-
Notifications
You must be signed in to change notification settings - Fork 16
/
Copy pathrun_mono_dpc.py
204 lines (175 loc) · 12.1 KB
/
run_mono_dpc.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
import torch
import sys
sys.path.insert(0,'..')
from data.kitti_loader import KittiLoaderPytorch
from train_mono import Train as Train_Mono
from validate import Validate, test_depth_and_reconstruction, test_trajectory
import models.stn as stn
import models.mono_model_joint as mono_model_joint
from utils.learning_helpers import *
from utils.custom_transforms import *
import losses
from vis import *
import numpy as np
import datetime
import time
from tensorboardX import SummaryWriter
import argparse
import scipy.io as sio
import torch.backends.cudnn as cudnn
import os
device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
parser = argparse.ArgumentParser(description='training arguments.')
parser.add_argument('--data_dir', type=str, default='/media/datasets/KITTI-dpc')
parser.add_argument('--date', type=str, default='0000000')
parser.add_argument('--train_seq', nargs='+', type=str, default=['all'])
parser.add_argument('--val_seq', nargs='+',type=str, default=['00'])
parser.add_argument('--test_seq', nargs='+', type=str, default=['05'])
parser.add_argument('--augment_motion', action='store_true', default=False) #skip samples to simulate faster motion
parser.add_argument('--augment_backwards', action='store_true', default=False) #reverse sequences to simulate backwards motion
parser.add_argument('--img_per_sample', type=int, default=2) ## 1 target image, and 'img_per_sample -1 source images.
parser.add_argument('--correction_rate', type=int, default=1) ## if not one, only perform corrections every 'correction_rate' frames (samples become {1,3},{3,5},{5,7} when rate is 2)
parser.add_argument('--skip', type=int, default=1) ## if not one, skip every 'skip' samples that are generated ({1,2}, {2,3}, {3,4} becomes {1,2}, {3,4})
parser.add_argument('--exploss', action='store_true', default=True) ## For using the explainability loss
parser.add_argument('--exp_weight', type=float, default=0.18)
parser.add_argument('--use_flow', action='store_true',default=True) ## For using optical flow as a network input
parser.add_argument('--normalize_img', action='store_true',default=True) ## Normalize images for network
parser.add_argument("--estimator_type", type=str, default='mono') ## mono or stereo (the type of VO estimator being corrected)
parser.add_argument('--minibatch', type=int, default=32) ## minibatch size
parser.add_argument('--wd', type=float, default=0) ## weight decay
parser.add_argument('--lr', type=float, default=9e-4) ## Learning rate
parser.add_argument('--num_epochs', type=int, default=20) ## Number of training epochs
parser.add_argument('--lr_decay_epoch', type=float, default=4) ## decay learning rate every X epochs
parser.add_argument('--dropout_prob', type=float, default=0.3) ## dropout for posenet
parser.add_argument('--save_results', action='store_true', default=False)
args = parser.parse_args()
config={
'num_frames': None, ## if used, only take num_frames from each sequence for training
}
for k in args.__dict__:
config[k] = args.__dict__[k]
dsets = {x: KittiLoaderPytorch(args.data_dir, config, [args.train_seq, args.val_seq, args.test_seq], mode=x, transform_img=get_data_transforms(config)[x], num_frames = config['num_frames'], \
augment=config['augment_motion'], skip=config['skip'], augment_backwards=config['augment_backwards']) for x in ['train', 'val']}
