-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
364 lines (296 loc) · 10.7 KB
/
train.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
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from torch.utils.data import Dataset, DataLoader
from tensorboardX import SummaryWriter
import torchvision
from networks import encoder as ENC
from networks import decoder as DEC
from networks import layers as LYR
from util.upb_dataset import *
from util.vis import *
from util.warp import *
from flow_rigid import *
import cv2
import argparse
import os
# define parser
parser = argparse.ArgumentParser()
parser.add_argument('--batch_size', type=int, default=12)
parser.add_argument('--num_layers', type=int, default=18)
parser.add_argument('--num_input_images', type=int, default=2)
parser.add_argument('--num_output_channels', type=int, default=2)
parser.add_argument('--num_vis', type=int, default=4)
parser.add_argument('--scales', type=list, default=[0, 1, 2, 3])
parser.add_argument('--height', type=int, default=256)
parser.add_argument('--width', type=int, default=512)
parser.add_argument('--lr', type=float, default=1e-4)
parser.add_argument('--log_int', type=int, default=100)
parser.add_argument('--vis_int', type=int, default=1000)
parser.add_argument('--num_epochs', type=int, default=20)
parser.add_argument('--scheduler_step_size', type=int, default=15)
parser.add_argument('--log_dir', type=str, default='./logs')
parser.add_argument('--checkpoint_dir', type=str, default='./snapshots')
parser.add_argument('--model_name', type=str, default='default')
parser.add_argument('--load_checkpoint', action='store_true')
parser.add_argument('--dataset', type=str, help="name of the dataset")
args = parser.parse_args()
# create directories
if not os.path.exists(args.log_dir):
os.mkdir(args.log_dir)
if not os.path.exists(args.checkpoint_dir):
os.mkdir(args.checkpoint_dir)
os.makedirs(os.path.join(args.checkpoint_dir, "imgs"))
os.makedirs(os.path.join(args.checkpoint_dir, "checkpoints"))
# define summary writer
writer = SummaryWriter(log_dir=args.log_dir)
# define device
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
# define encoder
encoder = ENC.ResnetEncoder(
num_layers=args.num_layers,
pretrained=True,
num_input_images=args.num_input_images
)
encoder.encoder.conv1 = nn.Conv2d(12, 64, kernel_size=7, stride=2, padding=3, bias=False)
encoder = encoder.to(device)
# define decoder
decoder = DEC.FlowDecoder(
num_ch_enc=encoder.num_ch_enc,
scales=args.scales,
num_output_channels=args.num_output_channels,
use_skips=True
).to(device)
# define masking network
encoder_mask = ENC.ResnetEncoder(
num_layers=args.num_layers,
pretrained=True,
num_input_images=args.num_input_images
)
encoder_mask.encoder.conv1 = nn.Conv2d(12, 64, kernel_size=7, stride=2, padding=3, bias=False)
encoder_mask.to(device)
decoder_mask = DEC.FlowDecoder(
num_ch_enc=encoder_mask.num_ch_enc,
scales=args.scales,
num_output_channels=1,
use_skips=True
).to(device)
# define ssim
ssim = LYR.SSIM().to(device)
# define rigid flow
rigid_flow = RigidFlow()
# define optimizer
params = list(encoder.parameters())
params += list(decoder.parameters())
params += list(encoder_mask.parameters())
params += list(decoder_mask.parameters())
optimizer = optim.Adam(params, args.lr)
scheduler = optim.lr_scheduler.StepLR(optimizer, args.scheduler_step_size, 0.1)
# define dataloader
with open(os.path.join("splits", args.dataset, "train_files.txt")) as fin:
train_filenames = fin.readlines()
