def _get_coordinate_map_key(
input: SparseTensor,
coordinates: torch.Tensor = None,
tensor_stride: StrideType = 1,
expand_coordinates: bool = False,
):
r"""Returns the coordinates map key."""
if coordinates is not None and not expand_coordinates:
assert isinstance(coordinates,
(CoordinateMapKey, torch.Tensor, SparseTensor))
if isinstance(coordinates, torch.Tensor):
assert coordinates.ndim == 2
coordinate_map_key = CoordinateMapKey(
convert_to_int_list(tensor_stride,
coordinates.size(1) - 1), "")
(
coordinate_map_key,
(unique_index, inverse_mapping),
) = input._manager.insert_and_map(coordinates,
*coordinate_map_key.get_key())
elif isinstance(coordinates, SparseTensor):
coordinate_map_key = coordinates.coordinate_map_key
else: # CoordinateMapKey type due to the previous assertion
coordinate_map_key = coordinates
else: # coordinates is None
coordinate_map_key = CoordinateMapKey(
input.coordinate_map_key.get_coordinate_size())
return coordinate_map_key
def forward(
ctx,
input_features: torch.Tensor,
pooling_mode: PoolingMode,
in_coordinate_map_key: CoordinateMapKey,
out_coordinate_map_key: CoordinateMapKey = None,
coordinate_manager: CoordinateManager = None,
):
if out_coordinate_map_key is None:
out_coordinate_map_key = CoordinateMapKey(
in_coordinate_map_key.get_coordinate_size())
input_features = input_features.contiguous()
ctx.input_features = input_features
ctx.in_coords_key = in_coordinate_map_key
ctx.out_coords_key = out_coordinate_map_key
ctx.coordinate_manager = coordinate_manager
ctx.pooling_mode = pooling_mode
fw_fn = get_minkowski_function("GlobalPoolingForward", input_features)
out_feat, num_nonzero = fw_fn(
input_features,
pooling_mode,
ctx.in_coords_key,
ctx.out_coords_key,
ctx.coordinate_manager._manager,
)
ctx.num_nonzero = num_nonzero
return out_feat
def forward(
ctx,
in_feat: torch.Tensor,
in_coords_key: CoordinateMapKey,
glob_coords_key: CoordinateMapKey = None,
coords_manager: CoordinateManager = None,
gpooling_mode=PoolingMode.GLOBAL_AVG_POOLING_KERNEL,
):
if glob_coords_key is None:
glob_coords_key = CoordinateMapKey(
in_coords_key.get_coordinate_size())
gpool_avg_forward = get_minkowski_function("GlobalPoolingForward",
in_feat)
broadcast_forward = get_minkowski_function("BroadcastForward", in_feat)
mean, num_nonzero = gpool_avg_forward(
in_feat,
gpooling_mode,
in_coords_key,
glob_coords_key,
coords_manager._manager,
)
# X - \mu
centered_feat = broadcast_forward(
in_feat,
-mean,
BroadcastMode.ELEMENTWISE_ADDITON,
in_coords_key,
glob_coords_key,
coords_manager._manager,
)
# Variance = 1/N \sum (X - \mu) ** 2
variance, num_nonzero = gpool_avg_forward(
centered_feat**2,
gpooling_mode,
in_coords_key,
glob_coords_key,
coords_manager._manager,
)
# norm_feat = (X - \mu) / \sigma
inv_std = 1 / (variance + 1e-8).sqrt()
norm_feat = broadcast_forward(
centered_feat,
inv_std,
BroadcastMode.ELEMENTWISE_MULTIPLICATION,
in_coords_key,
glob_coords_key,
coords_manager._manager,
)
ctx.saved_vars = (in_coords_key, glob_coords_key, coords_manager,
gpooling_mode)
# For GPU tensors, must use save_for_backward.
ctx.save_for_backward(inv_std, norm_feat)
return norm_feat
def forward(self, input: SparseTensor, mask: torch.Tensor):
r"""
Args:
:attr:`input` (:attr:`MinkowskiEnigne.SparseTensor`): a sparse tensor
to remove coordinates from.
