/
train.py
125 lines (101 loc) · 6.09 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
from glob import glob
import hydra
import torch
from models.model import PPFEncoder, PointEncoder
from utils.dataset import ShapeNetDataset
import numpy as np
import logging
import torch.optim as optim
import torch.nn as nn
from tqdm import tqdm
import torch.nn.functional as F
from utils.util import AverageMeter, typename2shapenetid
import os
@hydra.main(config_path='./config', config_name='config')
def main(cfg):
logger = logging.getLogger(__name__)
# load name
name_path = hydra.utils.to_absolute_path('data/shapenet_names/{}.txt'.format(cfg.category))
if os.path.exists(name_path):
shapenames = open(name_path).read().splitlines()
else:
shapenet_id = typename2shapenetid[cfg.category]
shapenames = os.listdir(os.path.join(hydra.utils.to_absolute_path(cfg.shapenet_root), '{}'.format(shapenet_id)))
shapenames = [shapenet_id + '/' + name for name in shapenames]
ds = ShapeNetDataset(cfg, shapenames)
df = torch.utils.data.DataLoader(ds, pin_memory=True, batch_size=cfg.batch_size, shuffle=True, num_workers=10)
assert cfg.batch_size == 1
point_encoder = PointEncoder(k=cfg.knn, spfcs=[32, 64, 32, 32], num_layers=1, out_dim=32).cuda()
ppf_encoder = PPFEncoder(ppffcs=[84, 32, 32, 16], out_dim=2 * cfg.tr_num_bins + 2 * cfg.rot_num_bins + 2 + 3).cuda()
opt = optim.Adam([*point_encoder.parameters(), *ppf_encoder.parameters()], lr=cfg.opt.lr, weight_decay=cfg.opt.weight_decay)
kldiv = nn.KLDivLoss(reduction='batchmean')
bcelogits = nn.BCEWithLogitsLoss()
logger.info('Train')
best_loss = np.inf
for epoch in range(cfg.max_epoch):
n = 0
loss_meter = AverageMeter()
loss_tr_meter = AverageMeter()
loss_up_meter = AverageMeter()
loss_up_aux_meter = AverageMeter()
loss_right_meter = AverageMeter()
loss_right_aux_meter = AverageMeter()
loss_scale_meter = AverageMeter()
point_encoder.train()
ppf_encoder.train()
with tqdm(df) as t:
for pcs, pc_normals, targets_tr, targets_rot, targets_rot_aux, targets_scale, point_idxs in t:
pcs, pc_normals, targets_tr, targets_rot, targets_rot_aux, targets_scale, point_idxs = \
pcs.cuda(), pc_normals.cuda(), targets_tr.cuda(), targets_rot.cuda(), targets_rot_aux.cuda(), targets_scale.cuda(), point_idxs.cuda()
opt.zero_grad()
with torch.no_grad():
dist = torch.cdist(pcs, pcs)
sprin_feat = point_encoder(pcs, pc_normals, dist)
preds = ppf_encoder(pcs, pc_normals, sprin_feat, idxs=point_idxs[0])
preds_tr = preds[..., :2 * cfg.tr_num_bins].reshape(-1, 2, cfg.tr_num_bins)
preds_up = preds[..., 2 * cfg.tr_num_bins:2 * cfg.tr_num_bins + cfg.rot_num_bins]
preds_right = preds[..., 2 * cfg.tr_num_bins + cfg.rot_num_bins:2 * cfg.tr_num_bins + 2 * cfg.rot_num_bins]
preds_up_aux = preds[..., -5]
preds_right_aux = preds[..., -4]
preds_scale = preds[..., -3:]
loss_tr = kldiv(F.log_softmax(preds_tr[:, 0], dim=-1), targets_tr[0, :, 0]) + kldiv(F.log_softmax(preds_tr[:, 1], dim=-1), targets_tr[0, :, 1])
loss_up = kldiv(F.log_softmax(preds_up[0], dim=-1), targets_rot[0, :, 0])
loss_up_aux = bcelogits(preds_up_aux[0], targets_rot_aux[0, :, 0])
loss_scale = F.mse_loss(preds_scale, targets_scale[:, None])
loss = loss_up + loss_tr + loss_up_aux + loss_scale
if cfg.regress_right:
loss_right = kldiv(F.log_softmax(preds_right[0], dim=-1), targets_rot[0, :, 1])
loss_right_aux = bcelogits(preds_right_aux[0], targets_rot_aux[0, :, 1])
loss += loss_right + loss_right_aux
loss_right_meter.update(loss_right.item())
loss_right_aux_meter.update(loss_right_aux.item())
loss.backward(retain_graph=False)
# torch.nn.utils.clip_grad_norm_([*point_encoder.parameters(), *ppf_encoder.parameters()], 1.)
opt.step()
loss_meter.update(loss.item())
loss_tr_meter.update(loss_tr.item())
loss_up_meter.update(loss_up.item())
loss_up_aux_meter.update(loss_up_aux.item())
loss_scale_meter.update(loss_scale.item())
n += 1
if cfg.regress_right:
t.set_postfix(loss=loss_meter.avg, loss_tr=loss_tr_meter.avg,
loss_up=loss_up_meter.avg, loss_right=loss_right_meter.avg,
loss_up_aux=loss_up_aux_meter.avg, loss_right_aux=loss_right_aux_meter.avg,
loss_scale=loss_scale_meter.avg)
else:
t.set_postfix(loss=loss_meter.avg, loss_tr=loss_tr_meter.avg,
loss_up=loss_up_meter.avg,
oss_up_aux=loss_up_aux_meter.avg,
loss_scale=loss_scale_meter.avg)
if epoch % 20 == 0:
torch.save(point_encoder.state_dict(), f'point_encoder_epoch{epoch}.pth')
torch.save(ppf_encoder.state_dict(), f'ppf_encoder_epoch{epoch}.pth')
if loss_meter.avg < best_loss:
best_loss = loss_meter.avg
torch.save(point_encoder.state_dict(), f'point_encoder_epochbest.pth')
torch.save(ppf_encoder.state_dict(), f'ppf_encoder_epochbest.pth')
logger.info('loss: {:.4f}, loss_tr: {:.4f}, loss_up: {:.4f}, loss_right: {:.4f}, loss_up_aux: {:.4f}, loss_right_aux: {:.4f}, loss_scale: {:.4f}'
.format(loss_meter.avg, loss_tr_meter.avg, loss_up_meter.avg, loss_right_meter.avg, loss_up_aux_meter.avg, loss_right_aux_meter.avg, loss_scale_meter.avg))
if __name__ == '__main__':
main()