Added hdr-nerfacto from nerfacto

nerfstudio/hdr_nerfacto/hdr_nerf_field.py

@@ -0,0 +1,301 @@
+# Copyright 2022 the Regents of the University of California, Nerfstudio Team and contributors. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Field for compound nerf model, adds scene contraction and image embeddings to instant ngp
+"""
+
+
+from typing import Dict, Literal, Optional, Tuple
+
+import torch
+from torch import Tensor, nn
+
+from nerfstudio.cameras.rays import RaySamples
+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.encodings import HashEncoding, NeRFEncoding, SHEncoding
+from nerfstudio.field_components.field_heads import (
+    FieldHeadNames,
+    PredNormalsFieldHead,
+    SemanticFieldHead,
+    TransientDensityFieldHead,
+    TransientRGBFieldHead,
+    UncertaintyFieldHead,
+)
+from nerfstudio.field_components.mlp import MLP
+from nerfstudio.field_components.spatial_distortions import SpatialDistortion
+from nerfstudio.fields.base_field import Field, get_normalized_directions
+
+
+class NerfactoField(Field):
+    """Compound Field that uses TCNN
+
+    Args:
+        aabb: parameters of scene aabb bounds
+        num_images: number of images in the dataset
+        num_layers: number of hidden layers
+        hidden_dim: dimension of hidden layers
+        geo_feat_dim: output geo feat dimensions
+        num_levels: number of levels of the hashmap for the base mlp
+        base_res: base resolution of the hashmap for the base mlp
+        max_res: maximum resolution of the hashmap for the base mlp
+        log2_hashmap_size: size of the hashmap for the base mlp
+        num_layers_color: number of hidden layers for color network
+        num_layers_transient: number of hidden layers for transient network
+        features_per_level: number of features per level for the hashgrid
+        hidden_dim_color: dimension of hidden layers for color network
+        hidden_dim_transient: dimension of hidden layers for transient network
+        appearance_embedding_dim: dimension of appearance embedding
+        transient_embedding_dim: dimension of transient embedding
+        use_transient_embedding: whether to use transient embedding
+        use_semantics: whether to use semantic segmentation
+        num_semantic_classes: number of semantic classes
+        use_pred_normals: whether to use predicted normals
+        use_average_appearance_embedding: whether to use average appearance embedding or zeros for inference
+        spatial_distortion: spatial distortion to apply to the scene
+    """
+
+    aabb: Tensor
+
+    def __init__(
+        self,
+        aabb: Tensor,
+        num_images: int,
+        num_layers: int = 2,
+        hidden_dim: int = 64,
+        geo_feat_dim: int = 15,
+        num_levels: int = 16,
+        base_res: int = 16,
+        max_res: int = 2048,
+        log2_hashmap_size: int = 19,
+        num_layers_color: int = 3,
+        num_layers_transient: int = 2,
+        features_per_level: int = 2,
+        hidden_dim_color: int = 64,
+        hidden_dim_transient: int = 64,
+        appearance_embedding_dim: int = 32,
+        transient_embedding_dim: int = 16,
+        use_transient_embedding: bool = False,
+        use_semantics: bool = False,
+        num_semantic_classes: int = 100,
+        pass_semantic_gradients: bool = False,
+        use_pred_normals: bool = False,
+        use_average_appearance_embedding: bool = False,
+        spatial_distortion: Optional[SpatialDistortion] = None,
+        implementation: Literal["tcnn", "torch"] = "tcnn",
+    ) -> None:
+        super().__init__()
+
+        self.register_buffer("aabb", aabb)
+        self.geo_feat_dim = geo_feat_dim
+
+        self.register_buffer("max_res", torch.tensor(max_res))
+        self.register_buffer("num_levels", torch.tensor(num_levels))
+        self.register_buffer("log2_hashmap_size", torch.tensor(log2_hashmap_size))
+
