-
Notifications
You must be signed in to change notification settings - Fork 4
/
util.py
248 lines (203 loc) · 7.93 KB
/
util.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
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
import sys
import os
import time
import numpy as np
from os.path import join,exists
import glob
from tqdm import trange, tqdm
import cv2
import math
import scipy
import torch
from torch.nn import functional as F
import json
def automkdir(path):
if not exists(path):
os.makedirs(path)
def automkdirs(path):
[automkdir(p) for p in path]
def compute_psnr_torch(img1, img2):
mse = torch.mean((img1 - img2) ** 2)
return 10 * torch.log10(1. / mse)
def compute_psnr(img1, img2):
mse=np.mean((img1 - img2) ** 2)
return 10 * np.log(1. / mse) / np.log(10)
def DUF_downsample(x, scale=4):
"""Downsamping with Gaussian kernel used in the DUF official code
Args:
x (Tensor, [B, T, C, H, W]): frames to be downsampled.
scale (int): downsampling factor: 2 | 3 | 4.
"""
assert scale in [2, 3, 4], 'Scale [{}] is not supported'.format(scale)
def gkern(kernlen=13, nsig=1.6):
import scipy.ndimage.filters as fi
inp = np.zeros((kernlen, kernlen))
# set element at the middle to one, a dirac delta
inp[kernlen // 2, kernlen // 2] = 1
# gaussian-smooth the dirac, resulting in a gaussian filter mask
return fi.gaussian_filter(inp, nsig)
B, T, C, H, W = x.size()
x = x.view(-1, 1, H, W)
filter_height, filter_width = 13, 13
pad_w, pad_h = (filter_width-1)//2, (filter_height-1)//2 # 6 is the pad of the gaussian filter
r_h, r_w = 0, 0
if scale == 3:
if H % 3 != 0:
r_h = 3 - (H % 3)
if W % 3 != 0:
r_w = 3 - (W % 3)
x = F.pad(x, (pad_w, pad_w + r_w, pad_h, pad_h + r_h), 'reflect')
gaussian_filter = torch.from_numpy(gkern(filter_height, 0.4 * scale)).type_as(x).unsqueeze(0).unsqueeze(0)
x = F.conv2d(x, gaussian_filter, stride=scale)
x = x.view(B, T, C, x.size(2), x.size(3))
return x
def makelr_fromhr_cuda(hr, scale=4, device=None, data_kind='single'):
if data_kind == 'double' or isinstance(hr, (tuple, list)):
return [i.to(device) for i in hr]
else:
hr = hr.to(device)
lr = DUF_downsample(hr, scale)
return lr, hr
def evaluation(model, eval_data, config):
model.eval()
start = torch.cuda.Event(enable_timing=True)
end = torch.cuda.Event(enable_timing=True)
psnr_all=[]
scale = config.model.scale
epoch = config.train.epoch
device = config.device
test_runtime=[]
in_h=128
in_w=240
bd=2
for iter_eval, (img_hq) in enumerate(tqdm(eval_data)):
img_hq = img_hq[:, :, :, bd * scale: (bd + in_h) * scale, bd * scale: (bd + in_w) * scale]
img_lq, img_hq = makelr_fromhr_cuda(img_hq, scale, device, config.data_kind)
# img_lq = img_lq[:, :, :, :in_h, :in_w]
# img_hq = img_hq[:, :, :, :in_h*scale, :in_w*scale]
B, C, T, H, W = img_lq.shape
start.record()
with torch.no_grad():
img_clean = model(img_lq)
end.record()
torch.cuda.synchronize()
test_runtime.append(start.elapsed_time(end) / T)
cleans = [_.permute(0,2,3,4,1) for _ in img_clean]
hr = img_hq.permute(0,2,3,4,1)
psnr_cleans, psnr_hr = cleans, hr
psnrs = [compute_psnr_torch(_, psnr_hr).cpu().numpy() for _ in psnr_cleans]
clean = (np.round(np.clip(cleans[0].cpu().numpy()[0, T // 2] * 255, 0, 255))).astype(np.uint8)
cv2_imsave(join(config.path.eval_result,'{:0>4}.png'.format(iter_eval )), clean)
psnr_all.append(psnrs)
psnrs = np.array(psnr_all)
psnr_avg = np.mean(psnrs, 0, keepdims = False)
with open(config.path.eval_file,'a+') as f:
eval_dict = {'Epoch': epoch, 'PSNR': psnr_avg.tolist()}
eval_json = json.dumps(eval_dict)
f.write(eval_json)
f.write('\n')
print(eval_json)
ave_runtime = sum(test_runtime) / len(test_runtime)
print(f'average time cost {ave_runtime} ms')
model.train()
return psnr_avg
def test_video(model, path, savepath, config):
model.eval()
automkdir(savepath)
scale = config.model.scale
device = config.device
# print(savepath)
prefix = os.path.split(path)[-1]
inp_type = 'truth' if config.data_kind == 'single' else f'input{config.model.scale}'
imgs=sorted(glob.glob(join(path, inp_type, '*.png')))
start = torch.cuda.Event(enable_timing=True)
end = torch.cuda.Event(enable_timing=True)
test_runtime=[]
if inp_type == 'truth':
img_hq = [cv2_imread(i) for i in imgs]
img_hq = torch.from_numpy(np.array(img_hq)/255.).float().permute(3,0,1,2).contiguous()
img_hq = img_hq.to(device)
img_lq = DUF_downsample(img_hq.unsqueeze(0), scale)
else:
img_lq = [cv2_imread(i) for i in imgs]
img_lq = torch.from_numpy(np.array(img_lq)).float().permute(3,0,1,2).contiguous()/255.
