-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathtrain.py
144 lines (120 loc) · 6.64 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
import argparse
import time
from model import *
from Dataset import *
from losses import *
from utils import *
class Trainer(object):
def __init__(self, args):
self.args = args
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
self.sess = tf.Session(config=config)
self._build_graph()
def _build_graph(self):
shared_keys = ['dataset_dir', 'batch_size', 'num_parallel_calls']
shared_args = {}
for key in shared_keys:
shared_args[key] = getattr(self.args, key)
# Input data
with tf.name_scope('Data'):
tset = BaseDataset_3D(train_or_val='train', **shared_args)
self.images = tset.get_element()
self.image0, self.image1 = self.images[:, 0], self.images[:, 1]
vset = BaseDataset_3D(train_or_val='val', **shared_args)
self.images_v, self.initializer_v = vset.get_element()
self.image0_v, self.image1_v = self.images_v[:, 0], self.images_v[:, 1]
self.num_batches = len(tset.samples) // self.args.batch_size
self.num_batches_v = len(vset.samples) // self.args.batch_size
# Model inference
model = BiLevelNet_train(num_levels=self.args.num_levels)
self.flows_final, self.y, self.flows, self.pypa_0, self.pypa_1 = model(self.image0, self.image1) # image0 image1 (fixed and moving)
self.flows_final_v, self.y_v, self.flows_v, self.pypa_0_v, self.pypa_1_v \
= model(self.image0_v, self.image1_v, reuse=True)
# Loss calculation
with tf.name_scope('Loss'):
_loss = multirobust_ncc(self.image0, self.image1, self.flows, self.args.weights, self.args.num_levels)
_loss_v = multirobust_ncc(self.image0_v, self.image1_v, self.flows_v, self.args.weights,
self.args.num_levels)
self.ncc = ncc(self.image0, self.y)
self.ncc_v = ncc(self.image0_v, self.y_v)
weights_l2 = Grad(self.flows_final)
weights_l2_v = Grad(self.flows_final_v)
kl_loss = multirobust_MAP(self.pypa_0, self.pypa_1, self.args.weights)
kl_loss_v = multirobust_MAP(self.pypa_0_v, self.pypa_1_v, self.args.weights)
self.loss = _loss + self.args.gamma * kl_loss + 15 * self.ncc + weights_l2 * 10
self.loss_v = _loss_v + self.args.gamma * kl_loss_v + 10 * self.ncc_v + weights_l2_v * 12
self.kl_loss, self.weights_l2 = kl_loss, weights_l2
self.kl_loss_v, self.weights_l2_v = kl_loss_v, weights_l2_v
# Gradient descent optimization
with tf.name_scope('Optimize'):
self.global_step = tf.train.get_or_create_global_step()
lr = self.args.lr
self.optimizer = tf.train.AdamOptimizer(learning_rate=lr).minimize(self.loss, var_list=model.vars)
with tf.control_dependencies([self.optimizer]):
self.optimizer = tf.assign_add(self.global_step, 1)
# Load learned model
self.saver = tf.train.Saver(model.vars, max_to_keep=20)
self.sess.run(tf.global_variables_initializer())
if self.args.resume is not None:
print(f'Loading learned model from checkpoint {self.args.resume}')
self.saver.restore(self.sess, self.args.resume)
tf.summary.FileWriter('./logs', graph=self.sess.graph)
def train(self):
train_start = time.time()
for e in range(self.args.num_epochs):
for i in range(self.num_batches):
time_s = time.time()
_, loss, ncc, reg_loss, kl_loss = self.sess.run(
[self.optimizer, self.loss, self.ncc, self.weights_l2, self.kl_loss])
if i % 100 == 0:
batch_time = time.time() - time_s
kwargs = {'loss': loss, 'ncc': ncc, 'reg_loss': reg_loss,
'kl_loss': kl_loss, 'batch time': batch_time}
show_progress(e + 1, i + 1, self.num_batches, **kwargs)
loss_vals, ncc_vals, reg_vals, kl_vals = [], [], [], []
self.sess.run([self.initializer_v])
for i in range(self.num_batches_v):
image0_v, image1_v, flows_val, loss_val, ncc_val, reg_val, kl_val \
= self.sess.run([self.image0_v, self.image1_v, self.flows_v,
self.loss_v, self.ncc_v, self.weights_l2_v, self.kl_loss_v])
loss_vals.append(loss_val)
ncc_vals.append(ncc_val)
reg_vals.append(reg_val)
kl_vals.append(kl_val)
g_step = self.sess.run(self.global_step)
print(
f'\r{e+1} epoch validation, loss: {np.mean(loss_vals)}, ncc: {np.mean(ncc_vals)}, reg_loss:{np.mean(reg_vals)}, '
f'kl_loss:{np.mean(kl_vals)}' \
+ f', global step: {g_step}, elapsed time: {time.time()-train_start} sec.')
if not os.path.exists('model/pro-model'):
os.mkdir('model/pro-model')
self.saver.save(self.sess, f'model/pro-model/model_{e + 1}.ckpt')
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('-dd', '--dataset_dir', type=str,
help='Directory containing train dataset')
parser.add_argument('-e', '--num_epochs', type=int, default=100,
help='# of epochs ')
parser.add_argument('-b', '--batch_size', type=int, default=1,
help='Batch size ')
parser.add_argument('-nc', '--num_parallel_calls', type=int, default=1,
help='# of parallel calls for data loading ')
parser.add_argument('--num_levels', type=int, default=2,
help='# of levels for feature extraction ')
parser.add_argument('--lr', type=float, default=1e-4,
help='Learning rate [1e-4]')
parser.add_argument('--weights', nargs='+', type=float,
default=[ 0.8, 3.2],
help='Weights for each pyramid loss')
parser.add_argument('--gamma', type=float, default=0.2,
help='Coefficient for weight decay [4e-4]')
parser.add_argument('-r', '--resume', type=str,
default= None,
help='Learned parameter checkpoint file [None]')
args = parser.parse_args()
for key, item in vars(args).items():
print(f'{key} : {item}')
os.environ['CUDA_VISIBLE_DEVICES'] = '0'
trainer = Trainer(args)
trainer.train()