ngp.py

from typing import Callable, List, Union

import torch
from torch.autograd import Function
from torch.cuda.amp import custom_bwd, custom_fwd

try:
    import tinycudann as tcnn
except ImportError as e:
    print(
        f"Error: {e}! "
        "Please install tinycudann by: "
        "pip install git+https://github.com/NVlabs/tiny-cuda-nn/#subdirectory=bindings/torch"
    )
    exit()


class _TruncExp(Function):  # pylint: disable=abstract-method
    # Implementation from torch-ngp:
    # https://github.com/ashawkey/torch-ngp/blob/93b08a0d4ec1cc6e69d85df7f0acdfb99603b628/activation.py
    @staticmethod
    @custom_fwd(cast_inputs=torch.float32)
    def forward(ctx, x):  # pylint: disable=arguments-differ
        ctx.save_for_backward(x)
        return torch.exp(x)

    @staticmethod
    @custom_bwd
    def backward(ctx, g):  # pylint: disable=arguments-differ
        x = ctx.saved_tensors[0]
        return g * torch.exp(torch.clamp(x, max=15))


trunc_exp = _TruncExp.apply


class NGPradianceField(torch.nn.Module):
    """Instance-NGP radiance Field"""

    def __init__(
        self,
        aabb: Union[torch.Tensor, List[float]],
        num_dim: int = 3,
        use_viewdirs: bool = True,
        density_activation: Callable = lambda x: trunc_exp(x - 1),
        unbounded: bool = False,
    ) -> None:
        super().__init__()
        if not isinstance(aabb, torch.Tensor):
            aabb = torch.tensor(aabb, dtype=torch.float32)
        self.register_buffer("aabb", aabb)
        self.num_dim = num_dim
        self.use_viewdirs = use_viewdirs
        self.density_activation = density_activation
        self.unbounded = unbounded

        self.geo_feat_dim = 15
        per_level_scale = 1.4472692012786865

        if self.use_viewdirs:
            self.direction_encoding = tcnn.Encoding(
                n_input_dims=num_dim,
                encoding_config={
                    "otype": "Composite",
                    "nested": [
                        {
                            "n_dims_to_encode": 3,
                            "otype": "SphericalHarmonics",
                            "degree": 4,
                        },
                        # {"otype": "Identity", "n_bins": 4, "degree": 4},
                    ],
                },
            )

        self.mlp_base = tcnn.NetworkWithInputEncoding(
            n_input_dims=num_dim,
            n_output_dims=1 + self.geo_feat_dim,
            encoding_config={
                "otype": "HashGrid",
                "n_levels": 16,
                "n_features_per_level": 2,
                "log2_hashmap_size": 19,
                "base_resolution": 16,
                "per_level_scale": per_level_scale,
            },
            network_config={
                "otype": "FullyFusedMLP",
                "activation": "ReLU",
                "output_activation": "None",
                "n_neurons": 64,
                "n_hidden_layers": 1,
            },
        )

        self.mlp_head = tcnn.Network(
            n_input_dims=(
                (
                    self.direction_encoding.n_output_dims
                    if self.use_viewdirs
                    else 0
                )
                + self.geo_feat_dim
            ),
            n_output_dims=3,
            network_config={
                "otype": "FullyFusedMLP",
                "activation": "ReLU",
                "output_activation": "Sigmoid",
                "n_neurons": 64,
                "n_hidden_layers": 2,
            },
        )

    def query_opacity(self, x, step_size):
        density = self.query_density(x)
        aabb_min, aabb_max = torch.split(self.aabb, self.num_dim, dim=-1)
        if self.unbounded:
            # TODO: [revisit] is this necessary?
            # 1.0 / derivative of tanh contraction
            x = (x - aabb_min) / (aabb_max - aabb_min)
            x = x - 0.5
            scaling = 1.0 / (
                torch.clamp(1.0 - torch.tanh(x) ** 2, min=1e6) * 0.5
            )
            scaling = scaling * (aabb_max - aabb_min)
        else:
            scaling = aabb_max - aabb_min
        step_size = step_size * scaling.norm(dim=-1, keepdim=True)
        # if the density is small enough those two are the same.
        # opacity = 1.0 - torch.exp(-density * step_size)
        opacity = density * step_size
        return opacity

    def query_density(self, x, return_feat: bool = False):
        if self.unbounded:
            # tanh contraction
            aabb_min, aabb_max = torch.split(self.aabb, self.num_dim, dim=-1)
            x = (x - aabb_min) / (aabb_max - aabb_min)
            x = x - 0.5
            x = (torch.tanh(x) + 1) * 0.5
        else:
            aabb_min, aabb_max = torch.split(self.aabb, self.num_dim, dim=-1)
            x = (x - aabb_min) / (aabb_max - aabb_min)
        selector = ((x > 0.0) & (x < 1.0)).all(dim=-1)
        x = (
            self.mlp_base(x.view(-1, self.num_dim))
            .view(list(x.shape[:-1]) + [1 + self.geo_feat_dim])
            .to(x)
        )
        density_before_activation, base_mlp_out = torch.split(
            x, [1, self.geo_feat_dim], dim=-1
        )
        density = (
            self.density_activation(density_before_activation)
            * selector[..., None]
        )
        if return_feat:
            return density, base_mlp_out
        else:
            return density

    def _query_rgb(self, dir, embedding):
        # tcnn requires directions in the range [0, 1]
        if self.use_viewdirs:
            dir = (dir + 1.0) / 2.0
            d = self.direction_encoding(dir.view(-1, dir.shape[-1]))
            h = torch.cat([d, embedding.view(-1, self.geo_feat_dim)], dim=-1)
        else:
            h = embedding.view(-1, self.geo_feat_dim)
        rgb = (
            self.mlp_head(h)
            .view(list(embedding.shape[:-1]) + [3])
            .to(embedding)
        )
        return rgb

    def forward(
        self,
        positions: torch.Tensor,
        directions: torch.Tensor = None,
    ):
        if self.use_viewdirs and (directions is not None):
            assert (
                positions.shape == directions.shape
            ), f"{positions.shape} v.s. {directions.shape}"
            density, embedding = self.query_density(positions, return_feat=True)
            rgb = self._query_rgb(directions, embedding=embedding)
        return rgb, density