nerfstudio/pyrad/nerf/graph/semantic_nerf.py

"""
Semantic NeRF implementation.
"""

from typing import Dict, Optional, Tuple

import torch
from torch import nn
from torchtyping import TensorType

from pyrad.nerf.field_modules.encoding import Encoding, Identity, NeRFEncoding
from pyrad.nerf.field_modules.field_heads import DensityFieldHead, FieldHeadNames, RGBFieldHead, SemanticFieldHead
from pyrad.nerf.field_modules.mlp import MLP
from pyrad.nerf.fields.base import Field
from pyrad.nerf.fields.nerf_field import NeRFField
from pyrad.nerf.graph.vanilla_nerf import NeRFGraph
from pyrad.nerf.renderers import SemanticRenderer
from pyrad.structures.rays import PointSamples, RayBundle
from pyrad.utils import writer


class SemanticNerfField(Field):
    """Semantic-NeRF field"""

    def __init__(
        self,
        num_semantic_classes: int,
        position_encoding: Encoding = Identity(in_dim=3),
        direction_encoding: Encoding = Identity(in_dim=3),
        base_mlp_num_layers: int = 8,
        base_mlp_layer_width: int = 256,
        head_mlp_num_layers: int = 2,
        head_mlp_layer_width: int = 128,
        skip_connections: Tuple[int] = (4,),
    ) -> None:
        super().__init__()
        self.num_semantic_classes = num_semantic_classes
        self.position_encoding = position_encoding
        self.direction_encoding = direction_encoding
        self.mlp_base = MLP(
            in_dim=self.position_encoding.get_out_dim(),
            num_layers=base_mlp_num_layers,
            layer_width=base_mlp_layer_width,
            skip_connections=skip_connections,
        )
        self.mlp_head = MLP(
            in_dim=self.mlp_base.get_out_dim() + self.direction_encoding.get_out_dim(),
            num_layers=head_mlp_num_layers,
            layer_width=head_mlp_layer_width,
        )
        self.mlp_semantic = MLP(
            in_dim=self.mlp_head.get_out_dim(),
            layer_width=self.mlp_head.layer_width // 2,
            num_layers=1,
            activation=nn.ReLU(),
        )
        self.field_head_density = DensityFieldHead(in_dim=self.mlp_base.get_out_dim())
        self.field_head_rgb = RGBFieldHead(in_dim=self.mlp_head.get_out_dim())
        self.field_head_semantic = SemanticFieldHead(
            in_dim=self.mlp_semantic.get_out_dim(), num_classes=self.num_semantic_classes
        )

    def get_density(self, point_samples: PointSamples):
        encoded_xyz = self.position_encoding(point_samples.positions)
        base_mlp_out = self.mlp_base(encoded_xyz)
        density = self.field_head_density(base_mlp_out)
        return density, base_mlp_out

    def get_outputs(
        self, point_samples: PointSamples, density_embedding: Optional[TensorType] = None
    ) -> Dict[FieldHeadNames, TensorType]:
        encoded_dir = self.direction_encoding(point_samples.directions)
        mlp_out = self.mlp_head(torch.cat([encoded_dir, density_embedding], dim=-1))
        outputs = {}
        # rgb
        outputs[self.field_head_rgb.field_head_name] = self.field_head_rgb(mlp_out)
        # semantic
        mlp_out_sem = self.mlp_semantic(mlp_out)
        outputs[self.field_head_semantic.field_head_name] = self.field_head_semantic(mlp_out_sem)
        return outputs


class SemanticNerfGraph(NeRFGraph):
    """Semantic-NeRF graph"""

    def __init__(self, intrinsics=None, camera_to_world=None, stuff_classes=None, stuff_colors=None, **kwargs) -> None:
        assert stuff_classes is not None
        self.stuff_classes = stuff_classes
        self.stuff_colors = stuff_colors
        super().__init__(intrinsics=intrinsics, camera_to_world=camera_to_world, **kwargs)

