Update NGP related models with nerfacc 0.5.2 (#1809)

* profiling

* set pose_optimizer to None if it's mode is "off"

* radial_and_tangential_undistort in nerfacc

* fix camera distortions

* update to nerfacc 0.5.1. Upgrade instant-ngp model.

* revert nerfacc to 0.5.1

* format

* allows to fall back to the pytorch undistortion

* format

* fix linting

* fix test

* fix test

* fix nerfplayer-ngp

* enable_collider=False for NGP

* nerfplayer verified.

* zero cone_angle for ngp bounded

* fix format

* fix pylint
This commit is contained in:
Ruilong Li(李瑞龙) authored and GitHub committed 2023-04-24 17:11:30 -07:00
1 parent c6abd3abcb
commit ea2993d5b1
12 files changed
+182 -135

No files matched your search

+23 -29
View File
@@ -21,7 +21,6 @@ from typing import Dict, Optional, Tuple
import numpy as np
import torch
from nerfacc import ContractionType, contract
from torch.nn.parameter import Parameter
from torchtyping import TensorType
@@ -30,6 +29,10 @@ from nerfstudio.data.scene_box import SceneBox
from nerfstudio.field_components.activations import trunc_exp
from nerfstudio.field_components.embedding import Embedding
from nerfstudio.field_components.field_heads import FieldHeadNames
from nerfstudio.field_components.spatial_distortions import (
SceneContraction,
SpatialDistortion,
)
from nerfstudio.fields.base_field import Field, shift_directions_for_tcnn
try:
@@ -69,16 +72,16 @@ class TCNNInstantNGPField(Field):
use_appearance_embedding: Optional[bool] = False,
num_images: Optional[int] = None,
appearance_embedding_dim: int = 32,
contraction_type: ContractionType = ContractionType.UN_BOUNDED_SPHERE,
num_levels: int = 16,
log2_hashmap_size: int = 19,
max_res: int = 2048,
spatial_distortion: Optional[SpatialDistortion] = SceneContraction(),
) -> None:
super().__init__()
self.aabb = Parameter(aabb, requires_grad=False)
self.geo_feat_dim = geo_feat_dim
self.contraction_type = contraction_type
self.spatial_distortion = spatial_distortion
self.use_appearance_embedding = use_appearance_embedding
if use_appearance_embedding:
@@ -88,7 +91,8 @@ class TCNNInstantNGPField(Field):
# TODO: set this properly based on the aabb
base_res: int = 16
per_level_scale = np.exp((np.log(max_res) - np.log(base_res)) / (num_levels - 1))
features_per_level: int = 2
growth_factor = np.exp((np.log(max_res) - np.log(base_res)) / (num_levels - 1))
self.direction_encoding = tcnn.Encoding(
n_input_dims=3,
@@ -104,10 +108,10 @@ class TCNNInstantNGPField(Field):
encoding_config={
"otype": "HashGrid",
"n_levels": num_levels,
"n_features_per_level": 2,
"n_features_per_level": features_per_level,
"log2_hashmap_size": log2_hashmap_size,
"base_resolution": base_res,
"per_level_scale": per_level_scale,
"per_level_scale": growth_factor,
},
network_config={
"otype": "FullyFusedMLP",
@@ -134,9 +138,16 @@ class TCNNInstantNGPField(Field):
)
def get_density(self, ray_samples: RaySamples) -> Tuple[TensorType, TensorType]:
positions = ray_samples.frustums.get_positions()
if self.spatial_distortion is not None:
positions = ray_samples.frustums.get_positions()
positions = self.spatial_distortion(positions)
positions = (positions + 2.0) / 4.0
else:
positions = SceneBox.get_normalized_positions(ray_samples.frustums.get_positions(), self.aabb)
# Make sure the tcnn gets inputs between 0 and 1.
selector = ((positions > 0.0) & (positions < 1.0)).all(dim=-1)
positions = positions * selector[..., None]
positions_flat = positions.view(-1, 3)
positions_flat = contract(x=positions_flat, roi=self.aabb, type=self.contraction_type)
h = self.mlp_base(positions_flat).view(*ray_samples.frustums.shape, -1)
density_before_activation, base_mlp_out = torch.split(h, [1, self.geo_feat_dim], dim=-1)
@@ -145,20 +156,18 @@ class TCNNInstantNGPField(Field):
# softplus, because it enables high post-activation (float32) density outputs
# from smaller internal (float16) parameters.
density = trunc_exp(density_before_activation.to(positions))
density = density * selector[..., None]
return density, base_mlp_out
def get_outputs(
self, ray_samples: RaySamples, density_embedding: Optional[TensorType] = None
) -> Dict[FieldHeadNames, TensorType]:
assert density_embedding is not None
directions = shift_directions_for_tcnn(ray_samples.frustums.directions)
directions_flat = directions.view(-1, 3)
d = self.direction_encoding(directions_flat)
if density_embedding is None:
positions = SceneBox.get_normalized_positions(ray_samples.frustums.get_positions(), self.aabb)
h = torch.cat([d, positions.view(-1, 3)], dim=-1)
else:
h = torch.cat([d, density_embedding.view(-1, self.geo_feat_dim)], dim=-1)
h = torch.cat([d, density_embedding.view(-1, self.geo_feat_dim)], dim=-1)
if self.use_appearance_embedding:
if ray_samples.camera_indices is None:
