-
Notifications
You must be signed in to change notification settings - Fork 0
/
models.py
262 lines (224 loc) · 8.78 KB
/
models.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
import torch.nn.functional as F
import time
import torch
import torch.nn as nn
import numpy as np
import torchvision
import util
import custom_layers
import geometry
import hyperlayers
class LightFieldModel(nn.Module):
def __init__(
self,
latent_dim,
parameterization="plucker",
network="relu",
fit_single=False,
conditioning="hyper",
depth=False,
alpha=False,
):
super().__init__()
self.latent_dim = latent_dim
self.num_hidden_units_phi = 256
self.fit_single = fit_single
self.parameterization = parameterization
self.conditioning = conditioning
self.depth = depth
self.alpha = alpha
out_channels = 3
if self.depth:
out_channels += 1
if self.alpha:
out_channels += 1
self.background = torch.ones((1, 1, 1, 3)).cuda()
if self.fit_single or conditioning in ["hyper", "low_rank"]:
if network == "relu":
self.phi = custom_layers.FCBlock(
hidden_ch=self.num_hidden_units_phi,
num_hidden_layers=6,
in_features=6,
out_features=out_channels,
outermost_linear=True,
norm="layernorm_na",
)
elif network == "siren":
omega_0 = 30.0
self.phi = custom_layers.Siren(
in_features=6,
hidden_features=256,
hidden_layers=8,
out_features=out_channels,
outermost_linear=True,
hidden_omega_0=omega_0,
first_omega_0=omega_0,
)
elif conditioning == "concat":
self.phi = nn.Sequential(
nn.Linear(6 + self.latent_dim, self.num_hidden_units_phi),
custom_layers.ResnetBlockFC(
size_in=self.num_hidden_units_phi,
size_out=self.num_hidden_units_phi,
size_h=self.num_hidden_units_phi,
),
custom_layers.ResnetBlockFC(
size_in=self.num_hidden_units_phi,
size_out=self.num_hidden_units_phi,
size_h=self.num_hidden_units_phi,
),
custom_layers.ResnetBlockFC(
size_in=self.num_hidden_units_phi,
size_out=self.num_hidden_units_phi,
size_h=self.num_hidden_units_phi,
),
nn.Linear(self.num_hidden_units_phi, 3),
)
if not self.fit_single:
if conditioning == "hyper":
self.hyper_phi = hyperlayers.HyperNetwork(
hyper_in_features=self.latent_dim,
hyper_hidden_layers=1,
hyper_hidden_features=self.latent_dim,
hypo_module=self.phi,
)
elif conditioning == "low_rank":
self.hyper_phi = hyperlayers.LowRankHyperNetwork(
hyper_in_features=self.latent_dim,
hyper_hidden_layers=1,
hyper_hidden_features=512,
hypo_module=self.phi,
nonlinearity="leaky_relu",
)
print(self.phi)
print(np.sum(np.prod(param.shape) for param in self.phi.parameters()))
def get_light_field_function(self, z=None):
if self.fit_single:
phi = self.phi
elif self.conditioning in ["hyper", "low_rank"]:
phi_weights = self.hyper_phi(z)
phi = lambda x: self.phi(x, params=phi_weights)
elif self.conditioning == "concat":
def phi(x):
b, n_pix = x.shape[:2]
z_rep = z.view(b, 1, self.latent_dim).repeat(1, n_pix, 1)
return self.phi(torch.cat((z_rep, x), dim=-1))
return phi
def get_query_cam(self, input):
query_dict = input["query"]
pose = util.flatten_first_two(query_dict["cam2world"])
intrinsics = util.flatten_first_two(query_dict["intrinsics"])
uv = util.flatten_first_two(query_dict["uv"].float())
return pose, intrinsics, uv
def forward(self, input, val=False, compute_depth=False, timing=False):
out_dict = {}
query = input["query"]
b, _ = query["uv"].shape[:2]
n_qry, n_pix = query["uv"].shape[1:3]
if not self.fit_single:
if "z" in input:
