gaussiant_sampler.py

"""
This is the cleaned up version of the original GaussianModel and GaussianSampler
Including it's core functionalities:
- Differentiable gaussian splatting
- Explicitly defined parameters?
- Image-based rendering?
- Cloning? Pruning? Splitting?
"""

from typing import TYPE_CHECKING
if TYPE_CHECKING:
    from easyvolcap.runners.volumetric_video_viewer import VolumetricVideoViewer

import torch
import numpy as np
from torch import nn
from typing import Literal
from torch.optim import Adam
from torch.nn import functional as F

from easyvolcap.engine import cfg, args
from easyvolcap.engine import SAMPLERS
from easyvolcap.engine.registry import call_from_cfg
from easyvolcap.utils.base_utils import dotdict

from easyvolcap.utils.console_utils import *
from easyvolcap.utils.console_utils import dotdict
from easyvolcap.utils.bound_utils import get_bounds
from easyvolcap.utils.chunk_utils import multi_gather, multi_scatter
from easyvolcap.utils.gaussian_utils import GaussianModel, in_frustrum
from easyvolcap.utils.net_utils import normalize, typed, update_optimizer_state
from easyvolcap.utils.data_utils import load_pts, export_pts, to_x, to_cuda, to_cpu, to_tensor, remove_batch

from easyvolcap.models.cameras.optimizable_camera import OptimizableCamera
from easyvolcap.models.samplers.point_planes_sampler import PointPlanesSampler
from easyvolcap.models.networks.volumetric_video_network import VolumetricVideoNetwork
from easyvolcap.dataloaders.datasets.volumetric_video_dataset import VolumetricVideoDataset


@SAMPLERS.register_module()
class GaussianTSampler(PointPlanesSampler):
    def __init__(self,
                 # Legacy APIs
                 network: VolumetricVideoNetwork = None,  # ignore this

                 # Initializations
                 sh_deg: int = 3,
                 init_occ: float = 0.75,
                 scale_min: float = 1e-4,  # 0.0001m should not be too small
                 scale_max: float = 1e1,  # 100m should not be too large
                 scale_mod: float = 1.0,

                 # Densify & pruning configs
                 densify_until_iter: int = 15000,
                 densify_from_iter: int = 500,
                 densification_interval: int = 100,
                 opacity_reset_interval: int = 3000e9,  # UNUSED:
                 sh_update_iter: int = 1000,
                 densify_grad_threshold: float = 0.0002,
                 percent_dense: float = 0.01,
                 size_threshold: float = None,  # UNUSED:
                 min_opacity: float = 0.005,
                 preload_gs: str = '',

                 # Housekeepings
                 **kwargs,
                 ):

        # Initialize parents and remove unwanted modules
        self.kwargs = dotdict(kwargs)
        call_from_cfg(super().__init__, kwargs, network=network)

        del self.pcd_embedder
        del self.xyz_embedder
        del self.resd_regressor
        del self.geo_regressor
        del self.dir_embedder
        del self.rgb_regressor

        # Gaussian models for every frame
        self.sh_deg = sh_deg
        self.scale_mod = scale_mod
        self.pcds: nn.ParameterList[GaussianModel] = nn.ParameterList([
            GaussianModel(
                pcd,
                self.rgbs[i] if len(self.rgbs) else None,
                self.occs[i] if len(self.occs) else init_occ,
                self.rads[i].expand(self.rads[i].shape[0], 3) if len(self.rads) else None,
                sh_deg,
                scale_min,
                scale_max
            )
            for i, pcd in enumerate(self.pcds)
        ])

        # Update parameters (densification & pruning)
        self.densify_until_iter = densify_until_iter
        self.densify_from_iter = densify_from_iter
        self.densification_interval = densification_interval
        self.opacity_reset_interval = opacity_reset_interval
        self.densify_grad_threshold = densify_grad_threshold
        self.sh_update_iter = sh_update_iter
        self.size_threshold = size_threshold
        self.percent_dense = percent_dense
        self.min_opacity = min_opacity
        self.last_output = None  # will only store the updates for one of the points

