-
Notifications
You must be signed in to change notification settings - Fork 45
/
train.py
334 lines (266 loc) · 14.3 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
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
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
import argparse
import csv
import time
import torch.backends.cudnn as cudnn
import torch.nn.functional as F
import torch.optim
import torch.utils.data
from tensorboardX import SummaryWriter
import custom_transforms
from convert import *
from logger import AverageMeter
from loss_functions import compute_errors_train
from models import PSNet as PSNet
from sequence_folders import SequenceFolder
from utils import tensor2array, save_checkpoint, save_path_formatter, adjust_learning_rate
parser = argparse.ArgumentParser(description='DeepSFM depth subnet train script',
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument('data', metavar='DIR',
help='path to dataset')
parser.add_argument('-j', '--workers', default=4, type=int, metavar='N',
help='number of data loading workers')
parser.add_argument('--epochs', default=10, type=int, metavar='N',
help='number of total epochs to run')
parser.add_argument('--epoch-size', default=0, type=int, metavar='N',
help='manual epoch size (will match dataset size if not set)')
parser.add_argument('-b', '--batch-size', default=6, type=int,
metavar='N', help='mini-batch size')
parser.add_argument('--lr', '--learning-rate', default=4e-5, type=float,
metavar='LR', help='initial learning rate')
parser.add_argument('--geo', '--geo-cost', default=True, type=bool,
metavar='GC', help='whether add geometry cost')
parser.add_argument('--momentum', default=0.9, type=float, metavar='M',
help='momentum for sgd, alpha parameter for adam')
parser.add_argument('--beta', default=0.999, type=float, metavar='M',
help='beta parameters for adam')
parser.add_argument('--weight-decay', '--wd', default=0, type=float,
metavar='W', help='weight decay')
parser.add_argument('--print-freq', default=10, type=int,
metavar='N', help='print frequency')
parser.add_argument('-e', '--evaluate', dest='evaluate', action='store_true',
help='evaluate model on validation set')
parser.add_argument('--pretrained-dps', dest='pretrained_dps',
default='', metavar='PATH',
help='path to pre-trained dispnet model')
parser.add_argument('--seed', default=0, type=int, help='seed for random functions, and network initialization')
parser.add_argument('--log-summary', default='progress_log_summary.csv', metavar='PATH',
help='csv where to save per-epoch train and valid stats')
parser.add_argument('--log-full', default='progress_log_full.csv', metavar='PATH',
help='csv where to save per-gradient descent train stats')
parser.add_argument('--log-output', action='store_true', help='will log dispnet outputs and warped imgs at validation step')
parser.add_argument('--ttype', default='train.txt', type=str, help='Text file indicates input data')
parser.add_argument('--ttype2', default='val.txt', type=str, help='Text file indicates input data')
parser.add_argument('-f', '--training-output-freq', type=int, help='frequence for outputting dispnet outputs and warped imgs at training for all scales if 0 will not output',
metavar='N', default=100)
parser.add_argument('--nlabel', type=int ,default=64, help='number of label')
parser.add_argument('--mindepth', type=float ,default=0.5, help='minimum depth')
parser.add_argument('--pose_init', default='demon', help='path to init pose')
parser.add_argument('--depth_init', default='demon', help='path to init depth')
n_iter = 0
def main():
global n_iter
args = parser.parse_args()
save_path = save_path_formatter(args, parser)
args.save_path = 'checkpoints'/save_path
print('=> will save everything to {}'.format(args.save_path))
args.save_path.makedirs_p()
torch.manual_seed(args.seed)
training_writer = SummaryWriter(args.save_path)
output_writers = []
if args.log_output:
for i in range(3):
output_writers.append(SummaryWriter(args.save_path/'valid'/str(i)))