+        self.spatial_distortion = spatial_distortion
+        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
+        self.use_transient_embedding = use_transient_embedding
+        self.use_semantics = use_semantics
+        self.use_pred_normals = use_pred_normals
+        self.pass_semantic_gradients = pass_semantic_gradients
+        self.base_res = base_res
+
+        self.direction_encoding = SHEncoding(
+            levels=4,
+            implementation=implementation,
+        )
+
+        self.position_encoding = NeRFEncoding(
+            in_dim=3, num_frequencies=2, min_freq_exp=0, max_freq_exp=2 - 1, implementation=implementation
+        )
+
+        self.mlp_base_grid = HashEncoding(
+            num_levels=num_levels,
+            min_res=base_res,
+            max_res=max_res,
+            log2_hashmap_size=log2_hashmap_size,
+            features_per_level=features_per_level,
+            implementation=implementation,
+        )
+        self.mlp_base_mlp = MLP(
+            in_dim=self.mlp_base_grid.get_out_dim(),
+            num_layers=num_layers,
+            layer_width=hidden_dim,
+            out_dim=1 + self.geo_feat_dim,
+            activation=nn.ReLU(),
+            out_activation=None,
+            implementation=implementation,
+        )
+        self.mlp_base = torch.nn.Sequential(self.mlp_base_grid, self.mlp_base_mlp)
+
+        # transients
+        if self.use_transient_embedding:
+            self.transient_embedding_dim = transient_embedding_dim
+            self.embedding_transient = Embedding(self.num_images, self.transient_embedding_dim)
+            self.mlp_transient = MLP(
+                in_dim=self.geo_feat_dim + self.transient_embedding_dim,
+                num_layers=num_layers_transient,
+                layer_width=hidden_dim_transient,
+                out_dim=hidden_dim_transient,
+                activation=nn.ReLU(),
+                out_activation=None,
+                implementation=implementation,
+            )
+            self.field_head_transient_uncertainty = UncertaintyFieldHead(in_dim=self.mlp_transient.get_out_dim())
+            self.field_head_transient_rgb = TransientRGBFieldHead(in_dim=self.mlp_transient.get_out_dim())
+            self.field_head_transient_density = TransientDensityFieldHead(in_dim=self.mlp_transient.get_out_dim())
+
+        # semantics
+        if self.use_semantics:
+            self.mlp_semantics = MLP(
+                in_dim=self.geo_feat_dim,
+                num_layers=2,
+                layer_width=64,
+                out_dim=hidden_dim_transient,
+                activation=nn.ReLU(),
+                out_activation=None,
+                implementation=implementation,
+            )
+            self.field_head_semantics = SemanticFieldHead(
+                in_dim=self.mlp_semantics.get_out_dim(), num_classes=num_semantic_classes
+            )
+
+        # predicted normals
+        if self.use_pred_normals:
+            self.mlp_pred_normals = MLP(
+                in_dim=self.geo_feat_dim + self.position_encoding.get_out_dim(),
+                num_layers=3,
+                layer_width=64,
+                out_dim=hidden_dim_transient,
+                activation=nn.ReLU(),
+                out_activation=None,
+                implementation=implementation,
+            )
+            self.field_head_pred_normals = PredNormalsFieldHead(in_dim=self.mlp_pred_normals.get_out_dim())
+
+        self.mlp_head = MLP(
+            in_dim=self.direction_encoding.get_out_dim() + self.geo_feat_dim + self.appearance_embedding_dim,
+            num_layers=num_layers_color,
+            layer_width=hidden_dim_color,
+            out_dim=3,
+            activation=nn.ReLU(),
+            out_activation=nn.Sigmoid(),
+            implementation=implementation,
+        )
+
+    def get_density(self, ray_samples: RaySamples) -> Tuple[Tensor, Tensor]:
+        """Computes and returns the densities."""
