-
Notifications
You must be signed in to change notification settings - Fork 1
/
train.py
184 lines (137 loc) · 8.34 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
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 os
import cv2
import numpy as np
import tensorflow as tf
import argparse
import progressbar
import signal
gentle_stop = np.array((1,), dtype=np.bool)
gentle_stop[0] = False
def on_ctrl_c(signal, frame):
gentle_stop[0] = True
signal.signal(signal.SIGINT, on_ctrl_c)
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'
num_gpus_used = len(os.environ['CUDA_VISIBLE_DEVICES'].split(','))
from utils import convert_batch_to_image_grid, add_summary, build_encoder, build_decoder, build_conditioned_block, build_dataset
from network import VectorQuantizerEMA, compute_reconstruction_losses
parser = argparse.ArgumentParser()
parser.add_argument('--training_data_type',default='tless_f18',\
choices=['tless_f18','camera_all','camera_bottle','camera_bowl','camera_camera','camera_can','camera_laptop','camera_mug'], \
help='Training dataset, tless_f18 for first 18 tless objects, camera_all for a combination of all 6 categories in CAMERA, and camera_category for the specified category in CAMERA.')
parser.add_argument('--dir_tfrecords',default='./dir/to/tfrecords',help='Path of directory for tfrecords.')
parser.add_argument('--dir_bg_imgs',default='./dir/to/background/images',help='Path of directory for background images (in jpg format).')
parser.add_argument('--normalize_encoder',action="store_true",help='Apply instance normalization in encoder.')
parser.add_argument('--decoder_branches', default='bgr',choices=['bgr','depth','bgr+depth'],help='Decoder branches.')
parser.add_argument('--tau',default=0.07, type=float,help='Temperature for shape loss.')
parser.add_argument('--ema_decay',default=0.9995,type=float,help='EMA decay for shape embedding.')
parser.add_argument('--lambda_shape_loss',default=0.004,type=float,help='Weight for shape loss')
parser.add_argument('--lambda_pose_loss',default=0.002,type=float,help='Weight for pose loss')
parser.add_argument('--lr',default=2e-4,type=float,help='Learning rate.')
parser.add_argument('--batch_size',default=64,type=int,help='Total batch size')
parser.add_argument('--num_ites',default=50000,type=int,help='Number of iterations for training')
parser.add_argument('--dir_checkpoints',default='./ws/my_exp/',help='path to save checkpoints and summary for tensorboard.')
args = parser.parse_args()
if not os.path.exists(args.dir_checkpoints):
os.makedirs(args.dir_checkpoints)
dir_train_figs=os.path.join(args.dir_checkpoints,'train_figures')
if not os.path.exists(dir_train_figs):
os.makedirs(dir_train_figs)
assert num_gpus_used==1, 'todo: for multiple gpu'
tf.reset_default_graph()
with tf.variable_scope('my_exp'):
with tf.variable_scope('data_queue'):
dataset,iterator=build_dataset(dataset_tag=args.training_data_type,\
dir_tfrecords=args.dir_tfrecords,
dir_bg_imgs=args.dir_bg_imgs,
batch_size=args.batch_size)
obj_id_to_embed_order = tf.convert_to_tensor(dataset.obj_id_to_embed_order, dtype=tf.int32)
next_element=dataset.next_element
if args.decoder_branches=='bgr':
enc_x,dec_gt_bgr,obj_label,hsh_code=next_element
dec_gt_depth=tf.constant(0.)
elif args.decoder_branches=='bgr+depth':
enc_x,dec_gt_bgr, dec_gt_depth,obj_label,hsh_code=next_element
elif args.decoder_branches=='depth':
enc_x,dec_gt_depth,obj_label,hsh_code=next_element
dec_gt_bgr=tf.constant(0.)
obj_label=tf.nn.embedding_lookup(obj_id_to_embed_order,obj_label)
with tf.variable_scope('encoder'):
encoder = build_encoder(inst_norm=args.normalize_encoder)
z = encoder(enc_x)
with tf.variable_scope('gprior'):
embedding_dim=encoder._latent_space_size
obj_embeds = VectorQuantizerEMA(embedding_dim=embedding_dim, num_embeddings=dataset.num_objs, name='obj_prior')
cond_block = build_conditioned_block()
obj_retr = obj_embeds(z['z_obj'], decay=args.ema_decay, temperature=args.tau, encoding_1nn_indices=obj_label,\
encodings=None, is_training=(num_gpus_used == 1))
cond_pose_code=cond_block(condition_code=tf.stop_gradient(z['z_obj']),src_code=tf.reshape(hsh_code,[-1,1,embedding_dim]))
cond_pose_code=tf.reshape(cond_pose_code,[-1,embedding_dim])
with tf.variable_scope('decoder'):
decoder_branches={}
for db in ['bgr','depth']:
if db in args.decoder_branches:
decoder_branches[db]=1.