# Data loading code
normalize = custom_transforms.Normalize(mean=[0.5, 0.5, 0.5],
std=[0.5, 0.5, 0.5])
train_transform = custom_transforms.Compose([
custom_transforms.RandomScaleCrop(),
custom_transforms.ArrayToTensor(),
normalize
])
valid_transform = custom_transforms.Compose([custom_transforms.ArrayToTensor(), normalize])
print("=> fetching scenes in '{}'".format(args.data))
train_set = SequenceFolder(
args.data,
transform=train_transform,
seed=args.seed,
ttype=args.ttype,
add_geo=args.geo,
depth_source=args.depth_init,
pose_source='%s_poses.txt'%args.pose_init if args.pose_init else 'poses.txt',
scale=False
)
val_set = SequenceFolder(
args.data,
transform=valid_transform,
seed=args.seed,
ttype=args.ttype2,
add_geo=args.geo,
depth_source=args.depth_init,
pose_source='%s_poses.txt' % args.pose_init if args.pose_init else 'poses.txt',
scale=False
)
print('{} samples found in {} train scenes'.format(len(train_set), len(train_set.scenes)))
print('{} samples found in {} valid scenes'.format(len(val_set), len(val_set.scenes)))
train_loader = torch.utils.data.DataLoader(
train_set, batch_size=args.batch_size, shuffle=True,
num_workers=args.workers, pin_memory=True)
if args.epoch_size == 0:
args.epoch_size = len(train_loader)
# create model
print("=> creating model")
depth_net = PSNet(args.nlabel, args.mindepth, add_geo_cost=args.geo,
depth_augment=False).cuda()
if args.pretrained_dps:
# for param in depth_net.feature_extraction.parameters():
# param.requires_grad = False
print("=> using pre-trained weights for DPSNet")
model_dict = depth_net.state_dict()
weights = torch.load(args.pretrained_dps)['state_dict']
pretrained_dict = {k: v for k, v in weights.items() if k in model_dict}
model_dict.update(pretrained_dict)
depth_net.load_state_dict(model_dict)
else:
depth_net.init_weights()
cudnn.benchmark = True
depth_net = torch.nn.DataParallel(depth_net)
print('=> setting adam solver')
optimizer = torch.optim.Adam(filter(lambda p: p.requires_grad, depth_net.parameters()), args.lr,
betas=(args.momentum, args.beta),
weight_decay=args.weight_decay)
with open(args.save_path/args.log_summary, 'w') as csvfile:
writer = csv.writer(csvfile, delimiter='\t')
writer.writerow(['train_loss', 'validation_loss'])
with open(args.save_path/args.log_full, 'w') as csvfile:
writer = csv.writer(csvfile, delimiter='\t')
writer.writerow(['train_loss'])
for epoch in range(args.epochs):
adjust_learning_rate(args, optimizer, epoch)
# train for one epoch
train_loss = train(args, train_loader, depth_net, optimizer, args.epoch_size, training_writer)
save_checkpoint(
args.save_path, {
'epoch': epoch + 1,
'state_dict': depth_net.module.state_dict()
},
epoch)
with open(args.save_path/args.log_summary, 'a') as csvfile:
writer = csv.writer(csvfile, delimiter='\t')
writer.writerow([train_loss])
def train(args, train_loader, depth_net, optimizer, epoch_size, train_writer):
global n_iter
batch_time = AverageMeter()
data_time = AverageMeter()
losses = AverageMeter(precision=4)
# switch to train mode
depth_net.train()
end = time.time()
for i, (tgt_img, ref_imgs, ref_poses, intrinsics, intrinsics_inv, tgt_depth, ref_depths) in enumerate(train_loader):
# measure data loading time
data_time.update(time.time() - end)
tgt_img_var = Variable(tgt_img.cuda())
ref_imgs_var = [Variable(img.cuda()) for img in ref_imgs]
ref_poses_var = [Variable(pose.cuda()) for pose in ref_poses]
ref_depths_var = [Variable(dep.cuda()) for dep in ref_depths]
intrinsics_var = Variable(intrinsics.cuda())
intrinsics_inv_var = Variable(intrinsics_inv.cuda())
tgt_depth_var = Variable(tgt_depth.cuda()).cuda()
# compute output
pose = torch.cat(ref_poses_var,1)
# get mask
mask = (tgt_depth_var <= args.nlabel*args.mindepth) & (tgt_depth_var >= args.mindepth) & (tgt_depth_var == tgt_depth_var)
mask.detach_()
depths = depth_net(tgt_img_var, ref_imgs_var, pose, intrinsics_var, intrinsics_inv_var, ref_depths_var)
disps = [args.mindepth*args.nlabel/(depth) for depth in depths]
loss = 0.
