-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmain.py
151 lines (106 loc) · 5.52 KB
/
main.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
#!/usr/bin/env python
# coding: utf-8
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
from utils import *
import numpy as np
import os
from flip_gradient import *
from config import *
from model.MDAN import MDAN
os.environ['KMP_DUPLICATE_LIB_OK']='True'
mnist = load_data(setname='mnist')
svhn = load_data(setname='svhn')
synth = load_data(setname='synth')
mnistm= load_data(setname='mnistm')
#the last is target domain; others are source domains
datasets = [svhn, mnist, synth, mnistm]
print('tar: ', datasets[-1].name)
combined_train, combined_train_label = get_train(datasets) # stacked properly.
tar_test = datasets[-1].test.data
tar_test_label = datasets[-1].test.label
# Compute pixel mean for normalizing data
pixel_mean = combined_train.mean((0, 1, 2))
num_test = test_n
sec = np.vstack([np.tile([1., 0.], [num_test // 6, 1]),
np.tile([0., 1.], [num_test // 6, 1])])
combined_test_domain = np.vstack([sec, sec, sec])
# Build the model graph
graph = tf.get_default_graph()
with graph.as_default():
model = MDAN(pixel_mean)
learning_rate = tf.placeholder(tf.float32, [])
domain_loss_ = tf.convert_to_tensor([
tf.reduce_mean(model.domain_loss_1),
tf.reduce_mean(model.domain_loss_2),
tf.reduce_mean(model.domain_loss_3)
])
pred_loss = tf.reduce_mean(model.pred_loss)#scalar
domain_loss = tf.reduce_mean(domain_loss_)
total_loss = tf.add(domain_loss_, pred_loss)
gamma = tf.constant(GAMMA)
total_loss = tf.multiply(tf.log(tf.reduce_sum(tf.exp(tf.multiply(gamma, total_loss))))/gamma, weight)
regular_train_op = tf.train.AdadeltaOptimizer(learning_rate).minimize(pred_loss)
dann_train_op = tf.train.MomentumOptimizer(learning_rate, 0.95).minimize(total_loss)
# Evaluation
correct_label_pred = tf.equal(tf.argmax(model.classify_labels, 1), tf.argmax(model.pred, 1))
label_acc = tf.reduce_mean(tf.cast(correct_label_pred, tf.float32))
correct_domain_pred_1 = tf.equal(tf.argmax(model.domain1, 1), tf.argmax(model.pred_1, 1))
correct_domain_pred_2 = tf.equal(tf.argmax(model.domain2, 1), tf.argmax(model.pred_2, 1))
correct_domain_pred_3 = tf.equal(tf.argmax(model.domain3, 1), tf.argmax(model.pred_3, 1))
domain_acc_1 = tf.reduce_mean(tf.cast(correct_domain_pred_1, tf.float32))
domain_acc_2 = tf.reduce_mean(tf.cast(correct_domain_pred_2, tf.float32))
domain_acc_3 = tf.reduce_mean(tf.cast(correct_domain_pred_3, tf.float32))
domain_acc = [domain_acc_1, domain_acc_2, domain_acc_3]
def train_and_evaluate(training_mode, graph, model, num_steps=10000, verbose=False):
"""Helper to run the model with different training modes."""
with tf.Session(graph=graph) as sess:
tf.global_variables_initializer().run()
gen_source_batch = batch_generator([combined_train, combined_train_label], batch_size)
domain_labels_single_batch_1 = np.vstack([np.tile([1., 0.], [batch_size // 6, 1]),
np.tile([0., 1.], [batch_size // 6, 1])])
domain_labels_single_batch_2 = np.vstack([np.tile([1., 0.], [batch_size // 6, 1]),
np.tile([0., 1.], [batch_size // 6, 1])])
domain_labels_single_batch_3 = np.vstack([np.tile([1., 0.], [batch_size // 6, 1]),
np.tile([0., 1.], [batch_size // 6, 1])])
# Training loop
for i in range(num_steps):
# Adaptation param and learning rate schedule as described in the paper
p = float(i) / num_steps
l = 2. / (1. + np.exp(-10. * p)) - 1
lr = head / (1. + 10 * p) ** 0.75
X, y = next(gen_source_batch)
_, batch_loss, dloss, ploss, d_acc, p_acc = sess.run(
[dann_train_op, total_loss, domain_loss, pred_loss, domain_acc, label_acc],
feed_dict={
model.X: X, model.y: y,
model.domain1: domain_labels_single_batch_1,
model.domain2: domain_labels_single_batch_2,
model.domain3: domain_labels_single_batch_3,
model.train: True, model.l: l, learning_rate: lr
})
if i % 50 == 1 and i > 50:
print('[test]testing...')
target_acc = sess.run(label_acc,
feed_dict={model.X: tar_test, model.y: tar_test_label,
model.train: False})
print('[test]target_label_acc: ', target_acc)
if verbose and i % 10 == 0:
print('i: {:5d} loss: {:10.6f} dloss:{:8.6f} ploss:{:8.6f} d_acc: {} p_acc: {:6f} p: {:.6f} l: {:.6f} lr: {:.6f}'.format(
i, batch_loss, dloss, ploss, d_acc, p_acc, p, l, lr))
target_acc = sess.run(label_acc,
feed_dict={model.X: tar_test, model.y: tar_test_label,
model.train: False})
print('[Final test]target_label_acc: ',target_acc)
test_emb = sess.run(model.feature, feed_dict={model.X: tar_test})
return target_acc, test_emb
def eval(n):
print('\nnum_steps: ', n, " Model: ", str(model))
target_acc, dann_emb = train_and_evaluate(
'dann', graph, model,
num_steps=n, verbose=True
)
print('Target (MNIST-M) accuracy:', target_acc)
print('----------------------------')
eval(step)