:attr:`mask` (:attr:`torch.BoolTensor`): mask vector that specifies
which one to keep. Coordinates with False will be removed.
Returns:
A :attr:`MinkowskiEngine.SparseTensor` with C = coordinates
corresponding to `mask == True` F = copy of the features from `mask ==
True`.
Example::
>>> # Define inputs
>>> input = SparseTensor(feats, coords=coords)
>>> # Any boolean tensor can be used as the filter
>>> mask = torch.rand(feats.size(0)) < 0.5
>>> pruning = MinkowskiPruning()
>>> output = pruning(input, mask)
"""
assert isinstance(input, SparseTensor)
out_coords_key = CoordinateMapKey(
input.coordinate_map_key.get_coordinate_size())
output = self.pruning.apply(input.F, mask, input.coordinate_map_key,
out_coords_key, input._manager)
return SparseTensor(output,
coordinate_map_key=out_coords_key,
coordinate_manager=input._manager)
def _binary_functor(self, other, binary_fn):
assert isinstance(other, (self.__class__, torch.Tensor))
if isinstance(other, self.__class__):
assert self._manager == other._manager, COORDINATE_MANAGER_DIFFERENT_ERROR
if self.coordinate_map_key == other.coordinate_map_key:
return self.__class__(
binary_fn(self._F, other.F),
coordinate_map_key=self.coordinate_map_key,
coordinate_manager=self._manager,
)
else:
# Generate union maps
out_key = CoordinateMapKey(
self.coordinate_map_key.get_coordinate_size())
union_maps = self.coordinate_manager.union_map(
[self.coordinate_map_key, other.coordinate_map_key],
out_key)
N_out = self.coordinate_manager.size(out_key)
out_F = torch.zeros((N_out, self._F.size(1)),
dtype=self.dtype,
device=self.device)
out_F[union_maps[0][1]] = self._F[union_maps[0][0]]
out_F[union_maps[1][1]] = binary_fn(out_F[union_maps[1][1]],
other._F[union_maps[1][0]])
return self.__class__(out_F,
coordinate_map_key=out_key,
coordinate_manager=self._manager)
else: # when it is a torch.Tensor
return self.__class__(
binary_fn(self._F, other),
coordinate_map_key=self.coordinate_map_key,
coordinate_manager=self._manager,
)
def forward(
self,
input,
coordinates: Union[torch.IntTensor, CoordinateMapKey,
SparseTensor] = None,
):
# Get a new coordinate map key or extract one from the coordinates
if isinstance(input, ME.TensorField):
in_coordinate_map_key = input.coordinate_field_map_key
out_coordinate_map_key = CoordinateMapKey(
input.coordinate_field_map_key.get_coordinate_size())
else:
in_coordinate_map_key = input.coordinate_map_key
out_coordinate_map_key = _get_coordinate_map_key(
input, coordinates)
output = self.pooling.apply(
input.F,
self.pooling_mode,
in_coordinate_map_key,
out_coordinate_map_key,
input._manager,
)
return SparseTensor(
output,
coordinate_map_key=out_coordinate_map_key,
coordinate_manager=input.coordinate_manager,
)
def forward(self, *inputs):
r"""
Args:
A variable number of :attr:`MinkowskiEngine.SparseTensor`'s.
Returns:
A :attr:`MinkowskiEngine.SparseTensor` with coordinates = union of all
input coordinates, and features = sum of all features corresponding to the
coordinate.
Example::
>>> # Define inputs
>>> input1 = SparseTensor(
>>> torch.rand(N, in_channels, dtype=torch.double), coords=coords)
>>> # All inputs must share the same coordinate manager
>>> input2 = SparseTensor(
>>> torch.rand(N, in_channels, dtype=torch.double),
>>> coords=coords + 1,
>>> coords_manager=input1.coordinate_manager, # Must use same coords manager
>>> force_creation=True # The tensor stride [1, 1] already exists.
