-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
232 lines (196 loc) · 7.76 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
import os
os.environ["CUDA_VISIBLE_DEVICES"] = GPU_NUM = "7"
import json
import time
import torch
import argparse
import traceback
from tqdm import tqdm
from functools import partial
from collections import OrderedDict
from torch.utils.data import DataLoader
from utils.seed_everything import seed_everything
from evaluate import evaluate as validate
from ramp.data_readers.TartanEvent import TartanEvent
from ramp.lietorch import SE3
from ramp.net import VONet
from ramp.utils import kabsch_umeyama
seed_everything(seed=1234)
try:
import wandb
log = True
except:
print("WARNING: wandb is not installed, cannot log results, please install wandb to log results")
log = False
def compute_losses(traj, so, train_config, patch_size):
loss = 0.0
for i, (v, x, y, P1, P2) in enumerate(traj):
e = (x - y).norm(dim=-1)
e = e.reshape(-1, patch_size**2)[(v > 0.5).reshape(-1)].min(dim=-1).values
N = P1.shape[1]
ii, jj = torch.meshgrid(torch.arange(N), torch.arange(N))
ii = ii.reshape(-1).cuda()
jj = jj.reshape(-1).cuda()
k = ii != jj
ii = ii[k]
jj = jj[k]
P1 = P1.inv()
P2 = P2.inv()
t1 = P1.matrix()[..., :3, 3]
t2 = P2.matrix()[..., :3, 3]
s = kabsch_umeyama(t2[0], t1[0]).detach().clamp(max=10.0)
P1 = P1.scale(s.view(1, 1))
dP = P1[:, ii].inv() * P1[:, jj] # relative predicted pose
dG = P2[:, ii].inv() * P2[:, jj] # ground truth
e1 = (dP * dG.inv()).log()
tr = e1[..., 0:3].norm(dim=-1)
ro = e1[..., 3:6].norm(dim=-1)
loss += train_config["flow_weight"] * e.mean()
if not so and i >= 2:
loss += train_config["pose_weight"] * (tr.mean() + ro.mean())
return loss, e, ro, tr
def train(args):
"""main training loop"""
config = json.load(open(args.config_path))
train_cfg = config["data_loader"]["train"]["args"]
log_results = args.log_results and log
# Initialize network, optimizer and scheduler
net = VONet(cfg=train_cfg)
patch_size = net.P
net.train()
net.to("cuda")
optimizer = torch.optim.AdamW(
net.parameters(), lr=train_cfg["lr"], weight_decay=train_cfg["weight_decay"]
)
scheduler = torch.optim.lr_scheduler.OneCycleLR(
optimizer,
max_lr=train_cfg["lr"],
total_steps=train_cfg["steps"],
pct_start=train_cfg["pct_start"],
cycle_momentum=False,
anneal_strategy="linear",
)
# Import the checkpoint data if provided
batch_idx, epoch, step = 1, 0, 0
resume_train = False
if args.ckpt is not None:
resume_train = True
checkpoint = torch.load(args.ckpt)
batch_idx = checkpoint["batch_idx"]
step = checkpoint["total_idx"]
model_state_dict = checkpoint["model_state_dict"]
epoch = checkpoint["epoch"] if checkpoint.get("epoch") else 0
optimizer.load_state_dict(checkpoint["optimizer_state_dict"])
scheduler.load_state_dict(checkpoint["scheduler_state_dict"])
new_state_dict = OrderedDict()
for k, v in model_state_dict.items():
new_state_dict[k.replace("module.", "")] = v
net.load_state_dict(new_state_dict, strict=False)
# instantiate the dataloader
Dataloader = partial(
DataLoader,
batch_size=train_cfg["batch_size"],
shuffle=train_cfg["shuffle"],
num_workers=args.workers,
prefetch_factor=1,
)
# Name your experiment through name argument or with config file
run_name = args.name if args.name is not None else config["experiment_name"]
if log_results:
wandb.init(
project="RAMP-VO",
name=run_name,
config=config,
id=run_name,
resume=resume_train,
)
wandb.watch(net, log="all")
for curr_epoch in range(10):
if curr_epoch < epoch:
continue
db = TartanEvent(
path=args.data_path, config=config, step=step, workers_n=args.workers
)
pbar = tqdm(
Dataloader(dataset=db), mininterval=10, ncols=50
)
for batch_idx, data_blob in enumerate(pbar):
step += 1
# skip the batches before the saved batch_idx
if data_blob == 0:
continue
events, images, poses, disps, K, supervision_mask = [x.to("cuda") for x in data_blob]
optimizer.zero_grad()
fix_repr_pose = step < 1000 and args.ckpt is None
traj = net(
STEPS=18,
disps=disps,
intrinsics=K,
poses=SE3(poses).inv(),
input_=(events, images, supervision_mask),
structure_only=fix_repr_pose,
)
loss, e, ro, tr = compute_losses(
traj=traj,
so=fix_repr_pose,
train_config=train_cfg,
patch_size=patch_size,
)
loss.backward()
torch.nn.utils.clip_grad_norm_(net.parameters(), train_cfg["clip"])
optimizer.step()
scheduler.step()
metrics = {
"loss": loss.item(),
"px1": (e < 0.25).float().mean().item(),
"ro": ro.float().mean().item(),
"tr": tr.float().mean().item(),
}
if log_results:
wandb.log(data=metrics, step=step)
if step % train_cfg["steps_to_save_ckpt"] == 0:
torch.cuda.empty_cache()
directory = os.path.join("checkpoints", run_name)
if not os.path.isdir(directory):
os.mkdir(directory)
PATH = directory + "/%s_%06d.pth" % (run_name, step)
torch.save(
{
"batch_idx": batch_idx,
"total_idx": step,
"epoch": curr_epoch,
"model_state_dict": net.state_dict(),
"optimizer_state_dict": optimizer.state_dict(),
"scheduler_state_dict": scheduler.state_dict(),
},
PATH,
)
steps_to_do_validation = 0
if train_cfg.get("steps_to_do_validation") is not None:
steps_to_do_validation = train_cfg["steps_to_do_validation"]
if step > steps_to_do_validation:
validation_results = None
try:
valid_start = time.time()
validation_results = validate(
dataset_path=args.data_path, eval_cfg=config, net=net
)
for k in validation_results:
print(k, validation_results[k])
print("\n Validation time: ", time.time() - valid_start, "s")
except Exception:
traceback.print_exc()
print("\n VALIDATION HASN'T WORKED")
if log_results and validation_results is not None:
wandb.log(data=validation_results, step=step)
torch.cuda.empty_cache()
net.train()
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("--data_path", type=str, help="Dataset path")
parser.add_argument("--name", type=str, default=None, help="name your experiment")
parser.add_argument("--ckpt", type=str, help="checkpoint to restore")
parser.add_argument("--config_path", type=str, help="config file path")
parser.add_argument("--log_results", action="store_true", default=False)
parser.add_argument("--workers", type=int, default=0)
train(args=parser.parse_args())