with open(os.path.join("splits", args.dataset, "test_files.txt")) as fin:
test_filenames = fin.readlines()
train_dataset = UPBRAWDataset(
data_path="./dataset",
filenames=train_filenames,
height=args.height,
width=args.width,
frame_idxs=[-1, 0],
num_scales=4,
is_train=True,
img_ext="png"
)
test_dataset = UPBRAWDataset(
data_path="./dataset",
filenames = test_filenames ,
height=args.height,
width=args.width,
frame_idxs=[-1, 0],
num_scales=4,
is_train=False,
img_ext="png"
)
train_dataloader = DataLoader(
train_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=4,
drop_last=True,
pin_memory=True
)
test_dataloader = DataLoader(
test_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=1,
drop_last=True,
pin_memory=True
)
test_iter = iter(test_dataloader)
def save_checkpoint(epoch: int, rloss: float):
state = {
'epoch': epoch,
'encoder': encoder.state_dict(),
'decoder': decoder.state_dict(),
'encoder_mask': encoder_mask.state_dict(),
'decoder_mask': decoder_mask.state_dict(),
'optimizer': optimizer.state_dict(),
'scheduler': scheduler,
'rloss': rloss
}
path = os.path.join(args.checkpoint_dir, 'checkpoints', args.model_name + ("_%d.pth" % (epoch)))
torch.save(state, path)
def load_checkpoint():
path = os.path.join(args.checkpoint_dir, 'checkpoints', args.model_name)
state = torch.load(path)
encoder.load_state_dict(state['encoder'])
decoder.load_state_dict(state['decoder'])
encoder_mask.load_state_dict(state['encoder_mask'])
decoder_mask.load_state_dict(state['decoder_mask'])
optimizer.load_state_dict(state['optimizer'])
scheduler = scheduler
return state['epoch'], state['rloss']
def get_rigid_flow(img1: torch.tensor, img2: torch.tensor):
B, W, H = img1.shape[0], rigid_flow.WIDTH, rigid_flow.HEIGHT
img1 = F.interpolate(img1, (H, W))
img2 = F.interpolate(img2, (H, W))
# get pix coords and then flow
pix_coords = rigid_flow.get_pix_coords(img1, img2, B)
rflow = rigid_flow.get_flow(pix_coords, B)
rflow = rflow.transpose(2, 3).transpose(1, 2)
rflow = F.interpolate(rflow, (args.height, args.width))
return rflow.float()
def test_sample():
global test_iter
encoder.eval()
encoder.eval()
try:
test_batch = next(test_iter)
except StopIteration:
test_iter = iter(test_dataloader)
test_batch = next(test_iter)
imgs1 = [data[('color_aug', -1, i)].to(device) for i in range(4)]
imgs2 = [data[('color_aug', 0, i)].to(device) for i in range(4)]
rflow = get_rigid_flow(imgs1[0], imgs2[0])
# compute warped image using rigid flow
wimg2_r = warp(imgs1[0], rflow)
input = torch.cat((imgs1[0], imgs2[0], wimg2_r), dim=1)
# compute reprojection loss
# for the warped image with rigid flow
ssim_loss = ssim(wimg2_r, imgs2[0]).mean(1, True)
l1_loss = torch.abs(wimg2_r - imgs2[0]).mean(1, True)
reprojection_loss = 0.85 * ssim_loss + 0.15 * l1_loss
with torch.no_grad():
enc_output = encoder(input, rflow, reprojection_loss)
dec_output = decoder(enc_output)
# compute warped image using rigid and dynamic flow
dflow = dec_output[('flow', 0)]
flow = dflow + rflow
wimg2_dr = warp(imgs1[0], flow)
# compute reprojection loss
# for the warped image using the rigid and dynamic flow
ssim_loss = ssim(wimg2_dr, imgs2[0]).mean(1, True)
l1_loss = torch.abs(wimg2_dr - imgs2[0]).mean(1, True)
reprojection_loss = 0.85 * ssim_loss + 0.15 * l1_loss
# compute mask
with torch.no_grad():
input = torch.cat([imgs1[0], imgs2[0], wimg2_dr], dim=1)
enc_mask_output = encoder_mask(input, flow, reprojection_loss)
dec_mask_output = decoder_mask(enc_mask_output)
mask = torch.sigmoid(dec_mask_output[('flow', 0)])
mask = mask.repeat(1, 3, 1, 1)