>>> )
>>> union = MinkowskiUnion()
>>> output = union(input1, iput2)
"""
assert isinstance(inputs,
(list, tuple)), "The input must be a list or tuple"
for s in inputs:
assert isinstance(s,
SparseTensor), "Inputs must be sparse tensors."
assert len(
inputs) > 1, "input must be a set with at least 2 SparseTensors"
# Assert the same coordinate manager
ref_coordinate_manager = inputs[0].coordinate_manager
for s in inputs:
assert (
ref_coordinate_manager == s.coordinate_manager
), "Invalid coordinate manager. All inputs must have the same coordinate manager."
in_coordinate_map_key = inputs[0].coordinate_map_key
coordinate_manager = inputs[0].coordinate_manager
out_coordinate_map_key = CoordinateMapKey(
in_coordinate_map_key.get_coordinate_size())
output = self.union.apply(
[input.coordinate_map_key for input in inputs],
out_coordinate_map_key,
coordinate_manager,
*[input.F for input in inputs],
)
return SparseTensor(
output,
coordinate_map_key=out_coordinate_map_key,
coordinate_manager=coordinate_manager,
)
def forward(
ctx,
input_features: torch.Tensor,
kernel_weights: torch.Tensor,
kernel_generator: KernelGenerator,
convolution_mode: ConvolutionMode,
in_coordinate_map_key: CoordinateMapKey,
out_coordinate_map_key: CoordinateMapKey = None,
coordinate_manager: CoordinateManager = None,
):
if out_coordinate_map_key is None:
out_coordinate_map_key = CoordinateMapKey(
in_coordinate_map_key.get_coordinate_size())
input_features = input_features.contiguous()
ctx.input_features = input_features
ctx.kernel_weights = kernel_weights
ctx.misc = [
kernel_generator,
convolution_mode,
in_coordinate_map_key,
out_coordinate_map_key,
coordinate_manager,
]
fw_fn = get_minkowski_function("ConvolutionForward", input_features)
return fw_fn(
ctx.input_features,
kernel_weights,
kernel_generator.kernel_size,
kernel_generator.kernel_stride,
kernel_generator.kernel_dilation,
kernel_generator.region_type,
kernel_generator.region_offsets,
kernel_generator.expand_coordinates,
convolution_mode,
in_coordinate_map_key,
out_coordinate_map_key,
coordinate_manager._manager,
)
def forward(
ctx,
input_features: torch.Tensor,
pooling_mode: PoolingMode,
kernel_generator: KernelGenerator,
in_coordinate_map_key: CoordinateMapKey,
out_coordinate_map_key: CoordinateMapKey = None,
coordinate_manager: CoordinateManager = None,
):
if out_coordinate_map_key is None:
out_coordinate_map_key = CoordinateMapKey(
in_coordinate_map_key.get_coordinate_size())
input_features = input_features.contiguous()
ctx.input_features = input_features
ctx = save_ctx(
ctx,
kernel_generator,
in_coordinate_map_key,
out_coordinate_map_key,
coordinate_manager,
)
ctx.pooling_mode = pooling_mode
fw_fn = get_minkowski_function("LocalPoolingTransposeForward",
input_features)
out_feat, num_nonzero = fw_fn(
ctx.input_features,
kernel_generator.kernel_size,
kernel_generator.kernel_stride,
kernel_generator.kernel_dilation,
kernel_generator.region_type,
kernel_generator.region_offsets,
kernel_generator.expand_coordinates,
pooling_mode,
ctx.in_coordinate_map_key,
ctx.out_coordinate_map_key,
ctx.coordinate_manager._manager,
)
ctx.num_nonzero = num_nonzero
return out_feat
def inverse_mapping(self, sparse_tensor_map_key: CoordinateMapKey):
if sparse_tensor_map_key not in self._inverse_mapping:
if not self._manager.exists_field_to_sparse(
self.coordinate_field_map_key, sparse_tensor_map_key
):
sparse_keys = self.coordinate_manager.field_to_sparse_keys(
self.coordinate_field_map_key
)
one_key = None
if len(sparse_keys) > 0:
for key in sparse_keys:
if np.prod(key.get_tensor_stride()) == 1:
one_key = key
else:
one_key = CoordinateMapKey(
[
1,
]
* self.D,
"",
)
if one_key not in self._inverse_mapping:
(
_,
self._inverse_mapping[one_key],
) = self._manager.get_field_to_sparse_map(
self.coordinate_field_map_key, one_key
)
_, stride_map = self.coordinate_manager.stride_map(
one_key, sparse_tensor_map_key
)
field_map = self._inverse_mapping[one_key]
self._inverse_mapping[sparse_tensor_map_key] = stride_map[field_map]
else:
# Extract the mapping
(
_,
self._inverse_mapping[sparse_tensor_map_key],
) = self._manager.get_field_to_sparse_map(
self.coordinate_field_map_key, sparse_tensor_map_key
)
return self._inverse_mapping[sparse_tensor_map_key]
def __init__(
self,
features: torch.Tensor,
coordinates: torch.Tensor = None,
# optional coordinate related arguments
tensor_stride: StrideType = 1,
coordinate_field_map_key: CoordinateMapKey = None,
coordinate_manager: CoordinateManager = None,
quantization_mode:
SparseTensorQuantizationMode = SparseTensorQuantizationMode.
