-
Notifications
You must be signed in to change notification settings - Fork 4
/
finetune.py
228 lines (161 loc) · 8.56 KB
/
finetune.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 os
import torch
import torch.nn as nn
from torch.utils import data
from datetime import datetime
import argparse
from dataset.data import VideoDataset
from dataset.transforms import get_train_transforms
from torch.utils.data import DataLoader
from model.model import VideoEncoder, VideoDecoder
import torch.optim as optim
from utils import adjust_lr, label_edge_prediction, visualize_prediction
from datetime import datetime
import smoothness
from torch.nn import functional as F
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
parser = argparse.ArgumentParser()
parser.add_argument('--epoch', type=int, default=20, help='epoch number')
parser.add_argument('--lr', type=float, default=1e-4, help='learning rate')
parser.add_argument('--batchsize', type=int, default=1, help='training batch size')
parser.add_argument('--size', type=int, default=256, help='training dataset size')
parser.add_argument('--clip', type=float, default=0.5, help='gradient clipping margin')
parser.add_argument('--decay_rate', type=float, default=0.9, help='decay rate of learning rate')
parser.add_argument('--decay_epoch', type=int, default=15, help='every n epochs decay learning rate')
parser.add_argument('--sm_loss_weight', type=float, default=0.3, help='weight for smoothness loss')
parser.add_argument('--edge_loss_weight', type=float, default=1.0, help='weight for edge loss')
args = parser.parse_args()
transforms = get_train_transforms(input_size=(args.size, args.size))
dataset = VideoDataset(root_dir="the dir of your dataset", trainingset_list=["DAVIS", "DAVSOD"], trainging=True, transforms=transforms)
train_dataloader = DataLoader(dataset=dataset, batch_size=args.batchsize, num_workers=6, shuffle=True, drop_last=True)
Encoder = VideoEncoder(output_stride=16, pretrained=True)
Decoder = VideoDecoder()
Encoder.load_state_dict(torch.load("./save_models/pretrained/scribble_Encoder_30.pth"))
Decoder_dict = Decoder.state_dict()
pretrained_dict = torch.load("./save_models/pretrained/scribble_Decoder_30.pth")
for k, v in pretrained_dict.items():
if (k in Decoder_dict):
print("load:%s" % k)
pretrained_dict = {k: v for k, v in pretrained_dict.items() if (k in Decoder_dict)}
Decoder_dict.update(pretrained_dict)
Decoder.load_state_dict(Decoder_dict)
Encoder = Encoder.cuda()
Decoder = Decoder.cuda()
print("network ready!!!")
optimizer_Encoder = optim.Adam(Encoder.parameters(), lr=args.lr)
optimizer_Decoder = optim.Adam(Decoder.parameters(), lr=args.lr)
CE = torch.nn.BCELoss()
smooth_loss = smoothness.smoothness_loss(size_average=True)
def get_one_hot(mask, gt):
a = torch.ones([1, 2, 4096]).cuda()
a[:, 0, mask.view(4096) == 0] = 0
a[:, 1, mask.view(4096) == 0] = 0
a[:, 0, gt.view(4096) == 1] = 1
a[:, 1, gt.view(4096) == 1] = 0
a[:, 0, (mask - gt).view(4096) == 1] = 0
a[:, 1, (mask - gt).view(4096) == 1] = 1
return a
def prior_loss(prior, masks, gts):
x1_mask = F.interpolate(masks[0], scale_factor=0.25, mode="nearest").view(1, 1, -1)
x2_mask = F.interpolate(masks[1], scale_factor=0.25, mode="nearest").view(1, 1, -1)
x3_mask = F.interpolate(masks[2], scale_factor=0.25, mode="nearest").view(1, 1, -1)
x4_mask = F.interpolate(masks[3], scale_factor=0.25, mode="nearest").view(1, 1, -1)
mask_11 = torch.matmul(x1_mask.permute(0, 2, 1), x1_mask)
mask_12 = torch.matmul(x1_mask.permute(0, 2, 1), x2_mask)
mask_13 = torch.matmul(x1_mask.permute(0, 2, 1), x3_mask)
mask_14 = torch.matmul(x1_mask.permute(0, 2, 1), x4_mask)
mask_22 = torch.matmul(x2_mask.permute(0, 2, 1), x2_mask)
mask_23 = torch.matmul(x2_mask.permute(0, 2, 1), x3_mask)
mask_24 = torch.matmul(x2_mask.permute(0, 2, 1), x4_mask)
mask_33 = torch.matmul(x3_mask.permute(0, 2, 1), x3_mask)
mask_34 = torch.matmul(x3_mask.permute(0, 2, 1), x4_mask)
mask_44 = torch.matmul(x4_mask.permute(0, 2, 1), x4_mask)
mask = torch.cat((mask_11, mask_12, mask_13, mask_14, mask_22, mask_23, mask_24, mask_33, mask_34, mask_44), dim=0)
