forked from yoon28/siamnet
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathsiamnet.py
168 lines (143 loc) · 8.12 KB
/
siamnet.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
import os
import sys
import cv2
import numpy as np
import tensorflow as tf
from omniLoader import OmniglotLoader as og
class SiameseNet():
eps = 1e-10
learn_rate = 1e-4
x_1 = tf.placeholder(tf.float32, shape=[None, og.im_size, og.im_size, og.im_channel])
x_2 = tf.placeholder(tf.float32, shape=[None, og.im_size, og.im_size, og.im_channel])
y = tf.placeholder(tf.float32, shape=[None])
model_param = {'filter_size':(10, 7, 4), 'feature_size':(64, 128, 128), 'n_stack':3}
def build_model(self, inputs, varscope, resuing=False):
param = self.model_param
input_channel = og.im_channel
layer = inputs
for l in range(param['n_stack']):
with tf.variable_scope('{}_conv_{}'.format(varscope, l), reuse=resuing):
filter_sz = param['filter_size'][l]
feature_sz = param['feature_size'][l]
filters = tf.get_variable('filter_{}'.format(l), [filter_sz, filter_sz, input_channel, feature_sz],
initializer=tf.contrib.layers.xavier_initializer())
bias = tf.get_variable('bias_{}'.format(l), [feature_sz], initializer=tf.constant_initializer(0.0))
Z = tf.nn.conv2d(layer, filters, strides=[1,1,1,1], padding='VALID')
activ = tf.nn.relu(Z + bias)
layer = tf.nn.max_pool(activ, ksize=[1,2,2,1], strides=[1,2,2,1], padding='VALID')
input_channel = feature_sz
last_layer = param['n_stack']
with tf.variable_scope('{}_conv_{}'.format(varscope, last_layer), reuse=resuing):
filters = tf.get_variable('filter_{}'.format(last_layer), [4, 4, param['feature_size'][last_layer-1], 256],
initializer=tf.contrib.layers.xavier_initializer())
bias = tf.get_variable('bias_{}'.format(last_layer), [256], initializer=tf.constant_initializer(0.0))
Z = tf.nn.conv2d(layer, filters, strides=[1,1,1,1], padding='VALID')
activ = tf.nn.relu(Z + bias)
layer = tf.contrib.layers.flatten(activ)
with tf.variable_scope('{}_dense'.format(varscope), reuse=resuing):
output = tf.layers.dense(layer, 4096, activation=tf.nn.sigmoid, use_bias=True,
kernel_initializer=tf.contrib.layers.xavier_initializer(),
name='fc_layer', reuse=resuing)
return output
def __init__(self, sharing=True):
varscope='x' if sharing else 'x1'
self.h1 = self.build_model(self.x_1, varscope, False)
varscope='x' if sharing else 'x2'
self.h2 = self.build_model(self.x_2, varscope, sharing)
l1dist = tf.abs(self.h1-self.h2)
with tf.variable_scope('dist_comp'):
self.alpha = tf.get_variable('alpha', [1, 4096],
initializer=tf.random_normal_initializer(mean=0.0, stddev=0.2))
with tf.variable_scope('loss_and_training'):
self.logits = tf.reduce_sum(tf.multiply(l1dist, self.alpha), 1, keepdims=True)
ys = tf.expand_dims(self.y, 1)
self.loss = tf.reduce_sum(tf.nn.sigmoid_cross_entropy_with_logits(labels=ys, logits=self.logits))
optim = tf.train.AdamOptimizer(learning_rate=self.learn_rate)
self.grad = optim.compute_gradients(self.loss)
self.training = optim.apply_gradients(self.grad)
with tf.variable_scope('accuracy'):
self.n_trues = tf.count_nonzero(self.y, dtype=tf.float32)
self.positives = tf.greater(tf.squeeze(tf.sigmoid(self.logits), axis=[1]), 0.5)
self.n_pos = tf.count_nonzero(self.positives, dtype=tf.float32)
self.tp = tf.logical_and(self.positives, tf.equal(self.y, 1.0))
self.n_tp = tf.count_nonzero(self.tp, dtype=tf.float32)
self.n_falses = tf.count_nonzero(tf.equal(self.y, 0.0), dtype=tf.float32)
self.fp = tf.logical_and(self.positives, tf.equal(self.y, 0.0))
