forked from csdwren/SelfDeblur
-
Notifications
You must be signed in to change notification settings - Fork 0
/
selfdeblur_levin.py
158 lines (126 loc) · 5.39 KB
/
selfdeblur_levin.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
from __future__ import print_function
import argparse
import datetime
import glob
import os
from skimage.io import imsave
from torch.optim.lr_scheduler import MultiStepLR
from tqdm import tqdm
from SSIM import SSIM
from networks.fcn import fcn
from networks.skip import skip
from utils.common_utils import *
parser = argparse.ArgumentParser()
parser.add_argument('--num_iter', type=int, default=5000, help='number of epochs of training')
parser.add_argument('--img_size', type=int, default=[256, 256], help='size of each image dimension')
parser.add_argument('--kernel_size', type=int, default=[21, 21], help='size of blur kernel [height, width]')
parser.add_argument('--data_path', type=str, default="datasets/levin/", help='path to blurry image')
parser.add_argument('--save_path', type=str, default="results/levin/", help='path to save results')
parser.add_argument('--save_frequency', type=int, default=100, help='lfrequency to save results')
opt = parser.parse_args()
# print(opt)
torch.backends.cudnn.enabled = False
torch.backends.cudnn.benchmark = False
dtype = torch.FloatTensor # torch.cuda.FloatTensor
# warnings.filterwarnings("ignore")
files_source = glob.glob(os.path.join(opt.data_path, '*.png'))
files_source.sort()
opt_save_path = opt.save_path + datetime.datetime.utcnow().strftime("%Y%m%d%H%M%S") + '/'
os.makedirs(opt_save_path, exist_ok=True)
# start #image
for f in files_source:
INPUT = 'noise'
pad = 'reflection'
LR = 0.01
num_iter = opt.num_iter
reg_noise_std = 0.001
path_to_image = f
imgname = os.path.basename(f)
imgname = os.path.splitext(imgname)[0]
if imgname.find('kernel1') != -1:
opt.kernel_size = [17, 17]
if imgname.find('kernel2') != -1:
opt.kernel_size = [15, 15]
if imgname.find('kernel3') != -1:
opt.kernel_size = [13, 13]
if imgname.find('kernel4') != -1:
opt.kernel_size = [27, 27]
if imgname.find('kernel5') != -1:
opt.kernel_size = [11, 11]
if imgname.find('kernel6') != -1:
opt.kernel_size = [19, 19]
if imgname.find('kernel7') != -1:
opt.kernel_size = [21, 21]
if imgname.find('kernel8') != -1:
opt.kernel_size = [21, 21]
_, imgs = get_image(path_to_image, -1) # load image and convert to np.
y = np_to_torch(imgs).type(dtype)
img_size = imgs.shape
print(imgname)
# ######################################################################
padh, padw = opt.kernel_size[0] - 1, opt.kernel_size[1] - 1
opt.img_size[0], opt.img_size[1] = img_size[1] + padh, img_size[2] + padw
'''
x_net:
'''
input_depth = 8
net_input = get_noise(input_depth, INPUT, (opt.img_size[0], opt.img_size[1])).type(dtype)
net = skip(input_depth, 1,
num_channels_down=[128, 128, 128, 128, 128],
num_channels_up=[128, 128, 128, 128, 128],
num_channels_skip=[16, 16, 16, 16, 16],
upsample_mode='bilinear',
need_sigmoid=True, need_bias=True, pad=pad, act_fun='LeakyReLU')
net = net.type(dtype)
'''
k_net:
'''
n_k = 200
net_input_kernel = get_noise(n_k, INPUT, (1, 1)).type(dtype)
net_input_kernel.squeeze_()
net_kernel = fcn(n_k, opt.kernel_size[0] * opt.kernel_size[1])
net_kernel = net_kernel.type(dtype)
# Losses
mse = torch.nn.MSELoss().type(dtype)
ssim = SSIM().type(dtype)
# optimizer
optimizer = torch.optim.Adam([{'params': net.parameters()}, {'params': net_kernel.parameters(), 'lr': 1e-4}], lr=LR)
scheduler = MultiStepLR(optimizer, milestones=[2000, 3000, 4000], gamma=0.5) # learning rates
# initilization inputs
net_input_saved = net_input.detach().clone()
net_input_kernel_saved = net_input_kernel.detach().clone()
### start SelfDeblur
for step in tqdm(range(num_iter)):
# input regularization
net_input = net_input_saved + reg_noise_std * torch.zeros(net_input_saved.shape).type_as(
net_input_saved.data).normal_()
# change the learning rate
scheduler.step(step)
optimizer.zero_grad()
# get the network output
out_x = net(net_input)
out_k = net_kernel(net_input_kernel)
out_k_m = out_k.view(-1, 1, opt.kernel_size[0], opt.kernel_size[1])
# print(out_k_m)
out_y = nn.functional.conv2d(out_x, out_k_m, padding=0, bias=None)
if step < 1000:
total_loss = mse(out_y, y)
else:
total_loss = 1 - ssim(out_y, y)
total_loss.backward()
optimizer.step()
if (step + 1) % opt.save_frequency == 0:
print('Iteration %05d' % (step + 1))
print('Loss: {:.7f}'.format(float(total_loss)))
save_path = os.path.join(opt_save_path, '{}_x_{:05.0f}.png'.format(imgname, step + 1))
out_x_np = torch_to_np(out_x)
out_x_np = out_x_np.squeeze()
out_x_np = out_x_np[padh // 2:padh // 2 + img_size[1], padw // 2:padw // 2 + img_size[2]]
imsave(save_path, out_x_np)
save_path = os.path.join(opt_save_path, '{}_k_{:05.0f}.png'.format(imgname, step + 1))
out_k_np = torch_to_np(out_k_m)
out_k_np = out_k_np.squeeze()
out_k_np /= np.max(out_k_np)
imsave(save_path, out_k_np)
torch.save(net, os.path.join(opt_save_path, "%s_xnet.pth" % imgname))
torch.save(net_kernel, os.path.join(opt_save_path, "%s_knet.pth" % imgname))