+        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]
+        self._sample_locations = positions
+        if not self._sample_locations.requires_grad:
+            self._sample_locations.requires_grad = True
+        positions_flat = positions.view(-1, 3)
+        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)
+        self._density_before_activation = density_before_activation
+
+        # Rectifying the density with an exponential is much more stable than a ReLU or
+        # 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[Tensor] = None
+    ) -> Dict[FieldHeadNames, Tensor]:
+        assert density_embedding is not None
+        outputs = {}
+        if ray_samples.camera_indices is None:
+            raise AttributeError("Camera indices are not provided.")
+        camera_indices = ray_samples.camera_indices.squeeze()
+        directions = get_normalized_directions(ray_samples.frustums.directions)
+        directions_flat = directions.view(-1, 3)
+        d = self.direction_encoding(directions_flat)
+
+        outputs_shape = ray_samples.frustums.directions.shape[:-1]
+
+        # appearance
+        if self.training:
+            embedded_appearance = self.embedding_appearance(camera_indices)
+        else:
+            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
+                )
+
+        # transients
+        if self.use_transient_embedding and self.training:
+            embedded_transient = self.embedding_transient(camera_indices)
+            transient_input = torch.cat(
+                [
+                    density_embedding.view(-1, self.geo_feat_dim),
+                    embedded_transient.view(-1, self.transient_embedding_dim),
+                ],
+                dim=-1,
+            )
+            x = self.mlp_transient(transient_input).view(*outputs_shape, -1).to(directions)
+            outputs[FieldHeadNames.UNCERTAINTY] = self.field_head_transient_uncertainty(x)
+            outputs[FieldHeadNames.TRANSIENT_RGB] = self.field_head_transient_rgb(x)
+            outputs[FieldHeadNames.TRANSIENT_DENSITY] = self.field_head_transient_density(x)
+
+        # semantics
+        if self.use_semantics:
+            semantics_input = density_embedding.view(-1, self.geo_feat_dim)
+            if not self.pass_semantic_gradients:
+                semantics_input = semantics_input.detach()
+
+            x = self.mlp_semantics(semantics_input).view(*outputs_shape, -1).to(directions)
+            outputs[FieldHeadNames.SEMANTICS] = self.field_head_semantics(x)
+
+        # predicted normals
+        if self.use_pred_normals:
+            positions = ray_samples.frustums.get_positions()
+
+            positions_flat = self.position_encoding(positions.view(-1, 3))
+            pred_normals_inp = torch.cat([positions_flat, density_embedding.view(-1, self.geo_feat_dim)], dim=-1)
+
+            x = self.mlp_pred_normals(pred_normals_inp).view(*outputs_shape, -1).to(directions)
+            outputs[FieldHeadNames.PRED_NORMALS] = self.field_head_pred_normals(x)
+
+        h = torch.cat(
+            [
+                d,
+                density_embedding.view(-1, self.geo_feat_dim),
+                embedded_appearance.view(-1, self.appearance_embedding_dim),
+            ],
+            dim=-1,
+        )
+        rgb = self.mlp_head(h).view(*outputs_shape, -1).to(directions)
+        outputs.update({FieldHeadNames.RGB: rgb})
+
+        return outputs

nerfstudio/hdr_nerfacto/hdr_nerf_model.py

@@ -0,0 +1,401 @@
+# Copyright 2022 the Regents of the University of California, Nerfstudio Team and contributors. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+NeRF implementation that combines many recent advancements.
