-
Notifications
You must be signed in to change notification settings - Fork 24
/
train_deform.py
218 lines (209 loc) · 11.1 KB
/
train_deform.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
211
212
213
214
215
216
217
218
import os
import time
import argparse
import random
import numpy as np
import torch
import torch.nn.functional as F
import tensorflow as tf
from lib.network import DeformNet
from lib.loss import Loss
from data.pose_dataset import PoseDataset
from lib.utils import setup_logger, compute_sRT_errors
from lib.align import estimateSimilarityTransform
parser = argparse.ArgumentParser()
parser.add_argument('--dataset', type=str, default='CAMERA', help='CAMERA or CAMERA+Real')
parser.add_argument('--data_dir', type=str, default='data', help='data directory')
parser.add_argument('--n_pts', type=int, default=1024, help='number of foreground points')
parser.add_argument('--n_cat', type=int, default=6, help='number of object categories')
parser.add_argument('--nv_prior', type=int, default=1024, help='number of vertices in shape priors')
parser.add_argument('--img_size', type=int, default=192, help='cropped image size')
parser.add_argument('--batch_size', type=int, default=32, help='batch size')
parser.add_argument('--num_workers', type=int, default=10, help='number of data loading workers')
parser.add_argument('--gpu', type=str, default='0', help='GPU to use')
parser.add_argument('--lr', type=float, default=0.0001, help='initial learning rate')
parser.add_argument('--start_epoch', type=int, default=1, help='which epoch to start')
parser.add_argument('--max_epoch', type=int, default=50, help='max number of epochs to train')
parser.add_argument('--resume_model', type=str, default='', help='resume from saved model')
parser.add_argument('--result_dir', type=str, default='results/camera', help='directory to save train results')
opt = parser.parse_args()
opt.decay_epoch = [0, 10, 20, 30, 40]
opt.decay_rate = [1.0, 0.6, 0.3, 0.1, 0.01]
opt.corr_wt = 1.0
opt.cd_wt = 5.0
opt.entropy_wt = 0.0001
opt.deform_wt = 0.01
def train_net():
# set result directory
if not os.path.exists(opt.result_dir):
os.makedirs(opt.result_dir)
tb_writer = tf.summary.FileWriter(opt.result_dir)
logger = setup_logger('train_log', os.path.join(opt.result_dir, 'log.txt'))
for key, value in vars(opt).items():
logger.info(key + ': ' + str(value))
os.environ['CUDA_VISIBLE_DEVICES'] = opt.gpu
# model & loss
estimator = DeformNet(opt.n_cat, opt.nv_prior)
estimator.cuda()
criterion = Loss(opt.corr_wt, opt.cd_wt, opt.entropy_wt, opt.deform_wt)
if opt.resume_model != '':
estimator.load_state_dict(torch.load(opt.resume_model))
# dataset
train_dataset = PoseDataset(opt.dataset, 'train', opt.data_dir, opt.n_pts, opt.img_size)
val_dataset = PoseDataset(opt.dataset, 'test', opt.data_dir, opt.n_pts, opt.img_size)
# start training
st_time = time.time()
train_steps = 1500
global_step = train_steps * (opt.start_epoch - 1)
n_decays = len(opt.decay_epoch)
assert len(opt.decay_rate) == n_decays
for i in range(n_decays):
if opt.start_epoch > opt.decay_epoch[i]:
decay_count = i
train_size = train_steps * opt.batch_size
indices = []
page_start = -train_size
for epoch in range(opt.start_epoch, opt.max_epoch + 1):
# train one epoch
logger.info('Time {0}'.format(time.strftime("%Hh %Mm %Ss", time.gmtime(time.time() - st_time)) + \
', ' + 'Epoch %02d' % epoch + ', ' + 'Training started'))
# create optimizer and adjust learning rate if needed
if decay_count < len(opt.decay_rate):
if epoch > opt.decay_epoch[decay_count]:
current_lr = opt.lr * opt.decay_rate[decay_count]
optimizer = torch.optim.Adam(estimator.parameters(), lr=current_lr)
decay_count += 1
# sample train subset
page_start += train_size
len_last = len(indices) - page_start
if len_last < train_size:
indices = indices[page_start:]
if opt.dataset == 'CAMERA+Real':
# CAMERA : Real = 3 : 1
camera_len = train_dataset.subset_len[0]
real_len = train_dataset.subset_len[1]
real_indices = list(range(camera_len, camera_len+real_len))
camera_indices = list(range(camera_len))
n_repeat = (train_size - len_last) // (4 * real_len) + 1
data_list = random.sample(camera_indices, 3*n_repeat*real_len) + real_indices*n_repeat
random.shuffle(data_list)
indices += data_list
else:
data_list = list(range(train_dataset.length))
for i in range((train_size - len_last) // train_dataset.length + 1):
random.shuffle(data_list)
indices += data_list
page_start = 0
train_idx = indices[page_start:(page_start+train_size)]
train_sampler = torch.utils.data.sampler.SubsetRandomSampler(train_idx)
train_dataloader = torch.utils.data.DataLoader(train_dataset, batch_size=opt.batch_size, sampler=train_sampler,
num_workers=opt.num_workers, pin_memory=True)
