-
Notifications
You must be signed in to change notification settings - Fork 0
/
train_patch_l2h.py
149 lines (123 loc) · 6.07 KB
/
train_patch_l2h.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
from __future__ import print_function
import argparse
import random
import torch.backends.cudnn as cudnn
import torch.optim as optim
# import time,datetime
from tensorboardX import SummaryWriter
from dataset_patch_l2h 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=1, help='input batch size')
parser.add_argument('--patchNum', type=int, default=48, help='input patch 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('--vipnet', type=str, default='vipnet_finetune.pt', help='model path')
parser.add_argument('--log', type=str, default='log_patch_l2h', help='log path')
parser.add_argument('--nepoch', type=int, default=15, help='number of epochs to train for')
parser.add_argument('--lr', type=float, default=5e-5, help='Initial learning rate.')
parser.add_argument('--weight_decay', type=float, default=1e-6, 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'
os.environ["CUDA_VISIBLE_DEVICES"] = opt.cuda
cudnn.benchmark = True
opt.manualSeed = random.randint(1, 10000) # fix seed
random.seed(opt.manualSeed)
torch.manual_seed(opt.manualSeed)
# 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)
# Creat train/val dataloader
dataset = ShapeNet(img_root=os.path.join(opt.dataRoot, 'renderingimg'),
vox256_root=os.path.join(opt.dataRoot, 'vox256'),
filelist_root=os.path.join(opt.dataRoot, 'train_val_test_list'),
cat=opt.cat, mode='train', patch_num=opt.patchNum, view_pick=False)
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'),
vox256_root=os.path.join(opt.dataRoot, 'vox256'),
filelist_root=os.path.join(opt.dataRoot, 'train_val_test_list'),
cat=opt.cat, mode='val', patch_num=opt.patchNum, view_pick=True)
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 network
vipnet = SVRNetworkOccGuide()
vipnet.cuda()
vipnet.load_state_dict(torch.load(os.path.join(model_path, opt.vipnet)))
vipnet.eval()
print("Previous vipnet weights loaded...")
vox_dim = 1
feat_high_dim = 16
feat_low_dim = 32
input_dim = vox_dim + feat_high_dim + feat_low_dim
network_patch_l2h = VoxelSuperResNetwork_16_64(input_dim=input_dim, gf_dim=64)
network_patch_l2h.cuda()
print('Defined patch_l2h network.')
# Create Loss Module
criterion = nn.BCEWithLogitsLoss()
# Create optimizer, just optimize the parameters of decoder
optimizer = optim.Adam(network_patch_l2h.parameters(), lr=opt.lr, weight_decay=opt.weight_decay)
it_step = 0
for epoch in range(1, opt.nepoch + 1):
# TRAIN MODE
network_patch_l2h.train()
for it, data in enumerate(dataloader, 0):
it_step += 1
optimizer.zero_grad()
img, vox64, idx, name, view_id = data
idx = idx[0]
img = img.cuda()
vox64 = vox64.cuda()
vox64 = vox64.contiguous().view(opt.patchNum, 1, 64, 64, 64)
vox_ref_cat_feat_high, feat_low = vipnet.predict_highres_input(img)
vox_in = torch.zeros(opt.patchNum, vox_dim+feat_high_dim, 16, 16, 16).type(torch.FloatTensor).cuda()
for i in range(0, opt.patchNum): # here batch_size = 1
vox_in[i] = vox_ref_cat_feat_high[0, :, dataset.begin_idx64[idx[i]][0]: dataset.begin_idx64[idx[i]][0] + 16,
dataset.begin_idx64[idx[i]][1]: dataset.begin_idx64[idx[i]][1] + 16,
dataset.begin_idx64[idx[i]][2]: dataset.begin_idx64[idx[i]][2] + 16]
del img, vox_ref_cat_feat_high
# torch.cuda.empty_cache()
feat_low = feat_low.repeat(opt.patchNum, 1, 1, 1, 1)
vox_in_cat = torch.cat([vox_in, feat_low], 1)
del feat_low, vox_in
# torch.cuda.empty_cache()
vox_pre, vox_pre_sigmoid = network_patch_l2h(vox_in_cat)
loss = criterion(vox_pre, vox64)
loss.backward()
optimizer.step()
if it % 10 == 0:
logger.add_scalar('train/loss', loss, it_step)
print('[%d: %d/%d] train loss: %f' % (epoch, it, len_dataset / opt.batchSize, loss.item()))
# del img, vox64, vox_feat, vox_ref, vox_in, vox_in_cat, vox_pre, vox_pre_sigmoid
# torch.cuda.empty_cache()
# #VALIDATION
# network_patch_l2h_img.eval()
# with torch.no_grad():
# for it, data in enumerate(dataloader_val, 0):
# vox16, vox64, name = data
# vox16 = vox16.cuda()
# vox16 = vox16.contiguous().view(opt.patchNum, 1, 16, 16, 16)
# vox64 = vox64.cuda()
# vox64 = vox64.contiguous().view(opt.patchNum, 1, 64, 64, 64)
#
# vox_pre, vox_pre_sigmoid = network_patch_l2h_img(vox16)
# loss = criterion(vox_pre, vox64)
# logger.add_scalar('val/loss', loss, it_step)
#
# print('[%d: %d/%d] val loss: %f' % (epoch, it, len_dataset_val / opt.batchSize, loss.item()))
# break # just validate one batch
torch.save(network_patch_l2h.state_dict(), os.path.join(model_path, 'patch_l2h.pt'))
print('save model')
print('Training done!')