# color flow
flow = flow.cpu()
colors = []
for j in range(args.batch_size):
color_flow = flow[j].numpy().transpose(1, 2, 0)
color_flow = flow_to_color(color_flow).transpose(2, 0, 1)
color_flow = torch.tensor(color_flow).unsqueeze(0).float() / 255
colors.append(color_flow)
colors = torch.cat(colors, dim=0)
img1 = imgs1[0][:args.num_vis].cpu()
img2 = imgs2[0][:args.num_vis].cpu()
wimg2_dr = wimg2_dr[:args.num_vis].cpu()
mask = mask[:args.num_vis].cpu()
colors = colors[:args.num_vis]
imgs = torch.cat([img1, img2, mask * wimg2_dr, 0.5 * (wimg2_dr + img2), 0.5 * (img1 + img2), colors, mask], dim=3)
imgs = torchvision.utils.make_grid(imgs, nrow=1, normalize=False)
imgs = (255 * imgs.numpy().transpose(1, 2, 0)).astype(np.uint8)
cv2.imwrite("./snapshots/imgs/%d.%d.png" % (epoch, i), imgs[..., ::-1])
encoder.train()
decoder.train()
if __name__ == "__main__":
rloss = None
start_epoch = 0
if args.load_checkpoint:
start_epoch, rloss = load_checkpoint()
for epoch in range(start_epoch, args.num_epochs):
for i, data in enumerate(train_dataloader):
# zero grad
optimizer.zero_grad()
# extract data
imgs1 = [data[('color_aug', -1, i)].to(device) for i in range(4)]
imgs2 = [data[('color_aug', 0, i)].to(device) for i in range(4)]
rflow = get_rigid_flow(imgs1[0], imgs2[0])
# compute warped imaged using rigid flow
wimg2_r = warp(imgs1[0], rflow)
input = torch.cat((imgs1[0], imgs2[0], wimg2_r), dim=1)
# compute reprojection loss
# for the warped image with rigid flow
ssim_loss = ssim(wimg2_r, imgs2[0]).mean(1, True)
l1_loss = torch.abs(wimg2_r - imgs2[0]).mean(1, True)
reprojection_loss = 0.85 * ssim_loss + 0.15 * l1_loss
# compute dynamic flow
enc_output = encoder(input, rflow, reprojection_loss)
dec_output = decoder(enc_output)
loss = 0
for j in args.scales:
img1 = imgs1[j]
img2 = imgs2[j]
# compute warped image using rigid + dynamic flow
scaled_dflow = dec_output[('flow', j)]
scaled_rflow = F.interpolate(rflow, (args.height//2**j, args.width//2**j))
scaled_flow = scaled_dflow + scaled_rflow
wimg2_dr = warp(img1, scaled_flow)
# compute reprojection loss
# for the warped image using the rigid and dynamic flow
ssim_loss = ssim(wimg2_dr, img2).mean(1, True)
l1_loss = torch.abs(wimg2_dr - img2).mean(1, True)
reprojection_loss = 0.85 * ssim_loss + 0.15 * l1_loss
# smooth loss
smooth_loss = get_smooth_loss(scaled_dflow[:, :1, :, :], img2)
smooth_loss += get_smooth_loss(scaled_dflow[:, 1:, :, :], img2)
# compute masks
if j == 0:
input = torch.cat((img1, img2, wimg2_dr), dim=1)
encoder_mask_output = encoder_mask(input, scaled_flow, reprojection_loss)
decoder_mask_output = decoder_mask(encoder_mask_output)
# compute maks loss
mask = decoder_mask_output[('flow', j)]
mask = torch.sigmoid(mask)
weighting_loss = nn.BCELoss()(mask, torch.ones(mask.shape).cuda())
# mask reprojection loss
reprojection_loss = mask * reprojection_loss
# total loss
loss += (reprojection_loss.mean() + 0.2 * weighting_loss + 0.01 * smooth_loss) / 2**j
# backward step
loss.backward()
optimizer.step()
# compute running loss
rloss = loss.item() if rloss is None else 0.99 * rloss + 0.01 * loss.item()
# log interval
if i % args.log_int == 0:
it = epoch * (len(train_dataset) // args.batch_size) + i
writer.add_scalar("Loss", rloss, it)
print("Epoch: %d, Batch: %d, Loss: %.4f" % (epoch, i, rloss))
# visualization interval
if i % args.vis_int == 0:
test_sample()
# scheduler step
scheduler.step()
# save model
save_checkpoint(epoch, rloss)
# export scalar data to JSON for external processing
writer.close()