-
Notifications
You must be signed in to change notification settings - Fork 7
/
run.py
165 lines (124 loc) · 7.57 KB
/
run.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
'''
This source code is licensed under the license found in the LICENSE file.
This is the implementation of the "Learning to deblur using light field generated and real defocus images" paper accepted to CVPR 2022.
Project GitHub repository: https://github.com/lingyanruan/DRBNet
Email: lyruanruan@gmail.com
Copyright (c) 2022-present, Lingyan Ruan
'''
import os
from options.test_options import TestOptions
from datetime import datetime
import torch
import torchvision.utils as vutils
from ptflops import get_model_complexity_info
from util.util import *
from pathlib import Path
import time
import sys
import lpips
from glob import glob
from natsort import natsorted
from skimage.metrics import peak_signal_noise_ratio as compute_psnr
from skimage.metrics import structural_similarity as compute_ssim
from models.DRBNet import *
#### metrics #################################
compute_lpips = lpips.LPIPS(net='alex').cuda()
opt = TestOptions().parse()
#### define time
folder_time = datetime.now().strftime('%Y-%m-%d_%H%M')
# results save position
opt.results_dir = opt.results_dir + '/' + opt.name + '/' + opt.eval_data + '/' + opt.net_mode +'/'+ folder_time
#### make directory ################################
Path(os.path.join(opt.results_dir, 'input' )).mkdir(parents=True, exist_ok=True)
Path(os.path.join(opt.results_dir, 'output')).mkdir(parents=True, exist_ok=True)
## evaluation values
PSNR_total,SSIM_total,LPIPS_total = 0,0,0
PSNR_score, SSIM_score, LPIPS_score , total_time= 0,0,0,0
######################################### Dataset List #################################################
input_c_file_path_list = []
if opt.eval_data == 'DPDD':
input_c_file_path_list = natsorted(glob(os.path.join(opt.dataroot_dpdd, 'test_c','source', '*.png')))
input_r_file_path_list = natsorted(glob(os.path.join(opt.dataroot_dpdd, 'test_r', 'source', '*.png')))
input_l_file_path_list = natsorted(glob(os.path.join(opt.dataroot_dpdd, 'test_l', 'source','*.png')))
gt_file_path_list = natsorted(glob(os.path.join(opt.dataroot_dpdd, 'test_c', 'target', '*.png')))
elif opt.eval_data == 'RealDOF':
input_c_file_path_list = natsorted(glob(os.path.join(opt.dataroot_rf, 'source', '*.png')))
gt_file_path_list = natsorted(glob(os.path.join(opt.dataroot_rf, 'target', '*.png')))
elif opt.eval_data == 'PixelDP':
input_c_file_path_list = natsorted(glob(os.path.join(opt.dataroot_pixeldp, 'test_c','source', '*.png')))
gt_file_path_list = None
elif opt.eval_data == 'CUHK':
input_c_file_path_list = natsorted(glob(os.path.join(opt.dataroot_cuhk,'*')))
gt_file_path_list = None
total_files = len(input_c_file_path_list)
assert total_files > 0, 'Wrong Dataset Name or No Dataset Exist, Please Check!!'
