-
Notifications
You must be signed in to change notification settings - Fork 3
/
Copy pathtrain.py
150 lines (134 loc) · 7.04 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
import os
import time
import torch
import random
import argparse
import numpy as np
import torch.nn as nn
import torch.optim as optim
from utils.tools import ERGAS
from torchinfo import summary
import torch.distributed as dist
from torch.autograd import Variable
import torch.backends.cudnn as cudnn
from model.u2net import U2Net as Net
from torch.utils.data import DataLoader
from utils.load_train_data import Dataset_Pro
SEED = 1
torch.manual_seed(SEED)
torch.cuda.manual_seed(SEED)
torch.cuda.manual_seed_all(SEED)
cudnn.benchmark = True
def save_checkpoint(args, model, epoch):
if not os.path.exists(args.weight_dir):
os.mkdir(args.weight_dir)
model_out_path = args.weight_dir + "{}.pth".format(epoch)
torch.save(model.state_dict(), model_out_path)
def prepare_training_data(args):
train_set = Dataset_Pro(args.train_data_path)
validate_set = Dataset_Pro(args.val_data_path)
training_data_loader = DataLoader(dataset=train_set, num_workers=0, batch_size=args.batch_size,
shuffle=True, pin_memory=True, drop_last=True)
validate_data_loader = DataLoader(dataset=validate_set, num_workers=0, batch_size=args.batch_size,
shuffle=True, pin_memory=True, drop_last=True)
return training_data_loader, validate_data_loader
def train(args, training_data_loader, validate_data_loader):
model = Net(args.channels, args.spa_channels, args.spe_channels, args.H, args.W, args.ratio).to(args.device)
summary(model, input_size=[(args.batch_size, args.spe_channels, 16, 16),
(args.batch_size, args.spa_channels, 64, 64)],
dtypes=[torch.float, torch.float])
if args.use_ergas is True:
criterion0 = nn.L1Loss(size_average=True).to(args.device)
criterion1 = ERGAS(args.ratio).to(args.device)
else:
criterion = nn.L1Loss(size_average=True).to(args.device)
optimizer = optim.Adam(model.parameters(), lr=args.lr, weight_decay=0)
lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer=optimizer, step_size=args.step, gamma=args.decay)
t_start = time.time()
print('Start training...')
# train
for epoch in range(0, args.epoch, 1):
epoch += 1
model.train()
epoch_train_loss = []
epoch_train_loss0 = []
epoch_train_loss1 = []
for iteration, batch in enumerate(training_data_loader, 1):
gt, pan, ms = batch[0].to(args.device), batch[3].to(args.device), batch[4].to(args.device)
optimizer.zero_grad()
sr = model(ms, pan)
if args.use_ergas is True:
loss0 = criterion0(sr, gt)
loss1 = criterion1(sr, gt)
loss = loss0 + args.ergas_hp * loss1
epoch_train_loss0.append(loss0.item())
epoch_train_loss1.append(loss1.item())
else:
loss = criterion(sr, gt)
epoch_train_loss.append(loss.item())
loss.backward()
optimizer.step()
lr_scheduler.step()
t_loss = np.nanmean(np.array(epoch_train_loss))
if args.use_ergas is True:
print('Epoch: {}/{} training loss: {:.7f} l1: {:.7f} ergas: {:.7f}'
.format(epoch, args.epoch, t_loss, np.nanmean(np.array(epoch_train_loss0)),
np.nanmean(np.array(epoch_train_loss1))))
else:
print('Epoch: {}/{} training loss: {:.7f}'.format(epoch, args.epoch, t_loss))
# validate
with torch.no_grad():
if epoch % 10 == 0:
model.eval()
epoch_val_loss = []
epoch_val_loss0 = []
epoch_val_loss1 = []
for iteration, batch in enumerate(validate_data_loader, 1):
gt, pan, ms = batch[0].to(args.device), batch[3].to(args.device), batch[4].to(args.device)
sr = model(ms, pan)
if args.use_ergas is True:
loss0 = criterion0(sr, gt)
loss1 = criterion1(sr, gt)
loss = loss0 + args.ergas_hp * loss1
epoch_val_loss0.append(loss0.item())
epoch_val_loss1.append(loss1.item())
else:
loss = criterion(sr, gt)
epoch_val_loss.append(loss.item())
v_loss = np.nanmean(np.array(epoch_val_loss))
t_end = time.time()
if args.use_ergas is True:
print('---------------validate loss: {:.7f} l1: {:.7f} ergas: {:.7f}----------------'
.format(v_loss, np.nanmean(np.array(epoch_val_loss0)), np.nanmean(np.array(epoch_val_loss1))))
else:
print('---------------validate loss: {:.7f}---------------'.format(v_loss))
print('-----------------total time cost: {:.4f}s--------------------'.format(t_end - t_start))
t_start = time.time()
# save weights
if epoch % args.ckpt == 0:
save_checkpoint(args, model, epoch)
else:
continue
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument('--ratio', type=int, default=4, help='Upsample ratio')
parser.add_argument('--H', type=int, default=64, help='Height of the high-resolution image')
parser.add_argument('--W', type=int, default=64, help='Width of the high-resolution image')
parser.add_argument('--channels', type=int, default=32, help='Feature channels')
parser.add_argument('--spa_channels', type=int, default=1, help='Spatial channels')
parser.add_argument('--spe_channels', type=int, default=8, help='Spectral channels')
parser.add_argument('--use_ergas', type=bool, default=False, help='Use ERGAS loss for training or not')
parser.add_argument('--ergas_hp', type=float, default=1e-4, help='Hyper-parameter for the ERGAS loss')
parser.add_argument('--epoch', type=int, default=500, help='Epochs')
parser.add_argument('--batch_size', type=int, default=32, help='Batch Size')
parser.add_argument('--lr', type=float, default=5e-4, help='Learning rate')
parser.add_argument('--step', type=int, default=200, help='Step number')
parser.add_argument('--decay', type=float, default=0.5, help='Learning rate decay')
parser.add_argument('--ckpt', type=int, default=20, help='Checkpoint')
parser.add_argument('--device', type=str, default='cuda')
parser.add_argument('--train_data_path', type=str, default='', help='Path of the training dataset.')
parser.add_argument('--val_data_path', type=str, default='', help='Path of the validation dataset.')
parser.add_argument('--weight_dir', type=str, default='weights/', help='Dir of the weight.')
args = parser.parse_args()
training_data_loader, validate_data_loader = prepare_training_data(args)
train(args, training_data_loader, validate_data_loader)