File size: 6,305 Bytes
904ef7d 6f790c0 904ef7d |
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 |
import torch
import torch.nn as nn
import torch.nn.functional as F
from activation import trunc_exp
from .renderer import NeRFRenderer
import numpy as np
from encoding import get_encoder
from .utils import safe_normalize
class MLP(nn.Module):
def __init__(self, dim_in, dim_out, dim_hidden, num_layers, bias=True):
super().__init__()
self.dim_in = dim_in
self.dim_out = dim_out
self.dim_hidden = dim_hidden
self.num_layers = num_layers
net = []
for l in range(num_layers):
net.append(nn.Linear(self.dim_in if l == 0 else self.dim_hidden, self.dim_out if l == num_layers - 1 else self.dim_hidden, bias=bias))
self.net = nn.ModuleList(net)
def forward(self, x):
for l in range(self.num_layers):
x = self.net[l](x)
if l != self.num_layers - 1:
x = F.relu(x, inplace=True)
return x
class NeRFNetwork(NeRFRenderer):
def __init__(self,
opt,
num_layers=3,
hidden_dim=64,
num_layers_bg=2,
hidden_dim_bg=64,
):
super().__init__(opt)
self.num_layers = num_layers
self.hidden_dim = hidden_dim
self.encoder, self.in_dim = get_encoder('tiledgrid', input_dim=3, desired_resolution=2048 * self.bound)
self.sigma_net = MLP(self.in_dim, 4, hidden_dim, num_layers, bias=True)
# background network
if self.bg_radius > 0:
self.num_layers_bg = num_layers_bg
self.hidden_dim_bg = hidden_dim_bg
# use a very simple network to avoid it learning the prompt...
# self.encoder_bg, self.in_dim_bg = get_encoder('tiledgrid', input_dim=2, num_levels=4, desired_resolution=2048)
self.encoder_bg, self.in_dim_bg = get_encoder('frequency', input_dim=2)
self.bg_net = MLP(self.in_dim_bg, 3, hidden_dim_bg, num_layers_bg, bias=True)
else:
self.bg_net = None
# add a density blob to the scene center
def gaussian(self, x):
# x: [B, N, 3]
d = (x ** 2).sum(-1)
g = 5 * torch.exp(-d / (2 * 0.2 ** 2))
return g
def common_forward(self, x):
# x: [N, 3], in [-bound, bound]
# sigma
h = self.encoder(x, bound=self.bound)
h = self.sigma_net(h)
sigma = trunc_exp(h[..., 0] + self.gaussian(x))
albedo = torch.sigmoid(h[..., 1:])
return sigma, albedo
# ref: https://github.com/zhaofuq/Instant-NSR/blob/main/nerf/network_sdf.py#L192
def finite_differnce_normal(self, x, epsilon=5e-4):
# x: [N, 3]
dx_pos, _ = self.common_forward((x + torch.tensor([[epsilon, 0.00, 0.00]], device=x.device)).clamp(-self.bound, self.bound))
dx_neg, _ = self.common_forward((x + torch.tensor([[-epsilon, 0.00, 0.00]], device=x.device)).clamp(-self.bound, self.bound))
dy_pos, _ = self.common_forward((x + torch.tensor([[0.00, epsilon, 0.00]], device=x.device)).clamp(-self.bound, self.bound))
dy_neg, _ = self.common_forward((x + torch.tensor([[0.00, -epsilon, 0.00]], device=x.device)).clamp(-self.bound, self.bound))
dz_pos, _ = self.common_forward((x + torch.tensor([[0.00, 0.00, epsilon]], device=x.device)).clamp(-self.bound, self.bound))
dz_neg, _ = self.common_forward((x + torch.tensor([[0.00, 0.00, -epsilon]], device=x.device)).clamp(-self.bound, self.bound))
normal = torch.stack([
0.5 * (dx_pos - dx_neg) / epsilon,
0.5 * (dy_pos - dy_neg) / epsilon,
0.5 * (dz_pos - dz_neg) / epsilon
], dim=-1)
return normal
def forward(self, x, d, l=None, ratio=1, shading='albedo'):
# x: [N, 3], in [-bound, bound]
# d: [N, 3], view direction, nomalized in [-1, 1]
# l: [3], plane light direction, nomalized in [-1, 1]
# ratio: scalar, ambient ratio, 1 == no shading (albedo only), 0 == only shading (textureless)
if shading == 'albedo':
# no need to query normal
sigma, color = self.common_forward(x)
normal = None
else:
# query normal
sigma, albedo = self.common_forward(x)
normal = self.finite_differnce_normal(x)
# with torch.enable_grad():
# x.requires_grad_(True)
# sigma, albedo = self.common_forward(x)
# # query gradient
# normal = - torch.autograd.grad(torch.sum(sigma), x, create_graph=True)[0] # [N, 3]
# normalize...
normal = safe_normalize(normal)
normal[torch.isnan(normal)] = 0
# light direction (random if not provided)
if l is None:
l = torch.randn(3, device=x.device, dtype=torch.float)
l = safe_normalize(l)
# lambertian shading
lambertian = ratio + (1 - ratio) * (normal @ -l).clamp(min=0) # [N,]
if shading == 'textureless':
color = lambertian.unsqueeze(-1).repeat(1, 3)
elif shading == 'normal':
color = (normal + 1) / 2
else: # 'lambertian'
color = albedo * lambertian.unsqueeze(-1)
return sigma, color, normal
def density(self, x):
# x: [N, 3], in [-bound, bound]
sigma, albedo = self.common_forward(x)
return {
'sigma': sigma,
'albedo': albedo,
}
def background(self, x, d):
# x: [N, 2], in [-1, 1]
h = self.encoder_bg(x) # [N, C]
h = self.bg_net(h)
# sigmoid activation for rgb
rgbs = torch.sigmoid(h)
return rgbs
# optimizer utils
def get_params(self, lr):
params = [
{'params': self.encoder.parameters(), 'lr': lr * 10},
{'params': self.sigma_net.parameters(), 'lr': lr},
]
if self.bg_radius > 0:
params.append({'params': self.encoder_bg.parameters(), 'lr': lr * 10})
params.append({'params': self.bg_net.parameters(), 'lr': lr})
return params |