-
Notifications
You must be signed in to change notification settings - Fork 4
/
train_cls.py
executable file
·188 lines (168 loc) · 9.48 KB
/
train_cls.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
'''
MIT License
Copyright (c) 2019 Wentao Yuan
Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.
'''
import argparse
import importlib
import numpy as np
import tensorflow as tf
from termcolor import colored
import os
import sys
BASE_DIR = os.path.dirname(os.path.abspath(__file__))
sys.path.append(os.path.join(BASE_DIR, 'classifier'))
sys.path.append(os.path.join(BASE_DIR, 'transformer'))
sys.path.append(os.path.join(BASE_DIR, 'util'))
from data_util import lmdb_dataflow
from log_util import create_log_dir
from visu_util import plot_iters
n_classes = 40
cat_names = ['airplane','bathtub','bed','bench','bookshelf',
'bottle','bowl','car','chair','cone',
'cup','curtain','desk','door','dresser',
'flower_pot','glass_box','guitar','keyboard','lamp',
'laptop','mantel','monitor','night_stand','person',
'piano','plant','radio','range_hood','sink',
'sofa','stairs','stool','table','tent',
'toilet','tv_stand','vase','wardrobe','xbox']
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--data_dir')
parser.add_argument('--log_dir')
parser.add_argument('--classifier', choices=['pointnet', 'dgcnn'])
parser.add_argument('--transformer', choices=['t_net', 'it_net', 'it_net_dgcnn'])
parser.add_argument('--regularize', action='store_true')
parser.add_argument('--n_iter', type=int, default=2)
parser.add_argument('--batch_size', type=int, default=32)
parser.add_argument('--num_points', type=int, default=1024)
parser.add_argument('--init_lr', type=float, default=0.001)
parser.add_argument('--lr_decay_rate', type=float, default=0.7)
parser.add_argument('--lr_decay_steps', type=int, default=400000)
parser.add_argument('--lr_clip', type=float, default=0.00001)
parser.add_argument('--init_bn_decay', type=float, default=0.5)
parser.add_argument('--bn_decay_decay_rate', type=float, default=0.5)
parser.add_argument('--bn_decay_decay_steps', type=int, default=400000)
parser.add_argument('--bn_decay_clip', type=float, default=0.99)
parser.add_argument('--grad_clip', type=float, default=30.0)
parser.add_argument('--max_epoch', type=int, default=60)
parser.add_argument('--epoch_per_save', type=int, default=10)
parser.add_argument('--steps_per_print', type=int, default=100)
parser.add_argument('--steps_per_eval', type=int, default=300)
parser.add_argument('--steps_per_plot', type=int, default=3000)
parser.add_argument('--val_steps_per_plot', type=int, default=2)
parser.add_argument('--plot_lim', type=float, default=0.7)
parser.add_argument('--plot_size', type=float, default=2)
parser.add_argument('--iter_plot_freq', type=int, default=1)
parser.add_argument('--restore', action='store_true')
parser.add_argument('--checkpoint')
args = parser.parse_args()
global_step = tf.Variable(0, trainable=False, name='global_step')
learning_rate = tf.train.exponential_decay(args.init_lr, global_step * args.batch_size,
args.lr_decay_steps, args.lr_decay_rate,
staircase=True, name='learning_rate')
learning_rate = tf.maximum(learning_rate, args.lr_clip)
bn_decay = 1 - tf.train.exponential_decay(args.init_bn_decay, global_step * args.batch_size,
args.bn_decay_decay_steps, args.bn_decay_decay_rate,
staircase=True, name='bn_decay')
bn_decay = tf.minimum(bn_decay, args.bn_decay_clip)
is_training_pl = tf.placeholder(tf.bool, (), 'is_training')
points_pl = tf.placeholder(tf.float32, (args.batch_size, args.num_points, 3), 'points')
labels_pl = tf.placeholder(tf.int32, (args.batch_size,), 'labels')
with tf.variable_scope('transformer', reuse=tf.AUTO_REUSE):