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass, field
+from typing import Dict, List, Literal, Tuple, Type
+
+import numpy as np
+import torch
+from torch.nn import Parameter
+from torchmetrics.functional import structural_similarity_index_measure
+from torchmetrics.image import PeakSignalNoiseRatio
+from torchmetrics.image.lpip import LearnedPerceptualImagePatchSimilarity
+
+from nerfstudio.cameras.rays import RayBundle, RaySamples
+from nerfstudio.engine.callbacks import TrainingCallback, TrainingCallbackAttributes, TrainingCallbackLocation
+from nerfstudio.field_components.field_heads import FieldHeadNames
+from nerfstudio.field_components.spatial_distortions import SceneContraction
+from nerfstudio.fields.density_fields import HashMLPDensityField
+from nerfstudio.fields.nerfacto_field import NerfactoField
+from nerfstudio.model_components.losses import (
+    MSELoss,
+    distortion_loss,
+    interlevel_loss,
+    orientation_loss,
+    pred_normal_loss,
+    scale_gradients_by_distance_squared,
+)
+from nerfstudio.model_components.ray_samplers import ProposalNetworkSampler, UniformSampler
+from nerfstudio.model_components.renderers import AccumulationRenderer, DepthRenderer, NormalsRenderer, RGBRenderer
+from nerfstudio.model_components.scene_colliders import NearFarCollider
+from nerfstudio.model_components.shaders import NormalsShader
+from nerfstudio.models.base_model import Model, ModelConfig
+from nerfstudio.utils import colormaps
+
+
+@dataclass
+class NerfactoModelConfig(ModelConfig):
+    """Nerfacto Model Config"""
+
+    _target: Type = field(default_factory=lambda: NerfactoModel)
+    near_plane: float = 0.05
+    """How far along the ray to start sampling."""
+    far_plane: float = 1000.0
+    """How far along the ray to stop sampling."""
+    background_color: Literal["random", "last_sample", "black", "white"] = "last_sample"
+    """Whether to randomize the background color."""
+    hidden_dim: int = 64
+    """Dimension of hidden layers"""
+    hidden_dim_color: int = 64
+    """Dimension of hidden layers for color network"""
+    hidden_dim_transient: int = 64
+    """Dimension of hidden layers for transient network"""
+    num_levels: int = 16
+    """Number of levels of the hashmap for the base mlp."""
+    base_res: int = 16
+    """Resolution of the base grid for the hashgrid."""
+    max_res: int = 2048
+    """Maximum resolution of the hashmap for the base mlp."""
+    log2_hashmap_size: int = 19
+    """Size of the hashmap for the base mlp"""
+    features_per_level: int = 2
+    """How many hashgrid features per level"""
+    num_proposal_samples_per_ray: Tuple[int, ...] = (256, 96)
+    """Number of samples per ray for each proposal network."""
+    num_nerf_samples_per_ray: int = 48
+    """Number of samples per ray for the nerf network."""
+    proposal_update_every: int = 5
+    """Sample every n steps after the warmup"""
+    proposal_warmup: int = 5000
+    """Scales n from 1 to proposal_update_every over this many steps"""
+    num_proposal_iterations: int = 2
+    """Number of proposal network iterations."""
+    use_same_proposal_network: bool = False
+    """Use the same proposal network. Otherwise use different ones."""
+    proposal_net_args_list: List[Dict] = field(
+        default_factory=lambda: [
+            {"hidden_dim": 16, "log2_hashmap_size": 17, "num_levels": 5, "max_res": 128, "use_linear": False},
+            {"hidden_dim": 16, "log2_hashmap_size": 17, "num_levels": 5, "max_res": 256, "use_linear": False},
+        ]
+    )
+    """Arguments for the proposal density fields."""
+    proposal_initial_sampler: Literal["piecewise", "uniform"] = "piecewise"
+    """Initial sampler for the proposal network. Piecewise is preferred for unbounded scenes."""
+    interlevel_loss_mult: float = 1.0
+    """Proposal loss multiplier."""
+    distortion_loss_mult: float = 0.002
+    """Distortion loss multiplier."""
+    orientation_loss_mult: float = 0.0001
+    """Orientation loss multiplier on computed normals."""
+    pred_normal_loss_mult: float = 0.001
+    """Predicted normal loss multiplier."""
+    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 = True
+    """Whether use single jitter or not for the proposal networks."""
+    predict_normals: bool = False
+    """Whether to predict normals or not."""
+    disable_scene_contraction: bool = False
+    """Whether to disable scene contraction or not."""
+    use_gradient_scaling: bool = False
+    """Use gradient scaler where the gradients are lower for points closer to the camera."""
+    implementation: Literal["tcnn", "torch"] = "tcnn"
+    """Which implementation to use for the model."""