img_lq = img_lq.to(device).unsqueeze(0)
B, C, T, H, W = img_lq.shape
files_info = [os.path.split(_)[-1] for _ in imgs]
start.record()
with torch.no_grad():
img_clean = model(img_lq)
end.record()
torch.cuda.synchronize()
test_runtime.append(start.elapsed_time(end)) # milliseconds
if isinstance(img_clean, tuple):
img_clean = img_clean[0]
sr = img_clean[0].permute(1,2,3,0)
sr = sr.cpu().numpy()
sr = (np.round(np.clip(sr * 255, 0, 255))).astype(np.uint8)
[cv2_imsave(join(savepath, files_info[i]), sr[i]) for i in range(T)]
print('Cost {} ms in average.\n'.format(np.mean(test_runtime) / T))
return
def save_checkpoint(model, epoch, model_folder):
model_out_path = os.path.join(model_folder , '{:0>4}.pth'.format(epoch))
state = {"epoch": epoch ,"model": model.state_dict()}
if not os.path.exists(model_folder):
os.makedirs(model_folder)
torch.save(state, model_out_path)
print("Checkpoint saved to {}".format(model_out_path))
return
def load_checkpoint(network=None, resume='', path='', weights_init=None, rank=0):
try:
num_resume = int(resume[resume.rfind('/')+1:resume.rfind('.')])
except Exception as e:
num_resume = 0
finally:
if num_resume < 0:
checkpointfile = sorted(glob.glob(join(path,'*')))
if len(checkpointfile)==0:
resume = 'nofile'
else:
resume = checkpointfile[-1]
if os.path.isfile(resume):
print("=> loading checkpoint '{}'".format(resume))
checkpoint = torch.load(resume, map_location=lambda storage, loc: storage.cuda(rank))
start_epoch = checkpoint["epoch"]
network.load_state_dict(checkpoint["model"])
else:
print("=> no checkpoint found at '{}'".format(resume))
if weights_init is not None:
network.apply(weights_init)
start_epoch = 0
return start_epoch
def adjust_learning_rate(init_lr, final_lr, epoch, epoch_decay, iteration, iter_per_epoch, optimizer, ifprint=False):
"""Sets the learning rate to the initial LR decayed by 10"""
lr = (init_lr-final_lr) * max((1 - (epoch + iteration / iter_per_epoch) / epoch_decay), 0)+final_lr
for i, param_group in enumerate(optimizer.param_groups):
param_group["lr"] = lr
if ifprint:
print("Epoch={}, lr={}".format(epoch, optimizer.param_groups[0]["lr"]))
return lr
def cv2_imsave(img_path, img):
img = np.squeeze(img)
if img.ndim == 3:
img = img[:, :, [2, 1, 0]]
cv2.imwrite(img_path, img)
def cv2_imread(img_path):
img=cv2.imread(img_path)
if img.ndim == 3:
img = img[:, :, [2, 1, 0]]
return img
class DICT2OBJ(object):
def __init__(self, obj, v=None):
# if not isinstance(obj, dict):
# setattr(self, obj, v)
# return
for k, v in obj.items():
if isinstance(v, dict):
# print('dict', k, v)
setattr(self, k, DICT2OBJ(v))
else:
# print('no dict', k, v)
setattr(self, k, v)
if __name__=='__main__':
pass