Spaces:
Running
on
Zero
Running
on
Zero
File size: 8,189 Bytes
f9cbc98 |
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 |
import torch
from torch import nn
import numpy as np
import math
# import tinycudann as tcnn
class SineLayer(nn.Module):
# See paper sec. 3.2, final paragraph, and supplement Sec. 1.5 for discussion of omega_0.
# If is_first=True, omega_0 is a frequency factor which simply multiplies the activations before the
# nonlinearity. Different signals may require different omega_0 in the first layer - this is a
# hyperparameter.
# If is_first=False, then the weights will be divided by omega_0 so as to keep the magnitude of
# activations constant, but boost gradients to the weight matrix (see supplement Sec. 1.5)
def __init__(self, in_features, out_features, bias=True,
is_first=False, omega_0=30):
super().__init__()
self.omega_0 = omega_0
self.is_first = is_first
self.in_features = in_features
self.linear = nn.Linear(in_features, out_features, bias=bias)
self.init_weights()
def init_weights(self):
with torch.no_grad():
if self.is_first:
self.linear.weight.uniform_(-1 / self.in_features,
1 / self.in_features)
else:
self.linear.weight.uniform_(-np.sqrt(6 / self.in_features) / self.omega_0,
np.sqrt(6 / self.in_features) / self.omega_0)
def forward(self, input):
return torch.sin(self.omega_0 * self.linear(input))
def forward_with_intermediate(self, input):
# For visualization of activation distributions
intermediate = self.omega_0 * self.linear(input)
return torch.sin(intermediate), intermediate
class Siren(nn.Module):
def __init__(self, in_features, hidden_features, hidden_layers, out_features, outermost_linear=False,
first_omega_0=30, hidden_omega_0=30.):
super().__init__()
self.net = []
self.net.append(SineLayer(in_features, hidden_features,
is_first=True, omega_0=first_omega_0))
for i in range(hidden_layers):
self.net.append(SineLayer(hidden_features, hidden_features,
is_first=False, omega_0=hidden_omega_0))
if outermost_linear:
final_linear = nn.Linear(hidden_features, out_features)
with torch.no_grad():
final_linear.weight.uniform_(-np.sqrt(6 / hidden_features) / hidden_omega_0,
np.sqrt(6 / hidden_features) / hidden_omega_0)
self.net.append(final_linear)
else:
self.net.append(SineLayer(hidden_features, out_features,
is_first=False, omega_0=hidden_omega_0))
self.net = nn.Sequential(*self.net)
def forward(self, coords):
output = self.net(coords)
return output
class Homography(nn.Module):
def __init__(self, in_features=1, hidden_features=256, hidden_layers=1):
super().__init__()
out_features = 8
self.net = []
self.net.append(nn.Linear(in_features, hidden_features))
self.net.append(nn.ReLU(inplace=True))
for i in range(hidden_layers):
self.net.append(nn.Linear(hidden_features, hidden_features))
self.net.append(nn.ReLU(inplace=True))
self.net.append(nn.Linear(hidden_features, out_features))
self.net = nn.Sequential(*self.net)
self.init_weights()
def init_weights(self):
with torch.no_grad():
self.net[-1].bias.copy_(torch.Tensor([1., 0., 0., 0., 1., 0., 0., 0.]))
def forward(self, coords):
output = self.net(coords)
return output
class Annealed(nn.Module):
def __init__(self, in_channels, annealed_step, annealed_begin_step=0, identity=True):
"""
Defines a function that embeds x to (x, sin(2^k x), cos(2^k x), ...)
in_channels: number of input channels (3 for both xyz and direction)
"""
super(Annealed, self).__init__()
self.N_freqs = 16
self.in_channels = in_channels
self.annealed = True
self.annealed_step = annealed_step
self.annealed_begin_step = annealed_begin_step
self.index = torch.linspace(0, self.N_freqs - 1, self.N_freqs)
self.identity = identity
self.index_2 = self.index.view(-1, 1).repeat(1, 2).view(-1)
def forward(self, x_embed, step):
"""
Embeds x to (x, sin(2^k x), cos(2^k x), ...)
Different from the paper, "x" is also in the output
See https://github.com/bmild/nerf/issues/12
Inputs:
x: (B, self.in_channels)
Outputs:
out: (B, self.out_channels)
"""
use_PE = False
if self.annealed_begin_step == 0:
# calculate the w for each freq bands
alpha = self.N_freqs * step / float(self.annealed_step)
else:
if step <= self.annealed_begin_step:
alpha = 0
else:
alpha = (self.N_freqs) * (step - self.annealed_begin_step) / float(
self.annealed_step)
w = (1 - torch.cos(math.pi * torch.clamp(alpha * torch.ones_like(self.index_2) - self.index_2, 0, 1))) / 2
if use_PE:
w[16:] = w[:16]
out = x_embed * w.to(x_embed.device)
return out
class BARF_PE(nn.Module):
def __init__(self, config):
super().__init__()
self.encoder = tcnn.Encoding(n_input_dims=2,
encoding_config=config["positional encoding"])
self.decoder = tcnn.Network(n_input_dims=self.encoder.n_output_dims +
2,
n_output_dims=3,
network_config=config["BARF network"])
def forward(self, x, step=0, aneal_func=None):
input = x
input = self.encoder(input)
if aneal_func is not None:
input = torch.cat([x, aneal_func(input,step)], dim=-1)
else:
input = torch.cat([x, input], dim=-1)
weight = torch.ones(input.shape[-1], device=input.device).cuda()
x = self.decoder(weight * input)
return x
class Deform_Hash3d(nn.Module):
def __init__(self, config):
super().__init__()
self.encoder = tcnn.Encoding(n_input_dims=3,
encoding_config=config["encoding_deform3d"])
self.decoder = nn.Sequential(nn.Linear(self.encoder.n_output_dims + 3, 256),
nn.ReLU(),
nn.Linear(256, 256),
nn.ReLU(),
nn.Linear(256, 256),
nn.ReLU(),
nn.Linear(256, 256),
nn.ReLU(),
nn.Linear(256, 256),
nn.ReLU(),
nn.Linear(256, 256),
nn.ReLU(),
nn.Linear(256, 2)
)
def forward(self, x, step=0, aneal_func=None):
input = x
input = self.encoder(input)
if aneal_func is not None:
input = torch.cat([x, aneal_func(input,step)], dim=-1)
else:
input = torch.cat([x, input], dim=-1)
weight = torch.ones(input.shape[-1], device=input.device).cuda()
x = self.decoder(weight * input) / 5
return x
class Deform_Hash3d_Warp(nn.Module):
def __init__(self, config):
super().__init__()
self.Deform_Hash3d = Deform_Hash3d(config)
def forward(self, xyt_norm, step=0,aneal_func=None):
x = self.Deform_Hash3d(xyt_norm,step=step, aneal_func=aneal_func)
return x |