+    appearance_embed_dim: int = 32
+    """Dimension of the appearance embedding."""
+
+
+class NerfactoModel(Model):
+    """Nerfacto model
+
+    Args:
+        config: Nerfacto configuration to instantiate model
+    """
+
+    config: NerfactoModelConfig
+
+    def populate_modules(self):
+        """Set the fields and modules."""
+        super().populate_modules()
+
+        if self.config.disable_scene_contraction:
+            scene_contraction = None
+        else:
+            scene_contraction = SceneContraction(order=float("inf"))
+
+        # Fields
+        self.field = NerfactoField(
+            self.scene_box.aabb,
+            hidden_dim=self.config.hidden_dim,
+            num_levels=self.config.num_levels,
+            max_res=self.config.max_res,
+            base_res=self.config.base_res,
+            features_per_level=self.config.features_per_level,
+            log2_hashmap_size=self.config.log2_hashmap_size,
+            hidden_dim_color=self.config.hidden_dim_color,
+            hidden_dim_transient=self.config.hidden_dim_transient,
+            spatial_distortion=scene_contraction,
+            num_images=self.num_train_data,
+            use_pred_normals=self.config.predict_normals,
+            use_average_appearance_embedding=self.config.use_average_appearance_embedding,
+            appearance_embedding_dim=self.config.appearance_embed_dim,
+            implementation=self.config.implementation,
+        )
+
+        self.density_fns = []
+        num_prop_nets = self.config.num_proposal_iterations
+        # Build the proposal network(s)
+        self.proposal_networks = torch.nn.ModuleList()
+        if self.config.use_same_proposal_network:
+            assert len(self.config.proposal_net_args_list) == 1, "Only one proposal network is allowed."
+            prop_net_args = self.config.proposal_net_args_list[0]
+            network = HashMLPDensityField(
+                self.scene_box.aabb,
+                spatial_distortion=scene_contraction,
+                **prop_net_args,
+                implementation=self.config.implementation,
+            )
+            self.proposal_networks.append(network)
+            self.density_fns.extend([network.density_fn for _ in range(num_prop_nets)])
+        else:
+            for i in range(num_prop_nets):
+                prop_net_args = self.config.proposal_net_args_list[min(i, len(self.config.proposal_net_args_list) - 1)]
+                network = HashMLPDensityField(
+                    self.scene_box.aabb,
+                    spatial_distortion=scene_contraction,
+                    **prop_net_args,
+                    implementation=self.config.implementation,
+                )
+                self.proposal_networks.append(network)
+            self.density_fns.extend([network.density_fn for network in self.proposal_networks])
+
+        # Samplers
+        def update_schedule(step):
+            return np.clip(
+                np.interp(step, [0, self.config.proposal_warmup], [0, self.config.proposal_update_every]),
+                1,
+                self.config.proposal_update_every,
+            )
+
+        # Change proposal network initial sampler if uniform
+        initial_sampler = None  # None is for piecewise as default (see ProposalNetworkSampler)
+        if self.config.proposal_initial_sampler == "uniform":
+            initial_sampler = UniformSampler(single_jitter=self.config.use_single_jitter)
+
+        self.proposal_sampler = ProposalNetworkSampler(
+            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_iterations,
+            single_jitter=self.config.use_single_jitter,
+            update_sched=update_schedule,
+            initial_sampler=initial_sampler,
+        )
+
+        # Collider
