def apply(self, rng_0, rng_1, origins, directions, viewdirs,
num_coarse_samples, num_fine_samples, use_viewdirs, near, far,
noise_std, net_depth, net_width, net_depth_condition,
net_width_condition, net_activation, skip_layer,
num_rgb_channels, num_sigma_channels, randomized, white_bkgd,
deg_point, deg_view, lindisp, rgb_activation, sigma_activation,
legacy_posenc_order):
"""Nerf Model.
Args:
rng_0: jnp.ndarray, random number generator for coarse model sampling.
rng_1: jnp.ndarray, random number generator for fine model sampling.
origins: jnp.ndarray(float32), [batch_size, 3], each ray origin.
directions: jnp.ndarray(float32), [batch_size, 3], each ray direction.
viewdirs: jnp.ndarray(float32), [batch_size, 3], the viewing direction for
each ray. This is only used if NDC rays are used, as otherwise
`directions` is equal to viewdirs.
num_coarse_samples: int, the number of samples for coarse nerf.
num_fine_samples: int, the number of samples for fine nerf.
use_viewdirs: bool, use viewdirs as a condition.
near: float, near clip.
far: float, far clip.
noise_std: float, std dev of noise added to regularize sigma output.
net_depth: int, the depth of the first part of MLP.
net_width: int, the width of the first part of MLP.
net_depth_condition: int, the depth of the second part of MLP.
net_width_condition: int, the width of the second part of MLP.
net_activation: function, the activation function used within the MLP.
skip_layer: int, add a skip connection to the output vector of every
skip_layer layers.
num_rgb_channels: int, the number of RGB channels.
num_sigma_channels: int, the number of density channels.
randomized: bool, use randomized stratified sampling.
white_bkgd: bool, use white background.
deg_point: degree of positional encoding for positions.
deg_view: degree of positional encoding for viewdirs.
lindisp: bool, sampling linearly in disparity rather than depth if true.
rgb_activation: function, the activation used to generate RGB.
sigma_activation: function, the activation used to generate density.
legacy_posenc_order: bool, keep the same ordering as the original tf code.
Returns:
ret: list, [(rgb_coarse, disp_coarse, acc_coarse), (rgb, disp, acc)]
"""
mlp_fn = functools.partial(model_utils.MLP,
net_depth=net_depth,
net_width=net_width,
net_depth_condition=net_depth_condition,
net_width_condition=net_width_condition,
net_activation=net_activation,
skip_layer=skip_layer,
num_rgb_channels=num_rgb_channels,
num_sigma_channels=num_sigma_channels)
# Stratified sampling along rays
key, rng_0 = random.split(rng_0)
z_vals, samples = model_utils.sample_along_rays(
key, origins, directions, num_coarse_samples, near, far,
randomized, lindisp)
samples_enc = model_utils.posenc(samples, deg_point,
legacy_posenc_order)
# Point attribute predictions
if use_viewdirs:
viewdirs_enc = model_utils.posenc(
viewdirs / jnp.linalg.norm(viewdirs, axis=-1, keepdims=True),
deg_view, legacy_posenc_order)
raw_rgb, raw_sigma = mlp_fn(samples_enc, viewdirs_enc)
else:
raw_rgb, raw_sigma = mlp_fn(samples_enc)
# Add noises to regularize the density predictions if needed
key, rng_0 = random.split(rng_0)
raw_sigma = model_utils.add_gaussian_noise(key, raw_sigma, noise_std,
randomized)
rgb = rgb_activation(raw_rgb)
sigma = sigma_activation(raw_sigma)
# Volumetric rendering.
comp_rgb, disp, acc, weights = model_utils.volumetric_rendering(
rgb,
sigma,
z_vals,
directions,
white_bkgd=white_bkgd,
)
ret = [
(comp_rgb, disp, acc),
]
# Hierarchical sampling based on coarse predictions
if num_fine_samples > 0:
z_vals_mid = .5 * (z_vals[Ellipsis, 1:] + z_vals[Ellipsis, :-1])
key, rng_1 = random.split(rng_1)
z_vals, samples = model_utils.sample_pdf(
key,
z_vals_mid,
weights[Ellipsis, 1:-1],
origins,
directions,
z_vals,
num_fine_samples,
randomized,
)
samples_enc = model_utils.posenc(samples, deg_point,
legacy_posenc_order)
if use_viewdirs:
raw_rgb, raw_sigma = mlp_fn(samples_enc, viewdirs_enc)
else:
raw_rgb, raw_sigma = mlp_fn(samples_enc)
key, rng_1 = random.split(rng_1)
raw_sigma = model_utils.add_gaussian_noise(key, raw_sigma,
noise_std, randomized)
rgb = rgb_activation(raw_rgb)
sigma = sigma_activation(raw_sigma)
comp_rgb, disp, acc, unused_weights = model_utils.volumetric_rendering(
rgb,
sigma,
z_vals,
directions,
white_bkgd=white_bkgd,
)
ret.append((comp_rgb, disp, acc))
return ret
def __call__(self, rng_0, rng_1, rays, randomized):
"""Nerf Model.
