-
Notifications
You must be signed in to change notification settings - Fork 1
/
model.py
324 lines (275 loc) · 11.9 KB
/
model.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
import torch
import torch.nn as nn
import torch.nn.functional as F
import torchvision.models as models
class SCNN(nn.Module):
def __init__(
self,
input_size,
ms_ks=9,
pretrained=True
):
"""
Argument
ms_ks: kernel size in message passing conv
"""
super(SCNN, self).__init__()
self.pretrained = pretrained
self.net_init(input_size, ms_ks)
if not pretrained:
self.weight_init()
self.scale_background = 0.4
self.scale_seg = 1.0
self.scale_exist = 0.1
self.ce_loss = nn.CrossEntropyLoss(weight=torch.tensor([self.scale_background, 1, 1, 1, 1]))
self.bce_loss = nn.BCELoss()
def forward(self, img, seg_gt=None, exist_gt=None, target=None):#
'''
seg_gt = img['segLabel']
target = img['target']
img = img['img']'''
x = self.backbone(img)
x = self.layer1(x)
x = self.message_passing_forward(x)
x = self.layer2(x)
seg_pred = F.interpolate(x, scale_factor=8, mode='bilinear', align_corners=True)
x = self.layer3(x)
x = x.view(-1, self.fc_input_feature)
# exist_pred = self.fc(x)
# if seg_gt is not None and exist_gt is not None:
# loss_seg = self.ce_loss(seg_pred, seg_gt)
# loss_exist = self.bce_loss(exist_pred, exist_gt)
# loss = loss_seg * self.scale_seg + loss_exist * self.scale_exist
# else:
# loss_seg = torch.tensor(0, dtype=img.dtype, device=img.device)
# loss_exist = torch.tensor(0, dtype=img.dtype, device=img.device)
# loss = torch.tensor(0, dtype=img.dtype, device=img.device)
if seg_gt is not None:
loss = self.ce_loss(seg_pred, seg_gt)
exist_pred = 0
loss_exist = 0
else:
# loss_seg = torch.tensor(0, dtype=img.dtype, device=img.device)
loss = torch.tensor(0, dtype=img.dtype, device=img.device)
return seg_pred, loss
def message_passing_forward(self, x):
Vertical = [True, True, False, False]
Reverse = [False, True, False, True]
for ms_conv, v, r in zip(self.message_passing, Vertical, Reverse):
x = self.message_passing_once(x, ms_conv, v, r)
return x
def message_passing_once(self, x, conv, vertical=True, reverse=False):
"""
Argument:
----------
x: input tensor
vertical: vertical message passing or horizontal
reverse: False for up-down or left-right, True for down-up or right-left
"""
nB, C, H, W = x.shape
if vertical:
slices = [x[:, :, i:(i + 1), :] for i in range(H)]
dim = 2
else:
slices = [x[:, :, :, i:(i + 1)] for i in range(W)]
dim = 3
if reverse:
slices = slices[::-1]
out = [slices[0]]
for i in range(1, len(slices)):
out.append(slices[i] + F.relu(conv(out[i - 1])))
if reverse:
out = out[::-1]
return torch.cat(out, dim=dim)
def net_init(self, input_size, ms_ks):
input_w, input_h = input_size
self.fc_input_feature = 5 * int(input_w/16) * int(input_h/16)
self.backbone = models.vgg16_bn(pretrained=self.pretrained).features
# ----------------- process backbone -----------------
for i in [34, 37, 40]:
conv = self.backbone._modules[str(i)]
dilated_conv = nn.Conv2d(
conv.in_channels, conv.out_channels, conv.kernel_size, stride=conv.stride,
padding=tuple(p * 2 for p in conv.padding), dilation=2, bias=(conv.bias is not None)
)