transformer = importlib.import_module(args.transformer)
transformed_points, T_out, Ts = transformer.get_model(points_pl, args.n_iter,
is_training_pl, bn_decay)
with tf.variable_scope('classifier'):
classifier = importlib.import_module(args.classifier)
logits = classifier.get_model(transformed_points, is_training_pl, bn_decay)
prediction = tf.argmax(logits, axis=1)
if args.regularize:
loss_op = classifier.get_loss_reg(logits, labels_pl, T_out[:, :3, :3])
else:
loss_op = classifier.get_loss(logits, labels_pl)
trainer = tf.train.AdamOptimizer(learning_rate)
grad, var = zip(*trainer.compute_gradients(loss_op, tf.trainable_variables()))
grad, global_norm = tf.clip_by_global_norm(grad, args.grad_clip)
train_op = trainer.apply_gradients(zip(grad, var), global_step)
avg_loss, loss_update = tf.metrics.mean(loss_op)
avg_acc, acc_update = tf.metrics.accuracy(labels_pl, prediction)
tf.summary.scalar('train/learning rate', learning_rate, collections=['train'])
tf.summary.scalar('train/gradient norm', global_norm, collections=['train'])
tf.summary.scalar('train/bn decay', bn_decay, collections=['train'])
tf.summary.scalar('valid/loss', avg_loss, collections=['valid'])
tf.summary.scalar('valid/accuracy', avg_acc, collections=['valid'])
update_ops = [loss_update, acc_update]
train_summary = tf.summary.merge_all('train')
valid_summary = tf.summary.merge_all('valid')
lmdb_train = os.path.join(args.data_dir, 'train.lmdb')
lmdb_valid = os.path.join(args.data_dir, 'valid.lmdb')
df_train, num_train = lmdb_dataflow(lmdb_train, args.batch_size, args.num_points,
shuffle=True, task='cls')
df_valid, num_valid = lmdb_dataflow(lmdb_valid, args.batch_size, args.num_points,
shuffle=False, task='cls')
train_gen = df_train.get_data()
valid_gen = df_valid.get_data()
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
config.allow_soft_placement = True
sess = tf.Session(config=config)
sess.run(tf.global_variables_initializer())
writer = create_log_dir(args, sess)
saver = tf.train.Saver()
step = sess.run(global_step)
epoch = step * args.batch_size // num_train + 1
next_epoch = epoch
while next_epoch <= args.max_epoch:
step += 1
epoch = step * args.batch_size // num_train + 1
_, points, labels, poses = next(train_gen)
_, loss, summary = sess.run([train_op, loss_op, train_summary],
feed_dict={is_training_pl: True,
points_pl: points,
labels_pl: labels})
writer.add_summary(summary, step)
if step % args.steps_per_print == 0:
print('Epoch %d Step %d Loss %f' % (epoch, step, loss))
if step % args.steps_per_eval == 0:
sess.run(tf.local_variables_initializer())
for i in range(num_valid // args.batch_size):
instance_id, points, labels, poses = next(valid_gen)
ts, pred, _ = sess.run([Ts, prediction, update_ops],
feed_dict={is_training_pl: False,
points_pl: points,
labels_pl: labels})
if step % args.steps_per_plot == 0 and (i + 1) % args.val_steps_per_plot == 0:
T = np.eye(4)
transforms = []
titles = []
for j in range(args.n_iter+1):
if j > 0:
T = np.dot(ts[j-1][0], T)
if j % args.iter_plot_freq == 0:
transforms.append(T)
titles.append('Iteration %d' % j)
suptitle = 'Predicted %s Actual %s' % (cat_names[pred[0]], cat_names[labels[0]])
figpath = '%s/plots/%s_step_%d.png' % (args.log_dir, instance_id[0], step)
plot_iters(figpath, points[0], transforms, titles, args.plot_lim, args.plot_size, suptitle)
loss, acc = sess.run([avg_loss, avg_acc])
print(colored('Loss %f Accuracy %.4f ' % (loss, acc), 'white', 'on_blue'))
summary = sess.run(valid_summary)
writer.add_summary(summary, step)
next_epoch = (step+1) * args.batch_size // num_train + 1
if epoch % args.epoch_per_save == 0 and epoch < next_epoch:
saver.save(sess, os.path.join(args.log_dir, 'model_epoch_%d' % epoch))