ValueError: Invalid shape for means3d: torch.Size([0, 3])

[chengluberkeley]

Describe the bug Trained splatfacto till 600 steps, then reports the error.

Traceback (most recent call last):
  File "/nerfstudio/scripts/train.py", line 294, in <module>
    entrypoint()
  File "/nerfstudio/scripts/train.py", line 285, in entrypoint
    main(
  File "/nerfstudio/scripts/train.py", line 270, in main
    launch(
  File "/nerfstudio/scripts/train.py", line 199, in launch
    main_func(local_rank=0, world_size=world_size, config=config)
  File "/nerfstudio/scripts/train.py", line 106, in train_loop
    trainer.train()
  File "/nerfstudio/engine/trainer.py", line 273, in train
    loss, loss_dict, metrics_dict = self.train_iteration(step)
  File "/nerfstudio/utils/profiler.py", line 130, in inner
    out = func(*args, **kwargs)
  File "/nerfstudio/engine/trainer.py", line 527, in train_iteration
    _, loss_dict, metrics_dict = self.pipeline.get_train_loss_dict(step=step)
  File "/nerfstudio/utils/profiler.py", line 130, in inner
    out = func(*args, **kwargs)
  File "/nerfstudio/pipelines/base_pipeline.py", line 326, in get_train_loss_dict
    model_outputs = self._model(
  File "/opt/python/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1518, in _wrapped_call_impl
    return self._call_impl(*args, **kwargs)
  File "/opt/python/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1527, in _call_impl
    return forward_call(*args, **kwargs)
  File "/nerfstudio/models/base_model.py", line 153, in forward
    return self.get_outputs(ray_bundle)
  File "/nerfstudio/models/splatfacto.py", line 838, in get_outputs
    self.xys, depths, self.radii, conics, comp, num_tiles_hit, cov3d = project_gaussians(  # type: ignore
  File "/opt/python/lib/python3.9/site-packages/gsplat/project_gaussians.py", line 61, in project_gaussians
    return _ProjectGaussians.apply(
  File "/opt/python/lib/python3.9/site-packages/torch/autograd/function.py", line 539, in apply
    return super().apply(*args, **kwargs)  # type: ignore[misc]
  File "/opt/python/lib/python3.9/site-packages/gsplat/project_gaussians.py", line 102, in forward
    raise ValueError(f"Invalid shape for means3d: {means3d.shape}")
ValueError: Invalid shape for means3d: torch.Size([0, 3])

Default training parameters are used.

NerfStudio: 1.0.3 gsplat: 0.1.6

Thank you for the help!

[ginazhouhuiwu]

Hi! What dataset are you training on? means3d: torch.Size([0, 3]) means there are 0 points hence the invalid shape error in projection step. I cannot recreate this error on my end, can you provide more details please?

[Filocava99]

Having the same issue, the same dataset used to work fine until today, now it started to give this error, using default parameters with splatfacto

[Filocava99]

Everytime it starts it downloads some weights for pytorch: have those been changed and are them causing the issue?

[jb-ye]

Could you upgrade your gsplat library?