-
Notifications
You must be signed in to change notification settings - Fork 5
/
Copy pathtrain.py
160 lines (128 loc) · 6.35 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
import os, numpy as np, random, torch, argparse, pdb
from utils.options import opts
from utils.diffusion_scheduler import DiffusionScheduler
from utils.get_lr_scheduler import get_lr_scheduler
from tqdm import tqdm
from collections import defaultdict
from utils.se_math import se3
from datasets.get_dataset import get_dataset
from utils.losses import compute_losses, compute_losses_diff
np.random.seed(opts.seed)
random.seed(opts.seed)
torch.manual_seed(opts.seed)
torch.backends.cudnn.benchmark=True
torch.backends.cudnn.enabled = False
def init_opts(opts):
opts.is_debug = False
opts.model_nm = "DiffusionReg"
opts.is_test = False
opts.is_normal = True
opts.n_cores = 6
opts.schedule_type = ["linear", "cosine"][1]
# dataset config
if opts.db_nm == "tudl":
opts.n_vids = 3
opts.n_epoches, opts.n_start_epoches, opts.batch_size = 20, 0, 32
opts.vid_infos = ["000001", "000002", "000003"]
else:
raise NotImplementedError
# diffusion configuration
opts.n_diff_steps = 200
opts.beta_1 = 1e-4
opts.beta_T = 0.05
opts.sigma_r = 0.05
opts.sigma_t = 0.03
diffusion_str = f"diffusion_{opts.n_diff_steps}_{opts.beta_1:.5f}_{opts.beta_T:.2f}_{opts.sigma_r:.2f}_{opts.sigma_t:.2f}"
opts.results_dir = f"./results/{opts.model_nm}-{opts.net_type}-{opts.db_nm}-{diffusion_str}-nvids{opts.n_vids}_{opts.schedule_type}"
os.makedirs(opts.results_dir, exist_ok=True)
print(opts.results_dir)
return opts
def main(opts):
opts = init_opts(opts)
## model setting
from modules.DCP.dcp import DCP
opts.vs = DiffusionScheduler(opts)
surrogate_model = DCP(opts)
if torch.cuda.device_count() > 1:
surrogate_model = torch.nn.DataParallel(surrogate_model, range(torch.cuda.device_count()))
surrogate_model = surrogate_model.to(opts.device)
train_loader, train_db = get_dataset(opts, db_nm=opts.db_nm, cls_nm=opts.vid_infos, partition="train",
batch_size=opts.batch_size, shuffle=True, drop_last=True, n_cores=opts.n_cores)
optimizer = torch.optim.Adam(surrogate_model.parameters(), lr=opts.lr, betas=(0.9, 0.999))
scheduler = get_lr_scheduler(opts, optimizer)
cal = lambda x: np.mean(x).item()
## training
for epoch_idx in range(opts.n_epoches):
# train
surrogate_model.train()
rcd = defaultdict(list)
for i, data in enumerate(tqdm(train_loader, 0)):
data = {k: v.to(opts.device) for k, v in data.items()}
# model prediction
X, X_normal = data["src_pcd"], data["src_pcd_normal"] # [B, N, 3]
Y, Y_normal = data["model_pcd"], data["model_pcd_normal"] # [B, M, 3]
Rs_gt, ts_gt = data['transform_gt'][:, :3, :3], data["transform_gt"][:, :3, 3] # [B, 3, 3], [B, 3]
B = Rs_gt.shape[0]
### SE(3) diffusion process
H_0 = torch.eye(4)[None].expand(B, -1, -1).to(opts.device)
H_0[:, :3, :3], H_0[:, :3, 3] = Rs_gt, ts_gt
H_T = torch.eye(4)[None].expand(B, -1, -1).to(opts.device)
taus = opts.vs.uniform_sample_t(B)
alpha_bars = opts.vs.alpha_bars[taus].to(opts.device)[:, None] # [B, 1]
H_t = se3.exp((1. - torch.sqrt(alpha_bars)) * se3.log(H_T @ torch.inverse(H_0))) @ H_0
### add noise
scale = torch.cat([torch.ones(3) * opts.sigma_r, torch.ones(3) * opts.sigma_t])[None].to(opts.device) # [1, 6]
noise = torch.sqrt(1. - alpha_bars) * scale * torch.randn(B, 6).to(opts.device) # [B, 6]
H_noise = se3.exp(noise)
H_t_noise = H_noise @ H_t # [B, 4, 4]
T_t_R = H_t_noise[:, :3, :3] # [B, 3, 3]
T_t_t = H_t_noise[:, :3, 3] # [B, 3]
X_t = (T_t_R @ X.transpose(2, 1) + T_t_t.unsqueeze(-1)).transpose(2, 1) # [B, N, 3]
X_normal_t = (T_t_R @ X_normal.transpose(2, 1)).transpose(2, 1) # [B, N, 3]
transform_gt = torch.eye(4)[None].expand(B, -1, -1).to(opts.device)
transform_gt[:, :3] = data['transform_gt']
input = {
"src_pcd": X_t,
"src_pcd_normal": X_normal_t,
"model_pcd": Y,
"model_pcd_normal": Y_normal,
}
Rs_pred_rot, ts_pred_rot = surrogate_model.forward(input)
pred_transforms = torch.cat([Rs_pred_rot, ts_pred_rot.unsqueeze(-1)], dim=2) # [B, 3, 4]
train_losses_diff = compute_losses_diff(opts, X, X_t, [pred_transforms], data['transform_gt'], loss_type="mae", reduction='mean')
loss = train_losses_diff['total']
# original loss
input = {
"src_pcd": X,
"src_pcd_normal": X_normal,
"model_pcd": Y,
"model_pcd_normal": Y_normal,
}
Rs_pred_rot1, ts_pred_rot1 = surrogate_model.forward(input)
pred_transforms1 = torch.cat([Rs_pred_rot1, ts_pred_rot1.unsqueeze(-1)], dim=2) # [B, 3, 4]
train_losses_origin = compute_losses(opts, X, [pred_transforms1], data['transform_gt'], loss_type="mae", reduction='mean')
loss += train_losses_origin["total"]
rcd["losses"].append(loss.item())
optimizer.zero_grad()
loss.backward()
optimizer.step()
print("=== Train. Epoch [%d], losses: %1.3f ===" % (epoch_idx, cal(rcd["losses"])))
if i > 0 and i % 200 == 0 and not opts.is_debug:
print("Save model. %s" % ('%s/model_epoch%d.pth' % (opts.results_dir, epoch_idx)))
torch.save(surrogate_model.state_dict(), '%s/model_epoch%d.pth' % (opts.results_dir, epoch_idx))
print(opts.results_dir)
# save model
if not opts.is_debug:
print("Save model. %s" % ('%s/model_epoch%d.pth' % (opts.results_dir, epoch_idx)))
torch.save(surrogate_model.state_dict(), '%s/model_epoch%d.pth' % (opts.results_dir, epoch_idx))
else:
print("Debug. Not save model.")
scheduler.step()
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--net_type', default="DiffusionDCP", type=str, choices=['DiffusionDCP'])
parser.add_argument('--db_nm', default="tudl", type=str, choices=["tudl"])
args = parser.parse_args()
opts.net_type = args.net_type
opts.db_nm = args.db_nm
main(opts)