-
Notifications
You must be signed in to change notification settings - Fork 9
/
train.py
174 lines (158 loc) · 7.64 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
# -*- coding:utf-8 -*-
# -----------------------------------------
# Filename: train.py
# Author : Qing Wu
# Email : [email protected]
# Date : 2021/9/19
# -----------------------------------------
import data
import torch
import model
import argparse
import time
from torch.utils.tensorboard import SummaryWriter
if __name__ == '__main__':
writer = SummaryWriter('./log')
# -----------------------
# parameters settings
# -----------------------
parser = argparse.ArgumentParser()
# about ArSSR model
parser.add_argument('-encoder_name', type=str, default='RDN', dest='encoder_name',
help='the type of encoder network, including RDN (default), ResCNN, and SRResnet.')
parser.add_argument('-decoder_depth', type=int, default=8, dest='decoder_depth',
help='the depth of the decoder network (default=8).')
parser.add_argument('-decoder_width', type=int, default=256, dest='decoder_width',
help='the width of the decoder network (default=256).')
parser.add_argument('-feature_dim', type=int, default=128, dest='feature_dim',
help='the dimension size of the feature vector (default=128)')
# about training and validation data
parser.add_argument('-hr_data_train', type=str, default='./data/hr_train', dest='hr_data_train',
help='the file path of HR patches for training')
parser.add_argument('-hr_data_val', type=str, default='./data/hr_val', dest='hr_data_val',
help='the file path of HR patches for validation')
# about training hyper-parameters
parser.add_argument('-lr', type=float, default=1e-4, dest='lr',
help='the initial learning rate')
parser.add_argument('-lr_decay_epoch', type=int, default=200, dest='lr_decay_epoch',
help='learning rate multiply by 0.5 per lr_decay_epoch .')
parser.add_argument('-epoch', type=int, default=2500, dest='epoch',
help='the total number of epochs for training')
parser.add_argument('-summary_epoch', type=int, default=200, dest='summary_epoch',
help='the current model will be saved per summary_epoch')
parser.add_argument('-bs', type=int, default=15, dest='batch_size',
help='the number of LR-HR patch pairs (i.e., N in Equ. 3)')
parser.add_argument('-ss', type=int, default=8000, dest='sample_size',
help='the number of sampled voxel coordinates (i.e., K in Equ. 3)')
parser.add_argument('-gpu', type=int, default=0, dest='gpu',
help='the number of GPU')
args = parser.parse_args()
encoder_name = args.encoder_name
decoder_depth = args.decoder_depth
decoder_width = args.decoder_width
feature_dim = args.feature_dim
hr_data_train = args.hr_data_train
hr_data_val = args.hr_data_val
lr = args.lr
lr_decay_epoch = args.lr_decay_epoch
epoch = args.epoch
summary_epoch = args.summary_epoch
batch_size = args.batch_size
sample_size = args.sample_size
gpu = args.gpu
# -----------------------
# display parameters
# -----------------------
print('Parameter Settings')
print('')
print('------------File------------')
print('hr_data_train: {}'.format(hr_data_train))
print('hr_data_val: {}'.format(hr_data_val))
print('------------Train-----------')
print('lr: {}'.format(lr))
print('batch_size_train: {}'.format(batch_size))
print('sample_size: {}'.format(sample_size))
print('gpu: {}'.format(gpu))
print('epochs: {}'.format(epoch))
print('summary_epoch: {}'.format(summary_epoch))
print('lr_decay_epoch: {}'.format(lr_decay_epoch))
print('------------Model-----------')
print('encoder_name : {}'.format(encoder_name))
print('decoder feature_dim: {}'.format(feature_dim))
print('decoder depth: {}'.format(decoder_depth))
print('decoder width: {}'.format(decoder_width))
for i in range(5):
print(i + 1, end="s,")
time.sleep(1)
# -----------------------
# load data
# -----------------------
train_loader = data.loader_train(in_path_hr=hr_data_train, batch_size=batch_size,
sample_size=sample_size, is_train=True)
val_loader = data.loader_train(in_path_hr=hr_data_val, batch_size=1,
sample_size=sample_size, is_train=False)
# -----------------------
# model & optimizer
# -----------------------
DEVICE = torch.device('cuda:{}'.format(str(gpu) if torch.cuda.is_available() else 'cpu'))
ArSSR = model.ArSSR(encoder_name=encoder_name, feature_dim=feature_dim,
decoder_depth=int(decoder_depth / 2), decoder_width=decoder_width).to(DEVICE)
loss_fun = torch.nn.L1Loss()
optimizer = torch.optim.Adam(params=ArSSR.parameters(), lr=lr)
# -----------------------
# training & validation
# -----------------------
for e in range(epoch):
ArSSR.train()
loss_train = 0
for i, (img_lr, xyz_hr, img_hr) in enumerate(train_loader):
# forward
img_lr = img_lr.unsqueeze(1).to(DEVICE).float() # N×1×h×w×d
img_hr = img_hr.to(DEVICE).float().view(batch_size, -1).unsqueeze(-1) # N×K×1 (K Equ. 3)
xyz_hr = xyz_hr.view(batch_size, -1, 3).to(DEVICE).float() # N×K×3
img_pre = ArSSR(img_lr, xyz_hr) # N×K×1
loss = loss_fun(img_pre, img_hr)
# backward
optimizer.zero_grad()
loss.backward()
optimizer.step()
# record and print loss
loss_train += loss.item()
current_lr = optimizer.state_dict()['param_groups'][0]['lr']
print('(TRAIN) Epoch[{}/{}], Steps[{}/{}], Lr:{}, Loss:{:.10f}'.format(e + 1,
epoch,
i + 1,
len(train_loader),
current_lr,
loss.item()))
writer.add_scalar('MES_train', loss_train / len(train_loader), e + 1)
# release memory
img_lr = None
img_hr = None
xyz_hr = None
img_pre = None
ArSSR.eval()
with torch.no_grad():
loss_val = 0
for i, (img_lr, xyz_hr, img_hr) in enumerate(val_loader):
img_lr = img_lr.unsqueeze(1).to(DEVICE).float() # N×1×h×w×d
xyz_hr = xyz_hr.view(1, -1, 3).to(DEVICE).float() # N×Q×3 (Q=H×W×D)
H, W, D = img_hr.shape[-3:]
img_hr = img_hr.to(DEVICE).float().view(1, -1).unsqueeze(-1) # N×Q×1 (Q=H×W×D)
img_pre = ArSSR(img_lr, xyz_hr) # N×Q×1 (Q=H×W×D)
loss_val += loss_fun(img_hr, img_pre)
# save validation
if (e + 1) % summary_epoch == 0:
# save model
torch.save(ArSSR.state_dict(), 'model/model_param_{}.pkl'.format(e + 1))
writer.add_scalar('MES_val', loss_val / len(val_loader), e + 1)
# release memory
img_lr = None
img_hr = None
xyz_hr = None
img_pre = None
# learning rate decays by half every some epochs.
if (e + 1) % lr_decay_epoch == 0:
for param_group in optimizer.param_groups:
param_group['lr'] *= 0.5
writer.flush()