forked from argman/EAST
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmultigpu_train.py
182 lines (148 loc) · 8.16 KB
/
multigpu_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
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
import time
import numpy as np
import tensorflow as tf
from tensorflow.contrib import slim
tf.app.flags.DEFINE_integer('input_size', 512, '')
tf.app.flags.DEFINE_integer('batch_size_per_gpu', 14, '')
tf.app.flags.DEFINE_integer('num_readers', 24, '')
tf.app.flags.DEFINE_float('learning_rate', 0.0001, '')
tf.app.flags.DEFINE_integer('max_steps', 20000, '')
tf.app.flags.DEFINE_float('moving_average_decay', 0.997, '')
tf.app.flags.DEFINE_string('gpu_list', '0,1,2,3', '')
tf.app.flags.DEFINE_string('checkpoint_path', '/home/xu/check_points_expanded_convs/', '')
tf.app.flags.DEFINE_boolean('restore', False, 'whether to resotre from checkpoint')
tf.app.flags.DEFINE_integer('save_checkpoint_steps', 2000, '')
tf.app.flags.DEFINE_integer('save_summary_steps', 100, '')
tf.app.flags.DEFINE_string('pretrained_model_path', '/home/xu/pre_trained/mobilenet_v2_1.0_224/mobilenet_v2_1.0_224.ckpt', '')
import model
import icdar
FLAGS = tf.app.flags.FLAGS
gpus = list(range(len(FLAGS.gpu_list.split(','))))
print(gpus)
def tower_loss(images, score_maps, geo_maps, training_masks, reuse_variables=None):
# Build inference graph
with tf.variable_scope(tf.get_variable_scope(), reuse=reuse_variables):
f_score, f_geometry = model.model(images, is_training=True)
model_loss = model.loss(score_maps, f_score,
geo_maps, f_geometry,
training_masks)
total_loss = tf.add_n([model_loss] + tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES))
# add summary
if reuse_variables is None:
tf.summary.image('input', images)
tf.summary.image('score_map', score_maps)
tf.summary.image('score_map_pred', f_score * 255)
tf.summary.image('geo_map_0', geo_maps[:, :, :, 0:1])
tf.summary.image('geo_map_0_pred', f_geometry[:, :, :, 0:1])
tf.summary.image('training_masks', training_masks)
tf.summary.scalar('model_loss', model_loss)
tf.summary.scalar('total_loss', total_loss)
return total_loss, model_loss
def average_gradients(tower_grads):
average_grads = []
for grad_and_vars in zip(*tower_grads):
grads = []
for g, _ in grad_and_vars:
expanded_g = tf.expand_dims(g, 0)
grads.append(expanded_g)
grad = tf.concat(grads, 0)
grad = tf.reduce_mean(grad, 0)
v = grad_and_vars[0][1]
grad_and_var = (grad, v)
average_grads.append(grad_and_var)
return average_grads
def main(argv=None):
import os
os.environ['CUDA_VISIBLE_DEVICES'] = FLAGS.gpu_list
if not tf.gfile.Exists(FLAGS.checkpoint_path):
tf.gfile.MkDir(FLAGS.checkpoint_path)
else:
if not FLAGS.restore:
tf.gfile.DeleteRecursively(FLAGS.checkpoint_path)
tf.gfile.MkDir(FLAGS.checkpoint_path)
input_images = tf.placeholder(tf.float32, shape=[None, 512, 512, 3], name='input_images')
input_score_maps = tf.placeholder(tf.float32, shape=[None, 128, 128, 1], name='input_score_maps')
if FLAGS.geometry == 'RBOX':
input_geo_maps = tf.placeholder(tf.float32, shape=[None, 128, 128, 5], name='input_geo_maps')
else:
input_geo_maps = tf.placeholder(tf.float32, shape=[None, 128, 128, 8], name='input_geo_maps')
input_training_masks = tf.placeholder(tf.float32, shape=[None, 128, 128, 1], name='input_training_masks')
global_step = tf.get_variable('global_step', [], initializer=tf.constant_initializer(0), trainable=False)