dset_loaders = {x: torch.utils.data.DataLoader(dsets[x], batch_size=config['minibatch'], shuffle=True, num_workers=4) for x in ['train', 'val']}
val_dset = KittiLoaderPytorch(args.data_dir, config, [args.train_seq, args.val_seq, args.test_seq], mode='val', transform_img=get_data_transforms(config)['val'])
val_dset_loaders = torch.utils.data.DataLoader(val_dset, batch_size=config['minibatch'], shuffle=False, num_workers=4)
test_dset = KittiLoaderPytorch(args.data_dir, config, [args.train_seq, args.val_seq, args.test_seq], mode='test', transform_img=get_data_transforms(config)['val'])
test_dset_loaders = torch.utils.data.DataLoader(test_dset, batch_size=config['minibatch'], shuffle=False, num_workers=4)
eval_dsets = {'val': val_dset_loaders, 'test':test_dset_loaders}
def main():
results = {}
results['estimator'] = config['estimator_type']
start = time.time()
now= datetime.datetime.now()
ts = '{}-{}-{}-{}-{}'.format(now.year, now.month, now.day, now.hour, now.minute)
criterion = losses.photometric_reconstruction_loss()
exp_loss = losses.explainability_loss()
Reconstructor = stn.Reconstructor().to(device)
loss = losses.Compute_Loss(Reconstructor, criterion, exp_loss, exp_weight = config['exp_weight'])
model = mono_model_joint.joint_model(num_img_channels=(6 + 2*config['use_flow']), output_exp=args.exploss, dropout_prob=config['dropout_prob']).to(device)
params = list(model.parameters())
optimizer = torch.optim.Adam(params, lr=config['lr'], weight_decay = config['wd'])
cudnn.benchmark = True
est_traj_stacked ={}
corr_pose_change_vecs_stacked = {}
corr_stacked = {}
losses_stacked = {}
best_val_loss, best_rot_seg_err, best_trans_err, most_loop_closure = {}, {}, {}, {}
best_rot_acc_epoch, best_trans_acc_epoch, best_loss_epoch, most_loop_closure_epoch = {}, {}, {}, {}
for key, dset in eval_dsets.items():
est_traj_stacked[key] = np.empty((0,eval_dsets[key].dataset.raw_gt_trials[0].shape[0], 4, 4))
losses_stacked[key] = np.empty((0, eval_dsets[key].dataset.raw_gt_trials[0].shape[0]-1))
corr_pose_change_vecs_stacked[key] = np.empty((0, eval_dsets[key].dataset.raw_gt_trials[0].shape[0]-1,6))
corr_stacked[key] = np.copy(corr_pose_change_vecs_stacked[key])
best_val_loss[key], best_rot_seg_err[key], best_trans_err[key], most_loop_closure[key] = 1e5, 1e5, 1e5, 0
for epoch in range(0,config['num_epochs']):
optimizer = exp_lr_scheduler(model, optimizer, epoch, lr_decay_epoch=config['lr_decay_epoch']) ## reduce learning rate as training progresses
train_loss = Train_Mono(device,model, Reconstructor, dset_loaders['train'], loss, optimizer, epoch)
val_loss = Validate(device, model, Reconstructor, dset_loaders['val'], loss)
#
if epoch == 0:
writer = SummaryWriter(comment="-val_seq-{}-test_seq-{}".format(args.val_seq[0], args.test_seq[0]))
writer.add_scalars('',{'train': train_loss, 'val': val_loss}, epoch+1)
for key, dset in eval_dsets.items():
print("{} Set, Epoch {}".format(key, epoch))
###plot images, depth map, explainability mask
img_array, disparity, exp_mask = test_depth_and_reconstruction(device, model, Reconstructor, dset)
img_array = plot_img_array(img_array)
depth = plot_disp(disparity[-1])
writer.add_image(key+'/imgs',img_array,epoch+1)
writer.add_image(key+'/depth', depth, epoch+1)
if args.exploss:
exp_mask = plot_img_array(exp_mask)
writer.add_image(key+'/exp_mask', exp_mask, epoch+1)
###plot trajectories
corr, gt_corr, corr_pose_change_vec, odom_pose_change_vec, gt_pose_change_vec, corr_traj, corr_traj_rot, est_traj, gt_traj, \
rot_seg_err, trans_err, cum_dist = test_trajectory(device, model, Reconstructor, dset, epoch)
corrections = plot_6_by_1(corr, title = 'Corrections')
