(self):
# input
# (self.feed('data_grid')
# .embed(self.num_vocabs, self.embedding_size, name='embedding', dropout=self.dropout))
# (self.feed('data_image')
# .conv(3, 3, 64, 1, 1, name='image_encoder1_1')
# .conv(3, 3, 128, 1, 1, name='image_encoder1_2'))
#
# # encoder
# (self.feed('embedding')
# .conv(3, 5, 64, 1, 1, name='encoder1_1')
# .conv(3, 5, 128, 1, 1, name='encoder1_2')
# .max_pool(2, 2, 2, 2, name='pool1')
# .conv(3, 5, 128, 1, 1, name='encoder2_1')
# .conv(3, 5, 256, 1, 1, name='encoder2_2')
# .max_pool(2, 2, 2, 2, name='pool2')
# .conv(3, 5, 256, 1, 1, name='encoder3_1')
# .conv(3, 5, 512, 1, 1, name='encoder3_2')
# .max_pool(2, 2, 2, 2, name='pool3')
# .conv(3, 5, 512, 1, 1, name='encoder4_1')
# .conv(3, 5, 512, 1, 1, name='encoder4_2'))
#
# # decoder
# (self.feed('encoder4_2')
# .up_conv(3, 5, 512, 1, 1, name='up1')
# .conv(3, 5, 256, 1, 1, name='decoder1_1')
# .conv(3, 5, 256, 1, 1, name='decoder1_2')
# .up_conv(3, 5, 256, 1, 1, name='up2')
# .conv(3, 5, 128, 1, 1, name='decoder2_1')
# .conv(3, 5, 128, 1, 1, name='decoder2_2')
# .up_conv(3, 5, 128, 1, 1, name='up3')
# .conv(3, 5, 64, 1, 1, name='decoder3_1')
# .conv(3, 5, 64, 1, 1, name='decoder3_2'))
#
# # classification
# (self.feed('decoder3_2')
# .conv(1, 1, self.num_classes, 1, 1, activation=self.activation, name='cls_logits')
# .softmax(name='softmax'))
pass
def disp_results(self, data_input, data_label, model_output, threshold):
data_input_flat = data_input.reshape([-1]) # [b * h * w]
labels = [] # [b * h * w, classes]
for item in data_label.reshape([-1]):
labels.append([i==item for i in range(self.num_classes)])
logits = model_output.reshape([-1, self.num_classes]) # [b * h * w, classes]
# ignore none word input
labels_flat = []
results_flat = []
for idx, item in enumerate(data_input_flat):
if item != 0:
labels_flat.extend(labels[idx])
results_flat.extend(logits[idx] > threshold)
num_p = sum(labels_flat)
num_n = sum([1-label for label in labels_flat])
num_all = len(results_flat)
num_correct = sum([True for i in range(num_all) if labels_flat[i] == results_flat[i]])
labels_flat_p = [label!=0 for label in labels_flat]
labels_flat_n = [label==0 for label in labels_flat]
num_tp = sum([labels_flat_p[i] * results_flat[i] for i in range(num_all)])
num_tn = sum([labels_flat_n[i] * (not results_flat[i]) for i in range(num_all)])
num_fp = num_n - num_tp
num_fn = num_p - num_tp
# accuracy, precision, recall
accuracy = num_correct / num_all
precision = num_tp / (num_tp + num_fp)
recall = num_tp / (num_tp + num_fn)
return accuracy, precision, recall
def inference(self):
return self.get_output('softmax') #cls_logits
def build_loss(self):
labels = self.get_output('gt_classes')
cls_logits = self.get_output('cls_logits')
cls_logits = tf.cond(self.use_ghm, lambda: cls_logits*self.get_output('ghm_weights'),
lambda: cls_logits, name="GradientHarmonizingMechanism")
cross_entropy = tf.nn.sparse_softmax_cross_entropy_with_logits(labels=labels, logits=cls_logits)
with tf.variable_scope('HardNegativeMining'):
labels = tf.reshape(labels, [-1])
cross_entropy = tf.reshape(cross_entropy, [-1])
fg_idx = tf.where(tf.not_equal(labels, 0))
fgs = tf.gather(cross_entropy, fg_idx)
bg_idx = tf.where(tf.equal(labels, 0))
bgs = tf.gather(cross_entropy, bg_idx)
num = self.hard_negative_ratio * tf.shape(fgs)[0]
num_bg = tf.cond(tf.shape(bgs)[0]<num, lambda:tf.shape(bgs)[0], lambda:num)
sorted_bgs, _ = tf.nn.top_k(tf.transpose(bgs), num_bg, sorted=True)
cross_entropy = fgs + sorted_bgs
# total loss
model_loss = tf.reduce_mean(cross_entropy)
regularization_loss = tf.add_n(tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES), name='regularization')
total_loss = model_loss + regularization_loss
tf.summary.scalar('model_loss', model_loss)
tf.summary.scalar('regularization_loss', regularization_loss)
tf.summary.scalar('total_loss', total_loss)
logits = self.get_output('cls_logits')
softmax_logits = self.get_output('softmax') #cls_logits
return model_loss, regularization_loss, total_loss, logits, softmax_logits