misc proposal network modifications (#578)

* misc improvements

* cleanup

* remove distortion

* dist loss

* only dist changed now

* changes

* remove comment
This commit is contained in:
Ethan Weber authored and GitHub committed 2022-09-29 16:31:22 -07:00
1 parent 5bc011e5fa
commit 81fa5f9ba5
7 files changed
+100 -28

No files matched your search

+11 -3
View File
@@ -70,6 +70,7 @@ class TCNNCompoundField(Field):
num_layers_color: number of hidden layers for color network
hidden_dim_color: dimension of hidden layers for color network
appearance_embedding_dim: dimension of appearance embedding
use_average_appearance_embedding: whether to use average appearance embedding or zeros for inference
spatial_distortion: spatial distortion to apply to the scene
"""
@@ -83,6 +84,7 @@ class TCNNCompoundField(Field):
num_layers_color: int = 3,
hidden_dim_color: int = 64,
appearance_embedding_dim: int = 32,
use_average_appearance_embedding: bool = False,
spatial_distortion: Optional[SpatialDistortion] = None,
) -> None:
super().__init__()
@@ -94,6 +96,7 @@ class TCNNCompoundField(Field):
self.num_images = num_images
self.appearance_embedding_dim = appearance_embedding_dim
self.embedding_appearance = Embedding(self.num_images, self.appearance_embedding_dim)
self.use_average_appearance_embedding = use_average_appearance_embedding
num_levels = 16
max_res = 1024
@@ -171,9 +174,14 @@ class TCNNCompoundField(Field):
if self.training:
embedded_appearance = self.embedding_appearance(camera_indices)
else:
embedded_appearance = torch.zeros(
(*directions.shape[:-1], self.appearance_embedding_dim), device=directions.device
)
if self.use_average_appearance_embedding:
embedded_appearance = torch.ones(
(*directions.shape[:-1], self.appearance_embedding_dim), device=directions.device
) * self.embedding_appearance.mean(dim=0)
else:
embedded_appearance = torch.zeros(
(*directions.shape[:-1], self.appearance_embedding_dim), device=directions.device
)
if density_embedding is None:
positions = SceneBounds.get_normalized_positions(ray_samples.frustums.get_positions(), self.aabb)
+4
View File
@@ -41,6 +41,10 @@ class Embedding(FieldModule):
def build_nn_modules(self) -> None:
self.embedding = torch.nn.Embedding(self.in_dim, self.out_dim)
def mean(self, dim=0):
"""Return the mean of the embedding weights along a dim."""
return self.embedding.weight.mean(dim)
def forward(self, in_tensor: TensorType[..., "input_dim"]) -> TensorType[..., "output_dim"]:
"""Call forward
+2 -4
View File
@@ -24,7 +24,6 @@ def distortion_loss(
ray_samples: RaySamples,
densities: TensorType["bs":..., "num_samples", 1] = None,
weights: TensorType["bs":..., "num_samples", 1] = None,
scale_factor: float = 1.0,
) -> TensorType["bs":..., 1]:
"""Ray based distortion loss proposed in MipNeRF-360. Returns distortion Loss.
@@ -40,7 +39,6 @@ def distortion_loss(
ray_samples: Ray samples to compute loss over
densities: Predicted sample densities
weights: Predicted weights from densities and sample locations
scale_factor: Scale the starts and ends by this factor
"""
if torch.is_tensor(densities):
assert not torch.is_tensor(weights), "Cannot use both densities and weights"
@@ -49,8 +47,8 @@ def distortion_loss(
if torch.is_tensor(weights):
assert not torch.is_tensor(densities), "Cannot use both densities and weights"
starts = ray_samples.frustums.starts * scale_factor
ends = ray_samples.frustums.ends * scale_factor
starts = ray_samples.spacing_starts
ends = ray_samples.spacing_ends
midpoints = (starts + ends) / 2.0 # (..., num_samples, 1)
+20 -4
View File
@@ -82,11 +82,13 @@ class SpacedSampler(Sampler):
spacing_fn_inv: Callable,
num_samples: Optional[int] = None,
train_stratified=True,
single_jitter=False,
density_field: Optional[DensityGrid] = None,
weight_threshold: float = 1e-2,
) -> None:
super().__init__(num_samples=num_samples, density_field=density_field, weight_threshold=weight_threshold)
self.train_stratified = train_stratified
self.single_jitter = single_jitter
self.spacing_fn = spacing_fn