print('\n\n================================= EVALUATION START ==================================================')
for i, filename in enumerate(input_c_file_path_list):
# Read Image
C = crop_image(read_image(input_c_file_path_list[i], 255.0))*2-1
C = torch.FloatTensor(C.transpose(0, 3, 1, 2).copy()).cuda()
filename = os.path.split(filename)[-1]
if opt.net_mode == 'dual':
R,L = crop_image(read_image(input_r_file_path_list[i], 255.0))*2-1, crop_image(read_image(input_l_file_path_list[i], 255.0))*2-1
R,L = torch.FloatTensor(R.transpose(0, 3, 1, 2).copy()).cuda(), torch.FloatTensor(L.transpose(0, 3, 1, 2).copy()).cuda()
if gt_file_path_list is not None:
GT = crop_image(read_image(gt_file_path_list[i], 255.0)) # here to [0,1]
GT = torch.FloatTensor(GT.transpose(0, 3, 1, 2).copy()).cuda()
##test resut
with torch.no_grad():
if opt.net_mode == 'single':
network = DRBNet_single().cuda()
opt.ckpt_path = './ckpts/single/single_image_defocus_deblurring.pth' #final one
network.load_state_dict(torch.load(opt.ckpt_path))
start_time = time.time()
output = network(C)
time_per = time.time() - start_time
else:
network = DeblurNet_dual().cuda()
opt.ckpt_path = './ckpts/dual/dual_images_defocus_deblurring.pth'
network.load_state_dict(torch.load(opt.ckpt_path))
start_time = time.time()
output = network(C,R,L)
time_per = time.time() - start_time
total_time = total_time + time_per
output_cpu = (output.cpu().numpy()[0].transpose(1, 2, 0) +1.0 )/2.0 # to [0,1] for psnr and ssim evaluation
if gt_file_path_list is not None:
GT_cpu = GT.cpu().numpy()[0].transpose(1, 2, 0)
PSNR_score = compute_psnr(output_cpu, GT_cpu,data_range=1.0)
SSIM_score = compute_ssim(output_cpu, GT_cpu,data_range=1.0,multichannel=True)
LPIPS_score = compute_lpips(output, GT * 2. - 1.).item()
if opt.save_images:
save_file_path_deblur_input = os.path.join(opt.results_dir, 'input', '{}'.format(filename))
save_file_path_deblur = os.path.join(opt.results_dir, 'output', '{}'.format(filename))
vutils.save_image((C+1.0)/2.0, '{}'.format(save_file_path_deblur_input), nrow=1, padding = 0, normalize = False)
vutils.save_image((output+1.0)/2.0, '{}'.format(save_file_path_deblur), nrow=1, padding = 0, normalize = False)
# Log
print('[EVAL on {}][{:02}/{}] {} PSNR: {:.5f}, SSIM: {:.5f}, LPIPS: {:.5f}, Time: {:.5f}sec'.format( opt.eval_data, i + 1, total_files, filename, PSNR_score, SSIM_score, LPIPS_score, time_per))
with open(os.path.join(opt.results_dir, 'score_{}.txt'.format(opt.eval_data)), 'w' if i == 0 else 'a') as file:
file.write('[EVAL][{:02}/{}] {} PSNR: {:.5f}, SSIM: {:.5f}, LPIPS: {:.5f}, Time: {:.5f}sec \n'.format( i + 1, total_files, filename, PSNR_score, SSIM_score, LPIPS_score, time_per))
file.close()
PSNR_total += PSNR_score
SSIM_total += SSIM_score
LPIPS_total += LPIPS_score
###=============================== network parameters info =======================================#######
PSNR_mean,SSIM_mean,LPIPS_mean,time_mean = PSNR_total / total_files,SSIM_total / total_files, LPIPS_total/total_files, total_time/total_files
def prepare_input(resolution):
input_blur_C = torch.FloatTensor(1, 3, 720, 1280).cuda()
input_blur_L = torch.FloatTensor(1, 3, 720, 1280).cuda()
input_blur_R = torch.FloatTensor(1, 3, 720, 1280).cuda()
return dict(C = input_blur_C, R=input_blur_L, L=input_blur_R)
### add network parameters info#######
if opt.net_mode == 'single':
Macs,params = get_model_complexity_info(network, (3, 720, 1280), as_strings=False)
print('\t{:<30} {:<8} B'.format('Computational complexity (Macs): ', Macs / 1000 ** 3 ))
print('\t{:<30} {:<8} M'.format('Number of parameters: ',params / 1000 ** 2, '\n'))
else:
Macs,params = get_model_complexity_info(network, (1,3, 720, 1280),input_constructor=prepare_input, as_strings=False,print_per_layer_stat=False)
print('\t{:<30} {:<8} B'.format('Computational complexity (Macs): ', Macs / 1000 ** 3 ))
print('\t{:<30} {:<8} M'.format('Number of parameters: ',params / 1000 ** 2, '\n'))
sys.stdout.write('\n[TOTAL |{}] PSNR: {:.5f} SSIM: {:.5f} LPIPS: {:.5f} ({:.5f}sec)'.format(opt.eval_data, PSNR_mean, SSIM_mean, LPIPS_mean, time_mean))
with open(os.path.join(opt.results_dir, 'score_{}.txt'.format(opt.eval_data)), 'a') as file:
file.write('\n[TOTAL ] PSNR: {:.5f} SSIM: {:.5f} LPIPS: {:.5f} ({:.5f}sec)'.format( PSNR_mean, SSIM_mean, LPIPS_mean, time_mean))
file.write('\n{:<30} {:<8} B'.format('Computational complexity (Macs): ', Macs / 1000 ** 3 ))
file.write('\n{:<30} {:<8} M'.format('Number of parameters: ', params / 1000 ** 2, '\n'))
file.close()