dilated_conv.load_state_dict(conv.state_dict())
self.backbone._modules[str(i)] = dilated_conv
self.backbone._modules.pop('33')
self.backbone._modules.pop('43')
# ----------------- SCNN part -----------------
self.layer1 = nn.Sequential(
nn.Conv2d(512, 1024, 3, padding=4, dilation=4, bias=False),
nn.BatchNorm2d(1024),
nn.ReLU(),
nn.Conv2d(1024, 128, 1, bias=False),
nn.BatchNorm2d(128),
nn.ReLU() # (nB, 128, 36, 100)
)
# ----------------- add message passing -----------------
self.message_passing = nn.ModuleList()
self.message_passing.add_module('up_down', nn.Conv2d(128, 128, (1, ms_ks), padding=(0, ms_ks // 2), bias=False))
self.message_passing.add_module('down_up', nn.Conv2d(128, 128, (1, ms_ks), padding=(0, ms_ks // 2), bias=False))
self.message_passing.add_module('left_right',
nn.Conv2d(128, 128, (ms_ks, 1), padding=(ms_ks // 2, 0), bias=False))
self.message_passing.add_module('right_left',
nn.Conv2d(128, 128, (ms_ks, 1), padding=(ms_ks // 2, 0), bias=False))
# (nB, 128, 36, 100)
# ----------------- SCNN part -----------------
self.layer2 = nn.Sequential(
nn.Dropout2d(0.1),
nn.Conv2d(128, 5, 1) # get (nB, 5, 36, 100)
)
self.layer3 = nn.Sequential(
nn.Softmax(dim=1), # (nB, 5, 36, 100)
nn.AvgPool2d(2, 2), # (nB, 5, 18, 50)
)
self.fc = nn.Sequential(
nn.Linear(self.fc_input_feature, 128),
nn.ReLU(),
nn.Linear(128, 4),
nn.Sigmoid()
)
def weight_init(self):
for m in self.modules():
if isinstance(m, nn.Conv2d):
m.reset_parameters()
elif isinstance(m, nn.BatchNorm2d):
m.weight.data[:] = 1.
m.bias.data.zero_()
class LaneNet(nn.Module):
def __init__(
self,
embed_dim=4,
delta_v=0.5,
delta_d=3.0,
scale_lane_line=1.0,
scale_var=1.0,
scale_dist=1.0,
pretrained=False,
**kwargs
):
super(LaneNet, self).__init__()
self.pretrained = pretrained
self.embed_dim = embed_dim
self.delta_v = delta_v
self.delta_d = delta_d
self.net_init()
self.scale_seg = scale_lane_line
self.scale_var = scale_var
self.scale_dist = scale_dist
self.scale_reg = 0
self.seg_loss = nn.CrossEntropyLoss(weight=torch.tensor([0.4, 1.]))
def net_init(self):
self.backbone = models.vgg16_bn(pretrained=self.pretrained).features
# ----------------- process backbone -----------------
for i in [34, 37, 40]:
conv = self.backbone._modules[str(i)]
dilated_conv = nn.Conv2d(
conv.in_channels, conv.out_channels, conv.kernel_size, stride=conv.stride,
padding=tuple(p * 2 for p in conv.padding), dilation=2, bias=(conv.bias is not None)
)
dilated_conv.load_state_dict(conv.state_dict())
self.backbone._modules[str(i)] = dilated_conv
self.backbone._modules.pop('33')
self.backbone._modules.pop('43')
# ----------------- additional conv -----------------
self.layer1 = nn.Sequential(
nn.Conv2d(512, 1024, 3, padding=4, dilation=4, bias=False),
nn.BatchNorm2d(1024),
nn.ReLU(),
nn.Conv2d(1024, 128, 3, padding=1, bias=False),
nn.BatchNorm2d(128),
nn.ReLU(),
nn.Conv2d(128, 32, 3, padding=1, bias=False),
nn.BatchNorm2d(32),
nn.ReLU(),
nn.Conv2d(32, 8, 3, padding=1, bias=False),
nn.BatchNorm2d(8),
nn.ReLU(),
nn.Upsample(scale_factor=8, mode='bilinear', align_corners=True)
)
# ----------------- embedding -----------------
self.embedding = nn.Sequential(
nn.Conv2d(8, 8, 1),
nn.BatchNorm2d(8),
nn.ReLU(),