    def populate_fields(self):
        """Set the fields."""
        position_encoding = NeRFEncoding(
            in_dim=3, num_frequencies=10, min_freq_exp=0.0, max_freq_exp=8.0, include_input=True
        )
        direction_encoding = NeRFEncoding(
            in_dim=3, num_frequencies=4, min_freq_exp=0.0, max_freq_exp=4.0, include_input=True
        )
        self.field_coarse = NeRFField(position_encoding=position_encoding, direction_encoding=direction_encoding)

        num_semantic_classes = len(self.stuff_classes)
        self.field_fine = SemanticNerfField(
            num_semantic_classes, position_encoding=position_encoding, direction_encoding=direction_encoding
        )

    def populate_misc_modules(self):
        super().populate_misc_modules()
        self.renderer_semantic = SemanticRenderer()
        self.cross_entropy_loss = nn.CrossEntropyLoss(reduction="mean")

    def get_outputs(self, ray_bundle: RayBundle):
        # uniform sampling
        ray_samples_uniform = self.sampler_uniform(ray_bundle)

        # coarse field
        field_outputs_coarse = self.field_coarse.forward(ray_samples_uniform.to_point_samples())
        weights_coarse = ray_samples_uniform.get_weights(field_outputs_coarse[FieldHeadNames.DENSITY])
        rgb_coarse = self.renderer_rgb(
            rgb=field_outputs_coarse[FieldHeadNames.RGB],
            weights=weights_coarse,
        )
        accumulation_coarse = self.renderer_accumulation(weights_coarse)
        depth_coarse = self.renderer_depth(weights_coarse, ray_samples_uniform.ts)

        # pdf sampling
        ray_samples_pdf = self.sampler_pdf(ray_bundle, ray_samples_uniform, weights_coarse)

        # fine field
        field_outputs_fine = self.field_fine.forward(ray_samples_pdf.to_point_samples())
        weights_fine = ray_samples_pdf.get_weights(field_outputs_fine[FieldHeadNames.DENSITY])
        rgb_fine = self.renderer_rgb(
            rgb=field_outputs_fine[FieldHeadNames.RGB],
            weights=weights_fine,
        )
        accumulation_fine = self.renderer_accumulation(weights_fine)
        depth_fine = self.renderer_depth(weights_fine, ray_samples_pdf.ts)
        semantic_fine = self.renderer_semantic(field_outputs_fine[FieldHeadNames.SEMANTICS], weights=weights_fine)

        outputs = {
            "rgb_coarse": rgb_coarse,
            "rgb_fine": rgb_fine,
            "accumulation_coarse": accumulation_coarse,
            "accumulation_fine": accumulation_fine,
            "depth_coarse": depth_coarse,
            "depth_fine": depth_fine,
            "semantic_fine": semantic_fine,
        }
        return outputs

    def get_loss_dict(self, outputs, batch):
        pixels = batch["pixels"]
        rgb_loss_coarse = self.rgb_loss(pixels, outputs["rgb_coarse"])
        rgb_loss_fine = self.rgb_loss(pixels, outputs["rgb_fine"])
        semantic_logits = outputs["semantic_fine"]
        semantic_classes = batch["semantics"][..., 0].long()
        semantic_loss_fine = self.cross_entropy_loss(semantic_logits, semantic_classes)
        loss_dict = {
            "rgb_loss_coarse": rgb_loss_coarse,
            "rgb_loss_fine": rgb_loss_fine,
            "semantic_loss_fine": semantic_loss_fine,
        }
        return loss_dict

    def log_test_image_outputs(self, image_idx, step, batch, outputs):
        super().log_test_image_outputs(image_idx, step, batch, outputs)
        semantic_logits = outputs["semantic_fine"]
        semantic_labels = torch.argmax(torch.nn.functional.softmax(semantic_logits, dim=-1), dim=-1)
        semantic_labels_image = self.stuff_colors[semantic_labels]
        writer.put_image(name=f"semantics/image_idx_{image_idx}", image=semantic_labels_image, step=step)

        mask = batch["mask"].repeat(1, 1, 3)
        writer.put_image(name=f"mask/image_idx_{image_idx}", image=mask, step=step)