-
Notifications
You must be signed in to change notification settings - Fork 6
/
train_multiple_tasks.py
212 lines (170 loc) · 9.48 KB
/
train_multiple_tasks.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
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
"""Training script for the DeepLab-ResNet network on the PASCAL VOC dataset
for semantic image segmentation.
This script trains the model using augmented PASCAL VOC,
which contains approximately 10000 images for training and 1500 images for validation.
"""
import argparse
from datetime import datetime
import os
import sys
import time
import tensorflow as tf
import numpy as np
from nets import vgg_16_deeplab_nddr, vgg_16_shortcut_deeplab_nddr
from nets import vgg_16_deeplab_mt, vgg_16_shortcut_deeplab_mt
from nets import vgg_arg_scope
from util.img_encoder_mt import ImageReader
from util.img_decoder import prepare_label
from util.input_arguments import arguments_mt_train
from ops.losses import get_seg_loss, get_normal_loss
from ops.create_train_ops import create_train_ops_mt
from ops.create_summary_ops import create_summary_op_mt
from ops.save_ckpt import save
from ops.load_ckpt import load_mt
slim = tf.contrib.slim
IMG_MEAN = np.array((104.00698793,116.66876762,122.67891434), dtype=np.float32)
def train():
"""Create the model and start the training."""
args = arguments_mt_train()
h, w = map(int, args.input_size.split(','))
input_size = (h, w)
init_weights = map(float, args.init_weights.split(','))
if args.use_random_seed:
tf.set_random_seed(args.random_seed)
# Create queue coordinator.
coord = tf.train.Coordinator()
# Load reader.
with tf.name_scope("create_inputs"):
reader = ImageReader(
args.data_dir,
args.data_list_1,
args.data_list_2,
input_size=input_size,
random_scale=args.random_scale,
random_mirror=args.random_mirror,
random_crop=args.random_crop,
ignore_label=args.ignore_label,
img_mean=IMG_MEAN,
coord=coord,
task_1=args.task_1,
task_2=args.task_2)
image_batch, label_batch_1, label_batch_2 = reader.dequeue(args.batch_size)
# Create network.
with slim.arg_scope(vgg_arg_scope(weight_decay=args.weight_decay, use_batch_norm=True, is_training=True)):
if args.network == 'vgg_16_deeplab_nddr':
_, _, end_points_1, end_points_2, _ = vgg_16_deeplab_nddr(image_batch,
num_classes_1=args.num_classes_1,
num_classes_2=args.num_classes_2,
is_training=True,
dropout_keep_prob=args.keep_prob,
init_method=args.init_method,
init_weights=init_weights)
elif args.network == 'vgg_16_shortcut_deeplab_nddr':
_, _, end_points_1, end_points_2, _ = vgg_16_shortcut_deeplab_nddr(image_batch,
num_classes_1=args.num_classes_1,
num_classes_2=args.num_classes_2,
is_training=True,
dropout_keep_prob=args.keep_prob,
init_method=args.init_method,
init_weights=init_weights)
elif args.network == 'vgg_16_deeplab_mt':
_, _, end_points_1, end_points_2, _ = vgg_16_deeplab_mt(image_batch,
num_classes_1=args.num_classes_1,
num_classes_2=args.num_classes_2,
is_training=True,
dropout_keep_prob=args.keep_prob)
elif args.network == 'vgg_16_shortcut_deeplab_mt':
_, _, end_points_1, end_points_2, _ = vgg_16_shortcut_deeplab_mt(image_batch,
num_classes_1=args.num_classes_1,
num_classes_2=args.num_classes_2,
is_training=True,
dropout_keep_prob=args.keep_prob)
else:
raise Exception('network name is not recognized!')
