torch.sparse¶
Warning
The PyTorch API of sparse tensors is in beta and may change in the near future. We highly welcome feature requests, bug reports and general suggestions as GitHub issues.
Why and when to use sparsity¶
By default, PyTorch stores torch.Tensor
elements contiguously in
physical memory. This leads to efficient implementations of various array
processing algorithms that require fast access to elements.
Now, some users might decide to represent data such as graph adjacency matrices, pruned weights or points clouds by Tensors whose elements are mostly zero valued. We recognize these are important applications and aim to provide performance optimizations for these use cases via sparse storage formats.
Various sparse storage formats such as COO, CSR/CSC, semistructured, LIL, etc. have been developed over the years. While they differ in exact layouts, they all compress data through efficient representation of zero valued elements. We call the uncompressed values specified in contrast to unspecified, compressed elements.
By compressing repeat zeros sparse storage formats aim to save memory and computational resources on various CPUs and GPUs. Especially for high degrees of sparsity or highly structured sparsity this can have significant performance implications. As such sparse storage formats can be seen as a performance optimization.
Like many other performance optimization sparse storage formats are not always advantageous. When trying sparse formats for your use case you might find your execution time to increase rather than decrease.
Please feel encouraged to open a GitHub issue if you analytically expected to see a stark increase in performance but measured a degradation instead. This helps us prioritize the implementation of efficient kernels and wider performance optimizations.
We make it easy to try different sparsity layouts, and convert between them, without being opinionated on what’s best for your particular application.
Functionality overview¶
We want it to be straightforward to construct a sparse Tensor from a given dense Tensor by providing conversion routines for each layout.
In the next example we convert a 2D Tensor with default dense (strided) layout to a 2D Tensor backed by the COO memory layout. Only values and indices of nonzero elements are stored in this case.
>>> a = torch.tensor([[0, 2.], [3, 0]])
>>> a.to_sparse()
tensor(indices=tensor([[0, 1],
[1, 0]]),
values=tensor([2., 3.]),
size=(2, 2), nnz=2, layout=torch.sparse_coo)
PyTorch currently supports COO, CSR, CSC, BSR, and BSC.
We also have a prototype implementation to support :ref: semistructured sparsity<sparsesemistructureddocs>. Please see the references for more details.
Note that we provide slight generalizations of these formats.
Batching: Devices such as GPUs require batching for optimal performance and thus we support batch dimensions.
We currently offer a very simple version of batching where each component of a sparse format itself is batched. This also requires the same number of specified elements per batch entry. In this example we construct a 3D (batched) CSR Tensor from a 3D dense Tensor.
>>> t = torch.tensor([[[1., 0], [2., 3.]], [[4., 0], [5., 6.]]])
>>> t.dim()
3
>>> t.to_sparse_csr()
tensor(crow_indices=tensor([[0, 1, 3],
[0, 1, 3]]),
col_indices=tensor([[0, 0, 1],
[0, 0, 1]]),
values=tensor([[1., 2., 3.],
[4., 5., 6.]]), size=(2, 2, 2), nnz=3,
layout=torch.sparse_csr)
Dense dimensions: On the other hand, some data such as Graph embeddings might be better viewed as sparse collections of vectors instead of scalars.
In this example we create a 3D Hybrid COO Tensor with 2 sparse and 1 dense dimension from a 3D strided Tensor. If an entire row in the 3D strided Tensor is zero, it is not stored. If however any of the values in the row are nonzero, they are stored entirely. This reduces the number of indices since we need one index one per row instead of one per element. But it also increases the amount of storage for the values. Since only rows that are entirely zero can be emitted and the presence of any nonzero valued elements cause the entire row to be stored.
>>> t = torch.tensor([[[0., 0], [1., 2.]], [[0., 0], [3., 4.]]])
