-
Notifications
You must be signed in to change notification settings - Fork 0
/
train_img2pc.py
143 lines (123 loc) · 5.23 KB
/
train_img2pc.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
from __future__ import print_function
import sys
import argparse
import random
import torch.backends.cudnn as cudnn
import torch.optim as optim
# import time,datetime
from tensorboardX import SummaryWriter
from dataset_img2pc import *
from model import *
parser = argparse.ArgumentParser()
parser.add_argument('--dataRoot', type=str, default='/data/dudong/ShapeNetCore.v1', help='data root path')
parser.add_argument('--batchSize', type=int, default=32, help='input batch size')
parser.add_argument('--workers', type=int, help='number of data loading workers', default=8)
parser.add_argument('--model', type=str, default='checkpoint', help='model path')
parser.add_argument('--log', type=str, default='log_img2pc', help='log path')
parser.add_argument('--nepoch', type=int, default=35, help='number of epochs to train for')
parser.add_argument('--lr', type=float, default=1e-4, help='Initial learning rate.')
parser.add_argument('--weight_decay', type=float, default=1e-5, help='Weight decay (L2 loss on parameters).')
parser.add_argument('--cat', type=str, default='03001627')
parser.add_argument('--cuda', type=str, default='0')
opt = parser.parse_args()
# cat_set: '03001627', '02691156', '02958343', '04090263', '04256520', '04379243'
# '04090263'(rifle) need lower learning rate(5e-5), --weight_decay(1e-6) and less training epochs(10)
os.environ["CUDA_VISIBLE_DEVICES"] = opt.cuda
opt.manualSeed = random.randint(1, 10000) # fix seed
random.seed(opt.manualSeed)
torch.manual_seed(opt.manualSeed)
cudnn.benchmark = True
# Creat train/val dataloader
dataset = ShapeNet(img_root=os.path.join(opt.dataRoot, 'renderingimg'),
pc_root=os.path.join(opt.dataRoot, 'pc'),
filelist_root=os.path.join(opt.dataRoot, 'train_val_test_list'),
cat=opt.cat, mode='train')
dataloader = torch.utils.data.DataLoader(dataset, batch_size=opt.batchSize,
shuffle=True, num_workers=int(opt.workers))
# dataset_val = ShapeNet(img_root=os.path.join(opt.dataRoot, 'renderingimg'),
# pc_root=os.path.join(opt.dataRoot, 'pc'),
# filelist_root=os.path.join(opt.dataRoot, 'train_val_test_list'),
# cat=opt.cat, mode='val')
# dataloader_val = torch.utils.data.DataLoader(dataset_val, batch_size=opt.batchSize,
# shuffle=False, num_workers=int(opt.workers))
len_dataset = len(dataset)
# len_dataset_val = len(dataset_val)
print('training set num', len_dataset)
# print('validation set num', len_dataset_val)
# create path
model_path = os.path.join(opt.model, opt.cat)
log_path = os.path.join(opt.log, opt.cat)
if not os.path.exists(model_path):
os.makedirs(model_path)
if not os.path.exists(log_path):
os.makedirs(log_path)
logger = SummaryWriter(log_path)
# Create network
network_img2pc = PointSetGenerationNetwork(z_dim=256, n_pc=1024)
network_img2pc.cuda()
# Create Loss Module
sys.path.append('./utils/chamferdistance/')
import dist_chamfer as ext
distChamfer = ext.chamferDist()
# Create optimizer
optimizer = optim.Adam(network_img2pc.parameters(), lr=opt.lr, weight_decay=opt.weight_decay)
it_step = 0
min_loss = 1e8
best_epoch = 0
for epoch in range(1, opt.nepoch + 1):
# TRAIN MODE
network_img2pc.train()
for i, data in enumerate(dataloader, 0):
it_step += 1
optimizer.zero_grad()
img, pc, name, view_id = data
img = img.cuda()
pc = pc.cuda()
pc_pre = network_img2pc(img)*10./9.
dist1, dist2 = distChamfer(pc, pc_pre) # here we sample from the mesh of occnet, where mesh vertex coodinates are belong to 10/9*[-0.5, 0.5]
loss = (torch.mean(dist1) + torch.mean(dist2))*10.
if it_step % 10 == 0:
logger.add_scalar('train/loss', loss, it_step)
print('[%d: %d/%d] train loss: %f' % (epoch, i, len_dataset / opt.batchSize, loss.item()))
loss.backward()
optimizer.step()
# #VALIDATION
# network_img2pc.eval()
#
# loss_val = 0
# loss_n = 0
# with torch.no_grad():
# for i, data in enumerate(dataloader_val, 0):
# img, pc, name, view_id = data
# img = img.cuda()
# pc = pc.cuda()
#
# pc_pre = network_img2pc(img)*10/.9.
# dist1, dist2 = distChamfer(pc, pc_pre)
# loss = (torch.mean(dist1) + torch.mean(dist2))*10.
#
# loss_val += loss
# loss_n += 1
#
# if i % 10 == 0:
# print('[%d: %d/%d] val loss: %f' % (epoch, i, len_dataset_val / opt.batchSize, loss.item()))
#
# loss_val /= loss_n
# logger.add_scalar('val/loss', loss_val, it_step)
#
# if loss_val < min_loss + 1e-4:
# min_loss = loss_val
# torch.save(network_img2pc.state_dict(), os.path.join(model_path, 'img2pc.pt'))
#
# best_epoch = epoch
# print('Best epoch is:', best_epoch)
#
# # else:
# # break
torch.save(network_img2pc.state_dict(), os.path.join(model_path, 'img2pc.pt'))
print('save model')
# if epoch % 10 == 0:
# torch.save(network_img2pc.state_dict(), '%s/img2pc_%s.pt' % (model_path, str(epoch)))
# print('save model')
print('Training done!')
# print('Best epoch is:', best_epoch)