self.spacing_fn_inv = spacing_fn_inv
@@ -117,7 +119,10 @@ class SpacedSampler(Sampler):
# TODO More complicated than it needs to be.
if self.train_stratified and self.training:
t_rand = torch.rand((num_rays, num_samples + 1), dtype=bins.dtype, device=bins.device)
if self.single_jitter:
t_rand = torch.rand((num_rays, 1), dtype=bins.dtype, device=bins.device)
else:
t_rand = torch.rand((num_rays, num_samples + 1), dtype=bins.dtype, device=bins.device)
bin_centers = (bins[..., 1:] + bins[..., :-1]) / 2.0
bin_upper = torch.cat([bin_centers, bins[..., -1:]], -1)
bin_lower = torch.cat([bins[..., :1], bin_centers], -1)
@@ -152,6 +157,7 @@ class UniformSampler(SpacedSampler):
self,
num_samples: Optional[int] = None,
train_stratified=True,
single_jitter=False,
density_field: Optional[DensityGrid] = None,
weight_threshold: float = 1e-2,
) -> None:
@@ -160,6 +166,7 @@ class UniformSampler(SpacedSampler):
spacing_fn=lambda x: x,
spacing_fn_inv=lambda x: x,
train_stratified=train_stratified,
single_jitter=single_jitter,
density_field=density_field,
weight_threshold=weight_threshold,
)
@@ -179,6 +186,7 @@ class LinearDisparitySampler(SpacedSampler):
self,
num_samples: Optional[int] = None,
train_stratified=True,
single_jitter=False,
density_field: Optional[DensityGrid] = None,
weight_threshold: float = 1e-2,
) -> None:
@@ -187,6 +195,7 @@ class LinearDisparitySampler(SpacedSampler):
spacing_fn=lambda x: 1 / x,
spacing_fn_inv=lambda x: 1 / x,
train_stratified=train_stratified,
single_jitter=single_jitter,
density_field=density_field,
weight_threshold=weight_threshold,
)
@@ -206,6 +215,7 @@ class SqrtSampler(SpacedSampler):
self,
num_samples: Optional[int] = None,
train_stratified=True,
single_jitter=False,
density_field: Optional[DensityGrid] = None,
weight_threshold: float = 1e-2,
) -> None:
@@ -214,6 +224,7 @@ class SqrtSampler(SpacedSampler):
spacing_fn=torch.sqrt,
spacing_fn_inv=lambda x: x**2,
train_stratified=train_stratified,
single_jitter=single_jitter,
density_field=density_field,
weight_threshold=weight_threshold,
)
@@ -233,6 +244,7 @@ class LogSampler(SpacedSampler):
self,
num_samples: Optional[int] = None,
train_stratified=True,
single_jitter=False,
density_field: Optional[DensityGrid] = None,
weight_threshold: float = 1e-2,
) -> None:
@@ -241,6 +253,7 @@ class LogSampler(SpacedSampler):
spacing_fn=torch.log,
spacing_fn_inv=torch.exp,
train_stratified=train_stratified,
single_jitter=single_jitter,
density_field=density_field,
weight_threshold=weight_threshold,
)
@@ -262,6 +275,7 @@ class UniformLinDispPiecewiseSampler(SpacedSampler):
self,
num_samples: Optional[int] = None,
train_stratified=True,
single_jitter=False,
density_field: Optional[DensityGrid] = None,
weight_threshold: float = 1e-2,
) -> None:
@@ -270,6 +284,7 @@ class UniformLinDispPiecewiseSampler(SpacedSampler):
spacing_fn=lambda x: torch.where(x < 1, x / 2, 1 - 1 / (2 * x)),
spacing_fn_inv=lambda x: torch.where(x < 0.5, 2 * x, 1 / (2 - 2 * x)),
train_stratified=train_stratified,
single_jitter=single_jitter,
density_field=density_field,
weight_threshold=weight_threshold,
)
@@ -531,9 +546,10 @@ class ProposalNetworkSampler(Sampler):
def __init__(
self,
num_proposal_samples_per_ray: int = 64,
num_proposal_samples_per_ray: Tuple[int] = (64,),
num_nerf_samples_per_ray: int = 32,
num_proposal_network_iterations: int = 2,
single_jitter: bool = True,
density_field: Optional[DensityGrid] = None,
weight_threshold: float = 1e-2,
) -> None:
@@ -543,7 +559,7 @@ class ProposalNetworkSampler(Sampler):
self.num_proposal_network_iterations = num_proposal_network_iterations
# samplers
self.initial_sampler = UniformLinDispPiecewiseSampler()
self.initial_sampler = UniformLinDispPiecewiseSampler(single_jitter=single_jitter)
self.pdf_sampler = PDFSampler(include_original=False)
self._anneal = 1.0
@@ -566,7 +582,7 @@ class ProposalNetworkSampler(Sampler):
n = self.num_proposal_network_iterations
for i_level in range(n + 1):
is_prop = i_level < n
num_samples = self.num_proposal_samples_per_ray if is_prop else self.num_nerf_samples_per_ray
