How to use PointToVoxel on GPU?

[KyleYueye]

I've noticed the WARNING: you MUST construct PointToVoxel AFTER set device. in class PointToVoxel in /spconv/pytorch/utils.py. What can i do to change the previous code which was written for VoxelGenerator to suit for PointToVoxel? It is running on cpu and the memory is no enough at all.

[fuyawangye]

init PointToVoxel after dataloader multiprocessing

class VoxelGen():
    def __init__(*args):
        self._is_init = False
    def init():
        self._is_init = True
        self.point2voxel = PointToVoxel(.....)
    def __call__(points):
        if not self._is_init:
            self.init()
        return self.point2voxel(points)

[KyleYueye]

init PointToVoxel after dataloader multiprocessing

def transform_points_to_voxels(self, data_dict=None, config=None, voxel_generator=None):
    if data_dict is None:
        voxel_generator = VoxelGen(
            vsize_xyz=config.VOXEL_SIZE,
            coors_range_xyz=self.point_cloud_range,
            num_point_features=config.NUM_POINT_FEATURES,
            max_num_points_per_voxel=config.MAX_POINTS_PER_VOXEL,
            max_num_voxels=config.MAX_NUMBER_OF_VOXELS[self.mode],
            device=torch.device('cuda')
        )
        grid_size = (self.point_cloud_range[3:6] - self.point_cloud_range[0:3]) / np.array(config.VOXEL_SIZE)
        self.grid_size = np.round(grid_size).astype(np.int64)
        self.voxel_size = config.VOXEL_SIZE
        return partial(self.transform_points_to_voxels, voxel_generator=voxel_generator)

    points = data_dict['points']
    points = torch.from_numpy(points).to(torch.device("cuda"))
    voxel_output = voxel_generator(points)
    .........

I used it in pcdet/datasets/processor/data_processor.py but the bug RuntimeError: Cannot re-initialize CUDA in forked subprocess. To use CUDA with multiprocessing, you must use the 'spawn' start method still occurrs. What might the problem be?