UNWEIGHTED_AVERAGE,
# optional manager related arguments
allocator_type: GPUMemoryAllocatorType = None,
minkowski_algorithm: MinkowskiAlgorithm = None,
requires_grad=None,
device=None,
):
r"""
Args:
:attr:`features` (:attr:`torch.FloatTensor`,
:attr:`torch.DoubleTensor`, :attr:`torch.cuda.FloatTensor`, or
:attr:`torch.cuda.DoubleTensor`): The features of a sparse
tensor.
:attr:`coordinates` (:attr:`torch.IntTensor`): The coordinates
associated to the features. If not provided, :attr:`coordinate_map_key`
must be provided.
:attr:`tensor_stride` (:attr:`int`, :attr:`list`,
:attr:`numpy.array`, or :attr:`tensor.Tensor`): The tensor stride
of the current sparse tensor. By default, it is 1.
:attr:`coordinate_field_map_key`
(:attr:`MinkowskiEngine.CoordinateMapKey`): When the coordinates
are already cached in the MinkowskiEngine, we could reuse the same
coordinate map by simply providing the coordinate map key. In most
case, this process is done automatically. When you provide a
`coordinate_field_map_key`, `coordinates` will be be ignored.
:attr:`coordinate_manager`
(:attr:`MinkowskiEngine.CoordinateManager`): The MinkowskiEngine
manages all coordinate maps using the `_C.CoordinateMapManager`. If
not provided, the MinkowskiEngine will create a new computation
graph. In most cases, this process is handled automatically and you
do not need to use this.
:attr:`quantization_mode`
(:attr:`MinkowskiEngine.SparseTensorQuantizationMode`): Defines how
continuous coordinates will be quantized to define a sparse tensor.
Please refer to :attr:`SparseTensorQuantizationMode` for details.
:attr:`allocator_type`
(:attr:`MinkowskiEngine.GPUMemoryAllocatorType`): Defines the GPU
memory allocator type. By default, it uses the c10 allocator.
:attr:`minkowski_algorithm`
(:attr:`MinkowskiEngine.MinkowskiAlgorithm`): Controls the mode the
minkowski engine runs, Use
:attr:`MinkowskiAlgorithm.MEMORY_EFFICIENT` if you want to reduce
the memory footprint. Or use
:attr:`MinkowskiAlgorithm.SPEED_OPTIMIZED` if you want to make it
run fasterat the cost of more memory.
:attr:`requires_grad` (:attr:`bool`): Set the requires_grad flag.
:attr:`device` (:attr:`torch.device`): Set the device the sparse
tensor is defined.
