-
Notifications
You must be signed in to change notification settings - Fork 0
/
adaptation.py
165 lines (146 loc) · 6 KB
/
adaptation.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
# coding: UTF-8
from __future__ import print_function
import argparse
from collections import deque
import logging
import os
from keras.preprocessing.image import ImageDataGenerator
import matplotlib.pyplot as plt
import numpy as np
from scipy.io import loadmat
import tensorflow as tf
from tqdm import tqdm
from models import discriminator
from models import source_cnn
from models import target_cnn
import util
def main(args):
util.config_logging()
# Settings
lr = args.lr
beta1 = args.beta1
batch_size = args.batch_size
iterations = args.iterations
snapshot = args.snapshot
stepsize = args.stepsize
display = args.display
path_source_cnn = './output/source_cnn'
output_dir = os.path.join('output', 'target_cnn')
save_path = os.path.join(output_dir, 'target_cnn.ckpt')
if not os.path.exists(output_dir):
os.makedirs(output_dir)
# Load source data
train_mat = loadmat('./data/svhn/train_32x32.mat')
train_images = train_mat['X'].transpose((3, 0, 1, 2))
train_images = train_images.astype(np.float32) / 255.
RGB2GRAY = np.array([0.2989, 0.5870, 0.1140], dtype=np.float32)
train_images = np.sum(
np.multiply(train_images, RGB2GRAY),
3, keepdims=True
)
# Load target data
target_images = util._read_images(
'./data/mnist/train-images-idx3-ubyte.gz')
# Data generator
idg = ImageDataGenerator()
source_data_gen = idg.flow(
train_images, batch_size=batch_size, shuffle=True
)
target_data_gen = idg.flow(
target_images, batch_size=batch_size, shuffle=True
)
# Define graph
nb_classes = 10
tf.reset_default_graph()
x_source = tf.placeholder(tf.float32, (None, 32, 32, 1))
x_source_resized = tf.image.resize_images(x_source, [28, 28])
x_target = tf.placeholder(tf.float32, (None, 28, 28, 1))
feature_src = source_cnn(
x_source_resized, nb_classes=nb_classes,
trainable=False, adapt=True)
feature_target = target_cnn(x_target, nb_classes, trainable=True)
d_logits_src = discriminator(feature_src)
d_logits_target = discriminator(feature_target, reuse=True)
# Loss: Discriminator
d_loss_src = tf.reduce_mean(
tf.nn.sigmoid_cross_entropy_with_logits(
logits=d_logits_src, labels=tf.ones_like(d_logits_src)))
d_loss_target = tf.reduce_mean(
tf.nn.sigmoid_cross_entropy_with_logits(
logits=d_logits_target, labels=tf.zeros_like(d_logits_target)))
d_loss = d_loss_src + d_loss_target
# Loss: target CNN
target_loss = tf.reduce_mean(
tf.nn.sigmoid_cross_entropy_with_logits(
logits=d_logits_target, labels=tf.ones_like(d_logits_target)))
t_vars = tf.trainable_variables()
target_vars = [var for var in tf.get_collection(
tf.GraphKeys.GLOBAL_VARIABLES, scope='target_cnn')]
d_vars = [var for var in t_vars if var.name.startswith('discriminator')]
src_vars = [
var for var in tf.get_collection(
tf.GraphKeys.GLOBAL_VARIABLES, scope='source_cnn')]
lr_var = tf.Variable(lr, name='learning_rate', trainable=False)
optimizer = tf.train.AdamOptimizer(lr_var, beta1)
target_train_op = optimizer.minimize(target_loss, var_list=target_vars)
d_train_op = optimizer.minimize(d_loss, var_list=d_vars)
# Train
source_saver = tf.train.Saver(var_list=src_vars)
target_saver = tf.train.Saver(var_list=target_vars)
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
target_losses = deque(maxlen=10)
d_losses = deque(maxlen=10)
bar = tqdm(range(iterations))
bar.set_description('(lr: {:.0e})'.format(lr))
bar.refresh()
losses = []
with tf.Session(config=config) as sess:
sess.run(tf.global_variables_initializer())
source_saver.restore(
sess,
tf.train.latest_checkpoint(path_source_cnn)
)
for i in bar:
batch_source = next(source_data_gen)
batch_target = next(target_data_gen)
target_loss_val, d_loss_val, _, _ = sess.run(
[target_loss, d_loss, target_train_op, d_train_op],
feed_dict={x_source: batch_source, x_target: batch_target}
)
target_losses.append(target_loss_val)
d_losses.append(d_loss_val)
losses.append([target_loss_val, d_loss_val])
if i % display == 0:
logging.info('{:20} Target: {:5.4f} (avg: {:5.4f})'
' Discriminator: {:5.4f} (avg: {:5.4f})'
.format('Iteration {}:'.format(i),
target_loss_val,
np.mean(target_losses),
d_loss_val,
np.mean(d_losses)))
if stepsize is not None and (i + 1) % stepsize == 0:
lr = sess.run(lr_var.assign(lr * 0.1))
logging.info('Changed learning rate to {:.0e}'.format(lr))
bar.set_description('(lr: {:.0e})'.format(lr))
if (i + 1) % snapshot == 0:
snapshot_path = target_saver.save(sess, save_path)
logging.info('Saved snapshot to {}'.format(snapshot_path))
# Save visualization of training losses
losses = np.array(losses)
plt.plot(losses.T[0], label='Target CNN Loss', alpha=0.5)
plt.plot(losses.T[1], label='Discriminator Loss', alpha=0.5)
plt.title('Training Losses')
plt.legend()
plt.savefig('./losses.png')
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--iterations', type=int, default=20000)
parser.add_argument('--batch_size', type=int, default=50)
parser.add_argument('--display', type=int, default=10)
parser.add_argument('--lr', type=float, default=1e-4)
parser.add_argument('--snapshot', type=int, default=5000)
parser.add_argument('--stepsize', type=int, default=None)
parser.add_argument('--beta1', type=float, default=0.5)
args = parser.parse_args()
main(args)