-
Notifications
You must be signed in to change notification settings - Fork 8
/
model.py
168 lines (140 loc) · 6.02 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
import torch
import torch.nn as nn
import torch.distributions as tdist
class SocialCellLocal(nn.Module):
def __init__(self,
spatial_input=2,
spatial_output=2,
temporal_input=8,
temporal_output=12):
super(SocialCellLocal, self).__init__()
#Spatial Section
self.feat = nn.Conv1d(spatial_input,
spatial_output,
3,
padding=1,
padding_mode='zeros')
self.feat_act = nn.ReLU()
self.highway_input = nn.Conv1d(spatial_input,
spatial_output,
1,
padding=0)
#Temporal Section
self.highway = nn.Conv1d(temporal_input, temporal_output, 1, padding=0)
self.tpcnn = nn.Conv1d(temporal_input,
temporal_output,
3,
padding=1,
padding_mode='zeros')
def forward(self, v):
v_shape = v.shape
#Spatial Section
v = v.permute(0, 3, 1,
2).reshape(v_shape[0] * v_shape[3], v_shape[1],
v_shape[2]) #= PED*batch, [x,y], TIME,
v_res = self.highway_input(v)
v = self.feat_act(self.feat(v)) + v_res
#Temporal Section
v = v.permute(0, 2, 1)
v_res = self.highway(v)
v = self.tpcnn(v) + v_res
#Final Output
v = v.permute(0, 2, 1).reshape(v_shape[0], v_shape[3], v_shape[1],
12).permute(0, 2, 3, 1)
return v
class SocialCellGlobal(nn.Module):
def __init__(self,
spatial_input=2,
spatial_output=2,
temporal_input=8,
temporal_output=12,
noise_w=None):
super(SocialCellGlobal, self).__init__()
#Spatial Section
self.feat = nn.Conv2d(spatial_input,
spatial_output,
3,
padding=1,
padding_mode='zeros')
self.feat_act = nn.ReLU()
self.highway_input = nn.Conv2d(spatial_input,
spatial_output,
1,
padding=0)
#Temporal Section
self.highway = nn.Conv2d(temporal_input, temporal_output, 1, padding=0)
self.tpcnn = nn.Conv2d(temporal_input,
temporal_output,
3,
padding=1,
padding_mode='zeros')
#Self Learning Weights
self.noise_w = nn.Parameter(torch.zeros(1), requires_grad=True)
self.noise_weights = noise_w # Used to scale the variance
self.global_w = nn.Parameter(torch.zeros(1), requires_grad=True)
self.local_w = nn.Parameter(torch.zeros(1), requires_grad=True)
#Local Stream
self.ped = SocialCellLocal(spatial_input=spatial_input,
spatial_output=spatial_output,
temporal_input=temporal_input,
temporal_output=temporal_output)
def forward(self, v, noise, weight_select=1):
#Combine Vectorized Noise
v = v + self.noise_w * self.noise_weights[weight_select] * noise
#Spatial Section
v_ped = self.ped(v)
v_res = self.highway_input(v)
v = self.feat_act(self.feat(v)) + v_res
#Temporal Section
v = v.permute(0, 2, 1, 3)
v_res = self.highway(v)
v = self.tpcnn(v) + v_res
#Fuse Local and Global Streams
v = v.permute(0, 2, 1, 3)
v = self.global_w * v + self.local_w * v_ped
return v
class SocialImplicit(nn.Module):
def __init__(self,
spatial_input=2,
spatial_output=2,
temporal_input=8,
temporal_output=12,
bins=[0, 0.01, 0.1, 1.2],
noise_weight=[0.05, 1, 4, 8]):
super(SocialImplicit, self).__init__()
self.bins = torch.Tensor(bins).cuda()
self.implicit_cells = nn.ModuleList([
SocialCellGlobal(spatial_input=spatial_input,
spatial_output=spatial_output,
temporal_input=temporal_input,
temporal_output=temporal_output,
noise_w=noise_weight)
for i in range(len(self.bins))
])
self.noise = tdist.multivariate_normal.MultivariateNormal(
torch.zeros(2), torch.Tensor([[1, 0], [0, 1]]))
def forward(self, v, obs_traj, KSTEPS=20):
noise = self.noise.sample((KSTEPS, )).unsqueeze(-1).unsqueeze(-1).to(
v.device).double().contiguous()
#Social-Zones Section
# Use max speed change(inf norm) to assign a zone
norm = torch.linalg.norm(v.permute(0, 3, 1, 2)[0, :, :, 0],
float('inf'),
dim=1)
displacment_indx = torch.bucketize(
norm,
self.bins,
right=True,
) - 1 #Used to set each vector to a zone
v_out = torch.zeros(KSTEPS, 2, 12, v.shape[-1]).double().to(
v.device).contiguous() #Stores results of each zone
#Per each Social-Zone, call the proper Social-Cell
for i in range(len(self.bins)):
select = displacment_indx == i
if torch.any(select):
v_out[...,
select] = self.implicit_cells[i](v[...,
select].contiguous(),
noise,
weight_select=i)
return v_out.contiguous()