forked from DmitryUlyanov/deep-image-prior
-
Notifications
You must be signed in to change notification settings - Fork 0
/
spatial_sr_video.py
310 lines (263 loc) · 11.1 KB
/
spatial_sr_video.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
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
from __future__ import print_function
from video_consistency_check import SSIM3D
from models import *
from models.skip_3d import skip_3d, skip_3d_mlp
from utils.sr_utils import *
from utils.wandb_utils import *
from utils.video_utils import VideoDataset, DownsamplingSequence
from utils.common_utils import np_cvt_color
from utils.preemption import CHECKPOINT_NAME, resume_run, graceful_exit_handler
import random
import signal
import torch.optim
import os
import wandb
import argparse
import numpy as np
import tqdm
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
random.seed(seed)
np.random.seed(seed)
torch.random.manual_seed(seed)
parser = argparse.ArgumentParser()
parser.add_argument('--gpu', default='0')
parser.add_argument('--input_vid_path', default='', type=str, required=True)
parser.add_argument('--input_index', default=0, type=int)
parser.add_argument('--learning_rate', default=0.01, type=float)
parser.add_argument('--num_freqs', default=8, type=int)
parser.add_argument('--batch_size', default=6, type=int)
args = parser.parse_args()
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu
imsize = -1
PLOT = True
sigma = 25
mode = ['2d', '3d'][0]
signal.signal(signal.SIGTERM, graceful_exit_handler)
def eval_video(val_dataset, model, epoch):
spatial_size = vid_dataset.get_cropped_video_dims()
img_for_video = np.zeros((val_dataset.n_frames, 3, *spatial_size), dtype=np.uint8)
img_for_psnr = np.zeros((val_dataset.n_frames, 3, *spatial_size), dtype=np.float32)
ssim_loss = SSIM3D(window_size=11)
val_dataset.init_batch_list()
with torch.no_grad():
while True:
batch_data = val_dataset.next_batch()
if batch_data is None:
break
batch_data = val_dataset.prepare_batch(batch_data)
net_out = model(batch_data['input_batch'])
if mode == '3d':
out = net_out.squeeze(0).transpose(0, 1)
else:
out = net_out # N x 3 x H x W
out_np = out.detach().cpu().numpy()
img_for_psnr[batch_data['cur_batch']] = out_np
out_rgb = np.array([np_cvt_color(o) for o in out_np])
img_for_video[batch_data['cur_batch']] = (out_rgb * 255).astype(np.uint8)
ignore_start_ind = vid_dataset_eval.n_batches * vid_dataset_eval.batch_size
psnr_whole_video = compare_psnr(val_dataset.get_all_gt(numpy=True)[:ignore_start_ind],
img_for_psnr[:ignore_start_ind])
ssim_whole_video = ssim_loss(
val_dataset.get_all_gt(numpy=False)[2:ignore_start_ind].permute(1, 0, 2, 3).unsqueeze(0),
torch.from_numpy(img_for_psnr[2:ignore_start_ind]).permute(1, 0, 2, 3).unsqueeze(0))
wandb.log({'Checkpoint (FPS=10)'.format(epoch): wandb.Video(img_for_video, fps=10, format='mp4'),
'Checkpoint (FPS=25)'.format(epoch): wandb.Video(img_for_video, fps=25, format='mp4'),
'Video PSNR': psnr_whole_video,
'Video 3D-SSIM': ssim_whole_video},
commit=True)
torch.save({
'epoch': epoch,
'model_state_dict': net.state_dict(),
'optimizer_state_dict': optimizer.state_dict(),
}, 'spatial_sr_checkpoint_{}.pth'.format(epoch))
INPUT = ['noise', 'fourier', 'meshgrid', 'infer_freqs'][args.input_index]
spatial_factor = 4
temporal_factor = 1
vid_dataset = VideoDataset(args.input_vid_path,
input_type=INPUT,
num_freqs=args.num_freqs,
task='spatial_sr',
sigma=sigma,
crop_shape=None,
batch_size=args.batch_size,
arch_mode=mode,
train=True,
temp_stride=temporal_factor,
spatial_factor=spatial_factor,
mode='cont')
vid_dataset_eval = VideoDataset(args.input_vid_path,
input_type=INPUT,
num_freqs=args.num_freqs,
task='spatial_sr',
crop_shape=None,
batch_size=args.batch_size,
arch_mode=mode,
train=False,
temp_stride=temporal_factor,
spatial_factor=spatial_factor,
mode='cont')
pad = 'reflection'
if INPUT == 'infer_freqs':
OPT_OVER = 'net,input'
else:
OPT_OVER = 'net'
train_input = True if ',' in OPT_OVER else False
reg_noise_std = 0 # 1. / 30. # set to 1./20. for sigma=50
LR = args.learning_rate
OPTIMIZER = 'adam' # 'LBFGS'
exp_weight = 0.99
if mode == '2d':
show_every = 300
n_epochs = 10000
num_iter = 1
figsize = 4
if INPUT == 'noise':
input_depth = vid_dataset.input_depth
net = skip_3d(input_depth, 3,
num_channels_down=[16, 32, 64, 128, 128, 128],
num_channels_up=[16, 32, 64, 128, 128, 128],
num_channels_skip=[4, 4, 4, 4, 4, 4],
filter_size_up=(3, 3, 3),
filter_size_down=(3, 3, 3),
filter_size_skip=(1, 1, 1),