"""
# Type checks
assert isinstance(features,
torch.Tensor), "Features must be a torch.Tensor"
assert (
features.ndim == 2
), f"The feature should be a matrix, The input feature is an order-{features.ndim} tensor."
assert isinstance(quantization_mode, SparseTensorQuantizationMode)
assert quantization_mode in [
SparseTensorQuantizationMode.UNWEIGHTED_AVERAGE,
SparseTensorQuantizationMode.UNWEIGHTED_SUM,
SparseTensorQuantizationMode.RANDOM_SUBSAMPLE,
SparseTensorQuantizationMode.MAX_POOL,
], "invalid quantization mode"
self.quantization_mode = quantization_mode
if coordinates is not None:
assert isinstance(coordinates, torch.Tensor)
if coordinate_field_map_key is not None:
assert isinstance(coordinate_field_map_key, CoordinateMapKey)
if coordinate_manager is not None:
assert isinstance(coordinate_manager, CoordinateManager)
if coordinates is None and (coordinate_field_map_key is None
or coordinate_manager is None):
raise ValueError(
"Either coordinates or (coordinate_field_map_key, coordinate_manager) pair must be provided."
)
Tensor.__init__(self)
# To device
if device is not None:
features = features.to(device)
if coordinates is not None:
# assertion check for the map key done later
coordinates = coordinates.to(device)
self._D = (coordinates.size(1) -
1 if coordinates is not None else coordinate_manager.D)
##########################
# Setup CoordsManager
##########################
if coordinate_manager is None:
# If set to share the coords man, use the global coords man
if (sparse_tensor_operation_mode() ==
SparseTensorOperationMode.SHARE_COORDINATE_MANAGER):
coordinate_manager = global_coordinate_manager()
if coordinate_manager is None:
coordinate_manager = CoordinateManager(
D=self._D,
coordinate_map_type=CoordinateMapType.CUDA
if coordinates.is_cuda else CoordinateMapType.CPU,
allocator_type=allocator_type,
minkowski_algorithm=minkowski_algorithm,
)
set_global_coordinate_manager(coordinate_manager)
else:
coordinate_manager = CoordinateManager(
D=coordinates.size(1) - 1,
coordinate_map_type=CoordinateMapType.CUDA
if coordinates.is_cuda else CoordinateMapType.CPU,
allocator_type=allocator_type,
minkowski_algorithm=minkowski_algorithm,
)
self._manager = coordinate_manager
##########################
# Initialize coords
##########################
# Coordinate Management
if coordinates is not None:
assert (
features.shape[0] == coordinates.shape[0]
), "The number of rows in features and coordinates must match."
assert (features.is_cuda == coordinates.is_cuda
), "Features and coordinates must have the same backend."
coordinate_field_map_key = CoordinateMapKey(
convert_to_int_list(tensor_stride, self._D), "")
coordinate_field_map_key = self._manager.insert_field(
coordinates.float(),
convert_to_int_list(tensor_stride, self._D), "")
else:
assert (coordinate_field_map_key.is_key_set()
), "The coordinate field map key must be valid."
if requires_grad is not None:
features.requires_grad_(requires_grad)
self._F = features
self._C = coordinates
self.coordinate_field_map_key = coordinate_field_map_key
self._batch_rows = None
self._inverse_mapping = {}
def field_to_sparse_keys(self, field_map_key: CoordinateMapKey):
return self._manager.field_to_sparse_keys(field_map_key.get_key())
def __init__(
self,
features: torch.Tensor,
coordinates: torch.Tensor = None,
# optional coordinate related arguments
tensor_stride: StrideType = 1,
coordinate_map_key: CoordinateMapKey = None,
coordinate_manager: CoordinateManager = None,
quantization_mode:
SparseTensorQuantizationMode = SparseTensorQuantizationMode.
RANDOM_SUBSAMPLE,
# optional manager related arguments
allocator_type: GPUMemoryAllocatorType = None,
minkowski_algorithm: MinkowskiAlgorithm = None,
device=None,
):
r"""
Args:
:attr:`features` (:attr:`torch.FloatTensor`,
:attr:`torch.DoubleTensor`, :attr:`torch.cuda.FloatTensor`, or
:attr:`torch.cuda.DoubleTensor`): The features of a sparse
tensor.
:attr:`coordinates` (:attr:`torch.IntTensor`): The coordinates
associated to the features. If not provided, :attr:`coordinate_map_key`
must be provided.
