ops.add_points_features_to_volume_densities_features Error

[hywkim-brian]

Running this code outputs the following error although I haven't specified any "grid_size" (in which case should be set to None and not raise any such errors). The parameter grid_size is not even documented in the API documentation. Could you look into this issue? The forward() function seems to be checking errors for all parameters, including those that are optional. (voxel_ds.cloud() in my function is a valid PointCloud datatype). Thanks!

features=torch.zeros(len(voxel_ds.cloud.points_list()), 10000, 5)
volume_densities, volume_features = ops.add_points_features_to_volume_densities_features(torch.stack(voxel_ds.cloud.points_list(), axis=0), features, torch.zeros(30, 1, 32, 32, 32),torch.zeros(30, 5, 32, 32, 32))

yield the error
/usr/local/lib/python3.7/dist-packages/pytorch3d/ops/points_to_volumes.py in forward(ctx, points_3d, points_features, volume_densities, volume_features, grid_sizes, point_weight, mask, align_corners, splat)
108 raise ValueError("Bad volume_features shape")
109 if grid_sizes.dtype != torch.int64:
--> 110 raise ValueError("grid_sizes must be int64")
111 N4, D1 = grid_sizes.shape
112 if N4 != N or D1 != 3:

ValueError: grid_sizes must be int64

[bottler]

Indeed this is a new bug when the optional grid_sizes argument is not provided.
To fix your code, you can provide grid_sizes = torch.full((30, 3), 32, dtype=torch.int64) to the call to add_points_features_to_volume_densities_features.

The fix to pytorch3d will be that

# init the volumetric grid sizes if uninitialized
if grid_sizes is None:
    # grid sizes shape (minibatch, 3)
    grid_sizes = (
        torch.LongTensor(list(volume_densities.shape[2:]))
        .to(volume_densities)
        .expand(volume_densities.shape[0], 3)
    )

should be

# init the volumetric grid sizes if uninitialized
if grid_sizes is None:
    # grid sizes shape (minibatch, 3)
    grid_sizes = (
        torch.LongTensor(list(volume_densities.shape[2:]))
        .to(volume_densities.device)
        .expand(volume_densities.shape[0], 3)
    )

in pytorch3d/ops/points_to_volumes.py around line 364.

[hywkim-brian]

Thanks for the update!

[hywkim-brian]

Possibly a similar issue for mask.. Running the code with grid_sizes = torch.full((30, 3), 32, dtype=torch.int64)
yields the error.

    112         if N4 != N or D1 != 3:
    113             raise ValueError("Bad grid_sizes.shape")
--> 114         if mask.dtype != torch.float32:
    115             raise ValueError("mask must be float32")
    116         N5, P2 = mask.shape

AttributeError: 'NoneType' object has no attribute 'dtype'```

[bottler]

Yes indeed. You need to supply mask = torch.ones(30,10000) in the call as well.