+        self.collider = NearFarCollider(near_plane=self.config.near_plane, far_plane=self.config.far_plane)
+
+        # renderers
+        self.renderer_rgb = RGBRenderer(background_color=self.config.background_color)
+        self.renderer_accumulation = AccumulationRenderer()
+        self.renderer_depth = DepthRenderer()
+        self.renderer_normals = NormalsRenderer()
+
+        # shaders
+        self.normals_shader = NormalsShader()
+
+        # losses
+        self.rgb_loss = MSELoss()
+
+        # metrics
+        self.psnr = PeakSignalNoiseRatio(data_range=1.0)
+        self.ssim = structural_similarity_index_measure
+        self.lpips = LearnedPerceptualImagePatchSimilarity(normalize=True)
+
+    def get_param_groups(self) -> Dict[str, List[Parameter]]:
+        param_groups = {}
+        param_groups["proposal_networks"] = list(self.proposal_networks.parameters())
+        param_groups["fields"] = list(self.field.parameters())
+        return param_groups
+
+    def get_training_callbacks(
+        self, training_callback_attributes: TrainingCallbackAttributes
+    ) -> List[TrainingCallback]:
+        callbacks = []
+        if self.config.use_proposal_weight_anneal:
+            # anneal the weights of the proposal network before doing PDF sampling
+            N = self.config.proposal_weights_anneal_max_num_iters
+
+            def set_anneal(step):
+                # https://arxiv.org/pdf/2111.12077.pdf eq. 18
+                train_frac = np.clip(step / N, 0, 1)
+
+                def bias(x, b):
+                    return b * x / ((b - 1) * x + 1)
+
+                anneal = bias(train_frac, self.config.proposal_weights_anneal_slope)
+                self.proposal_sampler.set_anneal(anneal)
+
+            callbacks.append(
+                TrainingCallback(
+                    where_to_run=[TrainingCallbackLocation.BEFORE_TRAIN_ITERATION],
+                    update_every_num_iters=1,
+                    func=set_anneal,
+                )
+            )
+            callbacks.append(
+                TrainingCallback(
+                    where_to_run=[TrainingCallbackLocation.AFTER_TRAIN_ITERATION],
+                    update_every_num_iters=1,
+                    func=self.proposal_sampler.step_cb,
+                )
+            )
+        return callbacks
+
+    def get_outputs(self, ray_bundle: RayBundle):
+        ray_samples: RaySamples
+        ray_samples, weights_list, ray_samples_list = self.proposal_sampler(ray_bundle, density_fns=self.density_fns)
+        field_outputs = self.field.forward(ray_samples, compute_normals=self.config.predict_normals)
+        if self.config.use_gradient_scaling:
+            field_outputs = scale_gradients_by_distance_squared(field_outputs, ray_samples)
+
+        weights = ray_samples.get_weights(field_outputs[FieldHeadNames.DENSITY])
+        weights_list.append(weights)
+        ray_samples_list.append(ray_samples)
+
+        rgb = self.renderer_rgb(rgb=field_outputs[FieldHeadNames.RGB], weights=weights)
+        depth = self.renderer_depth(weights=weights, ray_samples=ray_samples)
+        accumulation = self.renderer_accumulation(weights=weights)
+
+        outputs = {
+            "rgb": rgb,
+            "accumulation": accumulation,
+            "depth": depth,
+        }
+
+        if self.config.predict_normals:
+            normals = self.renderer_normals(normals=field_outputs[FieldHeadNames.NORMALS], weights=weights)
+            pred_normals = self.renderer_normals(field_outputs[FieldHeadNames.PRED_NORMALS], weights=weights)
+            outputs["normals"] = self.normals_shader(normals)
+            outputs["pred_normals"] = self.normals_shader(pred_normals)
+        # These use a lot of GPU memory, so we avoid storing them for eval.