Args:
rng_0: jnp.ndarray, random number generator for coarse model sampling.
rng_1: jnp.ndarray, random number generator for fine model sampling.
rays: util.Rays, a namedtuple of ray origins, directions, and viewdirs.
randomized: bool, use randomized stratified sampling.
Returns:
ret: list, [(rgb_coarse, disp_coarse, acc_coarse), (rgb, disp, acc)]
"""
# Stratified sampling along rays
key, rng_0 = random.split(rng_0)
z_vals, samples = model_utils.sample_along_rays(
key,
rays.origins,
rays.directions,
self.num_coarse_samples,
self.near,
self.far,
randomized,
self.lindisp,
)
samples_enc = model_utils.posenc(
samples,
self.min_deg_point,
self.max_deg_point,
self.legacy_posenc_order,
)
# Construct the "coarse" MLP.
coarse_mlp = model_utils.MLP(
net_depth=self.net_depth,
net_width=self.net_width,
net_depth_condition=self.net_depth_condition,
net_width_condition=self.net_width_condition,
net_activation=self.net_activation,
skip_layer=self.skip_layer,
num_rgb_channels=self.num_rgb_channels,
num_sigma_channels=self.num_sigma_channels)
# Point attribute predictions
if self.use_viewdirs:
viewdirs_enc = model_utils.posenc(
rays.viewdirs,
0,
self.deg_view,
self.legacy_posenc_order,
)
raw_rgb, raw_sigma = coarse_mlp(samples_enc, viewdirs_enc)
else:
raw_rgb, raw_sigma = coarse_mlp(samples_enc)
# Add noises to regularize the density predictions if needed
key, rng_0 = random.split(rng_0)
raw_sigma = model_utils.add_gaussian_noise(
key,
raw_sigma,
self.noise_std,
randomized,
)
rgb = self.rgb_activation(raw_rgb)
sigma = self.sigma_activation(raw_sigma)
# Volumetric rendering.
comp_rgb, disp, acc, weights = model_utils.volumetric_rendering(
rgb,
sigma,
z_vals,
rays.directions,
white_bkgd=self.white_bkgd,
)
ret = [
(comp_rgb, disp, acc),
]
# Hierarchical sampling based on coarse predictions
if self.num_fine_samples > 0:
z_vals_mid = .5 * (z_vals[Ellipsis, 1:] + z_vals[Ellipsis, :-1])
key, rng_1 = random.split(rng_1)
z_vals, samples = model_utils.sample_pdf(
key,
z_vals_mid,
weights[Ellipsis, 1:-1],
rays.origins,
rays.directions,
z_vals,
self.num_fine_samples,
randomized,
)
samples_enc = model_utils.posenc(
samples,
self.min_deg_point,
self.max_deg_point,
self.legacy_posenc_order,
)
# Construct the "fine" MLP.
fine_mlp = model_utils.MLP(
net_depth=self.net_depth,
net_width=self.net_width,
net_depth_condition=self.net_depth_condition,
net_width_condition=self.net_width_condition,
net_activation=self.net_activation,
skip_layer=self.skip_layer,
num_rgb_channels=self.num_rgb_channels,
num_sigma_channels=self.num_sigma_channels)
if self.use_viewdirs:
raw_rgb, raw_sigma = fine_mlp(samples_enc, viewdirs_enc)
else:
raw_rgb, raw_sigma = fine_mlp(samples_enc)
key, rng_1 = random.split(rng_1)
raw_sigma = model_utils.add_gaussian_noise(
key,
raw_sigma,
self.noise_std,
randomized,
)
rgb = self.rgb_activation(raw_rgb)
sigma = self.sigma_activation(raw_sigma)
comp_rgb, disp, acc, unused_weights = model_utils.volumetric_rendering(
rgb,
sigma,
z_vals,
rays.directions,
white_bkgd=self.white_bkgd,
)
ret.append((comp_rgb, disp, acc))
return ret
def apply(self, rng_0, rng_1, rays, num_coarse_samples, num_fine_samples,
use_viewdirs, near, far, noise_std, net_depth, net_width,
net_depth_condition, net_width_condition, net_activation,
skip_layer, num_rgb_channels, num_sigma_channels, randomized,
white_bkgd, deg_point, deg_view, lindisp, rgb_activation,
sigma_activation):
"""Nerf Model.
