souging/TRELLIS_TextTo3D

1import torch2import torch.nn as nn3from .. import SparseTensor4 5 6class SparseConv3d(nn.Module):7    def __init__(self, in_channels, out_channels, kernel_size, stride=1, dilation=1, bias=True, indice_key=None):8        super(SparseConv3d, self).__init__()9        if 'torchsparse' not in globals():10            import torchsparse11        self.conv = torchsparse.nn.Conv3d(in_channels, out_channels, kernel_size, stride, 0, dilation, bias)12 13    def forward(self, x: SparseTensor) -> SparseTensor:14        out = self.conv(x.data)15        new_shape = [x.shape[0], self.conv.out_channels]16        out = SparseTensor(out, shape=torch.Size(new_shape), layout=x.layout if all(s == 1 for s in self.conv.stride) else None)17        out._spatial_cache = x._spatial_cache18        out._scale = tuple([s * stride for s, stride in zip(x._scale, self.conv.stride)])19        return out20 21 22class SparseInverseConv3d(nn.Module):23    def __init__(self, in_channels, out_channels, kernel_size, stride=1, dilation=1, bias=True, indice_key=None):24        super(SparseInverseConv3d, self).__init__()25        if 'torchsparse' not in globals():26            import torchsparse27        self.conv = torchsparse.nn.Conv3d(in_channels, out_channels, kernel_size, stride, 0, dilation, bias, transposed=True)28 29    def forward(self, x: SparseTensor) -> SparseTensor:30        out = self.conv(x.data)        31        new_shape = [x.shape[0], self.conv.out_channels]32        out = SparseTensor(out, shape=torch.Size(new_shape), layout=x.layout if all(s == 1 for s in self.conv.stride) else None)33        out._spatial_cache = x._spatial_cache34        out._scale = tuple([s // stride for s, stride in zip(x._scale, self.conv.stride)])35        return out36 37 38 39