estimator.train()
for i, data in enumerate(train_dataloader, 1):
points, rgb, choose, cat_id, model, prior, sRT, nocs = data
points = points.cuda()
rgb = rgb.cuda()
choose = choose.cuda()
cat_id = cat_id.cuda()
model = model.cuda()
prior = prior.cuda()
sRT = sRT.cuda()
nocs = nocs.cuda()
assign_mat, deltas = estimator(points, rgb, choose, cat_id, prior)
loss, corr_loss, cd_loss, entropy_loss, deform_loss = criterion(assign_mat, deltas, prior, nocs, model)
optimizer.zero_grad()
loss.backward()
optimizer.step()
global_step += 1
# write results to tensorboard
summary = tf.Summary(value=[tf.Summary.Value(tag='learning_rate', simple_value=current_lr),
tf.Summary.Value(tag='train_loss', simple_value=loss),
tf.Summary.Value(tag='corr_loss', simple_value=corr_loss),
tf.Summary.Value(tag='cd_loss', simple_value=cd_loss),
tf.Summary.Value(tag='entropy_loss', simple_value=entropy_loss),
tf.Summary.Value(tag='deform_loss', simple_value=deform_loss)])
tb_writer.add_summary(summary, global_step)
if i % 10 == 0:
logger.info('Batch {0} Loss:{1:f}, corr_loss:{2:f}, cd_loss:{3:f}, entropy_loss:{4:f}, deform_loss:{5:f}'.format(
i, loss.item(), corr_loss.item(), cd_loss.item(), entropy_loss.item(), deform_loss.item()))
logger.info('>>>>>>>>----------Epoch {:02d} train finish---------<<<<<<<<'.format(epoch))
# evaluate one epoch
logger.info('Time {0}'.format(time.strftime("%Hh %Mm %Ss", time.gmtime(time.time() - st_time)) +
', ' + 'Epoch %02d' % epoch + ', ' + 'Testing started'))
val_loss = 0.0
total_count = np.zeros((opt.n_cat,), dtype=int)
strict_success = np.zeros((opt.n_cat,), dtype=int) # 5 degree and 5 cm
easy_success = np.zeros((opt.n_cat,), dtype=int) # 10 degree and 5 cm
iou_success = np.zeros((opt.n_cat,), dtype=int) # relative scale error < 0.1
# sample validation subset
val_size = 1500
val_idx = random.sample(list(range(val_dataset.length)), val_size)
val_sampler = torch.utils.data.sampler.SubsetRandomSampler(val_idx)
val_dataloader = torch.utils.data.DataLoader(val_dataset, batch_size=1, sampler=val_sampler,
num_workers=opt.num_workers, pin_memory=True)
estimator.eval()
for i, data in enumerate(val_dataloader, 1):
points, rgb, choose, cat_id, model, prior, sRT, nocs = data
points = points.cuda()
rgb = rgb.cuda()
choose = choose.cuda()
cat_id = cat_id.cuda()
model = model.cuda()
prior = prior.cuda()
sRT = sRT.cuda()
nocs = nocs.cuda()
assign_mat, deltas = estimator(points, rgb, choose, cat_id, prior)
loss, _, _, _, _ = criterion(assign_mat, deltas, prior, nocs, model)
# estimate pose and scale
inst_shape = prior + deltas
assign_mat = F.softmax(assign_mat, dim=2)
nocs_coords = torch.bmm(assign_mat, inst_shape)
nocs_coords = nocs_coords.detach().cpu().numpy()[0]
points = points.cpu().numpy()[0]
# use choose to remove repeated points
choose = choose.cpu().numpy()[0]
_, choose = np.unique(choose, return_index=True)
nocs_coords = nocs_coords[choose, :]
points = points[choose, :]
_, _, _, pred_sRT = estimateSimilarityTransform(nocs_coords, points)
# evaluate pose
cat_id = cat_id.item()
if pred_sRT is not None:
sRT = sRT.detach().cpu().numpy()[0]
R_error, T_error, IoU = compute_sRT_errors(pred_sRT, sRT)
if R_error < 5 and T_error < 0.05:
strict_success[cat_id] += 1
if R_error < 10 and T_error < 0.05:
easy_success[cat_id] += 1
if IoU < 0.1:
iou_success[cat_id] += 1
total_count[cat_id] += 1
val_loss += loss.item()
if i % 100 == 0:
logger.info('Batch {0} Loss:{1:f}'.format(i, loss.item()))
# compute accuracy
strict_acc = 100 * (strict_success / total_count)
easy_acc = 100 * (easy_success / total_count)
iou_acc = 100 * (iou_success / total_count)
for i in range(opt.n_cat):
logger.info('{} accuracies:'.format(val_dataset.cat_names[i]))
logger.info('5^o 5cm: {:4f}'.format(strict_acc[i]))
logger.info('10^o 5cm: {:4f}'.format(easy_acc[i]))
logger.info('IoU < 0.1: {:4f}'.format(iou_acc[i]))
strict_acc = np.mean(strict_acc)
easy_acc = np.mean(easy_acc)
iou_acc = np.mean(iou_acc)
val_loss = val_loss / val_size
summary = tf.Summary(value=[tf.Summary.Value(tag='val_loss', simple_value=val_loss),
tf.Summary.Value(tag='5^o5cm_acc', simple_value=strict_acc),
tf.Summary.Value(tag='10^o5cm_acc', simple_value=easy_acc),
tf.Summary.Value(tag='iou_acc', simple_value=iou_acc)])
tb_writer.add_summary(summary, global_step)
logger.info('Epoch {0:02d} test average loss: {1:06f}'.format(epoch, val_loss))
logger.info('Overall accuracies:')
logger.info('5^o 5cm: {:4f} 10^o 5cm: {:4f} IoU: {:4f}'.format(strict_acc, easy_acc, iou_acc))
logger.info('>>>>>>>>----------Epoch {:02d} test finish---------<<<<<<<<'.format(epoch))
# save model after each epoch
torch.save(estimator.state_dict(), '{0}/model_{1:02d}.pth'.format(opt.result_dir, epoch))
if __name__ == '__main__':
train_net()