-
Notifications
You must be signed in to change notification settings - Fork 6
/
sparsegrid.py
159 lines (121 loc) · 5.5 KB
/
sparsegrid.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
import torch
import torch.nn as nn
class SparseGrid(nn.Module):
def __init__(self, level_dim=2, x_resolution=300, y_resolution=300, t_resolution=600, upsample=False):
super().__init__()
self.level_dim = level_dim # latent dimension
self.x_resolution = x_resolution
self.y_resolution = y_resolution
self.t_resolution = t_resolution
self.embeddings = nn.Parameter(torch.empty(self.t_resolution, self.x_resolution, self.y_resolution, self.level_dim))
self.upsample = upsample
self.reset_parameters()
def reset_parameters(self):
std = 1e-4
self.embeddings.data.uniform_(-std, std)
def forward(self, inputs):
# inputs [0, 1]
if self.upsample:
# upsampling sparse positional features
tmp_embeddings = self.embeddings.permute(3, 0, 1, 2) # dim, T, H, W
tmp_embeddings = torch.nn.functional.interpolate(tmp_embeddings, scale_factor=2, mode='bilinear')
tmp_embeddings = tmp_embeddings.permute(1, 2, 3, 0)
tmp_shape = tmp_embeddings.shape
t_res = self.t_resolution
x_res = tmp_shape[1]
y_res = tmp_shape[2]
else:
t_res = self.t_resolution
x_res = self.x_resolution
y_res = self.y_resolution
tmp_embeddings = self.embeddings
# round down
t_coord = inputs[:, 0]
t_coord_float = ((t_res-1)*t_coord)
t_coord_idx = (t_coord_float+0.5).type(torch.int64)
t_coord_idx = torch.clamp(t_coord_idx, 0, t_res-1)
x_coord = inputs[:, 1]
x_coord_float = ((x_res-1)*x_coord)
x_coord_idx = (x_coord_float+0.5).type(torch.int64)
x_coord_idx = torch.clamp(x_coord_idx, 0, x_res-1)
y_coord = inputs[:, 2]
y_coord_float = ((y_res-1)*y_coord)
y_coord_idx = (y_coord_float+0.5).type(torch.int64)
y_coord_idx = torch.clamp(y_coord_idx, 0, y_res-1)
grid_features = None
unfold_list = [-1, 0, 1]
for i in unfold_list:
for j in unfold_list:
vx = torch.clamp(x_coord_idx+i, 0, x_res-1)
vy = torch.clamp(y_coord_idx+j, 0, y_res-1)
feat = (tmp_embeddings[t_coord_idx, vx, vy, :])
if grid_features == None:
grid_features = feat
else:
grid_features = torch.cat((grid_features, feat), dim=1)
return grid_features
def forward_inter(self, inputs):
# inputs [0, 1]
if self.upsample:
# upsampling sparse positional features
tmp_embeddings = self.embeddings.permute(3, 0, 1, 2) # dim, T, H, W
tmp_embeddings = torch.nn.functional.interpolate(tmp_embeddings, scale_factor=2, mode='bilinear')
tmp_embeddings = tmp_embeddings.permute(1, 2, 3, 0)
tmp_shape = tmp_embeddings.shape
t_res = self.t_resolution
x_res = tmp_shape[1]
y_res = tmp_shape[2]
else:
t_res = self.t_resolution
x_res = self.x_resolution
y_res = self.y_resolution
tmp_embeddings = self.embeddings
# round down
t_coord = inputs[:, 0]
t_coord_float = ((t_res-1)*t_coord)
t_coord_idx = (t_coord_float+0.5).type(torch.int64)
t_coord_idx = torch.clamp(t_coord_idx, 0, t_res-1)
t_coord_idx_lower = t_coord_float.type(torch.int64)
t_coord_idx_upper = (t_coord_float+1).type(torch.int64)
t_coord_idx_upper = torch.clamp(t_coord_idx_upper, 0, t_res-1)
upper_coeff = t_coord_float-t_coord_idx_lower
lower_coeff = t_coord_idx_upper-t_coord_float
upper_coeff = upper_coeff/(upper_coeff+lower_coeff)
lower_coeff = lower_coeff/(upper_coeff+lower_coeff)
x_coord = inputs[:, 1]
x_coord_float = ((x_res-1)*x_coord)
x_coord_idx = (x_coord_float+0.5).type(torch.int64)
x_coord_idx = torch.clamp(x_coord_idx, 0, x_res-1)
y_coord = inputs[:, 2]
y_coord_float = ((y_res-1)*y_coord)
y_coord_idx = (y_coord_float+0.5).type(torch.int64)
y_coord_idx = torch.clamp(y_coord_idx, 0, y_res-1)
unfold_list = [-1, 0, 1]
grid_features_1 = None
lower_coeff = lower_coeff.unsqueeze(1)
lower_coeff = lower_coeff.repeat(1, self.level_dim)
for i in unfold_list:
for j in unfold_list:
vx = torch.clamp(x_coord_idx+i, 0, x_res-1)
vy = torch.clamp(y_coord_idx+j, 0, y_res-1)
feat = (tmp_embeddings[t_coord_idx_lower, vx, vy, :])
feat *= lower_coeff
if grid_features_1 == None:
grid_features_1 = feat
else:
grid_features_1 = torch.cat((grid_features_1, feat), dim=1)
grid_features_2 = None
upper_coeff = upper_coeff.unsqueeze(1)
upper_coeff = upper_coeff.repeat(1, self.level_dim)
for i in unfold_list:
for j in unfold_list:
vx = torch.clamp(x_coord_idx+i, 0, x_res-1)
vy = torch.clamp(y_coord_idx+j, 0, y_res-1)
feat = (tmp_embeddings[t_coord_idx_upper, vx, vy, :])
feat *= upper_coeff
if grid_features_2 == None:
grid_features_2 = feat
else:
grid_features_2 = torch.cat((grid_features_2, feat), dim=1)
grid_features = grid_features_1+grid_features_2
return grid_features