learning_rate = tf.train.exponential_decay(FLAGS.learning_rate, global_step, decay_steps=10000, decay_rate=0.94, staircase=True)
# add summary
tf.summary.scalar('learning_rate', learning_rate)
opt = tf.train.AdamOptimizer(learning_rate)
# opt = tf.train.MomentumOptimizer(learning_rate, 0.9)
# split
input_images_split = tf.split(input_images, len(gpus))
input_score_maps_split = tf.split(input_score_maps, len(gpus))
input_geo_maps_split = tf.split(input_geo_maps, len(gpus))
input_training_masks_split = tf.split(input_training_masks, len(gpus))
tower_grads = []
reuse_variables = None
for i, gpu_id in enumerate(gpus):
with tf.device('/gpu:%d' % gpu_id):
with tf.name_scope('model_%d' % gpu_id) as scope:
iis = input_images_split[i]
isms = input_score_maps_split[i]
igms = input_geo_maps_split[i]
itms = input_training_masks_split[i]
total_loss, model_loss = tower_loss(iis, isms, igms, itms, reuse_variables)
batch_norm_updates_op = tf.group(*tf.get_collection(tf.GraphKeys.UPDATE_OPS, scope))
reuse_variables = True
grads = opt.compute_gradients(total_loss)
tower_grads.append(grads)
grads = average_gradients(tower_grads)
apply_gradient_op = opt.apply_gradients(grads, global_step=global_step)
summary_op = tf.summary.merge_all()
# save moving average
variable_averages = tf.train.ExponentialMovingAverage(
FLAGS.moving_average_decay, global_step)
variables_averages_op = variable_averages.apply(tf.trainable_variables())
# batch norm updates
with tf.control_dependencies([variables_averages_op, apply_gradient_op, batch_norm_updates_op]):
train_op = tf.no_op(name='train_op')
saver = tf.train.Saver(tf.global_variables())
summary_writer = tf.summary.FileWriter(FLAGS.checkpoint_path, tf.get_default_graph())
init = tf.global_variables_initializer()
if FLAGS.pretrained_model_path is not None:
variable_restore_op = slim.assign_from_checkpoint_fn(FLAGS.pretrained_model_path, slim.get_trainable_variables(),
ignore_missing_vars=True)
with tf.Session(config=tf.ConfigProto(allow_soft_placement=True)) as sess:
if FLAGS.restore:
print('continue training from previous checkpoint')
ckpt = tf.train.latest_checkpoint(FLAGS.checkpoint_path)
saver.restore(sess, ckpt)
else:
sess.run(init)
if FLAGS.pretrained_model_path is not None:
variable_restore_op(sess)
data_generator = icdar.get_batch(num_workers=FLAGS.num_readers,
input_size=FLAGS.input_size,
batch_size=FLAGS.batch_size_per_gpu * len(gpus))
start = time.time()
for step in range(FLAGS.max_steps):
data = next(data_generator)
ml, tl, _ = sess.run([model_loss, total_loss, train_op], feed_dict={input_images: data[0],
input_score_maps: data[2],
input_geo_maps: data[3],
input_training_masks: data[4]})
if np.isnan(tl):
print('Loss diverged, stop training')
break
if step % 10 == 0:
avg_time_per_step = (time.time() - start)/10
avg_examples_per_second = (10 * FLAGS.batch_size_per_gpu * len(gpus))/(time.time() - start)
start = time.time()
print('Step {:06d}, model loss {:.4f}, total loss {:.4f}, {:.2f} seconds/step, {:.2f} examples/second'.format(
step, ml, tl, avg_time_per_step, avg_examples_per_second))
if step % FLAGS.save_checkpoint_steps == 0:
saver.save(sess, FLAGS.checkpoint_path + 'model.ckpt', global_step=global_step)
if step % FLAGS.save_summary_steps == 0:
_, tl, summary_str = sess.run([train_op, total_loss, summary_op], feed_dict={input_images: data[0],
input_score_maps: data[2],
input_geo_maps: data[3],
input_training_masks: data[4]})
summary_writer.add_summary(summary_str, global_step=step)
if __name__ == '__main__':
tf.app.run()