:attr:`coordinate_map_key`
(:attr:`MinkowskiEngine.CoordinateMapKey`): When the coordinates
are already cached in the MinkowskiEngine, we could reuse the same
coordinate map by simply providing the coordinate map key. In most
case, this process is done automatically. When you provide a
`coordinate_map_key`, `coordinates` will be be ignored.
:attr:`coordinate_manager`
(:attr:`MinkowskiEngine.CoordinateManager`): The MinkowskiEngine
manages all coordinate maps using the `_C.CoordinateMapManager`. If
not provided, the MinkowskiEngine will create a new computation
graph. In most cases, this process is handled automatically and you
do not need to use this.
:attr:`quantization_mode`
(:attr:`MinkowskiEngine.SparseTensorQuantizationMode`): Defines how
continuous coordinates will be quantized to define a sparse tensor.
Please refer to :attr:`SparseTensorQuantizationMode` for details.
:attr:`tensor_stride` (:attr:`int`, :attr:`list`,
:attr:`numpy.array`, or :attr:`tensor.Tensor`): The tensor stride
of the current sparse tensor. By default, it is 1.
"""
# Type checks
assert isinstance(features,
torch.Tensor), "Features must be a torch.Tensor"
assert (
features.ndim == 2
), f"The feature should be a matrix, The input feature is an order-{features.ndim} tensor."
assert isinstance(quantization_mode, SparseTensorQuantizationMode)
self.quantization_mode = quantization_mode
if coordinates is not None:
assert isinstance(coordinates, torch.Tensor)
if coordinate_map_key is not None:
assert isinstance(coordinate_map_key, CoordinateMapKey)
if coordinate_manager is not None:
assert isinstance(coordinate_manager, CoordinateManager)
# To device
if device is not None:
features = features.to(device)
coordinates = coordinates.to(device)
# Coordinate Management
self._D = 0 # coordinate size - 1
if coordinates is None and (coordinate_map_key is None
or coordinate_manager is None):
raise ValueError(
"Either coordinates or (coordinate_map_key, coordinate_manager) pair must be provided."
)
elif coordinates is not None:
assert (
features.shape[0] == coordinates.shape[0]
), "The number of rows in features and coordinates must match."
assert (features.is_cuda == coordinates.is_cuda
), "Features and coordinates must have the same backend."
self._D = coordinates.size(1) - 1
coordinate_map_key = CoordinateMapKey(
convert_to_int_list(tensor_stride, self._D), "")
else:
# not (coordinate_map_key is None or coordinate_manager is None)
self._D = coordinate_manager.D
##########################
# Setup CoordsManager
##########################
if coordinate_manager is None:
# If set to share the coords man, use the global coords man
global _sparse_tensor_operation_mode, _global_coordinate_manager
if (_sparse_tensor_operation_mode ==
SparseTensorOperationMode.SHARE_COORDINATE_MANAGER):
if _global_coordinate_manager is None:
_global_coordinate_manager = CoordinateManager(
D=self._D,
coordinate_map_type=CoordinateMapType.CUDA
if coordinates.is_cuda else CoordinateMapType.CPU,
allocator_type=allocator_type,
minkowski_algorithm=minkowski_algorithm,
)
coordinate_manager = _global_coordinate_manager
else:
coordinate_manager = CoordinateManager(
D=coordinates.size(1) - 1,
coordinate_map_type=CoordinateMapType.CUDA
if coordinates.is_cuda else CoordinateMapType.CPU,
allocator_type=allocator_type,
minkowski_algorithm=minkowski_algorithm,
)
self._manager = coordinate_manager
##########################
# Initialize coords
##########################
if coordinates is not None:
coordinates, features, coordinate_map_key = self.initialize_coordinates(
coordinates, features, coordinate_map_key)
elif coordinate_map_key is not None:
assert (coordinate_map_key.is_key_set()
), "The coordinate key must be a valid key."
self.coordinate_map_key = coordinate_map_key
self._F = features
self._C = coordinates
self._batch_rows = None
