souging/TRELLIS_TextTo3D

1from typing import *2import torch3import torch.nn as nn4from . import SparseTensor5 6__all__ = [7    'SparseDownsample',8    'SparseUpsample',9    'SparseSubdivide'10]11 12 13class SparseDownsample(nn.Module):14    """15    Downsample a sparse tensor by a factor of `factor`.16    Implemented as average pooling.17    """18    def __init__(self, factor: Union[int, Tuple[int, ...], List[int]]):19        super(SparseDownsample, self).__init__()20        self.factor = tuple(factor) if isinstance(factor, (list, tuple)) else factor21 22    def forward(self, input: SparseTensor) -> SparseTensor:23        DIM = input.coords.shape[-1] - 124        factor = self.factor if isinstance(self.factor, tuple) else (self.factor,) * DIM25        assert DIM == len(factor), 'Input coordinates must have the same dimension as the downsample factor.'26 27        coord = list(input.coords.unbind(dim=-1))28        for i, f in enumerate(factor):29            coord[i+1] = coord[i+1] // f30 31        MAX = [coord[i+1].max().item() + 1 for i in range(DIM)]32        OFFSET = torch.cumprod(torch.tensor(MAX[::-1]), 0).tolist()[::-1] + [1]33        code = sum([c * o for c, o in zip(coord, OFFSET)])34        code, idx = code.unique(return_inverse=True)35 36        new_feats = torch.scatter_reduce(37            torch.zeros(code.shape[0], input.feats.shape[1], device=input.feats.device, dtype=input.feats.dtype),38            dim=0,39            index=idx.unsqueeze(1).expand(-1, input.feats.shape[1]),40            src=input.feats,41            reduce='mean'42        )43        new_coords = torch.stack(44            [code // OFFSET[0]] +45            [(code // OFFSET[i+1]) % MAX[i] for i in range(DIM)],46            dim=-147        )48        out = SparseTensor(new_feats, new_coords, input.shape,)49        out._scale = tuple([s // f for s, f in zip(input._scale, factor)])50        out._spatial_cache = input._spatial_cache51 52        out.register_spatial_cache(f'upsample_{factor}_coords', input.coords)53        out.register_spatial_cache(f'upsample_{factor}_layout', input.layout)54        out.register_spatial_cache(f'upsample_{factor}_idx', idx)55 56        return out57 58 59class SparseUpsample(nn.Module):60    """61    Upsample a sparse tensor by a factor of `factor`.62    Implemented as nearest neighbor interpolation.63    """64    def __init__(self, factor: Union[int, Tuple[int, int, int], List[int]]):65        super(SparseUpsample, self).__init__()66        self.factor = tuple(factor) if isinstance(factor, (list, tuple)) else factor67 68    def forward(self, input: SparseTensor) -> SparseTensor:69        DIM = input.coords.shape[-1] - 170        factor = self.factor if isinstance(self.factor, tuple) else (self.factor,) * DIM71        assert DIM == len(factor), 'Input coordinates must have the same dimension as the upsample factor.'72 73        new_coords = input.get_spatial_cache(f'upsample_{factor}_coords')74        new_layout = input.get_spatial_cache(f'upsample_{factor}_layout')75        idx = input.get_spatial_cache(f'upsample_{factor}_idx')76        if any([x is None for x in [new_coords, new_layout, idx]]):77            raise ValueError('Upsample cache not found. SparseUpsample must be paired with SparseDownsample.')78        new_feats = input.feats[idx]79        out = SparseTensor(new_feats, new_coords, input.shape, new_layout)80        out._scale = tuple([s * f for s, f in zip(input._scale, factor)])81        out._spatial_cache = input._spatial_cache82        return out83    84class SparseSubdivide(nn.Module):85    """86    Upsample a sparse tensor by a factor of `factor`.87    Implemented as nearest neighbor interpolation.88    """89    def __init__(self):90        super(SparseSubdivide, self).__init__()91 92    def forward(self, input: SparseTensor) -> SparseTensor:93        DIM = input.coords.shape[-1] - 194        # upsample scale=2^DIM95        n_cube = torch.ones([2] * DIM, device=input.device, dtype=torch.int)96        n_coords = torch.nonzero(n_cube)97        n_coords = torch.cat([torch.zeros_like(n_coords[:, :1]), n_coords], dim=-1)98        factor = n_coords.shape[0]99        assert factor == 2 ** DIM100        # print(n_coords.shape)101        new_coords = input.coords.clone()102        new_coords[:, 1:] *= 2103        new_coords = new_coords.unsqueeze(1) + n_coords.unsqueeze(0).to(new_coords.dtype)104        105        new_feats = input.feats.unsqueeze(1).expand(input.feats.shape[0], factor, *input.feats.shape[1:])106        out = SparseTensor(new_feats.flatten(0, 1), new_coords.flatten(0, 1), input.shape)107        out._scale = input._scale * 2108        out._spatial_cache = input._spatial_cache109        return out110 111