nn.Conv2d(8, self.embed_dim, 1)
)
# ----------------- binary segmentation -----------------
self.binary_seg = nn.Sequential(
nn.Conv2d(8, 8, 1),
nn.BatchNorm2d(8),
nn.ReLU(),
nn.Conv2d(8, 2, 1)
)
def forward(self, img, segLabel=None, target=None):#
'''
segLabel=img['segLabel']
target=img['target']
img=img['img']'''
input_size = img.size()
x = self.backbone(img)
x = self.layer1(x)
embedding = self.embedding(x)
binary_seg = self.binary_seg(x)
if segLabel is not None:
var_loss, dist_loss, reg_loss = self.discriminative_loss(embedding, segLabel)
seg_loss = self.seg_loss(binary_seg, torch.gt(segLabel, 0).type(torch.long))
else:
var_loss = torch.tensor(0, dtype=img.dtype, device=img.device)
dist_loss = torch.tensor(0, dtype=img.dtype, device=img.device)
seg_loss = torch.tensor(0, dtype=img.dtype, device=img.device)
loss = seg_loss * self.scale_seg + var_loss * self.scale_var + dist_loss * self.scale_dist
output = {
"embedding": embedding,
"binary_seg": binary_seg,
"seg_loss": seg_loss,
"var_loss": var_loss,
"dist_loss": dist_loss,
"reg_loss": reg_loss,
"loss": loss
}
return output["binary_seg"], output["loss"]
def discriminative_loss(self, embedding, seg_gt):
batch_size = embedding.shape[0]
var_loss = torch.tensor(0, dtype=embedding.dtype, device=embedding.device)
dist_loss = torch.tensor(0, dtype=embedding.dtype, device=embedding.device)
reg_loss = torch.tensor(0, dtype=embedding.dtype, device=embedding.device)
for b in range(batch_size):
embedding_b = embedding[b] # (embed_dim, H, W)
seg_gt_b = seg_gt[b]
labels = torch.unique(seg_gt_b)
labels = labels[labels!=0]
num_lanes = len(labels)
if num_lanes==0:
# please refer to issue here: https://github.com/harryhan618/LaneNet/issues/12
_nonsense = embedding.sum()
_zero = torch.zeros_like(_nonsense)
var_loss = var_loss + _nonsense * _zero
dist_loss = dist_loss + _nonsense * _zero
reg_loss = reg_loss + _nonsense * _zero
continue
centroid_mean = []
for lane_idx in labels:
seg_mask_i = (seg_gt_b == lane_idx)
if not seg_mask_i.any():
continue
embedding_i = embedding_b[:, seg_mask_i]
mean_i = torch.mean(embedding_i, dim=1)
centroid_mean.append(mean_i)
# ---------- var_loss -------------
var_loss = var_loss + torch.mean( F.relu(torch.norm(embedding_i-mean_i.reshape(self.embed_dim,1), dim=0) - self.delta_v)**2 ) / num_lanes
centroid_mean = torch.stack(centroid_mean) # (n_lane, embed_dim)
if num_lanes > 1:
centroid_mean1 = centroid_mean.reshape(-1, 1, self.embed_dim)
centroid_mean2 = centroid_mean.reshape(1, -1, self.embed_dim)
dist = torch.norm(centroid_mean1-centroid_mean2, dim=2) # shape (num_lanes, num_lanes)
dist = dist + torch.eye(num_lanes, dtype=dist.dtype, device=dist.device) * self.delta_d # diagonal elements are 0, now mask above delta_d
# divided by two for double calculated loss above, for implementation convenience
dist_loss = dist_loss + torch.sum(F.relu(-dist + self.delta_d)**2) / (num_lanes * (num_lanes-1)) / 2
# reg_loss is not used in original paper
# reg_loss = reg_loss + torch.mean(torch.norm(centroid_mean, dim=1))
var_loss = var_loss / batch_size
dist_loss = dist_loss / batch_size
reg_loss = reg_loss / batch_size
return var_loss, dist_loss, reg_loss