forked from DmitryUlyanov/deep-image-prior
-
Notifications
You must be signed in to change notification settings - Fork 0
/
denoising_example.py
244 lines (197 loc) · 7.14 KB
/
denoising_example.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
from __future__ import print_function
from models import *
from utils.common_utils import *
from utils.denoising_utils import *
# from utils.wandb_utils import *
import torch
import torch.optim
import matplotlib.pyplot as plt
import os
# import wandb
import argparse
import numpy as np
# from skimage.measure import compare_psnr
from skimage.metrics import peak_signal_noise_ratio as compare_psnr
torch.backends.cudnn.enabled = True
torch.backends.cudnn.benchmark =True
dtype = torch.cuda.FloatTensor
# Fix seeds
seed = 0
np.random.seed(seed)
torch.random.manual_seed(seed)
parser = argparse.ArgumentParser()
parser.add_argument('--gpu', default='2')
parser.add_argument('--index', default=0, type=int)
parser.add_argument('--input_index', default=0, type=int)
args = parser.parse_args()
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu
imsize = -1
PLOT = True
sigma = 25
sigma_ = sigma/255.
fnames = ['data/denoising/F16_GT.png', 'data/inpainting/kate.png', 'data/inpainting/vase.png', 'data/sr/zebra_GT.png']
fname = fnames[args.index]
if fname == 'data/denoising/snail.jpg':
img_noisy_pil = crop_image(get_image(fname, imsize)[0], d=32)
img_noisy_np = pil_to_np(img_noisy_pil)
# As we don't have ground truth
img_pil = img_noisy_pil
img_np = img_noisy_np
if PLOT:
plot_image_grid([img_np], 4, 5)
elif fname in fnames:
# Add synthetic noise
img_pil = crop_image(get_image(fname, imsize)[0], d=32)
img_np = pil_to_np(img_pil)
img_noisy_pil, img_noisy_np = get_noisy_image(img_np, sigma_)
# if PLOT:
# plot_image_grid([img_np, img_noisy_np], 4, 6)
else:
assert False
INPUT = ['noise', 'fourier', 'meshgrid', 'infer_freqs'][args.input_index]
pad = 'reflection'
OPT_OVER = 'net' # 'net'
train_input = True if ',' in OPT_OVER else False
reg_noise_std = 1. / 30. # set to 1./20. for sigma=50
# LR = 0.01
LR = 0.001
OPTIMIZER = 'adam' # 'LBFGS'
show_every = 100
exp_weight = 0.99
if fname == 'data/denoising/snail.jpg':
num_iter = 2400
input_depth = 3
figsize = 5
net = skip(
input_depth, 3,
num_channels_down=[8, 16, 32, 64, 128],
num_channels_up=[8, 16, 32, 64, 128],
num_channels_skip=[0, 0, 0, 4, 4],
upsample_mode='bilinear',
need_sigmoid=True, need_bias=True, pad=pad, act_fun='LeakyReLU')
net = net.type(dtype)
elif fname in fnames:
num_iter = 8000
input_depth = 32
figsize = 4
freq_dict = {
'method': 'log',
'max': 64,
'n_freqs': 8
}
net = get_net(input_depth, 'skip', pad,
skip_n33d=128,
skip_n33u=128,
skip_n11=4,
num_scales=5,
upsample_mode='bilinear').type(dtype)
else:
assert False
net_input = get_input(input_depth, INPUT, (img_pil.size[1], img_pil.size[0]), freq_dict=freq_dict).type(dtype)
# Compute number of parameters
s = sum([np.prod(list(p.size())) for p in net.parameters()])
print('Number of params: %d' % s)
# Loss
mse = torch.nn.MSELoss().type(dtype)
img_noisy_torch = np_to_torch(img_noisy_np).type(dtype)
if train_input:
net_input_saved = net_input
else:
net_input_saved = net_input.detach().clone()
noise = torch.rand_like(net_input) if INPUT == 'infer_freqs' else net_input.detach().clone()
# if INPUT == 'fourier':
# indices = sample_indices(input_depth, net_input_saved)
out_avg = None
