-
Notifications
You must be signed in to change notification settings - Fork 3
/
train.py
190 lines (152 loc) · 7.22 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
import os
import time
import torch
from utils import config_parser, load_fragments, load_idx, lr_decay, write_video, mse2psnr
from dataset.dataset import nerfDataset, ScanDataset, DTUDataset, TTDataset, toyDeskDataset
from model.renderer import Renderer
import matplotlib.pyplot as plt
import torch.optim as optim
from backup_utils import backup_terminal_outputs, backup_code, set_seed
from torch.utils.tensorboard import SummaryWriter
from piqa import SSIM, PSNR
import lpips
parser = config_parser()
args = parser.parse_args()
set_seed(1023)
back_path = os.path.join('logs', time.strftime("%y%m%d-%H%M%S-" + f'{args.expname}'))
os.makedirs(back_path)
backup_terminal_outputs(back_path)
backup_code(back_path, ignored_in_current_folder=['logs_pc_opt','logs_edit','ckpt','data','.git','pytorch_rasterizer.egg-info','build','logs','__pycache__'])
print(back_path)
logger = SummaryWriter(back_path)
video_path = os.path.join(back_path, 'video')
os.makedirs(video_path)
if __name__ == '__main__':
if args.dataset == 'nerf':
train_set = nerfDataset(args, 'train', 'render')
test_set = nerfDataset(args, 'test', 'render')
elif args.dataset == 'scan':
train_set = ScanDataset(args, 'train', 'render')
test_set = ScanDataset(args, 'test', 'render')
elif args.dataset == 'dtu':
train_set = DTUDataset(args, 'train', 'render')
test_set = DTUDataset(args, 'test', 'render')
elif args.dataset == 'tt':
train_set = TTDataset(args, 'train', 'render')
test_set = TTDataset(args, 'test', 'render')
elif args.dataset == 'toy':
train_set = toyDeskDataset(args, 'train', 'render')
test_set = toyDeskDataset(args, 'test', 'render')
else:
assert False
train_loader = torch.utils.data.DataLoader(train_set, batch_size=args.batch, shuffle=True)
test_loader = torch.utils.data.DataLoader(test_set, batch_size=1)
renderer = Renderer(args)
edge = args.edge_mask
# Optimizer
opt_para = []
opt_para.append({"params": renderer.unet.parameters(), "lr": args.u_lr})
opt_para.append({"params": renderer.afnet.parameters(), "lr": args.mlp_lr})
opt = optim.Adam(opt_para)
fn_psnr = PSNR().to(args.device)
# fn_lpips = LPIPS('vgg').to(args.device)
loss_lpips = lpips.LPIPS(net='vgg').to(args.device)
fn_ssim = SSIM().to(args.device)
train_buf, test_buf = load_fragments(args) # cpu 100 800 800 k
if args.ckpt is not None:
print(f'load model from {args.ckpt}')
renderer.load_state_dict(torch.load(args.ckpt))
it = 0
epoch = 0
best_psnr = 0
training_time = 0.
while True:
# if epoch in [0, 2, 9, 24, 70, 140] or epoch % 300 == 0:
if epoch % args.test_freq == 0:
print('TEST BEGIN!!!')
# video_it_path = os.path.join(video_path, str(epoch))
# os.makedirs(video_it_path)
test_psnr = 0
test_lpips = 0
test_ssim = 0
renderer.eval()
with torch.autograd.no_grad():
for i, batch in enumerate(test_loader):
idx = [int(id) for id in batch['idx']]
ray = batch['ray'] # b h w 7
img_gt = batch['rgb'].permute(0,3,1,2) # b 3 h w
zbuf = test_buf[idx].to(args.device) # b h w 1
output = renderer(zbuf, ray, gt=None, mask_gt=None, isTrain=False)
img_pre = torch.clamp(output['img'], 0, 1)
if edge > 0:
# for ScanNet_0000, since the artifacts at edges.
psnr = fn_psnr(img_pre[...,edge:-edge,edge:-edge], img_gt[...,edge:-edge,edge:-edge])
ssim = fn_ssim(img_pre[...,edge:-edge,edge:-edge], img_gt[...,edge:-edge,edge:-edge])
lpips_ = loss_lpips(img_pre[...,edge:-edge,edge:-edge], img_gt[...,edge:-edge,edge:-edge], normalize=True)
else:
psnr = fn_psnr(img_pre, img_gt)
ssim = fn_ssim(img_pre, img_gt)
lpips_ = loss_lpips(img_pre, img_gt, normalize=True)
test_lpips += lpips_.item()
test_psnr += psnr.item()
test_ssim += ssim.item()
# if epoch % args.vid_freq == 0:
img_pre = img_pre.squeeze(0).permute(1,2,0)
img_pre = img_pre.cpu().numpy()
plt.imsave(os.path.join(video_path, str(i).rjust(3,'0') + '.png'), img_pre)
# torch.cuda.empty_cache()
test_lpips = test_lpips / len(test_set)
test_psnr = test_psnr / len(test_set)
test_ssim = test_ssim / len(test_set)
logger.add_scalar('test/psnr', test_psnr, it)
logger.add_scalar('test/lpips', test_lpips, it)
logger.add_scalar('test/ssim', test_ssim, it)
if test_psnr > best_psnr:
best_psnr = test_psnr
ckpt = os.path.join(back_path, 'model.pkl')
torch.save(renderer.state_dict(), ckpt)
print(f'Model Saved! Best PSNR: {best_psnr:{4}.{4}}')
print(f'Test PSNR! Epoch:{epoch} Training_time: {training_time:{4}.{4}} hours, current: {test_psnr:{4}.{4}}, best: {best_psnr:{4}.{4}}')
renderer.train()
t1 = time.time()
epoch += 1
for batch in train_loader:
it += 1
idx = [int(id) for id in batch['idx']]
ray = batch['ray'] # b h w 7
img_gt = batch['rgb'] # b h w 3
zbuf = train_buf[idx].to(args.device) # b h w 1
# if args.dataset == 'dtu':
# mask_gt = batch['mask'] # b h w 1
# else:
# mask_gt = None
output = renderer(zbuf, ray, img_gt, mask_gt=None, isTrain=True)
img_pre = output['img']
if output['gt'].min() == 1:
# print('None img, skip')
# torch.cuda.empty_cache()
continue
opt.zero_grad()
loss_l2 = torch.mean((img_pre - output['gt']) ** 2)
if args.vgg_l > 0:
loss_vgg = torch.mean(loss_lpips(img_pre, output['gt'], normalize=True))
loss = loss_l2 + args.vgg_l * loss_vgg
else:
loss = loss_l2
loss.backward()
opt.step()
if it % 100 == 0:
psnr = mse2psnr(loss_l2)
logger.add_scalar('train/psnr', psnr.item(), it)
if it % 400 == 0:
if args.vgg_l > 0:
print('[{}]-it:{}, psnr:{:.4f}, l2_loss:{:.4f}, vgg_loss:{:.4f}'.format(back_path, it, psnr.item(), loss_l2.item(), loss_vgg.item()))
else:
print('[{}]-it:{}, psnr:{:.4f}, l2_loss:{:.4f}'.format(back_path, it, psnr.item(), loss.item()))
img_pre[img_pre>1] = 1.
img_pre[img_pre<0] = 0.
logger.add_image('train/img_fine', img_pre[0].permute(1,2,0), global_step=it, dataformats='HWC')
# torch.cuda.empty_cache()
lr_decay(opt)
t2 = time.time()
training_time += (t2 - t1) / 3600