decoder = build_decoder(branches=list(decoder_branches.keys()))
dec_out = decoder(pose_code=z['z_pose'], obj_code=z['z_obj'], is_training=True)
with tf.variable_scope('loss'):
summaries = {}
total_loss = 0.
dec_gt={'bgr':dec_gt_bgr,'depth':dec_gt_depth}
recon_error = compute_reconstruction_losses(dec_out,dec_gt, decoder_branches)
total_loss += recon_error['reconst_loss']
for k in decoder_branches.keys():
summaries['reconst_loss_{:s}'.format(k)] = recon_error['reconst_loss_{:s}'.format(k)]
total_loss += obj_retr["loss"]*args.lambda_shape_loss
summaries['shape_loss'] = obj_retr["loss"]
pose_loss = tf.reduce_mean((tf.nn.l2_normalize(cond_pose_code, axis=1)-tf.nn.l2_normalize(z['z_pose'], axis=1))**2)
total_loss+=pose_loss*args.lambda_pose_loss/2
summaries['pose_loss'] = pose_loss
#pose_cosine= tf.reduce_mean(tf.multiply(tf.nn.l2_normalize(cond_pose_code, axis=1),tf.nn.l2_normalize(z['z_pose'], axis=1)))
#summaries['pose_check']=(1-pose_cosine*128.)-pose_loss*128./2
add_summary([summaries])
with tf.variable_scope('update_grads'):
global_step = tf.Variable(0, dtype=tf.int64, trainable=False, name='global_step')
optim = tf.train.AdamOptimizer(learning_rate=args.lr)
train_op = tf.contrib.training.create_train_op(total_loss, optim, global_step=global_step)
saver = tf.train.Saver(save_relative_paths=True, max_to_keep=50)
merged_summary = tf.summary.merge_all()
widgets = ['Training: ', progressbar.Percentage(),
' ', progressbar.Bar(),
' ', progressbar.Counter(), ' / {:d}'.format(args.num_ites),
' ', progressbar.ETA(), ' ']
bar = progressbar.ProgressBar(maxval=args.num_ites, widgets=widgets)
gpu_options = tf.GPUOptions(allow_growth=True, per_process_gpu_memory_fraction=0.9)
config = tf.ConfigProto(gpu_options=gpu_options)
with tf.Session(config=config) as sess:
sess.run(tf.global_variables_initializer())
sess.run(iterator.initializer)
if not (dataset.bg_img_init is None):
sess.run(dataset.bg_img_init.initializer)
sess.graph.finalize()
summary_writer = tf.summary.FileWriter(args.dir_checkpoints, sess.graph)
bar.start()
for i in range(0, args.num_ites):
bar.update(i)
if i % 1000 == 0:
results = sess.run([merged_summary,enc_x,dec_gt,dec_out,train_op, global_step])
gs = results[-1]
summary_writer.add_summary(results[0], gs)
cv2.imwrite(os.path.join(dir_train_figs, '{:05d}_x.png'.format(i)), (convert_batch_to_image_grid(results[1]) * 255.).astype(np.uint8))
if 'bgr' in args.decoder_branches:
cv2.imwrite(os.path.join(dir_train_figs, '{:05d}_y_gt.png'.format(i)), (convert_batch_to_image_grid(results[2]['bgr']) * 255.).astype(np.uint8))
cv2.imwrite(os.path.join(dir_train_figs, '{:05d}_y_out.png'.format(i)), (convert_batch_to_image_grid(results[3]['x_bgr']) * 255.).astype(np.uint8))
if 'depth' in args.decoder_branches:
cv2.imwrite(os.path.join(dir_train_figs, '{:05d}_yd_gt.png'.format(i)), (convert_batch_to_image_grid(results[2]['depth']) * 255.).astype(np.uint8))
cv2.imwrite(os.path.join(dir_train_figs, '{:05d}_yd_out.png'.format(i)), (convert_batch_to_image_grid(results[3]['x_depth']) * 255.).astype(np.uint8))
elif i%100==0:
results=sess.run([merged_summary,train_op,global_step])
gs = results[-1]
summary_writer.add_summary(results[0], gs)
else:
results = sess.run([train_op, global_step])
gs = results[-1]
if (i + 1) % 10000 == 0 or i == 0:
path_ckpt=os.path.join(args.dir_checkpoints,'chkpt')
saver.save(sess, path_ckpt, global_step=gs+1)
if gentle_stop[0]:
break