for l, depth in enumerate(depths):
output = torch.squeeze(depth,1)
loss += F.smooth_l1_loss(output[mask], tgt_depth_var[mask], size_average=True) * pow(0.7, len(depths)-l-1)
if i > 0 and n_iter % args.print_freq == 0:
train_writer.add_scalar('total_loss', loss.data[0], n_iter)
if n_iter > 0 and n_iter % 5000 == 0:
save_checkpoint(
args.save_path, {
'epoch': n_iter + 1,
'state_dict': depth_net.module.state_dict()
},
n_iter)
if args.training_output_freq > 0 and n_iter % args.training_output_freq == 0:
train_writer.add_image('train Input', tensor2array(tgt_img[0]), n_iter)
depth_to_show = tgt_depth_var.data[0].cpu()
depth_to_show[depth_to_show > args.nlabel*args.mindepth] = args.nlabel*args.mindepth
disp_to_show = (args.nlabel*args.mindepth/depth_to_show)
disp_to_show[disp_to_show > args.nlabel] = 0
train_writer.add_image('train Dispnet GT Normalized',
tensor2array(disp_to_show, max_value=args.nlabel, colormap='bone'),
n_iter)
train_writer.add_image('train Depth GT Normalized',
tensor2array(depth_to_show, max_value=args.nlabel*args.mindepth*0.3),
n_iter)
for k,scaled_depth in enumerate(depths):
train_writer.add_image('train Dispnet Output Normalized {}'.format(k),
tensor2array(disps[k].data[0].cpu(), max_value=args.nlabel, colormap='bone'),
n_iter)
train_writer.add_image('train Depth Output Normalized {}'.format(k),
tensor2array(depths[k].data[0].cpu(), max_value=args.nlabel*args.mindepth*0.3),
n_iter)
# record loss and EPE
losses.update(loss.data[0], args.batch_size)
# compute gradient and do Adam step
optimizer.zero_grad()
loss.backward()
optimizer.step()
# measure elapsed time
batch_time.update(time.time() - end)
end = time.time()
with open(args.save_path/args.log_full, 'a') as csvfile:
writer = csv.writer(csvfile, delimiter='\t')
writer.writerow([loss.data[0]])
if i % args.print_freq == 0:
print('Train {}: Time {} Data {} Loss {}'.format(i, batch_time, data_time, losses))
if i >= epoch_size - 1:
break
n_iter += 1
return losses.avg[0]
def validate_with_gt(args, val_loader, depth_net, epoch, output_writers=[]):
batch_time = AverageMeter()
error_names = ['abs_rel', 'abs_diff', 'sq_rel', 'a1', 'a2', 'a3']
errors = AverageMeter(i=len(error_names))
log_outputs = len(output_writers) > 0
# switch to evaluate mode
depth_net.eval()
end = time.time()
for i, (tgt_img, ref_imgs, ref_poses, intrinsics, intrinsics_inv, tgt_depth, ref_depths) in enumerate(val_loader):
tgt_img_var = Variable(tgt_img.cuda(), volatile=True)
ref_imgs_var = [Variable(img.cuda(), volatile=True) for img in ref_imgs]
ref_poses_var = [Variable(pose.cuda(), volatile=True) for pose in ref_poses]
intrinsics_var = Variable(intrinsics.cuda(), volatile=True)
intrinsics_inv_var = Variable(intrinsics_inv.cuda(), volatile=True)
tgt_depth_var = Variable(tgt_depth.cuda(), volatile=True)
ref_depths_var = [Variable(dep.cuda()) for dep in ref_depths]
pose = torch.cat(ref_poses_var,1)
output_depth = depth_net(tgt_img_var, ref_imgs_var, pose, intrinsics_var, intrinsics_inv_var, ref_depths_var)
output_disp = args.nlabel*args.mindepth/(output_depth)
mask = (tgt_depth <= args.nlabel*args.mindepth) & (tgt_depth >= args.mindepth) & (tgt_depth == tgt_depth)
output = torch.squeeze(output_depth.data.cpu(),1)
if log_outputs and i % 100 == 0 and i/100 < len(output_writers):
index = int(i//100)
if epoch == 0:
output_writers[index].add_image('val Input', tensor2array(tgt_img[0]), 0)
depth_to_show = tgt_depth_var.data[0].cpu()
depth_to_show[depth_to_show > args.nlabel*args.mindepth] = args.nlabel*args.mindepth
disp_to_show = (args.nlabel*args.mindepth/depth_to_show)
disp_to_show[disp_to_show > args.nlabel] = 0
output_writers[index].add_image('val target Disparity Normalized', tensor2array(disp_to_show, max_value=args.nlabel, colormap='bone'), epoch)
output_writers[index].add_image('val target Depth Normalized', tensor2array(depth_to_show, max_value=args.nlabel*args.mindepth*0.3), epoch)
output_writers[index].add_image('val Dispnet Output Normalized', tensor2array(output_disp.data[0].cpu(), max_value=args.nlabel, colormap='bone'), epoch)
output_writers[index].add_image('val Depth Output', tensor2array(output_depth.data[0].cpu(), max_value=args.nlabel*args.mindepth*0.3), epoch)
errors.update(compute_errors_train(tgt_depth, output, mask))
# measure elapsed time
batch_time.update(time.time() - end)
end = time.time()
if i % args.print_freq == 0:
print('valid: Time {} Abs Error {:.4f} ({:.4f})'.format(batch_time, errors.val[0], errors.avg[0]))
return errors.avg, error_names
if __name__ == '__main__':
main()