        # Debug options
        self.scale_mult = 1.0
        self.alpha_mult = 1.0

        # Test time controls
        self.post_handle = self.register_load_state_dict_post_hook(self._load_state_dict_post_hook)

        if preload_gs:
            self.load_from_file(preload_gs)

    def render_imgui(self, viewer: 'VolumetricVideoViewer', batch: dotdict):
        from imgui_bundle import imgui
        self.scale_mult = imgui.slider_float(f'Scale multiplier', self.scale_mult, 0.1, 5.0)[1]  # 0.1mm
        self.alpha_mult = imgui.slider_float(f'Alpha multiplier', self.alpha_mult, 0.1, 5.0)[1]  # 0.1mm

        for i, pcd in enumerate(self.pcds):
            imgui.text(f'Number of points: {len(pcd._xyz)}')

    def render_gaussians(self, xyz: torch.Tensor, sh: torch.Tensor, scale3: torch.Tensor, rot4: torch.Tensor, occ1: torch.Tensor, batch: dotdict):
        # Lazy imports
        from diff_gauss import rasterize_gaussians, GaussianRasterizationSettings, GaussianRasterizer
        from easyvolcap.utils.gaussian_utils import prepare_gaussian_camera

        # Remove batch dimension
        xyz, sh, scale3, rot4, occ1 = remove_batch([xyz, sh, scale3, rot4, occ1])

        # Prepare the camera transformation for Gaussian
        gaussian_camera = to_x(prepare_gaussian_camera(batch), torch.float)

        # is_in_frustrum = in_frustrum(xyz, gaussian_camera.full_proj_transform)
        # print('Number of points to render:', is_in_frustrum.sum().item())

        # Prepare rasterization settings for gaussian
        raster_settings = GaussianRasterizationSettings(
            image_height=gaussian_camera.image_height,
            image_width=gaussian_camera.image_width,
            tanfovx=gaussian_camera.tanfovx,
            tanfovy=gaussian_camera.tanfovy,
            bg=torch.full([3], self.bg_brightness if hasattr(self, 'bg_brightness') else 0.0, device=xyz.device),  # GPU
            scale_modifier=self.scale_mod if hasattr(self, 'bg_brightness') else 1.0,
            viewmatrix=gaussian_camera.world_view_transform,
            projmatrix=gaussian_camera.full_proj_transform,
            sh_degree=self.sh_deg if hasattr(self, 'sh_deg') else 0,
            campos=gaussian_camera.camera_center,
            prefiltered=False,
            debug=self.debug if hasattr(self, 'debug') else False,
        )

        # Rasterize visible Gaussians to image, obtain their radii (on screen).
        scr = torch.zeros_like(xyz, requires_grad=True) + 0  # gradient magic
        if scr.requires_grad: scr.retain_grad()
        rasterizer = GaussianRasterizer(raster_settings=raster_settings)
        rendered_image, rendered_depth, rendered_alpha, radii = typed(torch.float, torch.float)(rasterizer)(
            means3D=xyz,
            means2D=scr,
            shs=sh.mT,
            colors_precomp=None,
            opacities=occ1,
            scales=scale3,
            rotations=rot4,
            cov3D_precomp=None,
        )

        rgb = rendered_image[None].permute(0, 2, 3, 1)
        acc = rendered_alpha[None].permute(0, 2, 3, 1)
        dpt = rendered_depth[None].permute(0, 2, 3, 1)
        batch.output.rad = radii[None]  # Store radii for later use
        batch.output.scr = scr  # Store screen space points for later use, # !: BATCH
        return rgb, acc, dpt

    def render_mips(self, xyz: torch.Tensor, sh: torch.Tensor, scale3: torch.Tensor, rot4: torch.Tensor, occ1: torch.Tensor, batch: dotdict):
        # Lazy imports
        from diff_mip_rasterization import rasterize_mips, MipRasterizationSettings, MipRasterizer
        from easyvolcap.utils.gaussian_utils import prepare_gaussian_camera

        # Remove batch dimension
        xyz, sh, scale3, rot4, occ1 = remove_batch([xyz, sh, scale3, rot4, occ1])

        # Prepare the camera transformation for Gaussian
        gaussian_camera = to_x(prepare_gaussian_camera(batch), torch.float)

        if self.training:
            subpixel_offset = torch.rand((int(gaussian_camera.image_height), int(gaussian_camera.image_width), 2), dtype=xyz.dtype, device=xyz.device)
        else:
            subpixel_offset = xyz.new_zeros((int(gaussian_camera.image_height), int(gaussian_camera.image_width), 2))

        # Prepare rasterization settings for gaussian
        raster_settings = MipRasterizationSettings(
            image_height=gaussian_camera.image_height,
            image_width=gaussian_camera.image_width,
            tanfovx=gaussian_camera.tanfovx,
            tanfovy=gaussian_camera.tanfovy,
            bg=torch.full([3], self.bg_brightness if hasattr(self, 'bg_brightness') else 0.0, device=xyz.device),  # GPU
            scale_modifier=self.scale_mod if hasattr(self, 'scale_mod') else 1.0,
            viewmatrix=gaussian_camera.world_view_transform,
            projmatrix=gaussian_camera.full_proj_transform,
            sh_degree=self.sh_deg if hasattr(self, 'sh_deg') else 0,
            campos=gaussian_camera.camera_center,
            prefiltered=False,
            kernel_size=self.kernel_size if hasattr(self, 'kernel_size') else 0.1,
            subpixel_offset=subpixel_offset,
            debug=self.debug if hasattr(self, 'debug') else False,
        )

