forked from ashawkey/torch-ngp
-
Notifications
You must be signed in to change notification settings - Fork 0
/
network.py
206 lines (161 loc) · 6.21 KB
/
network.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
import torch
import torch.nn as nn
import torch.nn.functional as F
from encoding import get_encoder
from activation import trunc_exp
from .renderer import NeRFRenderer
class NeRFNetwork(NeRFRenderer):
def __init__(self,
encoding="hashgrid",
encoding_dir="sphere_harmonics",
encoding_bg="hashgrid",
num_layers=2,
hidden_dim=64,
geo_feat_dim=15,
num_layers_color=3,
hidden_dim_color=64,
num_layers_bg=2,
hidden_dim_bg=64,
bound=1,
**kwargs,
):
super().__init__(bound, **kwargs)
# sigma network
self.num_layers = num_layers
self.hidden_dim = hidden_dim
self.geo_feat_dim = geo_feat_dim
self.encoder, self.in_dim = get_encoder(encoding, desired_resolution=2048 * bound)
sigma_net = []
for l in range(num_layers):
if l == 0:
in_dim = self.in_dim
else:
in_dim = hidden_dim
if l == num_layers - 1:
out_dim = 1 + self.geo_feat_dim # 1 sigma + 15 SH features for color
else:
out_dim = hidden_dim
sigma_net.append(nn.Linear(in_dim, out_dim, bias=False))
self.sigma_net = nn.ModuleList(sigma_net)
# color network
self.num_layers_color = num_layers_color
self.hidden_dim_color = hidden_dim_color
self.encoder_dir, self.in_dim_dir = get_encoder(encoding_dir)
color_net = []
for l in range(num_layers_color):
if l == 0:
in_dim = self.in_dim_dir + self.geo_feat_dim
else:
in_dim = hidden_dim
if l == num_layers_color - 1:
out_dim = 3 # 3 rgb
else:
out_dim = hidden_dim
color_net.append(nn.Linear(in_dim, out_dim, bias=False))
self.color_net = nn.ModuleList(color_net)
# background network
if self.bg_radius > 0:
self.num_layers_bg = num_layers_bg
self.hidden_dim_bg = hidden_dim_bg
self.encoder_bg, self.in_dim_bg = get_encoder(encoding_bg, input_dim=2, num_levels=4, log2_hashmap_size=19, desired_resolution=2048) # much smaller hashgrid
bg_net = []
for l in range(num_layers_bg):
if l == 0:
in_dim = self.in_dim_bg + self.in_dim_dir
else:
in_dim = hidden_dim_bg
if l == num_layers_bg - 1:
out_dim = 3 # 3 rgb
else:
out_dim = hidden_dim_bg
bg_net.append(nn.Linear(in_dim, out_dim, bias=False))
self.bg_net = nn.ModuleList(bg_net)
else:
self.bg_net = None
def forward(self, x, d):
# x: [N, 3], in [-bound, bound]
# d: [N, 3], nomalized in [-1, 1]
# sigma
x = self.encoder(x, bound=self.bound)
h = x
for l in range(self.num_layers):
h = self.sigma_net[l](h)
if l != self.num_layers - 1:
h = F.relu(h, inplace=True)
#sigma = F.relu(h[..., 0])
sigma = trunc_exp(h[..., 0])
geo_feat = h[..., 1:]
# color
d = self.encoder_dir(d)
h = torch.cat([d, geo_feat], dim=-1)
for l in range(self.num_layers_color):
h = self.color_net[l](h)
if l != self.num_layers_color - 1:
h = F.relu(h, inplace=True)
# sigmoid activation for rgb
color = torch.sigmoid(h)
return sigma, color
def density(self, x):
# x: [N, 3], in [-bound, bound]
x = self.encoder(x, bound=self.bound)
h = x
for l in range(self.num_layers):
h = self.sigma_net[l](h)
if l != self.num_layers - 1:
h = F.relu(h, inplace=True)
#sigma = F.relu(h[..., 0])
sigma = trunc_exp(h[..., 0])
geo_feat = h[..., 1:]
return {
'sigma': sigma,
'geo_feat': geo_feat,
}
def background(self, x, d):
# x: [N, 2], in [-1, 1]
h = self.encoder_bg(x) # [N, C]
d = self.encoder_dir(d)
h = torch.cat([d, h], dim=-1)
for l in range(self.num_layers_bg):
h = self.bg_net[l](h)
if l != self.num_layers_bg - 1:
h = F.relu(h, inplace=True)
# sigmoid activation for rgb
rgbs = torch.sigmoid(h)
return rgbs
# allow masked inference
def color(self, x, d, mask=None, geo_feat=None, **kwargs):
# x: [N, 3] in [-bound, bound]
# mask: [N,], bool, indicates where we actually needs to compute rgb.
if mask is not None:
rgbs = torch.zeros(mask.shape[0], 3, dtype=x.dtype, device=x.device) # [N, 3]
# in case of empty mask
if not mask.any():
return rgbs
x = x[mask]
d = d[mask]
geo_feat = geo_feat[mask]
d = self.encoder_dir(d)
h = torch.cat([d, geo_feat], dim=-1)
for l in range(self.num_layers_color):
h = self.color_net[l](h)
if l != self.num_layers_color - 1:
h = F.relu(h, inplace=True)
# sigmoid activation for rgb
h = torch.sigmoid(h)
if mask is not None:
rgbs[mask] = h.to(rgbs.dtype) # fp16 --> fp32
else:
rgbs = h
return rgbs
# optimizer utils
def get_params(self, lr):
params = [
{'params': self.encoder.parameters(), 'lr': lr},
{'params': self.sigma_net.parameters(), 'lr': lr},
{'params': self.encoder_dir.parameters(), 'lr': lr},
{'params': self.color_net.parameters(), 'lr': lr},
]
if self.bg_radius > 0:
params.append({'params': self.encoder_bg.parameters(), 'lr': lr})
params.append({'params': self.bg_net.parameters(), 'lr': lr})
return params