correction_errors = plot_6_by_1(np.abs(corr - gt_corr), title='6x1 Errors')
est_traj_img = plot_multi_traj(est_traj, 'Odom.', gt_traj, 'GT', key+' Set')
corr_traj_img = plot_multi_traj(corr_traj, 'corr.', gt_traj, 'GT', key+' Set')
corr_traj_rot_img = plot_multi_traj(corr_traj_rot, 'corr.', gt_traj, 'GT', key+' Set')
writer.add_image(key+'/corr_traj', corr_traj_img, epoch+1)
writer.add_image(key+'/corr_traj (Rot. only)', corr_traj_rot_img, epoch+1)
writer.add_image(key+'/est_traj', est_traj_img, epoch+1)
writer.add_image(key+'/correction_errors', correction_errors, epoch+1)
corr_stacked[key] = np.vstack((corr_stacked[key], corr.reshape((1,-1,6))))
est_traj_stacked[key] = np.vstack((est_traj_stacked[key], corr_traj_rot.reshape((1,-1,4,4))))
corr_pose_change_vecs_stacked[key] = np.vstack((corr_pose_change_vecs_stacked[key], corr_pose_change_vec.reshape((1,-1,6))))
results[key] = {'val_seq': args.val_seq,
'test_seq': args.test_seq,
'epochs': epoch+1,
'est_traj': est_traj_stacked[key],
'corrections': corr_stacked[key],
'est_traj_reconstruction_loss': losses_stacked[key],
'corr_pose_vecs': corr_pose_change_vecs_stacked[key],
'odom_pose_vecs': odom_pose_change_vec,
'gt_traj': gt_traj,
}
if args.save_results:
os.makedirs('results/{}'.format(config['date']), exist_ok=True)
if config['estimator_type'] == 'mono':
_, num_loop_closure, _ = find_loop_closures(corr_traj, cum_dist)
if config['estimator_type'] == 'stereo':
_, num_loop_closure, _ = find_loop_closures(corr_traj_rot, cum_dist)
print("{} Loop Closures detected".format(num_loop_closure))
##Save the best models
if val_loss < best_val_loss[key]:
best_val_loss[key] = val_loss
best_loss_epoch[key] = epoch
state_dict_loss = model.state_dict()
print("Lowest validation loss (saving model)")
if key == 'val':
torch.save(state_dict_loss, 'results/{}/{}-best-loss-val_seq-{}-test_seq-{}.pth'.format(config['date'], ts, args.val_seq[0], args.test_seq[0]))
if rot_seg_err < best_rot_seg_err[key]:
best_rot_seg_err[key] = rot_seg_err
best_rot_acc_epoch[key] = epoch
state_dict_acc = model.state_dict()
print("Lowest error (saving model)")
if key == 'val':
torch.save(state_dict_acc, 'results/{}/{}-best_rot_acc-val_seq-{}-test_seq-{}.pth'.format(config['date'], ts, args.val_seq[0], args.test_seq[0]))
if trans_err < best_trans_err[key]:
best_trans_err[key] = trans_err
best_trans_acc_epoch[key] = epoch
state_dict_trans = model.state_dict()
print("Lowest position error (saving model)")
if key == 'val':
torch.save(state_dict_trans, 'results/{}/{}-best_trans_acc-val_seq-{}-test_seq-{}.pth'.format(config['date'], ts, args.val_seq[0], args.test_seq[0]))
if num_loop_closure >= most_loop_closure[key]:
most_loop_closure[key] = num_loop_closure
most_loop_closure_epoch[key] = epoch
state_dict_loop_closure = model.state_dict()
print("Most Loop Closures detected ({})".format(most_loop_closure[key]))
if key == 'val':
torch.save(state_dict_loop_closure, 'results/{}/{}-most_loop_closures-val_seq-{}-test_seq-{}.pth'.format(config['date'], ts, args.val_seq[0], args.test_seq[0]))
results[key]['best_rot_acc_epoch'] = best_rot_acc_epoch[key]
results[key]['best_trans_acc_epoch'] = best_trans_acc_epoch[key]
results[key]['best_loss_epoch'] = best_loss_epoch[key]
results[key]['best_loop_closure_epoch'] = most_loop_closure_epoch[key]
sio.savemat('results/{}/{}-results-val_seq-{}-test_seq-{}.mat'.format(config['date'], ts, args.val_seq[0], args.test_seq[0]),results)
f = open("results/{}/{}-config.txt".format(config['date'],ts),"w")
f.write( str(config) )
f.close()
duration = timeSince(start)
print("Training complete (duration: {})".format(duration))
main()