        # Rasterize visible Gaussians to image, obtain their radii (on screen).
        scr = torch.zeros_like(xyz, requires_grad=True) + 0  # gradient magic
        if scr.requires_grad: scr.retain_grad()
        rasterizer = MipRasterizer(raster_settings=raster_settings)
        rendered_image, radii = typed(torch.float, torch.float)(rasterizer)(
            means3D=xyz,
            means2D=scr,
            shs=sh.mT,
            colors_precomp=None,
            opacities=occ1,
            scales=scale3,
            rotations=rot4,
            cov3D_precomp=None,
        )

        rgb = rendered_image[None].permute(0, 2, 3, 1)
        batch.output.rad = radii[None]  # Store radii for later use
        batch.output.scr = scr  # Store screen space points for later use, # !: BATCH
        return rgb, rgb[..., :1], rgb[..., :1]

    def render_radius(self, xyz: torch.Tensor, sh: torch.Tensor, radius: torch.Tensor, occ1: torch.Tensor, batch: dotdict):
        # Lazy imports
        from diff_point_rasterization import rasterize_points, PointRasterizationSettings, PointRasterizer
        from easyvolcap.utils.gaussian_utils import prepare_gaussian_camera
        assert sh.ndim == 4, 'Should input 4 dim SH: B, N, C, SH'

        # Remove batch dimension
        xyz, sh, radius, occ1 = remove_batch([xyz, sh, radius, occ1])

        # Prepare the camera transformation for Gaussian
        gaussian_camera = to_x(prepare_gaussian_camera(batch), torch.float)

        # Prepare rasterization settings for gaussian
        raster_settings = PointRasterizationSettings(
            image_height=gaussian_camera.image_height,
            image_width=gaussian_camera.image_width,
            tanfovx=gaussian_camera.tanfovx,
            tanfovy=gaussian_camera.tanfovy,
            bg=torch.full([3], self.bg_brightness if hasattr(self, 'bg_brightness') else 0.0, device=xyz.device),  # GPU
            scale_modifier=self.scale_mod if hasattr(self, 'scale_mod') else 1.0,
            viewmatrix=gaussian_camera.world_view_transform,
            projmatrix=gaussian_camera.full_proj_transform,
            sh_degree=self.sh_deg if hasattr(self, 'sh_deg') else 0,
            campos=gaussian_camera.camera_center,
            prefiltered=False,
            debug=self.debug if hasattr(self, 'debug') else False,
        )

        # Rasterize visible Gaussians to image, obtain their radii (on screen).
        scr = torch.zeros_like(xyz, requires_grad=True) + 0  # gradient magic
        if scr.requires_grad: scr.retain_grad()
        rasterizer = PointRasterizer(raster_settings=raster_settings)
        rendered_image, rendered_depth, rendered_alpha, radii = typed(torch.float, torch.float)(rasterizer)(
            means3D=xyz,
            means2D=scr,
            shs=sh.mT,
            colors_precomp=None,
            opacities=occ1,
            radius=radius,
        )

        rgb = rendered_image[None].permute(0, 2, 3, 1)
        acc = rendered_alpha[None].permute(0, 2, 3, 1)
        dpt = rendered_depth[None].permute(0, 2, 3, 1)
        batch.output.rad = radii[None]  # Store radii for later use
        batch.output.scr = scr  # Store screen space points for later use, # !: BATCH
        return rgb, acc, dpt

    @torch.no_grad()
    def update_gaussians(self, batch: dotdict):
        if not self.training: return

        # Prepare global variables
        iter: int = batch.meta.iter  # controls whether we're to update in this iteration
        output = self.last_output  # contains necessary information for updating gaussians
        optimizer: Adam = cfg.runner.optimizer

        # Update for all frames
        if iter > 0 and iter < self.densify_until_iter and iter % self.sh_update_iter == 0:
            for pcd in self.pcds:
                pcd.oneupSHdegree()

        # Update only the rendered frame
        if iter > 0 and iter < self.densify_until_iter and output is not None:

            # The dictionary for controlling the optimizer
            optimizer_state = dotdict()

            # Update all rendered gaussians in the batch
            for i, pcd in enumerate(output.pcd):  # removing batch dim
                pcd: GaussianModel

