def test_voxelgrids_dim(self, device, dtype):
# The dimension of voxelgrids should be 4 (batched).
with pytest.raises(ValueError,
match="Expected voxelgrids to have 4 dimensions "
"but got 3 dimensions."):
voxelgrids = torch.ones([6, 6, 6], device=device, dtype=dtype)
vg.downsample(voxelgrids, [2, 2, 2])
def test_scale_val_2(self, device, dtype):
# Every element in the scale should be greater or equal to one.
with pytest.raises(ValueError,
match="Downsample ratio must be at least 1 "
"along every dimension but got -1 at "
"index 0."):
voxelgrids = torch.ones([2, 6, 6, 6], device=device, dtype=dtype)
vg.downsample(voxelgrids, [-1, 3, 2])
def test_scale_val_1(self, device, dtype):
# The scale should be smaller or equal to the size of the input.
with pytest.raises(
ValueError,
match="Downsample ratio must be less than voxelgrids "
"shape of 6 at index 2, but got 7."):
voxelgrids = torch.ones([2, 6, 6, 6], device=device, dtype=dtype)
vg.downsample(voxelgrids, [1, 2, 7])
def test_scale_dim(self, device, dtype):
# The dimension of scale should be 3 if it is a list.
with pytest.raises(ValueError,
match="Expected scale to have 3 dimensions "
"but got 2 dimensions."):
voxelgrids = torch.ones([2, 6, 6, 6], device=device, dtype=dtype)
vg.downsample(voxelgrids, [2, 2])
with pytest.raises(TypeError,
match="Expected scale to be type list or int "
"but got <class 'str'>."):
voxelgrids = torch.ones([2, 6, 6, 6], device=device, dtype=dtype)
vg.downsample(voxelgrids, "2")
def test_bool_input(self, device, dtype):
if dtype != torch.bool:
pytest.skip("This test is only for torch.bool.")
voxelgrids = torch.ones((2, 4, 4, 4), device=device, dtype=dtype)
voxelgrids[:, :, 1, :] = 0
voxelgrids[:, :, 3, :] = 0
output = vg.downsample(voxelgrids, 2)
expected_dtype = torch.half if device == "cuda" else torch.float
expected = torch.ones(
(2, 2, 2, 2), device=device, dtype=expected_dtype) * 0.5
assert torch.equal(output, expected)
def test_output_batch(self, device, dtype):
if dtype == torch.bool:
pytest.skip("This test won't work for torch.bool.")
# The size of the batched input shoud be correct.
# For example, if the input size is [2, 6, 6, 6],
# Scale is [3, 3, 3], the output size should be [2, 2, 2, 2]
# Also, test the function is numerically correct
voxelgrid1 = torch.ones([4, 4, 4], device=device, dtype=dtype)
voxelgrid2 = torch.ones((4, 4, 4), device=device, dtype=dtype)
voxelgrid2[1, :2] = 0.8
voxelgrid2[1, 2:] = 0.4
voxelgrid2[3] = 0
batched_voxelgrids = torch.stack((voxelgrid1, voxelgrid2))
output = vg.downsample(batched_voxelgrids, [2, 2, 2])
expected1 = torch.ones((2, 2, 2), device=device, dtype=dtype)
expected2 = torch.tensor(
[[[0.9, 0.9], [0.7, 0.7]], [[0.5000, 0.5000], [0.5000, 0.5000]]],
device=device,
dtype=dtype)
expected = torch.stack((expected1, expected2))
assert torch.allclose(output, expected)
def test_output_size(self, device, dtype):
# The size of the output should be input.shape / scale
voxelgrids = torch.ones([3, 6, 6, 6], device=device, dtype=dtype)
output = vg.downsample(voxelgrids, [1, 2, 3])
assert (output.shape == torch.Size([3, 6, 3, 2]))