last_net = None
psrn_noisy_last = 0
psnr_gt_list = []
i = 0
def closure():
global i, out_avg, psrn_noisy_last, last_net, net_input, psnr_gt_list
if INPUT == 'noise':
if reg_noise_std > 0:
net_input = net_input_saved + (noise.normal_() * reg_noise_std)
else:
net_input = net_input_saved
elif INPUT == 'fourier':
# net_input = net_input_saved[:, indices, :, :]
net_input = net_input_saved
elif INPUT == 'infer_freqs':
if reg_noise_std > 0:
net_input_ = net_input_saved + (noise.normal_() * reg_noise_std)
else:
net_input_ = net_input_saved
net_input = generate_fourier_feature_maps(net_input_, (img_pil.size[1], img_pil.size[0]), dtype)
else:
net_input = net_input_saved
out = net(net_input)
# Smoothing
if out_avg is None:
out_avg = out.detach()
else:
out_avg = out_avg * exp_weight + out.detach() * (1 - exp_weight)
total_loss = mse(out, img_noisy_torch)
total_loss.backward()
psrn_noisy = compare_psnr(img_noisy_np, out.detach().cpu().numpy()[0])
psrn_gt = compare_psnr(img_np, out.detach().cpu().numpy()[0])
psrn_gt_sm = compare_psnr(img_np, out_avg.detach().cpu().numpy()[0])
# Note that we do not have GT for the "snail" example
# So 'PSRN_gt', 'PSNR_gt_sm' make no sense
if PLOT and i % show_every == 0:
print('Iteration %05d Loss %f PSNR_noisy: %f PSRN_gt: %f PSNR_gt_sm: %f' % (
i, total_loss.item(), psrn_noisy, psrn_gt, psrn_gt_sm))
psnr_gt_list.append(psrn_gt)
# if train_input:
# log_inputs(net_input)
# wandb.log({'psnr_gt': psrn_gt, 'psnr_noisy': psrn_noisy}, commit=False)
# Backtracking
if i % show_every:
if psrn_noisy - psrn_noisy_last < -5:
print('Falling back to previous checkpoint.')
for new_param, net_param in zip(last_net, net.parameters()):
net_param.data.copy_(new_param.cuda())
return total_loss * 0
else:
last_net = [x.detach().cpu() for x in net.parameters()]
psrn_noisy_last = psrn_noisy
i += 1
# wandb.log({'training loss': total_loss.item()}, commit=True)
return total_loss
log_config = {
"learning_rate": LR,
"epochs": num_iter,
'optimizer': OPTIMIZER,
'loss': type(mse).__name__,
'input depth': input_depth,
'input type': INPUT,
'Train input': train_input
}
log_config.update(**freq_dict)
filename = os.path.basename(fname).split('.')[0]
# run = wandb.init(project="Fourier features DIP",
# entity="impliciteam",
# tags=['{}'.format(INPUT), 'depth:{}'.format(input_depth), filename],
# name='{}_depth_{}_{}'.format(filename, input_depth, '{}'.format(INPUT)),
# job_type='train',
# group='Denoising',
# mode='online',
# save_code=True,
# config=log_config,
# notes='Input type {} - {} random projected to depth {}'.format(
# INPUT, freq_dict['n_freqs'], input_depth))
# wandb.run.log_code(".")
p = get_params(OPT_OVER, net, net_input)
optimize(OPTIMIZER, p, closure, LR, num_iter)
if INPUT in ['fourier']:
# net_input = net_input_saved[:, indices, :, :]
net_input = net_input_saved
elif INPUT == 'infer_freqs':
net_input = generate_fourier_feature_maps(net_input_saved, (img_pil.size[1], img_pil.size[0]), dtype)
else:
net_input = net_input_saved
out_np = torch_to_np(net(net_input))
# log_images(np.array([np.clip(out_np, 0, 1), img_np]), num_iter, task='Denoising')
q = plot_image_grid([np.clip(out_np, 0, 1), img_np], factor=13)
plt.plot(psnr_gt_list)
plt.title('max: {}\nlast: {}'.format(max(psnr_gt_list), psnr_gt_list[-1]))
plt.show()