self.n_fp = tf.count_nonzero(self.fp, dtype=tf.float32)
def outputzero() : return 0.0
def comp_pr(): return tf.div(self.n_tp, self.n_pos)
def comp_rcll(): return tf.div(self.n_tp, self.n_trues)
def comp_fpr(): return tf.div(self.n_fp, self.n_falses)
self.precision = tf.cond(tf.greater(self.n_pos, 0.0), true_fn=comp_pr, false_fn=outputzero)
self.recall = tf.cond(tf.greater(self.n_trues, 0.0), true_fn=comp_rcll, false_fn=outputzero)
self.fpr = tf.cond(tf.greater(self.n_falses, 0.0), true_fn=comp_fpr, false_fn=outputzero)
with tf.variable_scope('20_way_oneshot_classification'):
self.pred = tf.squeeze(tf.sigmoid(self.logits), axis=[1])
self.pred = tf.expand_dims(self.pred, 0)
self.loc = tf.squeeze(tf.where(tf.equal(self.y, 1.0)), axis=[1])
self.correct = tf.squeeze(tf.nn.in_top_k(self.pred, self.loc, k=1))
tf.summary.scalar('loss', self.loss)
if __name__ == '__main__':
os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
loader = og(0)
model = SiameseNet(sharing=True)
save_name = 'model/siamnet.ckpt'
saver = tf.train.Saver()
session = tf.Session()
if tf.train.checkpoint_exists(save_name):
print('loading')
saver.restore(session, save_name)
else:
print('init')
session.run(tf.global_variables_initializer())
merged = tf.summary.merge_all()
log_writer = tf.summary.FileWriter('logs', session.graph)
step = 0
min_loss = float('inf')
while True:
x1, x2, y = loader.getTrainSample()
[_, summary, h1, h2, logits, loss, grad, precision, recall, fpr] = session.run([model.training, merged, model.h1, model.h2, model.logits, model.loss, model.grad, model.precision, model.recall, model.fpr],
feed_dict={model.x_1:x1, model.x_2:x2, model.y:y})
step += 1
if (step %50) == 0:
print('{}\n train: loss={:.5f}, lr={}, pr={:.2%}, rcll={:.2%}, fpr={:.2%}'.format(step, loss, model.learn_rate, precision, recall, fpr))
x1_t, x2_t, y_t = loader.getTestSample(batch_sz=128)
[pr_t, rcll_t, fpr_t] = session.run([model.precision, model.recall, model.fpr], feed_dict={model.x_1:x1_t, model.x_2:x2_t, model.y:y_t})
print(' test: pr={:.2%}, rcll={:.2%}, fpr={:.2%}'.format(pr_t, rcll_t, fpr_t))
acc_oneshot = []
n_try = 150
for i in range(n_try):
x1_20, x2_20, y_20 = loader.getTestSample_20way()
[pred_20, loc_20, corr_20] = session.run([model.pred, model.loc, model.correct],
feed_dict={model.x_1: x1_20, model.x_2: x2_20, model.y: y_20})
acc_oneshot.append(corr_20)
acc_20 = float(sum(acc_oneshot))/float(n_try)
print(' 20-way: acc={:.2%}'.format(acc_20))
log_writer.add_summary(summary, global_step=step)
log_writer.flush()
if min_loss > loss and (step % 100) == 0:
min_loss = loss
saver.save(session, save_name, global_step=step)
# if not (np.sum(np.float32(np.float32(np.squeeze(1/(1+np.exp(-logits)))>0.5) == correct)) == y.shape[0]):
# print('error 1')
if (step % 500) == 0 and model.learn_rate > 1e-6:
model.learn_rate *= 0.98
# sanity check code
if (step % 1000) == 0:
ck_n_trues = np.sum(np.float32(y == 1))
ck_n_false = np.sum(np.float32(y == 0))
ck_pred = np.squeeze(1/(1+np.exp(-logits)))
ck_pos = ck_pred > 0.5
ck_n_pos = np.sum(np.float32(ck_pos))
ck_tp = np.logical_and(ck_pos, (y==1))
ck_fp = np.logical_and(ck_pos, (y==0))
ck_prc = np.sum(ck_tp)/ck_n_pos if ck_n_pos>0.0 else 0.0
ck_rcll = np.sum(ck_tp)/ck_n_trues if ck_n_trues>0.0 else 0.0
ck_fpr = np.sum(ck_fp)/ck_n_false if ck_n_false>0.0 else 0.0
if np.abs(ck_prc-precision) > 1e-5:
print('error precision')
if np.abs(ck_rcll-recall) > 1e-5:
print('error recall')
if np.abs(ck_fpr-fpr) > 1e-5:
print('error fpr')