prior = mask * prior
size = prior.shape[0] * prior.shape[1] * prior.shape[2]
ratio = size / mask.sum()
x1_gt = F.interpolate(gts[0], scale_factor=0.25, mode="nearest").view(1, 1, -1)
x1_onehot = get_one_hot(x1_mask, x1_gt)
x2_gt = F.interpolate(gts[1], scale_factor=0.25, mode="nearest").view(1, 1, -1)
x2_onehot = get_one_hot(x2_mask, x2_gt)
x3_gt = F.interpolate(gts[2], scale_factor=0.25, mode="nearest").view(1, 1, -1)
x3_onehot = get_one_hot(x3_mask, x3_gt)
x4_gt = F.interpolate(gts[3], scale_factor=0.25, mode="nearest").view(1, 1, -1)
x4_onehot = get_one_hot(x4_mask, x4_gt)
x11_sim = torch.matmul(x1_onehot.permute(0, 2, 1), x1_onehot)
x12_sim = torch.matmul(x1_onehot.permute(0, 2, 1), x2_onehot)
x13_sim = torch.matmul(x1_onehot.permute(0, 2, 1), x3_onehot)
x14_sim = torch.matmul(x1_onehot.permute(0, 2, 1), x4_onehot)
x22_sim = torch.matmul(x2_onehot.permute(0, 2, 1), x2_onehot)
x23_sim = torch.matmul(x2_onehot.permute(0, 2, 1), x3_onehot)
x24_sim = torch.matmul(x2_onehot.permute(0, 2, 1), x4_onehot)
x33_sim = torch.matmul(x3_onehot.permute(0, 2, 1), x3_onehot)
x34_sim = torch.matmul(x3_onehot.permute(0, 2, 1), x4_onehot)
x44_sim = torch.matmul(x4_onehot.permute(0, 2, 1), x4_onehot)
sim = torch.cat((x11_sim, x12_sim, x13_sim, x14_sim, x22_sim, x23_sim, x24_sim, x33_sim, x34_sim, x44_sim), dim=0)
loss = ratio * CE(prior, sim)
return loss
def train(train_loader, Encoder, Decoder, optimizer_Encoder, optimizer_Decoder, epoch):
Encoder.train()
Decoder.train()
total_step = len(train_dataloader)
for i, packs in enumerate(train_dataloader):
optimizer_Encoder.zero_grad()
optimizer_Decoder.zero_grad()
loss = 0
rgb_features = []
flow_features = []
edge_maps = []
masks = []
gts = []
for pack in packs:
images, flow = pack["image"], pack["flow"]
images, flow = images.cuda(), flow.cuda()
rgb_feature, flow_feature = Encoder(images, flow)
edge_maps.append(rgb_feature[-1])
rgb_features.append(rgb_feature)
flow_features.append(flow_feature)
masks.append(pack["mask"].cuda())
gts.append(pack["gt"].cuda())
sal1s, sal2s, pred_prior = Decoder(rgb_features, flow_features)
pri_loss = prior_loss(pred_prior, masks, gts)
loss = 0
for k in range(len(sal1s)):
sal1 = sal1s[k]
sal2 = sal2s[k]
mask = packs[k]["mask"].cuda()
gt = packs[k]["gt"].cuda()
grey = packs[k]["grey"].cuda()
edge_map = edge_maps[k]
img_size = sal1.size(2) * sal1.size(3) * sal1.size(0)
ratio = img_size / torch.sum(mask)
sal1_prob = torch.sigmoid(sal1)
sal1_prob = sal1_prob * mask
sal2_prob = torch.sigmoid(sal2)
sal2_prob = sal2_prob * mask
edges_gt = label_edge_prediction(torch.sigmoid(sal2).detach())
edge_loss = CE(torch.sigmoid(edge_map), edges_gt)
sal_loss1 = ratio * CE(sal1_prob, gt*mask) + 0.3 * smooth_loss(torch.sigmoid(sal1), grey)
sal_loss2 = ratio * CE(sal2_prob, gt*mask) + 0.3 * smooth_loss(torch.sigmoid(sal2), grey)
loss += sal_loss1 + sal_loss2 + edge_loss
loss = loss + pri_loss
loss.backward()
optimizer_Encoder.step()
optimizer_Decoder.step()
if i % 10 == 0 or i == total_step:
print('{} Epoch [{:03d}/{:03d}], Step [{:04d}/{:04d}], loss: {:0.4f}, sal1 loss: {:0.4f}, edge loss: {:0.4f}, sal2 loss: {:0.4f}, prior loss: {:0.4f}'.
format(datetime.now(), epoch, args.epoch, i, total_step, loss.data, sal_loss1.data, edge_loss.data, sal_loss2.data, pri_loss.data))
visualize_prediction(torch.sigmoid(sal1), './show/', "sal1")
visualize_prediction(torch.sigmoid(edge_map), './show/', "edge")
visualize_prediction(torch.sigmoid(sal2), './show/', "sal2")
save_path = 'save_models/finetune/'
if not os.path.exists(save_path):
os.makedirs(save_path)
if epoch % 1 == 0:
torch.save(Encoder.state_dict(), save_path + 'scribble_Encoder' + '_%d' % epoch + '.pth')
torch.save(Decoder.state_dict(), save_path + 'scribble_Decoder' + '_%d' % epoch + '.pth')
print("start training!!!")
for epoch in range(1, args.epoch+1):
adjust_lr(optimizer_Encoder, epoch, args.decay_rate, args.decay_epoch)
adjust_lr(optimizer_Decoder, epoch, args.decay_rate, args.decay_epoch)
train(train_dataloader, Encoder, Decoder, optimizer_Encoder, optimizer_Decoder, epoch)