-
Notifications
You must be signed in to change notification settings - Fork 24
/
train.py
147 lines (111 loc) · 4.42 KB
/
train.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
from pathlib import Path
import random
import numpy as np
import scipy.io
import scipy.misc
import tensorflow as tf
import tensorflow.contrib.slim as slim
import slim_net
IMAGE_DIR = Path("./data/images_data_crop")
MASK_DIR = Path("./data/images_mask")
BATCH_SIZE = 2
HEIGHT = 800
WIDTH = 600
LEARNING_RATE = 1e-4
NUM_CLASSES = 2
PRETRAINED_MODEL = "fcn8s-heavy-pascal.mat"
LAYER_ID_MAP = {
'conv1/conv1_1': [2, True],
'conv1/conv1_2': [4, True],
'conv2/conv2_1': [7, True],
'conv2/conv2_2': [9, True],
'conv3/conv3_1': [12, True],
'conv3/conv3_2': [14, True],
'conv3/conv3_3': [16, True],
'conv4/conv4_1': [20, True],
'conv4/conv4_2': [22, True],
'conv4/conv4_3': [24, True],
'conv5/conv5_1': [28, True],
'conv5/conv5_2': [30, True],
'conv5/conv5_3': [32, True],
'fc6': [35, True],
'fc7': [37, True],
'score_fr': [39, True],
'upscore2': [40, False],
'score_pool4': [42, True],
'upscore_pool4': [45, False],
'score_pool3': [47, True],
'upscore8': [50, False],
}
def assign_default_value(sess, output_dim):
caffe_mat = np.load(PRETRAINED_MODEL, encoding="latin1")
graph = tf.get_default_graph()
for layer_name, idxs in LAYER_ID_MAP.items():
idx, bias_term = idxs
weight = caffe_mat[idx][1][0].transpose((2, 3, 1, 0))
if bias_term:
bias = caffe_mat[idx][1][1]
if layer_name.startswith('upscore'):
weight = weight[:, :, :output_dim, :output_dim]
bias = bias[:output_dim]
if layer_name.startswith('score'):
weight = weight[:, :, :, :output_dim]
bias = bias[:output_dim]
weight_tensor_name = f"fcn8s/{layer_name}/weights:0"
sess.run(tf.assign(graph.get_tensor_by_name(weight_tensor_name), weight))
if bias_term:
bias_tensor_name = f"fcn8s/{layer_name}/biases:0"
sess.run(tf.assign(graph.get_tensor_by_name(bias_tensor_name), bias))
def build_image(filename):
MEAN_VALUES = np.array([104.00698793, 116.66876762, 122.67891434])
MEAN_VALUES = MEAN_VALUES.reshape((1, 1, 1, 3))
img = scipy.misc.imread(filename, mode="RGB")[:, :, ::-1]
height, width, _ = img.shape
img = np.reshape(img, (1, height, width, 3)) - MEAN_VALUES
return img
def train():
inputs = tf.placeholder(tf.float32, [BATCH_SIZE, HEIGHT, WIDTH, 3])
with slim.arg_scope(slim_net.fcn8s_arg_scope()):
logits, _ = slim_net.fcn8s(inputs, NUM_CLASSES)
label = tf.placeholder(tf.uint8, shape=[BATCH_SIZE, HEIGHT, WIDTH])
cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(
logits=tf.reshape(logits, [-1, 2]),
labels=tf.stop_gradient(tf.one_hot(tf.reshape(label, [-1]), NUM_CLASSES))))
with tf.Session() as sess:
global_step = tf.Variable(0, name="global_step", trainable=False)
saver = tf.train.Saver(tf.global_variables())
model_file = tf.train.latest_checkpoint("./model/")
if model_file:
print(f"Restore from {model_file}")
saver.restore(sess, model_file)
else:
print("Initialize from pre-trained model")
sess.run(tf.global_variables_initializer())
assign_default_value(sess, NUM_CLASSES)
print("Start to train")
optimizer = tf.train.GradientDescentOptimizer(LEARNING_RATE)
train_op = optimizer.minimize(cost, global_step=global_step)
all_images = list(IMAGE_DIR.glob("*.jpg"))
while True:
image_mat = []
label_mat = []
images = random.sample(all_images, BATCH_SIZE)
for image_fullpath in images:
mask_filename = f"{image_fullpath.stem}_mask.mat"
mask_fullpath = MASK_DIR / mask_filename
image_mat.append(build_image(image_fullpath))
label_mat.append(scipy.io.loadmat(mask_fullpath)["mask"])
feed_dict = {
inputs: np.concatenate(image_mat),
label: np.stack(label_mat)
}
_, loss, step = sess.run([train_op, cost, global_step], feed_dict=feed_dict)
if step % 1 == 0:
print(f"[Step {step}] Loss: {loss}")
if step % 500 == 0:
saver.save(sess, "./model/PortraitFCN", global_step=step)
print(f"[Step {step}] Saved")
if step >= 100000:
break
if __name__ == "__main__":
train()