-
Notifications
You must be signed in to change notification settings - Fork 1
/
train_fcn8_vgg.py
executable file
·161 lines (132 loc) · 6.64 KB
/
train_fcn8_vgg.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
#!/usr/bin/env python
import scipy as scp
import scipy.misc
import numpy as np
import logging
import tensorflow as tf
import sys
import fcn8_vgg
import utils
logging.basicConfig(format='%(asctime)s %(levelname)s %(message)s',
level=logging.INFO,
stream=sys.stdout)
# Load images
ind_img = [2, 8, 10, 11, 12, 15]
num_img = len(ind_img)
img = np.zeros((num_img, 2016, 2016, 3))
lab = np.zeros((num_img, 2016, 2016, 1))
for i, ind in enumerate(ind_img):
img[i, :, :, :] = 1.0 * scp.misc.imread("data/%02i_original.png" % ind)
lab[i, :, :, 0] = scp.misc.imread("data/%02i_label_dna.png" % ind)[:, :, 0] // 255
lab[i, :, :, 0] += 2 * (scp.misc.imread("data/%02i_label_nucleosome_new.png" % ind)[:, :, 0] // 255)
img_train = img[[0, 1, 2, 4], :, :, :]
lab_train = lab[[0, 1, 2, 4], :, :, :]
img_val = img[[3, 5], :, :, :]
lab_val = lab[[3, 5], :, :, :]
num_train = 4
num_val = 2
# Specify training parameter
num_iter = 1200
size_random_crop = [800, 800, 4]
starter_learning_rate = 1e-4
learning_decay_rate = 0.5
decay_every = 200
pred_every = 10
val_every = 10
save_every = 200
file_name_log_train = 'loss_train.csv'
file_name_log_val = 'loss_val.csv'
# Layers get trained progressively
# Define which layer will be trained beginning at which iteration
# 1: Originally fc layers, upscoring layers
# 2: Scoring layer pool 4
# 3: Scoring layer pool 3
# 4: Convolutional layers
train_step_start = [200, 400, 600]
# Open output file
log_file_train = open(file_name_log_train, 'w', 1)
log_file_train.write('iteration,loss,accuracy\n')
log_file_val = open(file_name_log_val, 'w', 1)
log_file_val.write('iteration,loss,accuracy\n')
with tf.Session() as sess:
# Preprocess images
image = tf.placeholder(tf.float32)
label = tf.placeholder(tf.int32)
img_lab = tf.concat([image, tf.cast(label, tf.float32)], axis = 2)
img_lab = tf.random_crop(img_lab, size_random_crop)
img_lab = tf.image.random_flip_left_right(img_lab)
img_lab = tf.image.random_flip_up_down(img_lab)
img_lab = tf.cond(tf.reshape(tf.random_uniform([1]), []) > 0.5, lambda: img_lab,
lambda: tf.image.transpose_image(img_lab))
img_proc, lab_proc = tf.split(tf.expand_dims(img_lab, 0), [3, 1], 3)
lab_proc = tf.cast(tf.reshape(lab_proc, tf.shape(lab_proc)[0:-1]), tf.int32)
# Build network
vgg_fcn = fcn8_vgg.FCN8VGG(vgg16_npy_path='vgg16.npy', load_semantic_net=False)
with tf.name_scope("content_vgg"):
vgg_fcn.build(img_proc, debug=True, num_classes=3, train=True)
# Determine loss
loss = tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(labels=lab_proc, logits=vgg_fcn.upscore32))
# Declare optimizer
global_step = tf.Variable(0, trainable=False)
learning_rate = tf.train.exponential_decay(starter_learning_rate, global_step, decay_every, learning_decay_rate, staircase=False)
var_list_conv = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, "conv*")
var_list_fc = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, "fc*")
var_list_upscore = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, "upscore*")
var_list_fr = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, "score_fr*")
var_list_p4 = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, "score_pool4*")
var_list_p3 = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, "score_pool3*")
var_list1 = var_list_fc + var_list_upscore + var_list_fr
train_step1 = tf.train.AdamOptimizer(learning_rate).minimize(loss, global_step=global_step, var_list=var_list1)
var_list2 = var_list_fc + var_list_upscore + var_list_fr + var_list_p4
train_step2 = tf.train.AdamOptimizer(learning_rate).minimize(loss, global_step=global_step, var_list=var_list2)
var_list3 = var_list_fc + var_list_upscore + var_list_fr + var_list_p4 + var_list_p3
train_step3 = tf.train.AdamOptimizer(learning_rate).minimize(loss, global_step=global_step, var_list=var_list3)
var_list4 = var_list_fc + var_list_upscore + var_list_fr + var_list_p4 + var_list_p3 + var_list_conv
train_step4 = tf.train.AdamOptimizer(learning_rate).minimize(loss, global_step=global_step, var_list=var_list4)
print('Finished building Network.')
logging.warning("Score weights are initialized random.")
logging.warning("Do not expect meaningful results.")
logging.info("Start Initializing Variabels.")
init = tf.global_variables_initializer()
sess.run(init)
print('Running the Network')
# Do FCN training
for i in range(num_iter):
if i < train_step_start[0]:
train_step = train_step1
elif i < train_step_start[1]:
train_step = train_step2
elif i < train_step_start[2]:
train_step = train_step3
else:
train_step = train_step4
if (i + 1) % pred_every == 0:
ts_out, loss_np, pred, img_orig, lab_orig = sess.run(
[train_step, loss, vgg_fcn.pred_up, img_proc, lab_proc],
feed_dict={image: img_train[i % num_train, :, :, :], label: lab_train[i % num_train, :, :, :]})
acc = np.mean(pred[0] == lab_orig[0])
log_file_train.write(str(i + 1) + ',' + str(loss_np) + ',' + str(acc) + '\n')
pred_color = utils.color_image(pred[0], num_classes=3)
orig_color = utils.color_image(lab_orig[0], num_classes=3)
scp.misc.imsave('train_pred_%04i.png' % (i + 1), pred_color)
scp.misc.imsave('train_lab_orig_%04i.png' % (i + 1), orig_color)
scp.misc.imsave('train_img_orig_%04i.png' % (i + 1), img_orig[0])
else:
ts_out = sess.run(train_step,
feed_dict={image: img_train[i % num_train, :, :, :], label: lab_train[i % num_train, :, :, :]})
if (i + 1) % val_every == 0:
loss_np, pred, img_orig, lab_orig = sess.run(
[loss, vgg_fcn.pred_up, img_proc, lab_proc],
feed_dict={image: img_val[(i // val_every) % num_val, :, :, :], label: lab_val[(i // val_every) % num_val, :, :, :]})
acc = np.mean(pred[0] == lab_orig[0])
log_file_val.write(str(i + 1) + ',' + str(loss_np) + ',' + str(acc) + '\n')
pred_color = utils.color_image(pred[0], num_classes=3)
orig_color = utils.color_image(lab_orig[0], num_classes=3)
scp.misc.imsave('val_pred_%04i.png' % (i + 1), pred_color)
scp.misc.imsave('val_lab_orig_%04i.png' % (i + 1), orig_color)
scp.misc.imsave('val_img_orig_%04i.png' % (i + 1), img_orig[0])
if (i + 1) % save_every == 0:
vgg_fcn.save(sess, file_name='vgg16_%04i.npy' % (i + 1))
# Close output files
log_file_train.close()
log_file_val.close()