Args:
rng_0: jnp.ndarray, random number generator for coarse model sampling.
rng_1: jnp.ndarray, random number generator for fine model sampling.
rays: jnp.ndarray(float32), [batch_size, 6/9], each ray is a 6-d vector
where the first 3 dimensions represent the ray origin and the last 3
dimensions represent the unormalized ray direction. Note that if ndc
rays are used, rays are 9-d where the extra 3-dimensional vector is the
view direction before transformed to ndc rays.
num_coarse_samples: int, the number of samples for coarse nerf.
num_fine_samples: int, the number of samples for fine nerf.
use_viewdirs: bool, use viewdirs as a condition.
near: float, near clip.
far: float, far clip.
noise_std: float, std dev of noise added to regularize sigma output.
net_depth: int, the depth of the first part of MLP.
net_width: int, the width of the first part of MLP.
net_depth_condition: int, the depth of the second part of MLP.
net_width_condition: int, the width of the second part of MLP.
net_activation: function, the activation function used within the MLP.
skip_layer: int, add a skip connection to the output vector of every
skip_layer layers.
num_rgb_channels: int, the number of RGB channels.
num_sigma_channels: int, the number of density channels.
randomized: bool, use randomized stratified sampling.
white_bkgd: bool, use white background.
deg_point: degree of positional encoding for positions.
deg_view: degree of positional encoding for viewdirs.
lindisp: bool, sampling linearly in disparity rather than depth if true.
rgb_activation: function, the activation used to generate RGB.
sigma_activation: function, the activation used to generate density.
Returns:
ret: list, [(rgb_coarse, disp_coarse, acc_coarse), (rgb, disp, acc)]
"""
# Extract viewdirs from the ray array
if rays.shape[-1] > 6: # viewdirs different from rays_d
viewdirs = rays[Ellipsis, -3:]
rays = rays[Ellipsis, :-3]
else: # viewdirs are normalized rays_d
viewdirs = rays[Ellipsis, 3:6]
# Stratified sampling along rays
key, rng_0 = random.split(rng_0)
z_vals, samples = model_utils.sample_along_rays(
key, rays, num_coarse_samples, near, far, randomized, lindisp)
samples = model_utils.posenc(samples, deg_point)
# Point attribute predictions
if use_viewdirs:
norms = jnp.linalg.norm(viewdirs, axis=-1, keepdims=True)
viewdirs = viewdirs / norms
viewdirs = model_utils.posenc(viewdirs, deg_view)
raw_rgb, raw_sigma = model_utils.MLP(
samples,
viewdirs,
net_depth=net_depth,
net_width=net_width,
net_depth_condition=net_depth_condition,
net_width_condition=net_width_condition,
net_activation=net_activation,
skip_layer=skip_layer,
num_rgb_channels=num_rgb_channels,
num_sigma_channels=num_sigma_channels,
)
else:
raw_rgb, raw_sigma = model_utils.MLP(
samples,
net_depth=net_depth,
net_width=net_width,
net_depth_condition=net_depth_condition,
net_width_condition=net_width_condition,
net_activation=net_activation,
skip_layer=skip_layer,
num_rgb_channels=num_rgb_channels,
num_sigma_channels=num_sigma_channels,
)
# Add noises to regularize the density predictions if needed
key, rng_0 = random.split(rng_0)
raw_sigma = model_utils.add_gaussian_noise(key, raw_sigma, noise_std,
randomized)
rgb = rgb_activation(raw_rgb)
sigma = sigma_activation(raw_sigma)
# Volumetric rendering.
comp_rgb, disp, acc, weights = model_utils.volumetric_rendering(
rgb,
sigma,
z_vals,
rays[Ellipsis, 3:6],
white_bkgd=white_bkgd,
)
ret = [
(comp_rgb, disp, acc),
]
# Hierarchical sampling based on coarse predictions
if num_fine_samples > 0:
z_vals_mid = .5 * (z_vals[Ellipsis, 1:] + z_vals[Ellipsis, :-1])
key, rng_1 = random.split(rng_1)
z_vals, samples = model_utils.sample_pdf(
key,
z_vals_mid,
weights[Ellipsis, 1:-1],
rays,
z_vals,
num_fine_samples,
randomized,
)
samples = model_utils.posenc(samples, deg_point)
if use_viewdirs:
raw_rgb, raw_sigma = model_utils.MLP(samples, viewdirs)
else:
raw_rgb, raw_sigma = model_utils.MLP(samples)
key, rng_1 = random.split(rng_1)
raw_sigma = model_utils.add_gaussian_noise(key, raw_sigma,
noise_std, randomized)
rgb = rgb_activation(raw_rgb)
sigma = sigma_activation(raw_sigma)
comp_rgb, disp, acc, unused_weights = model_utils.volumetric_rendering(
rgb,
sigma,
z_vals,
rays[Ellipsis, 3:6],
white_bkgd=white_bkgd,
)
ret.append((comp_rgb, disp, acc))
return ret