@@ -183,20 +192,5 @@ class TCNNInstantNGPField(Field):
step_size: the step size to use for the opacity evaluation.
"""
density = self.density_fn(positions)
## TODO: We should scale step size based on the distortion. Currently it uses too much memory.
# aabb_min, aabb_max = self.aabb[0], self.aabb[1]
# if self.contraction_type is not ContractionType.AABB:
# x = (positions - aabb_min) / (aabb_max - aabb_min)
# x = x * 2 - 1 # aabb is at [-1, 1]
# mag = x.norm(dim=-1, keepdim=True)
# mask = mag.squeeze(-1) > 1
# dev = (2 * mag - 1) / mag**2 + 2 * x**2 * (1 / mag**3 - (2 * mag - 1) / mag**4)
# dev[~mask] = 1.0
# dev = torch.clamp(dev, min=1e-6)
# step_size = step_size / dev.norm(dim=-1, keepdim=True)
# else:
# step_size = step_size * (aabb_max - aabb_min)
opacity = density * step_size
return opacity
+1 -1
View File
@@ -103,7 +103,7 @@ class TCNNNerfactoField(Field):
pass_semantic_gradients: bool = False,
use_pred_normals: bool = False,
use_average_appearance_embedding: bool = False,
spatial_distortion: SpatialDistortion = None,
spatial_distortion: Optional[SpatialDistortion] = None,
) -> None:
super().__init__()
+19 -15
View File
@@ -20,7 +20,6 @@ NeRFPlayer (https://arxiv.org/abs/2210.15947) field implementations with Instant
from typing import Dict, Optional, Tuple
import torch
from nerfacc import ContractionType, contract
from torch.nn.parameter import Parameter
from torchtyping import TensorType
@@ -29,6 +28,10 @@ from nerfstudio.data.scene_box import SceneBox
from nerfstudio.field_components.activations import trunc_exp
from nerfstudio.field_components.embedding import Embedding
from nerfstudio.field_components.field_heads import FieldHeadNames
from nerfstudio.field_components.spatial_distortions import (
SceneContraction,
SpatialDistortion,
)
from nerfstudio.field_components.temporal_grid import TemporalGridEncoder
from nerfstudio.fields.base_field import Field, shift_directions_for_tcnn
@@ -82,13 +85,13 @@ class NerfplayerNGPField(Field):
disable_viewing_dependent: bool = False,
num_images: Optional[int] = None,
appearance_embedding_dim: int = 32,
contraction_type: ContractionType = ContractionType.UN_BOUNDED_SPHERE,
spatial_distortion: Optional[SpatialDistortion] = SceneContraction(),
) -> None:
super().__init__()
self.aabb = Parameter(aabb, requires_grad=False)
self.geo_feat_dim = geo_feat_dim
self.contraction_type = contraction_type
self.spatial_distortion = spatial_distortion
self.use_appearance_embedding = use_appearance_embedding
if use_appearance_embedding:
@@ -144,9 +147,16 @@ class NerfplayerNGPField(Field):
)
def get_density(self, ray_samples: RaySamples) -> Tuple[TensorType, TensorType]:
positions = ray_samples.frustums.get_positions()
if self.spatial_distortion is not None:
positions = ray_samples.frustums.get_positions()
positions = self.spatial_distortion(positions)
positions = (positions + 2.0) / 4.0
else:
positions = SceneBox.get_normalized_positions(ray_samples.frustums.get_positions(), self.aabb)
# Make sure the tcnn gets inputs between 0 and 1.
selector = ((positions > 0.0) & (positions < 1.0)).all(dim=-1)
positions = positions * selector[..., None]
positions_flat = positions.view(-1, 3)
positions_flat = contract(x=positions_flat, roi=self.aabb, type=self.contraction_type)
assert ray_samples.times is not None, "Time should be included in the input for NeRFPlayer"
times_flat = ray_samples.times.view(-1, 1)
@@ -158,28 +168,22 @@ class NerfplayerNGPField(Field):
# softplus, because it enables high post-activation (float32) density outputs
# from smaller internal (float16) parameters.
density = trunc_exp(density_before_activation.to(positions))
density = density * selector[..., None]
return density, base_mlp_out
def get_outputs(
self, ray_samples: RaySamples, density_embedding: Optional[TensorType] = None
) -> Dict[FieldHeadNames, TensorType]:
assert density_embedding is not None
directions = shift_directions_for_tcnn(ray_samples.frustums.directions)
directions_flat = directions.view(-1, 3)
if self.direction_encoding is not None:
d = self.direction_encoding(directions_flat)
if density_embedding is None:
positions = SceneBox.get_normalized_positions(ray_samples.frustums.get_positions(), self.aabb)
h = torch.cat([d, positions.view(-1, 3)], dim=-1)
else:
h = torch.cat([d, density_embedding.view(-1, self.geo_feat_dim)], dim=-1)
h = torch.cat([d, density_embedding.view(-1, self.geo_feat_dim)], dim=-1)
else:
# viewing direction is disabled
if density_embedding is None:
positions = SceneBox.get_normalized_positions(ray_samples.frustums.get_positions(), self.aabb)
h = positions.view(-1, 3)
else:
h = density_embedding.view(-1, self.geo_feat_dim)
h = density_embedding.view(-1, self.geo_feat_dim)
if self.use_appearance_embedding:
if ray_samples.camera_indices is None: