-
Notifications
You must be signed in to change notification settings - Fork 0
/
inference_secan.py
198 lines (178 loc) · 6.63 KB
/
inference_secan.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
191
192
193
194
195
196
197
198
import argparse
import cv2
import glob
import numpy as np
import os
import torch
from torch.nn import functional as F
from secan.archs.secan_arch import SECAN
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--input', type=str, default='datasets/Set5/LRbicx4', help='input test image folder')
parser.add_argument('--output', type=str, default='results/SECAN/Set5', help='output folder')
parser.add_argument(
'--task',
type=str,
default='classical_sr',
help='classical_sr, lightweight_sr, real_sr, gray_dn, color_dn, jpeg_car')
# dn: denoising; car: compression artifact removal
# TODO: it now only supports sr, need to adapt to dn and jpeg_car
parser.add_argument('--patch_size', type=int, default=64, help='training patch size')
parser.add_argument('--scale', type=int, default=4, help='scale factor: 1, 2, 3, 4, 8') # 1 for dn and jpeg car
parser.add_argument('--noise', type=int, default=15, help='noise level: 15, 25, 50')
parser.add_argument('--jpeg', type=int, default=40, help='scale factor: 10, 20, 30, 40')
parser.add_argument('--large_model', action='store_true', help='Use large model, only used for real image sr')
parser.add_argument(
'--model_path',
type=str,
default='experiments/pretrained_models/SECANx2_973000.pth')
args = parser.parse_args()
os.makedirs(args.output, exist_ok=True)
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
# set up model
model = define_model(args)
model.eval()
model = model.to(device)
if args.task == 'jpeg_car':
window_size = 7
else:
window_size = 8
for idx, path in enumerate(sorted(glob.glob(os.path.join(args.input, '*')))):
# read image
imgname = os.path.splitext(os.path.basename(path))[0]
print('Testing', idx, imgname) # 2023.6.13为方便观看结果,注释
# read image
img = cv2.imread(path, cv2.IMREAD_COLOR).astype(np.float32) / 255.
img = torch.from_numpy(np.transpose(img[:, :, [2, 1, 0]], (2, 0, 1))).float()
img = img.unsqueeze(0).to(device)
# inference
with torch.no_grad():
# pad input image to be a multiple of window_size
mod_pad_h, mod_pad_w = 0, 0
_, _, h, w = img.size()
if h % window_size != 0:
mod_pad_h = window_size - h % window_size
if w % window_size != 0:
mod_pad_w = window_size - w % window_size
img = F.pad(img, (0, mod_pad_w, 0, mod_pad_h), 'reflect')
output = model(img)
_, _, h, w = output.size()
output = output[:, :, 0:h - mod_pad_h * args.scale, 0:w - mod_pad_w * args.scale]
# save image
output = output.data.squeeze().float().cpu().clamp_(0, 1).numpy()
if output.ndim == 3:
output = np.transpose(output[[2, 1, 0], :, :], (1, 2, 0))
output = (output * 255.0).round().astype(np.uint8)
cv2.imwrite(os.path.join(args.output, f'{imgname}_SECAN.png'), output)
def define_model(args):
# 001 classical image sr
if args.task == 'classical_sr':
model = SECAN(
upscale=args.scale,
in_chans=3,
img_size=args.patch_size,
window_size=8,
img_range=1.,
depths=[6, 6, 6, 6, 6, 6],
embed_dim=180,
num_heads=[6, 6, 6, 6, 6, 6],
mlp_ratio=2,
upsampler='pixelshuffle',
resi_connection='1conv')
# 002 lightweight image sr
# use 'pixelshuffledirect' to save parameters
elif args.task == 'lightweight_sr':
model = SECAN(
upscale=args.scale,
in_chans=3,
img_size=64,
window_size=8,
img_range=1.,
depths=[6, 6, 6, 6],
embed_dim=60,
num_heads=[6, 6, 6, 6],
mlp_ratio=2,
upsampler='pixelshuffledirect',
resi_connection='1conv')
# 003 real-world image sr
elif args.task == 'real_sr':
if not args.large_model:
# use 'nearest+conv' to avoid block artifacts
model = SECAN(
upscale=4,
in_chans=3,
img_size=64,
window_size=8,
img_range=1.,
depths=[6, 6, 6, 6, 6, 6],
embed_dim=180,
num_heads=[6, 6, 6, 6, 6, 6],
mlp_ratio=2,
upsampler='nearest+conv',
resi_connection='1conv')
else:
# larger model size; use '3conv' to save parameters and memory; use ema for GAN training
model = SECAN(
upscale=4,
in_chans=3,
img_size=64,
window_size=8,
img_range=1.,
depths=[6, 6, 6, 6, 6, 6, 6, 6, 6],
embed_dim=248,
num_heads=[8, 8, 8, 8, 8, 8, 8, 8, 8],
mlp_ratio=2,
upsampler='nearest+conv',
resi_connection='3conv')
# 004 grayscale image denoising
elif args.task == 'gray_dn':
model = SECAN(
upscale=1,
in_chans=1,
img_size=128,
window_size=8,
img_range=1.,
depths=[6, 6, 6, 6, 6, 6],
embed_dim=180,
num_heads=[6, 6, 6, 6, 6, 6],
mlp_ratio=2,
upsampler='',
resi_connection='1conv')
# 005 color image denoising
elif args.task == 'color_dn':
model = SECAN(
upscale=1,
in_chans=3,
img_size=128,
window_size=8,
img_range=1.,
depths=[6, 6, 6, 6, 6, 6],
embed_dim=180,
num_heads=[6, 6, 6, 6, 6, 6],
mlp_ratio=2,
upsampler='',
resi_connection='1conv')
# 006 JPEG compression artifact reduction
# use window_size=7 because JPEG encoding uses 8x8; use img_range=255 because it's slightly better than 1
elif args.task == 'jpeg_car':
model = SECAN(
upscale=1,
in_chans=1,
img_size=126,
window_size=7,
img_range=255.,
depths=[6, 6, 6, 6, 6, 6],
embed_dim=180,
num_heads=[6, 6, 6, 6, 6, 6],
mlp_ratio=2,
upsampler='',
resi_connection='1conv')
loadnet = torch.load(args.model_path)
if 'params_ema' in loadnet:
keyname = 'params_ema'
else:
keyname = 'params'
model.load_state_dict(loadnet[keyname], strict=True)
return model
if __name__ == '__main__':
main()