def test_label(self):
N = 16575
ignore_label = 255
coords = (np.random.rand(N, 3) * 100).astype(np.int32)
feats = np.random.rand(N, 4)
labels = np.floor(np.random.rand(N) * 3)
labels = labels.astype(np.int32)
# Make duplicates
coords[:3] = 0
labels[:3] = 2
mapping, colabels = MEB.quantize_label_np(coords, labels, ignore_label)
print('Unique labels and counts:',
np.unique(colabels, return_counts=True))
print('N unique:', len(mapping), 'N:', N)
mapping, colabels = MEB.quantize_label_th(
torch.from_numpy(coords), torch.from_numpy(labels), ignore_label)
print('Unique labels and counts:',
np.unique(colabels, return_counts=True))
print('N unique:', len(mapping), 'N:', N)
qcoords, qfeats, qlabels = sparse_quantize(coords, feats, labels,
ignore_label)
self.assertTrue(len(mapping) == len(qcoords))
def quantize_label(coords, labels, ignore_label):
assert isinstance(coords, np.ndarray) or isinstance(
coords, torch.Tensor), "Invalid coords type"
if isinstance(coords, np.ndarray):
assert isinstance(labels, np.ndarray)
assert (coords.dtype == np.int32
), f"Invalid coords type {coords.dtype} != np.int32"
assert (labels.dtype == np.int32
), f"Invalid label type {labels.dtype} != np.int32"
return MEB.quantize_label_np(coords, labels, ignore_label)
else:
assert isinstance(labels, torch.Tensor)
# Type check done inside
return MEB.quantize_label_th(coords, labels.int(), ignore_label)
def test_label_np(self):
N = 16575
coords = (np.random.rand(N, 3) * 100).astype(np.int32)
labels = np.floor(np.random.rand(N) * 3).astype(np.int32)
# Make duplicates
coords[:3] = 0
labels[:3] = 2
mapping, inverse_mapping, colabel = MEB.quantize_label_np(
coords, labels, -1)
self.assertTrue(
np.sum(np.abs(coords[mapping][inverse_mapping] - coords)) == 0)
self.assertTrue(np.sum(colabel < 0) > 3)
def sparse_quantize(
coordinates,
features=None,
labels=None,
ignore_label=-100,
return_index=False,
return_inverse=False,
return_maps_only=False,
quantization_size=None,
device="cpu",
):
r"""Given coordinates, and features (optionally labels), the function
generates quantized (voxelized) coordinates.
Args:
:attr:`coordinates` (:attr:`numpy.ndarray` or :attr:`torch.Tensor`): a
matrix of size :math:`N \times D` where :math:`N` is the number of
points in the :math:`D` dimensional space.
:attr:`features` (:attr:`numpy.ndarray` or :attr:`torch.Tensor`, optional): a
matrix of size :math:`N \times D_F` where :math:`N` is the number of
points and :math:`D_F` is the dimension of the features. Must have the
same container as `coords` (i.e. if `coords` is a torch.Tensor, `feats`
must also be a torch.Tensor).
:attr:`labels` (:attr:`numpy.ndarray` or :attr:`torch.IntTensor`,
optional): integer labels associated to eah coordinates. Must have the
same container as `coords` (i.e. if `coords` is a torch.Tensor,
`labels` must also be a torch.Tensor). For classification where a set
of points are mapped to one label, do not feed the labels.
:attr:`ignore_label` (:attr:`int`, optional): the int value of the
IGNORE LABEL.
:attr:`torch.nn.CrossEntropyLoss(ignore_index=ignore_label)`
:attr:`return_index` (:attr:`bool`, optional): set True if you want the
indices of the quantized coordinates. False by default.
:attr:`return_inverse` (:attr:`bool`, optional): set True if you want
the indices that can recover the discretized original coordinates.
False by default. `return_index` must be True when `return_reverse` is True.
:attr:`return_maps_only` (:attr:`bool`, optional): if set, return the
unique_map or optionally inverse map, but not the coordinates. Can be
used if you don't care about final coordinates or if you use
device==cuda and you don't need coordinates on GPU. This returns either
unique_map alone or (unique_map, inverse_map) if return_inverse is set.
:attr:`quantization_size` (attr:`float`, optional): if set, will use
the quanziation size to define the smallest distance between
coordinates.
:attr:`device` (attr:`str`, optional): Either 'cpu' or 'cuda'.