# Predictions.
raw_output_1 = end_points_1['vgg_16_1_6/fc8_voc12']
raw_output_2 = end_points_2['vgg_16_2_6/fc8_voc12']
if args.task_1 == 'seg' and args.task_2 == 'normal':
seg_output = raw_output_1
seg_classes = args.num_classes_1
seg_label = prepare_label(label_batch_1, tf.stack(raw_output_1.get_shape()[1:3]), num_classes=seg_classes,
one_hot=False, task='seg')
seg_label_batch = label_batch_1
seg_loss_scale = args.loss_scale_1
normal_output = raw_output_2
normal_classes = args.num_classes_2
normal_label = prepare_label(label_batch_2, tf.stack(raw_output_2.get_shape()[1:3]), num_classes=normal_classes,
one_hot=False, task='normal')
normal_label_batch = label_batch_2
normal_loss_scale = args.loss_scale_2
elif args.task_1 == 'normal' and args.task2 == 'seg':
seg_output = raw_output_2
seg_classes = args.num_classes_2
seg_label = prepare_label(label_batch_2, tf.stack(raw_output_2.get_shape()[1:3]), num_classes=seg_classes,
one_hot=False, task='seg')
seg_label_batch = label_batch_2
seg_loss_scale = args.loss_scale_2
normal_output = raw_output_1
normal_classes = args.num_classes_1
normal_label = prepare_label(label_batch_1, tf.stack(raw_output_1.get_shape()[1:3]), num_classes=normal_classes,
one_hot=False, task='normal')
normal_label_batch = label_batch_1
normal_loss_scale = args.loss_scale_1
else:
raise Exception('check the tasks!')
seg_loss = get_seg_loss(seg_output, seg_label, seg_classes, args.ignore_label)
normal_loss = get_normal_loss(normal_output, normal_label, normal_classes, args.ignore_label)
l2_losses = tf.add_n([args.weight_decay * tf.nn.l2_loss(v) for v in tf.trainable_variables() if 'weights' in v.name])
reduced_loss = tf.reduce_mean(seg_loss) * seg_loss_scale + tf.reduce_mean(normal_loss) * normal_loss_scale + l2_losses * 0.5
# Train Ops
train_op, step_ph, learning_rate = create_train_ops_mt(reduced_loss, args)
# Summaries
seg_pred = tf.image.resize_bilinear(seg_output, tf.shape(image_batch)[1:3, ])
seg_pred = tf.argmax(seg_pred, axis=3)
seg_pred = tf.expand_dims(seg_pred, dim=3)
normal_pred = tf.image.resize_bilinear(normal_output, tf.shape(image_batch)[1:3, ])
if 'nddr' in args.network:
nddr_l2_losses = tf.add_n([args.weight_decay * tf.nn.l2_loss(v) for v in tf.trainable_variables() if 'weights' in v.name and 'nddr' in v.name])
else:
nddr_l2_losses = tf.constant(0.0)
summary_op = create_summary_op_mt(image_batch, seg_label_batch, normal_label_batch, seg_pred, normal_pred, IMG_MEAN,
reduced_loss, seg_loss, normal_loss, l2_losses, nddr_l2_losses,
learning_rate, seg_classes, normal_classes, args.save_num_images, args.network)
summary_writer = tf.summary.FileWriter(args.snapshot_dir,
graph=tf.get_default_graph())
# Set up tf session and initialize variables.
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
init = tf.global_variables_initializer()
sess.run(init)
# Saver for storing checkpoints of the model.
saver = tf.train.Saver(var_list=tf.global_variables(), max_to_keep=10)
# Load variables if the checkpoint is provided.
load_mt(sess, args)
# Start queue threads.
threads = tf.train.start_queue_runners(coord=coord, sess=sess)
# Iterate over training steps.
for step in range(args.num_steps):
start_time = time.time()
feed_dict = {step_ph: step}
if step % args.save_pred_every == 0:
loss_value, seg_loss_value, normal_loss_value, loss_weight_value, loss_nddr_value, images, labels_1, \
labels_2, seg_pred_value, normal_pred_value, summary, _ = \
sess.run([reduced_loss, seg_loss, normal_loss, l2_losses, nddr_l2_losses, image_batch, label_batch_1,
label_batch_2, seg_pred, normal_pred, summary_op, train_op], feed_dict=feed_dict)
save(saver, sess, args.snapshot_dir, step)
else:
loss_value, seg_loss_value, normal_loss_value, loss_weight_value, loss_nddr_value, summary, _ = \
sess.run([reduced_loss, seg_loss, normal_loss, l2_losses, nddr_l2_losses, summary_op, train_op],
feed_dict=feed_dict)
summary_writer.add_summary(summary, step)
duration = time.time() - start_time
print('step {:d} \t loss = {:.3f} \t seg_loss = {:.3f}, seg_loss_scale = {:.1f} \t normal_loss = {:.3f}, '
'normal_loss_scale = {:.1f} \t loss_w = {:.3f}, loss_nddr_w = {:.3f} \t nddr_mult = {:.1f}'
.format(step, loss_value, seg_loss_value, seg_loss_scale, normal_loss_value, normal_loss_scale,
loss_weight_value, loss_nddr_value, args.nddr_mult))
coord.request_stop()
coord.join(threads)
if __name__ == '__main__':
train()