/
main.py
259 lines (219 loc) · 10 KB
/
main.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
248
249
250
251
252
253
254
255
256
257
258
259
import os
import argparse
import glob
import json
import cv2
import torch
import numpy as np
import pandas as pd
import loss
import models
import config_parser
from utils.common_utils import *
from utils.denoising_utils import *
from torch.utils.tensorboard import SummaryWriter
# beta version code
import additional_utils
def get_net(img_np, noise_np, args):
net = models.get_net(args)
if args.dip_type in ["dip_sure", "eSURE", "NCV_y", "eSURE_fixed", 'eSURE_new', 'eSURE_alpha', "eSURE_uniform", "eSURE_clip","eSURE_real", "no_div", "PURE", "PURE_dc", "dip_sure_new"]:
net_input = cv2_to_torch(noise_np, dtype)
print("[*] input_type : noisy image")
else:
INPUT = 'noise'
input_depth = 1 if args.gray else 3
# For SR, the get_noise should be same as img_np
net_input = get_noise(input_depth, INPUT, (img_np.shape[1], img_np.shape[2])).type(dtype).detach()
print("[*] input_type : noise")
return net, net_input
def get_optim(name, net, lr, beta):
if name == "adam":
print("[*] optim_type : Adam")
return torch.optim.Adam(net.parameters(), lr, beta)
elif name == "adamw":
print("[*] optim_type : AdamW (wd : 1e-2)")
return torch.optim.AdamW(net.parameters(), lr, beta) # default weight decay is 1e-2.
elif name == "RAdam":
return additional_utils.RAdam(net.parameters(), lr, beta)
else:
raise NotImplementedError
def image_restorazation(file, args):
# MAIN
stat = {}
task_type = args.task_type
# Step 1. prepare clean & degradation(noisy) pair
img_np, noisy_np = load_image_pair(file, task_type, args)
if args.GT_noise:
args.sigma = (img_np.astype(np.float) - noisy_np.astype(np.float)).std()
# np_to_torch function from utils.common_utils.
# _np : C,H,W [0, 255] -> _torch : C,H,W [0,1] scale
img_torch = cv2_to_torch(img_np, args.dtype)
noise_torch = cv2_to_torch(noisy_np, args.dtype)
# For PSNR measure.
noisy_clip_np = np.clip(noisy_np, 0, 255)
# Step 2. make model and model input
net, net_input = get_net(img_np, noisy_np, args)
net.train()
# Step 3. set loss function.
cal_loss = loss.get_loss(net, net_input, args)
optimizer = get_optim(args.optim, net, args.lr, (args.beta1, args.beta2))
if args.force_steplr:
scheduler = torch.optim.lr_scheduler.StepLR(optimizer, gamma=.9, step_size=300)
else:
scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer, milestones=[2000, 3000], gamma=0.5)
# Step 4. optimization and inference.
# Hyper_param for Learning
psnr_noisy_last = 0
psnr_gt_running = 0
save_dir = args.save_dir
# Ensemble methods.
running_avg = None
running_avg_ratio = args.running_avg_ratio
image_name = file.split("/")[-1][:-4]
np_save_dir = os.path.join(args.save_dir, image_name)
os.makedirs(np_save_dir, exist_ok=True)
stat["max_psnr"] = 0
stat["max_ssim"] = 0
stat["NUM_Backtracking"] = 0
args.writer = SummaryWriter(log_dir="runs/%s/%s" % (args.exp_tag, args.desc + image_name))
for ep in range(args.epoch):
optimizer.zero_grad()
total_loss, out = cal_loss(net_input, noise_torch)
with torch.no_grad():
mse_loss = torch.nn.functional.mse_loss(out, img_torch).item()
diff_loss = total_loss.item() - mse_loss
args.writer.add_scalar("loss/used_loss", total_loss.item(), global_step=ep)
args.writer.add_scalar("loss/MSE_loss", mse_loss, global_step=ep)
args.writer.add_scalar("loss/diff", diff_loss, global_step=ep)
# _torch : C,H,W [0,1] scale => _np : C,H,W [0, 255]
#out = torch_to_cv2(net(net_input))
out = torch_to_cv2(out)
psnr_noisy = calculate_psnr(noisy_clip_np, out)
psnr_gt = calculate_psnr(img_np, out)
lpips_noisy = calculate_lpips(noisy_clip_np, out, args.lpips)
lpips_gt = calculate_lpips(img_np, out, args.lpips)
args.writer.add_scalar("psnr/noisy_to_out", psnr_noisy, global_step=ep)
args.writer.add_scalar("psnr/clean_to_out", psnr_gt, global_step=ep)
args.writer.add_scalar("lpips/noisy_to_out", lpips_noisy, global_step=ep)
args.writer.add_scalar("lpips/clean_to_out", lpips_gt, global_step=ep)
if total_loss < 0:
print('\nLoss is less than 0')
for new_param, net_param in zip(last_net, net.parameters()):
net_param.data.copy_(new_param.cuda())
break
if (psnr_noisy - psnr_noisy_last < -5) and (ep > 5) :
print('\nFalling back to previous checkpoint.')