Example::
>>> unique_map, inverse_map = sparse_quantize(discrete_coords, return_index=True, return_inverse=True)
>>> unique_coords = discrete_coords[unique_map]
>>> print(unique_coords[inverse_map] == discrete_coords) # True
:attr:`quantization_size` (:attr:`float`, :attr:`list`, or
:attr:`numpy.ndarray`, optional): the length of the each side of the
hyperrectangle of of the grid cell.
Example::
>>> # Segmentation
>>> criterion = torch.nn.CrossEntropyLoss(ignore_index=-100)
>>> coords, feats, labels = MinkowskiEngine.utils.sparse_quantize(
>>> coords, feats, labels, ignore_label=-100, quantization_size=0.1)
>>> output = net(MinkowskiEngine.SparseTensor(feats, coords))
>>> loss = criterion(output.F, labels.long())
>>>
>>> # Classification
>>> criterion = torch.nn.CrossEntropyLoss(ignore_index=-100)
>>> coords, feats = MinkowskiEngine.utils.sparse_quantize(coords, feats)
>>> output = net(MinkowskiEngine.SparseTensor(feats, coords))
>>> loss = criterion(output.F, labels.long())
"""
assert isinstance(
coordinates,
(np.ndarray,
torch.Tensor)), "Coords must be either np.array or torch.Tensor."
use_label = labels is not None
use_feat = features is not None
assert (
coordinates.ndim == 2
), "The coordinates must be a 2D matrix. The shape of the input is " + str(
coordinates.shape)
if return_inverse:
assert return_index, "return_reverse must be set with return_index"
if use_feat:
assert features.ndim == 2
assert coordinates.shape[0] == features.shape[0]
if use_label:
assert coordinates.shape[0] == len(labels)
dimension = coordinates.shape[1]
# Quantize the coordinates
if quantization_size is not None:
if isinstance(quantization_size, (Sequence, np.ndarray, torch.Tensor)):
assert (len(quantization_size) == dimension
), "Quantization size and coordinates size mismatch."
if isinstance(coordinates, np.ndarray):
quantization_size = np.array([i for i in quantization_size])
discrete_coordinates = np.floor(coordinates /
quantization_size)
else:
quantization_size = torch.Tensor(
[i for i in quantization_size])
discrete_coordinates = (coordinates /
quantization_size).floor()
elif np.isscalar(quantization_size): # Assume that it is a scalar
if quantization_size == 1:
discrete_coordinates = coordinates
else:
discrete_coordinates = np.floor(coordinates /
quantization_size)
else:
raise ValueError("Not supported type for quantization_size.")
else:
discrete_coordinates = coordinates
discrete_coordinates = np.floor(discrete_coordinates)
if isinstance(coordinates, np.ndarray):
discrete_coordinates = discrete_coordinates.astype(np.int32)
else:
discrete_coordinates = discrete_coordinates.int()
if device == "cpu":
manager = MEB.CoordinateMapManagerCPU()
elif "cuda" in device:
manager = MEB.CoordinateMapManagerGPU_c10()
else:
raise ValueError("Invalid device. Only `cpu` or `cuda` supported.")
# Return values accordingly
if use_label:
if isinstance(coordinates, np.ndarray):
unique_map, inverse_map, colabels = MEB.quantize_label_np(
discrete_coordinates, labels, ignore_label)
else:
assert (not discrete_coordinates.is_cuda()
), "Quantization with label requires cpu tensors."
assert not labels.is_cuda(
), "Quantization with label requires cpu tensors."
unique_map, inverse_map, colabels = MEB.quantize_label_th(
discrete_coordinates, labels, ignore_label)
return_args = [discrete_coordinates[unique_map]]
if use_feat:
return_args.append(features[unique_map])
# Labels
return_args.append(colabels)
# Additional return args
if return_index:
return_args.append(unique_map)
if return_inverse:
return_args.append(inverse_map)
if len(return_args) == 1:
return return_args[0]
else:
return tuple(return_args)
else:
tensor_stride = [1 for i in range(discrete_coordinates.shape[1] - 1)]
discrete_coordinates = (
discrete_coordinates.to(device) if isinstance(
discrete_coordinates, torch.Tensor) else
torch.from_numpy(discrete_coordinates).to(device))
_, (unique_map,
inverse_map) = manager.insert_and_map(discrete_coordinates,
tensor_stride, "")
if return_maps_only:
if return_inverse:
return unique_map, inverse_map
else:
return unique_map
return_args = [discrete_coordinates[unique_map]]
if use_feat:
return_args.append(features[unique_map])
if return_index:
return_args.append(unique_map)
if return_inverse:
return_args.append(inverse_map)
if len(return_args) == 1:
return return_args[0]
else:
return tuple(return_args)