downsample_mode='stride',
need1x1_up=True, need_sigmoid=True, need_bias=True, pad='reflection',
act_fun='LeakyReLU').type(dtype)
else:
input_depth = args.num_freqs * 6 # 4 * F for spatial encoding, 4 * F for temporal encoding
if mode == '3d':
net = skip_3d_mlp(input_depth, 3,
num_channels_down=[256, 256, 256, 256, 256, 256],
num_channels_up=[256, 256, 256, 256, 256, 256],
num_channels_skip=[8, 8, 8, 8, 8, 8],
filter_size_up=(1, 1, 1),
filter_size_down=(1, 1, 1),
filter_size_skip=(1, 1, 1),
downsample_mode='stride',
need1x1_up=True, need_sigmoid=True, need_bias=True, pad='reflection',
act_fun='LeakyReLU').type(dtype)
else:
net = skip(input_depth, 3,
num_channels_down=[256, 256, 256, 256, 256, 256],
num_channels_up=[256, 256, 256, 256, 256, 256],
num_channels_skip=[8, 8, 8, 8, 8, 8],
filter_size_up=1,
filter_size_down=1,
filter_skip_size=1,
upsample_mode='bilinear',
downsample_mode='stride',
need1x1_up=True, need_sigmoid=True, need_bias=True, pad='reflection',
act_fun='LeakyReLU').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)
last_net = None
psrn_noisy_last = 0
psnr_gt_list = []
best_psnr_gt = -1.0
best_iter = 0
best_img = None
i = 0
def train_batch(batch_data):
global j
net_input_saved = batch_data['input_batch']
# noise = net_input_saved.detach().clone()
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
net_out = net(net_input)
if mode == '3d':
out = net_out.squeeze(0).transpose(0, 1) # N x 3 x H x W
else:
out = net_out
out_lr = downsampler.downsmaple_sequence(out)
total_loss = mse(out_lr, batch_data['img_degraded_batch'])
# total_loss = total_loss / accum_iter
total_loss.backward()
out_lr_np = out_lr.detach().cpu().numpy()
out_hr_np = out.detach().cpu().numpy()
psrn_noisy = compare_psnr(batch_data['img_degraded_batch'].cpu().numpy(), out_lr_np)
psrn_gt = compare_psnr(batch_data['gt_batch'].numpy(), out_hr_np)
wandb.log({'batch loss': total_loss.item(), 'psnr_noisy': psrn_noisy, 'psnr_gt': psrn_gt}, commit=True)
return total_loss, out_hr_np, psrn_gt
p = get_params(OPT_OVER, net, net_input=vid_dataset.input)
optimizer = torch.optim.Adam(p, lr=LR)
log_config = {
"learning_rate": LR,
"iteration per batch": num_iter,
'Epochs': n_epochs,
'optimizer': OPTIMIZER,
'loss': type(mse).__name__,
'input depth': input_depth,
'input type': INPUT,
'Train input': train_input,
'Reg. Noise STD': reg_noise_std,
'Sequence length': vid_dataset.batch_size,
'Video length': vid_dataset.n_frames,
'# of sequences': vid_dataset.n_batches,
'save every': show_every
}
log_config.update(**vid_dataset.freq_dict)
filename = os.path.basename(args.input_vid_path).split('.')[0]
if os.path.isfile(CHECKPOINT_NAME):
net, optimizer, start_epoch, wandb_id = resume_run(net, optimizer)
else:
start_epoch = 0
wandb_id = wandb.util.generate_id()
run = wandb.init(project="Fourier features DIP",
entity="impliciteam",
tags=['{}'.format(INPUT), 'depth:{}'.format(input_depth), filename, vid_dataset.freq_dict['method'],
'{}-PIP'.format(mode)],
name='{}_depth_{}_{}_{}_spatial_factor_{}_temporal_factor_{}'.format(
filename, input_depth, '{}'.format(INPUT), mode, spatial_factor, temporal_factor),
job_type='{}_{}'.format(INPUT, LR),
group='Spatial SR - Video',
mode='online',
save_code=True,
id=wandb_id,
resume="allow",
config=log_config,
notes=''
)
log_input_video(vid_dataset.get_all_gt(numpy=True),
vid_dataset.get_all_degraded(numpy=True))
wandb.run.log_code(".", exclude_fn=lambda path: path.find('venv') != -1)
print(net)
n_batches = vid_dataset.n_batches
img_idx = []
downsampler = DownsamplingSequence(factor=spatial_factor)
downsampler.set_dtype(dtype)
for epoch in tqdm.tqdm(range(start_epoch, n_epochs), desc='Epoch', position=0):
running_psnr = 0.
running_loss = 0.
vid_dataset.init_batch_list()
for batch_cnt in tqdm.tqdm(range(n_batches), desc="Batch", position=1, leave=False):
batch_data = vid_dataset.next_batch()
batch_data = vid_dataset.prepare_batch(batch_data)
for j in range(num_iter):
optimizer.zero_grad()
loss, out_sequence, psnr_gt = train_batch(batch_data)
running_psnr += psnr_gt
running_loss += loss.item()
optimizer.step()
denom = n_batches
# Log metrics for each epoch
wandb.log({'epoch loss': running_loss / denom, 'epoch psnr': running_psnr / denom}, commit=False)
# log_images(np.array([np_cvt_color(o) for o in out_sequence]), epoch, 'Video-Denoising',
# commit=False)
# Infer video:
if epoch % show_every == 0:
eval_video(vid_dataset_eval, net, epoch)
# Infer video at the end:
eval_video(vid_dataset_eval, net, epoch)