/
deep_decoder.py
226 lines (195 loc) · 7.42 KB
/
deep_decoder.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
import argparse
import os
import shutil
import warnings
import numpy as np
import torch
import torch.nn.functional as F
from torch.utils.tensorboard import SummaryWriter
from tqdm import tqdm, trange
from forward_model import GaussianCompressiveSensing, NoOp
from model.deep_decoder import DeepDecoder
from utils import dict_to_str, load_target_image, psnr_from_mse
warnings.filterwarnings("ignore")
def _deep_decoder_recover(
x,
forward_model,
optimizer,
num_filters,
depth,
lr,
img_size,
steps,
run_dir,
run_name,
disable_tqdm,
**kwargs,
):
# Keep batch_size = 1
batch_size = 1
if (isinstance(forward_model, GaussianCompressiveSensing)):
n_pixel_bora = 64 * 64 * 3
n_pixel = np.prod(x.shape)
noise = torch.randn(batch_size,
forward_model.n_measure,
device=x.device)
noise *= 0.1 * torch.sqrt(
torch.tensor(n_pixel / forward_model.n_measure / n_pixel_bora))
# z is a fixed latent vector
start_imsize = int(np.log2(img_size)) - depth + 1
z = torch.randn(batch_size,
num_filters,
start_imsize,
start_imsize,
device=x.device)
# make a fresh DD model for every run
model = DeepDecoder(num_filters=num_filters,
img_size=img_size,
depth=depth).to(x.device)
if optimizer == 'adam':
optimizer = torch.optim.Adam(model.parameters(), lr=lr)
save_img_every_n = 50
elif optimizer == 'lbfgs':
optimizer = torch.optim.LBFGS(model.parameters(), lr=lr)
save_img_every_n = 2
else:
raise NotImplementedError()
if run_name is not None:
logdir = os.path.join('recovery_tensorboard_logs', run_dir, run_name)
if os.path.exists(logdir):
print("Overwriting pre-existing logs!")
shutil.rmtree(logdir)
writer = SummaryWriter(logdir)
else:
writer = None
# Save original and distorted image
if run_name is not None:
writer.add_image("Original/Clamp", x.clamp(0, 1))
if forward_model.viewable:
writer.add_image(
"Distorted/Clamp",
forward_model(x.unsqueeze(0).clamp(0, 1)).squeeze(0))
# Make noisy gaussian measurements
x = x.expand(batch_size, *x.shape)
y_observed = forward_model(x)
if (isinstance(forward_model, GaussianCompressiveSensing)):
y_observed += noise
def closure():
optimizer.zero_grad()
x_hat = model.forward(z)
loss = F.mse_loss(forward_model(x_hat), y_observed)
loss.backward()
return loss
for j in trange(steps, desc='Fit', leave=False):
optimizer.step(closure)
with torch.no_grad():
x_hat = model.forward(z)
train_mse_clamped = F.mse_loss(
forward_model(x_hat.detach().clamp(0, 1)), y_observed)
if writer is not None:
writer.add_scalar('TRAIN_MSE', train_mse_clamped, j + 1)
writer.add_scalar('TRAIN_PSNR', psnr_from_mse(train_mse_clamped),
j + 1)
orig_mse_clamped = F.mse_loss(x_hat.detach().clamp(0, 1), x)
writer.add_scalar('ORIG_MSE', orig_mse_clamped, j + 1)
writer.add_scalar('ORIG_PSNR', psnr_from_mse(orig_mse_clamped),
j + 1)
if j % save_img_every_n == 0:
writer.add_image('Recovered',
x_hat.squeeze().clamp(0, 1), j + 1)
if writer is not None:
writer.add_image('Final', x_hat.squeeze().clamp(0, 1))
return x_hat.squeeze(), forward_model(x).squeeze(), train_mse_clamped
def deep_decoder_recover(
x,
forward_model,
optimizer='lbfgs',
num_filters=64,
depth=6, # TODO
lr=1,
img_size=64,
steps=50,
restarts=1,
run_dir=None,
run_name=None,
disable_tqdm=False,
**kwargs):
best_psnr = -float("inf")
best_return_val = None
for i in trange(restarts,
desc='Restarts',
leave=False,
disable=disable_tqdm):
if run_name is not None:
current_run_name = f'{run_name}_{i}'
else:
current_run_name = None
return_val = _deep_decoder_recover(x=x,
forward_model=forward_model,
optimizer=optimizer,
num_filters=num_filters,
depth=depth,
lr=lr,
img_size=img_size,
steps=steps,
run_dir=run_dir,
run_name=current_run_name,
disable_tqdm=disable_tqdm,
**kwargs)
p = psnr_from_mse(return_val[2])
if p > best_psnr:
best_psnr = p
best_return_val = return_val
return best_return_val
if __name__ == '__main__':
DEVICE = 'cuda:0' if torch.cuda.is_available() else 'cpu'
a = argparse.ArgumentParser()
a.add_argument('--img_dir', required=True)
a.add_argument('--disable_tqdm', default=False)
a.add_argument('--run_name_suffix', default='')
args = a.parse_args()
params_64 = {
'depth': 5,
'num_filters': 250,
'lr': 1e-2,
'steps': 5000,
'restarts': 1,
'optimizer': 'adam'
}
params_128 = {
'depth': 6,
'num_filters': 700,
'lr': 1e-2,
'steps': 5000,
'restarts': 1,
'optimizer': 'adam'
}
for img_size, n_measures, params in tqdm([(64, [600, 2000], params_64),
(128, [2400, 8000], params_128)],
desc='ImgSizes',
leave=True,
disable=args.disable_tqdm):
for n_measure in tqdm(n_measures,
desc='N_measures',
leave=False,
disable=args.disable_tqdm):
img_shape = (3, img_size, img_size)
forward_model = GaussianCompressiveSensing(n_measure=n_measure,
img_shape=img_shape)
# forward_model = NoOp()
for img_name in tqdm(os.listdir(args.img_dir),
desc='Images',
leave=False,
disable=args.disable_tqdm):
orig_img = load_target_image(
os.path.join(args.img_dir, img_name), img_size).to(DEVICE)
img_basename, _ = os.path.splitext(img_name)
x_hat, x_degraded, _ = deep_decoder_recover(
orig_img,
forward_model,
run_dir='deep_decoder',
run_name=(img_basename + args.run_name_suffix + '.' +
dict_to_str(params) +
f'.n_measure={n_measure}.img_size={img_size}'),
img_size=img_size,
**params)