num_samples = self.num_proposal_samples_per_ray[i_level] if is_prop else self.num_nerf_samples_per_ray
if i_level == 0:
# Uniform sampling because we need to start with some samples
ray_samples = self.initial_sampler(ray_bundle, num_samples=num_samples)
+14 -12
View File
@@ -34,10 +34,9 @@ from nerfactory.fields.density_field import DensityField
from nerfactory.fields.modules.field_heads import FieldHeadNames
from nerfactory.fields.modules.spatial_distortions import SceneContraction
from nerfactory.models.base import Model, ModelConfig
from nerfactory.models.modules.ray_losses import distortion_loss
from nerfactory.models.modules.ray_sampler import ProposalNetworkSampler
from nerfactory.models.modules.scene_colliders import NearFarCollider
from nerfactory.optimizers.loss import MSELoss, interlevel_loss
from nerfactory.optimizers.loss import MSELoss, distortion_loss, interlevel_loss
from nerfactory.renderers.renderers import (
AccumulationRenderer,
DepthRenderer,
@@ -62,7 +61,7 @@ class ProposalModelConfig(ModelConfig):
"""How far along the ray to stop sampling."""
randomize_background: bool = True
"""Whether to randomize the background color."""
num_proposal_samples_per_ray: int = 64
num_proposal_samples_per_ray: Tuple[int] = (64,)
"""Number of samples per ray for the proposal network."""
num_nerf_samples_per_ray: int = 64
"""Number of samples per ray for the nerf network."""
@@ -72,16 +71,20 @@ class ProposalModelConfig(ModelConfig):
"""Use the same proposal network. Otherwise use different ones."""
interlevel_loss_mult: float = 1.0
"""Proposal loss multiplier."""
distortion_loss_mult: float = 0.01
distortion_loss_mult: float = 0.001
"""Distortion loss multiplier."""
use_appearance_conditioning: bool = True
"""Whether to use appearance conditioning."""
use_proposal_weight_anneal: bool = True
"""Whether to use proposal weight annealing."""
use_average_appearance_embedding: bool = True
"""Whether to use average appearance embedding or zeros for inference."""
proposal_weights_anneal_slope: float = 10.0
"""Slope of the annealing function for the proposal weights."""
proposal_weights_anneal_max_num_iters: int = 1000
"""Max num iterations for the annealing function."""
use_single_jitter: bool = False
"""Whether use single jitter or not for first proposal network."""
class ProposalModel(Model):
@@ -102,7 +105,10 @@ class ProposalModel(Model):
# Fields
if self.config.use_appearance_conditioning:
self.field = TCNNCompoundField(
self.scene_bounds.aabb, spatial_distortion=scene_contraction, num_images=self.num_train_data
self.scene_bounds.aabb,
spatial_distortion=scene_contraction,
num_images=self.num_train_data,
use_average_appearance_embedding=self.config.use_average_appearance_embedding,
)
else:
raise NotImplementedError("Only appearance conditioning is supported.")
@@ -127,6 +133,7 @@ class ProposalModel(Model):
num_nerf_samples_per_ray=self.config.num_nerf_samples_per_ray,
num_proposal_samples_per_ray=self.config.num_proposal_samples_per_ray,
num_proposal_network_iterations=self.config.num_proposal_network_iterations,
single_jitter=self.config.use_single_jitter,
)
# renderers
@@ -197,6 +204,7 @@ class ProposalModel(Model):
metrics_dict = {}
image = batch["image"].to(self.device)
metrics_dict["psnr"] = self.psnr(outputs["rgb"], image)
metrics_dict["distortion"] = distortion_loss(outputs["weights_list"], outputs["ray_samples_list"])
return metrics_dict
def get_loss_dict(self, outputs, batch, metrics_dict=None):
@@ -206,13 +214,7 @@ class ProposalModel(Model):
loss_dict["interlevel_loss"] = self.config.interlevel_loss_mult * interlevel_loss(
outputs["weights_list"], outputs["ray_samples_list"]
)
loss_dict["distortion_loss"] = self.config.distortion_loss_mult * torch.mean(
distortion_loss(
outputs["ray_samples_list"][-1],
weights=outputs["weights_list"][-1],
scale_factor=1.0 / (self.config.far_plane - self.config.near_plane),
)
)
loss_dict["distortion_loss"] = self.config.distortion_loss_mult * metrics_dict["distortion"]