z = input["z"]
else:
z = self.get_z(input)
out_dict["z"] = z
z = z.view(b * n_qry, self.latent_dim)
query_pose, query_intrinsics, query_uv = self.get_query_cam(input)
if self.parameterization == "plucker":
light_field_coords = geometry.plucker_embedding(
query_pose, query_uv, query_intrinsics
)
else:
ray_origin = query_pose[:, :3, 3][:, None, :]
ray_dir = geometry.get_ray_directions(
query_uv, query_pose, query_intrinsics
)
intsec_1, intsec_2 = geometry.ray_sphere_intersect(
ray_origin, ray_dir, radius=100
)
intsec_1 = F.normalize(intsec_1, dim=-1)
intsec_2 = F.normalize(intsec_2, dim=-1)
light_field_coords = torch.cat((intsec_1, intsec_2), dim=-1)
out_dict["intsec_1"] = intsec_1
out_dict["intsec_2"] = intsec_2
out_dict["ray_dir"] = ray_dir
out_dict["ray_origin"] = ray_origin
light_field_coords.requires_grad_(True)
out_dict["coords"] = light_field_coords.view(b * n_qry, n_pix, 6)
lf_function = self.get_light_field_function(None if self.fit_single else z)
out_dict["lf_function"] = lf_function
if timing:
t0 = time.time()
lf_out = lf_function(out_dict["coords"])
if timing:
t1 = time.time()
total_n = t1 - t0
print(f"{total_n}")
rgb = lf_out[..., :3]
if self.depth:
depth = lf_out[..., 3:4]
out_dict["depth"] = depth.view(b, n_qry, n_pix, 1)
rgb = rgb.view(b, n_qry, n_pix, 3)
if self.alpha:
alpha = lf_out[..., -1:].view(b, n_qry, n_pix, 1)
weight = 1 - torch.exp(-torch.abs(alpha))
rgb = weight * rgb + (1 - weight) * self.background
out_dict["alpha"] = weight
if compute_depth:
with torch.enable_grad():
lf_function = self.get_light_field_function(z)
depth = util.light_field_depth_map(
light_field_coords, query_pose, lf_function
)["depth"]
depth = depth.view(b, n_qry, n_pix, 1)
out_dict["depth"] = depth
out_dict["rgb"] = rgb
return out_dict
class LFAutoDecoder(LightFieldModel):
def __init__(self, latent_dim, num_instances, parameterization="plucker", **kwargs):
super().__init__(
latent_dim=latent_dim, parameterization=parameterization, **kwargs
)
self.num_instances = num_instances
self.latent_codes = nn.Embedding(num_instances, self.latent_dim)
nn.init.normal_(self.latent_codes.weight, mean=0, std=0.01)
def get_z(self, input, val=False):
instance_idcs = input["query"]["instance_idx"].long()
z = self.latent_codes(instance_idcs)
return z
def normalize_imagenet(x):
"""Normalize input images according to ImageNet standards.
Args:
x (tensor): input images
"""
x = x.clone()
x[:, 0] = (x[:, 0] - 0.485) / 0.229
x[:, 1] = (x[:, 1] - 0.456) / 0.224
x[:, 2] = (x[:, 2] - 0.406) / 0.225
return x
class Resnet18(nn.Module):
r"""ResNet-18 encoder network for image input.
Args:
c_dim (int): output dimension of the latent embedding
normalize (bool): whether the input images should be normalized
use_linear (bool): whether a final linear layer should be used
"""
def __init__(self, c_dim, normalize=True, use_linear=True):
super().__init__()
self.normalize = normalize
self.use_linear = use_linear
self.features = torchvision.models.resnet18(pretrained=True)
self.features.fc = nn.Sequential()
if use_linear:
self.fc = nn.Linear(512, c_dim)
self.fc.apply(custom_layers.init_weights_normal)
elif c_dim == 512:
self.fc = nn.Sequential()
else:
raise ValueError("c_dim must be 512 if use_linear is False")
def forward(self, input):
x = (input + 1) / 2
if self.normalize:
x = normalize_imagenet(x)
net = self.features(x)
out = self.fc(net)
return out