Invalid automatic min_coordinate calculation for SparseTensor.dense

[matthewdm0816]

Document on SparseTensor indicates that, min_coordinate will be automatically computed if 0 is given. But it seems that the first assertion will throw error for any non-torch tensor inputs. There are correct codes below but will never be used under the first assert.

MinkowskiEngine/MinkowskiEngine/MinkowskiSparseTensor.py

Lines 488 to 526 in 02fc608

	if min_coordinate is not None:
	assert isinstance(min_coordinate, torch.IntTensor)
	assert min_coordinate.numel() == self._D
	if shape is not None:
	assert isinstance(shape, torch.Size)
	assert len(shape) == self._D + 2 # batch and channel
	if shape[1] != self._F.size(1):
	shape = torch.Size([shape[0], self._F.size(1), *[s for s in shape[2:]]])
	# Exception handling for empty tensor
	if self.__len__() == 0:
	assert shape is not None, "shape is required to densify an empty tensor"
	return (
	torch.zeros(shape, dtype=self.dtype, device=self.device),
	torch.zeros(self._D, dtype=torch.int32, device=self.device),
	self.tensor_stride,
	)
	# Use int tensor for all operations
	tensor_stride = torch.IntTensor(self.tensor_stride).to(self.device)
	# New coordinates
	batch_indices = self.C[:, 0]
	if min_coordinate is None:
	min_coordinate, _ = self.C.min(0, keepdim=True)
	min_coordinate = min_coordinate[:, 1:]
	if not torch.all(min_coordinate >= 0):
	raise ValueError(
	f"Coordinate has a negative value: {min_coordinate}. Please provide min_coordinate argument"
	)
	coords = self.C[:, 1:]
	elif isinstance(min_coordinate, int) and min_coordinate == 0:
	coords = self.C[:, 1:]
	else:
	min_coordinate = min_coordinate.to(self.device)
	if min_coordinate.ndim == 1:
	min_coordinate = min_coordinate.unsqueeze(0)
	coords = self.C[:, 1:] - min_coordinate