return loss_dict
def get_image_metrics_and_images(
+47 -3
View File
@@ -18,6 +18,7 @@ Collection of Losses.
import torch
from torch import nn
from torchtyping import TensorType
L1Loss = nn.L1Loss
MSELoss = nn.MSELoss
@@ -27,10 +28,12 @@ LOSSES = {"L1": L1Loss, "MSE": MSELoss}
EPS = 1.0e-7
def searchsorted(a, v):
def searchsorted(a: TensorType[..., "N"], v: TensorType[..., "M"]) -> TensorType[..., "M"]:
"""
https://github.com/kakaobrain/NeRF-Factory/blob/f61bb8744a5cb4820a4d968fb3bfbed777550f4a/src/model/mipnerf360/helper.py#L109
https://github.com/google-research/multinerf/blob/b02228160d3179300c7d499dca28cb9ca3677f32/internal/stepfun.py#L30
"""
i = torch.arange(a.shape[-1], device=a.device)
v_ge_a = v[..., None, :] >= a[..., :, None]
@@ -39,10 +42,17 @@ def searchsorted(a, v):
return idx_lo, idx_hi
def inner_outer(t0, t1, y1):
def inner_outer(
t0: TensorType[..., "num_samples+1"], t1: TensorType[..., "num_samples+1"], y1: TensorType[..., "num_samples"]
):
"""
https://github.com/kakaobrain/NeRF-Factory/blob/f61bb8744a5cb4820a4d968fb3bfbed777550f4a/src/model/mipnerf360/helper.py#L117
https://github.com/google-research/multinerf/blob/b02228160d3179300c7d499dca28cb9ca3677f32/internal/stepfun.py#L64
Args:
t0: interval edges
t1: interval edges
y1: weights
"""
cy1 = torch.cat([torch.zeros_like(y1[..., :1]), torch.cumsum(y1, dim=-1)], dim=-1)
@@ -61,10 +71,21 @@ def inner_outer(t0, t1, y1):
return y0_inner, y0_outer
def lossfun_outer(t, w, t_env, w_env):
def lossfun_outer(
t: TensorType[..., "num_samples+1"],
w: TensorType[..., "num_samples"],
t_env: TensorType[..., "num_samples+1"],
w_env: TensorType[..., "num_samples"],
):
"""
https://github.com/kakaobrain/NeRF-Factory/blob/f61bb8744a5cb4820a4d968fb3bfbed777550f4a/src/model/mipnerf360/helper.py#L136
https://github.com/google-research/multinerf/blob/b02228160d3179300c7d499dca28cb9ca3677f32/internal/stepfun.py#L80
Args:
t: interval edges
w: weights
t_env: interval edges of the upper bound enveloping historgram
w_env: weights that should upper bound the inner (t,w) histogram
"""
_, w_outer = inner_outer(t, t_env, w_env)
return torch.clip(w - w_outer, min=0) ** 2 / (w + EPS)
@@ -93,3 +114,26 @@ def interlevel_loss(weights_list, ray_samples_list):
wp = weights[..., 0] # (num_rays, num_samples)
loss_interlevel += torch.mean(lossfun_outer(c, w, cp, wp))
return loss_interlevel
# Verified
def lossfun_distortion(t, w):
"""
https://github.com/kakaobrain/NeRF-Factory/blob/f61bb8744a5cb4820a4d968fb3bfbed777550f4a/src/model/mipnerf360/helper.py#L142
https://github.com/google-research/multinerf/blob/b02228160d3179300c7d499dca28cb9ca3677f32/internal/stepfun.py#L266
"""
ut = (t[..., 1:] + t[..., :-1]) / 2
dut = torch.abs(ut[..., :, None] - ut[..., None, :])
loss_inter = torch.sum(w * torch.sum(w[..., None, :] * dut, dim=-1), dim=-1)
loss_intra = torch.sum(w**2 * (t[..., 1:] - t[..., :-1]), dim=-1) / 3
return loss_inter + loss_intra
def distortion_loss(weights_list, ray_samples_list):
"""From mipnerf360"""
c = ray_samples_to_sdist(ray_samples_list[-1])
w = weights_list[-1][..., 0]
loss = torch.mean(lossfun_distortion(c, w))
return loss
+2 -2
View File
@@ -131,8 +131,8 @@ class Pipeline(nn.Module):
self.datamanager.eval_sampler.set_epoch(step)
ray_bundle, batch = self.datamanager.next_eval(step)
model_outputs = self.model(ray_bundle, batch)
loss_dict = self.model.get_loss_dict(model_outputs, batch)
metrics_dict = self.model.get_metrics_dict(model_outputs, batch)
loss_dict = self.model.get_loss_dict(model_outputs, batch, metrics_dict)
self.train()
return model_outputs, loss_dict, metrics_dict
@@ -271,8 +271,8 @@ class VanillaPipeline(Pipeline):
if "valid_mask" in model_outputs:
valid_mask = model_outputs["valid_mask"]
batch = get_masked_dict(batch, valid_mask)
loss_dict = self.model.get_loss_dict(model_outputs, batch)
metrics_dict = self.model.get_metrics_dict(model_outputs, batch)
loss_dict = self.model.get_loss_dict(model_outputs, batch, metrics_dict)
self.train()
return model_outputs, loss_dict, metrics_dict