for new_param, net_param in zip(last_net, net.parameters()):
net_param.data.copy_(new_param.cuda())
stat["NUM_Backtracking"] += 1
if stat["NUM_Backtracking"] > 10:
break
# continue
else:
# Running ensemble
if True: #(ep % 50 == 0) and
if running_avg is None:
running_avg = out
else:
running_avg = running_avg * running_avg_ratio + out * (1 - running_avg_ratio)
psnr_gt_running = calculate_psnr(img_np, running_avg)
lpips_gt_running = calculate_lpips(img_np, running_avg, args.lpips, color="BGR")
args.writer.add_scalar("psnr/clean_to_avg", psnr_gt_running, global_step=ep)
args.writer.add_scalar("lpips/clean_to_avg", lpips_gt_running, global_step=ep)
if (stat["max_psnr"] <= psnr_gt):
stat["max_step"] = ep
stat["max_psnr"] = psnr_gt
stat["max_psnr_avg"] = psnr_gt_running
stat["max_lpips_avg"] = lpips_gt_running
stat["max_lpips"] = lpips_gt
max_out, maxavg_out = out.copy(),running_avg.copy()
#save file
if args.save_np:
state_dict = net.state_dict()
torch.save(state_dict, os.path.join(np_save_dir, "max_psnr_state_dict.pth"))
if (ep == 200 or ep == 10) and (psnr_gt_running < psnr_gt):
running_avg = None
# args.writer.add_image("result/gt_noise_out_avg", np.concatenate([img_np, noisy_np, out, running_avg], axis=2), ep)
print('Iteration %05d total loss / MSE / diff %f / %f / %f PSNR_noisy: %f psnr_gt: %f PSNR_gt_sm: %f' % (
ep, total_loss.item(), mse_loss, diff_loss, psnr_noisy, psnr_gt, psnr_gt_running), end='\r')
last_net = [x.detach().cpu() for x in net.parameters()]
psnr_noisy_last=psnr_noisy
total_loss.backward()
optimizer.step()
scheduler.step()
torch.cuda.empty_cache()
if args.optim_init > 0:
if ep % args.optim_init == 0:
additional_utils.init_optim(net, optimizer)
stat["final_ep"] = ep
stat["final_psnr"] = psnr_gt
stat["final_psnr_avg"] = psnr_gt_running
stat["final_lpips_avg"]= lpips_gt_running
stat["final_lpips"] = lpips_gt
# Make final images
if True:
save_CHW_np(save_dir + "/%s.png" % (image_name), out)
save_CHW_np(save_dir + "/%s_avg.png" % (image_name), running_avg)
save_CHW_np(save_dir + "/%s_max.png" % (image_name), max_out)
save_CHW_np(save_dir + "/%s_max_avg.png" % (image_name), maxavg_out)
if args.gray:
stat["final_ssim"] = calculate_ssim(img_np, out)
stat["final_ssim_avg"] = calculate_ssim(img_np, running_avg)
stat["max_ssim"] = calculate_ssim(img_np, max_out)
stat["max_ssim_avg"] = calculate_ssim(img_np, maxavg_out)
log_file = open(save_dir + "/%s_log.txt" % (image_name), "w")
print(stat, file=log_file)
print("%s psnr clean_out : %.2f, %.2f noise_out : %.2f, max %.2f, %.2f" % (
image_name, psnr_gt_running, lpips_gt_running, psnr_noisy, stat["max_psnr"], stat["max_lpips"]), " " * 100)
print(stat)
args.writer.close()
torch.cuda.empty_cache()
return stat
def read_dataset_file_list(eval_data):
dataset_dir = "./testset/%s/" % eval_data
file_list1 = glob.glob(dataset_dir + "*.tif")
file_list2 = glob.glob(dataset_dir + "*.png")
file_list3 = glob.glob(dataset_dir + "*.JPG")
file_list = file_list1 + file_list2 + file_list3
return file_list
if __name__ == "__main__":
# For REPRODUCIBILITY
print("[*] reproduce mode On")
torch.manual_seed(0)
np.random.seed(0)
if torch.cuda.is_available():
torch.backends.cudnn.enabled = True
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
dtype = torch.cuda.FloatTensor
lpips = get_lpips("cuda")
else:
dtype = torch.FloatTensor
lpips = get_lpips("cpu")
args = config_parser.main_parser()
args.save_dir = "./result/%s/%s/%s" % (args.task_type, args.exp_tag, args.dip_type + args.desc)
os.makedirs(args.save_dir, exist_ok = True)
# default epoch setup.
if args.task_type == "denoising":
args.epoch = 3000 if args.epoch == 0 else args.epoch
args.save_point = [1, 10, 100, 500, 1000, 2000, 3000, 4000]
elif args.task_type == "poisson":
args.epoch = 3000 if args.epoch == 0 else args.epoch
args.save_point = [1, 10, 100, 500, 1000, 2000, 3000, 4000]
with open(os.path.join(args.save_dir, 'args.json'), 'w') as f:
json.dump(args.__dict__, f, indent=2)
args.dtype = dtype
args.lpips = lpips
# file_list.
file_list = read_dataset_file_list(args.eval_data)
file_list = sorted(file_list)
stat_list = []
for file in file_list:
print("[*] process image file : %s" % file)
stat = image_restorazation(file, args)
stat_list.append(stat)
data = pd.DataFrame(stat_list, index= [i.split("/")[-1] for i in file_list])
os.makedirs("./csv/%s/%s/" % (args.task_type, args.exp_tag), exist_ok=True)
data.to_csv("./csv/%s/%s/%s.csv" % ( args.task_type, args.exp_tag ,args.dip_type+args.desc))
print("experiment done")
print(data)