Index lost with 3D sparse convolution

[Pandenotium]

Hi! I have spotted a problem while using the latest version of the library. Everytime when I apply a 3D sparse convolution with stride greater than 1 to a voxel, like the following:

rand_img = torch.randn(1, 3, 100, 100, 100)

mask = torch.randn(1, 1, 100, 100, 100)

rand_img[mask.expand_as(rand_img) < 1.5] = 0

voxels = Voxels.from_dense(rand_img, voxel_size=0.05).to("cuda:0")

conv = SparseConv3d(
    3,
    64,
    3,
    stride=2,
    out_code_backend="hashmap"
).to("cuda:0")

out = conv(voxels)

it always yields the error:

AssertionError: Offsets tensor([ 0, 53226], dtype=torch.int32) does not match tensors torch.Size([66909, 64])

It seems that some index are lost. This error continues to appear with greater batch size and voxels initialized with real point clouds, and index lost happens in every batch. In fact, this error never appeared before the commit ef7a0ca with exactly the same code. Could this be a misuse of the interface or a bug?

Thanks!

[chrischoy]

Try

B, C, H, W, D = 2, 3, 100, 100, 100
rand_img = torch.randn(B, C, H, W, D)
mask = torch.rand(B, 1, H, W, D) < 0.5
rand_img[mask.expand_as(rand_img)] = 0
voxels = Voxels.from_dense(rand_img, voxel_size=0.05).to("cuda:0")

on the latest commit.