+        if self.training:
+            outputs["weights_list"] = weights_list
+            outputs["ray_samples_list"] = ray_samples_list
+
+        if self.training and self.config.predict_normals:
+            outputs["rendered_orientation_loss"] = orientation_loss(
+                weights.detach(), field_outputs[FieldHeadNames.NORMALS], ray_bundle.directions
+            )
+
+            outputs["rendered_pred_normal_loss"] = pred_normal_loss(
+                weights.detach(),
+                field_outputs[FieldHeadNames.NORMALS].detach(),
+                field_outputs[FieldHeadNames.PRED_NORMALS],
+            )
+
+        for i in range(self.config.num_proposal_iterations):
+            outputs[f"prop_depth_{i}"] = self.renderer_depth(weights=weights_list[i], ray_samples=ray_samples_list[i])
+
+        return outputs
+
+    def get_metrics_dict(self, outputs, batch):
+        metrics_dict = {}
+        gt_rgb = batch["image"].to(self.device)  # RGB or RGBA image
+        gt_rgb = self.renderer_rgb.blend_background(gt_rgb)  # Blend if RGBA
+        predicted_rgb = outputs["rgb"]
+        metrics_dict["psnr"] = self.psnr(predicted_rgb, gt_rgb)
+
+        if self.training:
+            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):
+        loss_dict = {}
+        image = batch["image"].to(self.device)
+        pred_rgb, gt_rgb = self.renderer_rgb.blend_background_for_loss_computation(
+            pred_image=outputs["rgb"],
+            pred_accumulation=outputs["accumulation"],
+            gt_image=image,
+        )
+
+        loss_dict["rgb_loss"] = self.rgb_loss(gt_rgb, pred_rgb)
+        if self.training:
+            loss_dict["interlevel_loss"] = self.config.interlevel_loss_mult * interlevel_loss(
+                outputs["weights_list"], outputs["ray_samples_list"]
+            )
+            assert metrics_dict is not None and "distortion" in metrics_dict
+            loss_dict["distortion_loss"] = self.config.distortion_loss_mult * metrics_dict["distortion"]
+            if self.config.predict_normals:
+                # orientation loss for computed normals
+                loss_dict["orientation_loss"] = self.config.orientation_loss_mult * torch.mean(
+                    outputs["rendered_orientation_loss"]
+                )
+
+                # ground truth supervision for normals
+                loss_dict["pred_normal_loss"] = self.config.pred_normal_loss_mult * torch.mean(
+                    outputs["rendered_pred_normal_loss"]
+                )
+        return loss_dict
+
+    def get_image_metrics_and_images(
+        self, outputs: Dict[str, torch.Tensor], batch: Dict[str, torch.Tensor]
+    ) -> Tuple[Dict[str, float], Dict[str, torch.Tensor]]:
+        gt_rgb = batch["image"].to(self.device)
+        predicted_rgb = outputs["rgb"]  # Blended with background (black if random background)
+        gt_rgb = self.renderer_rgb.blend_background(gt_rgb)
+        acc = colormaps.apply_colormap(outputs["accumulation"])
+        depth = colormaps.apply_depth_colormap(
+            outputs["depth"],
+            accumulation=outputs["accumulation"],
+        )
+
+        combined_rgb = torch.cat([gt_rgb, predicted_rgb], dim=1)
+        combined_acc = torch.cat([acc], dim=1)
+        combined_depth = torch.cat([depth], dim=1)
+
+        # Switch images from [H, W, C] to [1, C, H, W] for metrics computations
+        gt_rgb = torch.moveaxis(gt_rgb, -1, 0)[None, ...]
+        predicted_rgb = torch.moveaxis(predicted_rgb, -1, 0)[None, ...]
+
+        psnr = self.psnr(gt_rgb, predicted_rgb)
+        ssim = self.ssim(gt_rgb, predicted_rgb)
+        lpips = self.lpips(torch.clamp(gt_rgb, 0, 1), torch.clamp(predicted_rgb, 0, 1))
+
+        # all of these metrics will be logged as scalars
+        metrics_dict = {"psnr": float(psnr.item()), "ssim": float(ssim)}  # type: ignore
+        metrics_dict["lpips"] = float(lpips)
+
+        images_dict = {"img": combined_rgb, "accumulation": combined_acc, "depth": combined_depth}
+
+        for i in range(self.config.num_proposal_iterations):
+            key = f"prop_depth_{i}"
+            prop_depth_i = colormaps.apply_depth_colormap(
+                outputs[key],
+                accumulation=outputs["accumulation"],
+            )
+            images_dict[key] = prop_depth_i
+
+        return metrics_dict, images_dict