-
Notifications
You must be signed in to change notification settings - Fork 0
/
train_proposed.py
123 lines (106 loc) · 6.07 KB
/
train_proposed.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
import torch
import os
from args import parse_args
from utils import initialize_params, get_psfs
from dataset import MyDataset
from torch.utils.data import DataLoader
from conv_deconv import conv_fn, loss_fn, sensor_noise, feat_extract, wiener
from torch.utils.tensorboard import SummaryWriter
"""
只有FE的网络
"""
def train(args):
params = initialize_params(args)
params['phase_type'] = 'hyperboloid_learn'
parameters_to_save = {}
if (params['phase_type'] == 'hyperboloid' or params['phase_type'] == 'cubic' or params['phase_type'] == 'log_asphere'):
fs = torch.tensor([2.5E-3 * 511 / 452]*9, device=params['device'])
elif (params['phase_type'] == 'hyperboloid_learn' or params['phase_type'] == 'cubic_learn'):
fs = torch.tensor([2.5E-3]*9, device=params['device'])
fs = torch.nn.Parameter(fs)
phase_optimizer = torch.optim.Adam([fs], lr=args.phase_lr)
phase_scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(phase_optimizer, mode='min', factor=0.1, patience=5)
parameters_to_save.update({'phase_params': fs, 'phase_optimizer': phase_optimizer.state_dict()})
trainset = MyDataset(args.train_dir,810)
evalset = MyDataset(args.eval_dir, 810)
trainloader = DataLoader(trainset, batch_size=args.train_batch_size, shuffle=True)
evalloader = DataLoader(evalset, batch_size=args.train_batch_size, shuffle=False)
nn = feat_extract(27).to(params['device'])
nn_optimizer = torch.optim.Adam(nn.parameters(), lr=args.nn_lr)
nn_scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(nn_optimizer, mode='min', factor=0.1, patience=5)
parameters_to_save.update({'nn_params': nn.state_dict(), 'nn_optimizer': nn_optimizer.state_dict()})
if not os.path.exists(os.path.join(args.save_dir, args.exp_name)):
os.makedirs(os.path.join(args.save_dir, args.exp_name))
writer = SummaryWriter(os.path.join(args.log_dir, args.exp_name))
train_stage_flag = 'optimize_nn'
train_stage_iter = 0
log_iter = 0
best_eval_loss = 1E6
for i in range(args.epochs):
# Train
loss_val_epoch = 0
for j, img in enumerate(trainloader):
train_stage_iter += 1
log_iter += 1
rand_depth = torch.rand(1, device=params['device']) * (params['ub'] - params['lb']) + params['lb']
psf = get_psfs(fs, rand_depth, params) # (27,810,810)
psf = psf.reshape(9, 3, psf.size(-2),psf.size(-1)) # (9,3,810,810)
img = img.to(params['device'])
blur = conv_fn(img, psf) # (b,9,3,810,810)
blur = sensor_noise(blur, params['b_sqrt']) # (b,9,3,810,810)
deconv_result = wiener(blur, psf) # (b,9,3,810,810)
deconv_result = deconv_result.reshape(blur.size(0), -1, blur.size(-2), blur.size(-1)) # (b,27,810,810)
deconv_result = nn(deconv_result) # (b,3,810,810)
loss_val = loss_fn(deconv_result, img, params)
loss_val_epoch += loss_val * img.size(0)
loss_val.backward()
writer.add_scalar("train loss vs iters", loss_val, log_iter)
print("epoch:", i, " ---- ", "iters(/{}): ".format(len(trainloader)), j, " ---- ", "loss_val: ", loss_val.item())
if (train_stage_flag == 'optimize_nn'):
nn_optimizer.step()
nn_optimizer.zero_grad()
if (train_stage_iter % args.nn_iters == 0):
train_stage_flag = 'optimize_phase'
train_stage_iter = 0
elif (train_stage_flag == 'optimize_phase'):
phase_optimizer.step()
phase_optimizer.zero_grad()
if (train_stage_iter % args.phase_iters == 0):
train_stage_flag = 'optimize_nn'
train_stage_iter = 0
loss_val_epoch /= len(trainset)
nn_scheduler.step(loss_val_epoch)
phase_scheduler.step(loss_val_epoch)
writer.add_scalar('train loss vs epochs', loss_val_epoch, i)
writer.add_scalar('phase learning rate', phase_optimizer.param_groups[0]['lr'], i)
writer.add_scalar('nn learning rate', nn_optimizer.param_groups[0]['lr'], i)
# Eval
loss_val_epoch = 0
if (i % args.log_freq == 0):
for j, img in enumerate(evalloader):
with torch.no_grad():
linear_depth = ((j+1) / len(evalloader)) * (params['ub'] - params['lb']) + params['lb']
psf = get_psfs(fs, linear_depth, params) # (27,810,810)
psf = psf.reshape(9, 3, psf.size(-2),psf.size(-1)) # (9,3,810,810)
img = img.to(params['device'])
blur = conv_fn(img, psf) # (b,9,3,810,810)
blur = sensor_noise(blur, params['b_sqrt']) # (b,9,3,810,810)
deconv_result = wiener(blur, psf) # (b,9,3,810,810)
deconv_result = deconv_result.reshape(blur.size(0), -1, blur.size(-2), blur.size(-1)) # (b,27,810,810)
deconv_result = nn(deconv_result) # (b,3,810,810)
loss_val = loss_fn(deconv_result, img, params)
loss_val_epoch += loss_val * img.size(0)
loss_val_epoch /= len(evalset)
writer.add_scalar('eval loss vs epochs', loss_val_epoch, i)
if (loss_val_epoch < best_eval_loss):
pt_save_path = os.path.join(args.save_dir, args.exp_name, 'paramters_best.pt'.format(i))
torch.save(parameters_to_save, pt_save_path)
best_eval_loss = loss_val_epoch
if (i % args.save_freq == 0):
pt_save_path = os.path.join(args.save_dir, args.exp_name, 'parameters.pt'.format(i))
torch.save(parameters_to_save, pt_save_path)
def main():
args = parse_args()
train(args)
if __name__ == '__main__':
main()