                # Preparing optimizer states for update
                for name, params in pcd.named_parameters():
                    if params.requires_grad:
                        optimizer_state[params] = dotdict(
                            name=name,
                            old_keep=torch.ones_like(params, dtype=torch.bool, requires_grad=False),
                            new_keep=torch.ones_like(params, dtype=torch.bool, requires_grad=False),
                            new_params=None,
                        )

                # Preparing gaussian stats for update
                radii = output.rad[i]
                visibility_filter = radii > 0
                viewspace_point_tensor = output.scr  # no indexing, otherwise no grad # !: BATCH
                if output.scr.grad is None: continue  # previous rendering was an evaluation
                pcd.max_radii2D[visibility_filter] = torch.max(pcd.max_radii2D[visibility_filter], radii[visibility_filter])
                pcd.add_densification_stats(viewspace_point_tensor, visibility_filter)

                # Perform densification and pruning
                if iter > self.densify_from_iter and iter % self.densification_interval == 0:
                    pcd.densify_and_prune(self.densify_grad_threshold, self.min_opacity, self.scale_mod, self.size_threshold, self.percent_dense, optimizer_state)
                    log(yellow_slim('Densification and pruning done! ' +
                                    f'min opacity: {pcd.get_opacity.min().item():.4f} ' +
                                    f'max opacity: {pcd.get_opacity.max().item():.4f} ' +
                                    f'number of points: {pcd.get_xyz.shape[0]}'))

                # Perform opacity reset
                if iter > self.densify_from_iter and iter % self.opacity_reset_interval == 0:
                    pcd.reset_opacity(optimizer_state)
                    log(yellow_slim('Resetting opacity done! ' +
                                    f'min opacity: {pcd.get_opacity.min().item():.4f} ' +
                                    f'max opacity: {pcd.get_opacity.max().item():.4f}'))

            # Update the actual optimizer states
            update_optimizer_state(optimizer, optimizer_state)

    @torch.no_grad()
    def _load_state_dict_pre_hook(self, state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs):
        # Supports loading points and features with different shapes
        if hasattr(self, 'pcds'):
            pcd_keys = []
            for f, pcd in enumerate(self.pcds):
                for name, params in pcd.named_parameters():
                    params.data = params.data.new_empty(state_dict[f'{prefix}pcds.{f}.{name}'].shape)
                    pcd_keys.append(f'{prefix}pcds.{f}.{name}')
            self.points_aligned = True  # need aligned when new check points are loaded
            self.points_expanded = True  # need aligned when new check points are loaded

            keys = list(state_dict.keys())
            for key in keys:
                if key.startswith(f'{prefix}pcds.') and key not in pcd_keys:
                    del state_dict[key]

        # Historical reason
        if f'{prefix}occs.0' in state_dict:
            del state_dict[f'{prefix}occs.0']

        # Historical reason
        if f'{prefix}rgbs.0' in state_dict:
            del state_dict[f'{prefix}rgbs.0']

        # Historical reason
        if f'{prefix}rads.0' in state_dict:
            del state_dict[f'{prefix}rads.0']

    @torch.no_grad()
    def _load_state_dict_post_hook(self, module, incompatible_keys):
        # Load tighter bounds from the trained models
        dataset: VolumetricVideoDataset = cfg.runner.val_dataloader.dataset
        dataset.vhull_bounds = [dataset.bounds for _ in range(len(self.pcds))]
        for i in range(len(self.pcds)):
            if self.pcds[i] is not None:
                dataset.vhull_bounds[i] = get_bounds(self.pcds[i].get_xyz[None], padding=0.01)[0].cpu()  # MARK: SYNC

    def forward(self, batch: dotdict):
        # Initialization & densification & pruning
        self.init_points(batch)
        self.update_gaussians(batch)

        # Construct renderable parameters
        index, time = self.sample_index_time(batch)
        xyz = torch.stack([self.pcds[l].get_xyz for l in index])  # B, N, 3
        scale3 = torch.stack([self.pcds[l].get_scaling for l in index])  # B, N, 3
        rot4 = torch.stack([self.pcds[l].get_rotation for l in index])  # B, N, 4
        alpha = torch.stack([self.pcds[l].get_opacity for l in index])  # B, N, 1
        sh = torch.stack([self.pcds[l].get_features for l in index]).mT  # B, N, C, SH

        # Perform points rendering
        rgb, acc, dpt = self.render_gaussians(xyz, sh, scale3 * self.scale_mult, rot4, alpha * self.alpha_mult, batch)  # B, HW, C

        # Prepare output
        batch.output.pcd = [self.pcds[l] for l in index]
        self.store_output(None, xyz, rgb, acc, dpt, batch)
        self.last_output = batch.output  # retain gradients after updates