>>> t.to_sparse(sparse_dim=2)
tensor(indices=tensor([[0, 1],
[1, 1]]),
values=tensor([[1., 2.],
[3., 4.]]),
size=(2, 2, 2), nnz=2, layout=torch.sparse_coo)
Operator overview¶
Fundamentally, operations on Tensor with sparse storage formats behave the same as operations on Tensor with strided (or other) storage formats. The particularities of storage, that is the physical layout of the data, influences the performance of an operation but should not influence the semantics.
We are actively increasing operator coverage for sparse tensors. Users should not expect support same level of support as for dense Tensors yet. See our operator documentation for a list.
>>> b = torch.tensor([[0, 0, 1, 2, 3, 0], [4, 5, 0, 6, 0, 0]])
>>> b_s = b.to_sparse_csr()
>>> b_s.cos()
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
RuntimeError: unsupported tensor layout: SparseCsr
>>> b_s.sin()
tensor(crow_indices=tensor([0, 3, 6]),
col_indices=tensor([2, 3, 4, 0, 1, 3]),
values=tensor([ 0.8415, 0.9093, 0.1411, 0.7568, 0.9589, 0.2794]),
size=(2, 6), nnz=6, layout=torch.sparse_csr)
As shown in the example above, we don’t support nonzero preserving unary operators such as cos. The output of a nonzero preserving unary operation will not be able to take advantage of sparse storage formats to the same extent as the input and potentially result in a catastrophic increase in memory. We instead rely on the user to explicitly convert to a dense Tensor first and then run the operation.
>>> b_s.to_dense().cos()
tensor([[ 1.0000, 0.4161],
[0.9900, 1.0000]])
We are aware that some users want to ignore compressed zeros for operations such as cos instead of preserving the exact semantics of the operation. For this we can point to torch.masked and its MaskedTensor, which is in turn also backed and powered by sparse storage formats and kernels.
Also note that, for now, the user doesn’t have a choice of the output layout. For example, adding a sparse Tensor to a regular strided Tensor results in a strided Tensor. Some users might prefer for this to stay a sparse layout, because they know the result will still be sufficiently sparse.
>>> a + b.to_sparse()
tensor([[0., 3.],
[3., 0.]])
We acknowledge that access to kernels that can efficiently produce different output layouts can be very useful. A subsequent operation might significantly benefit from receiving a particular layout. We are working on an API to control the result layout and recognize it is an important feature to plan a more optimal path of execution for any given model.
Sparse SemiStructured Tensors¶
Warning
Sparse semistructured tensors are currently a prototype feature and subject to change. Please feel free to open an issue to report a bug or if you have feedback to share.
SemiStructured sparsity is a sparse data layout that was first introduced in NVIDIA’s Ampere architecture. It is also referred to as finegrained structured sparsity or 2:4 structured sparsity.
This sparse layout stores n elements out of every 2n elements, with n being determined by the width of the Tensor’s data type (dtype). The most frequently used dtype is float16, where n=2, thus the term “2:4 structured sparsity.”
Semistructured sparsity is explained in greater detail in this NVIDIA blog post.
In PyTorch, semistructured sparsity is implemented via a Tensor subclass.
By subclassing, we can override __torch_dispatch__
, allowing us to use faster sparse kernels when performing matrix multiplication.
We can also store the tensor in it’s compressed form inside the subclass to reduce memory overhead.
In this compressed form, the sparse tensor is stored by retaining only the specified elements and some metadata, which encodes the mask.
Note
The specified elements and metadata mask of a semistructured sparse tensor are stored together in a single flat compressed tensor. They are appended to each other to form a contiguous chunk of memory.
compressed tensor = [ specified elements of original tensor  metadata_mask ]
For an original tensor of size (r, c) we expect the first m * k // 2 elements to be the kept elements and the rest of the tensor is metadata.
In order to make it easier for the user to view the specified elements
and mask, one can use .indices()
and .values()
to access the mask and specified elements respectively.
.values()
returns the specified elements in a tensor of size (r, c//2) and with the same dtype as the dense matrix..indices()
returns the metadata_mask in a tensor of size (r, c//2 ) and with element typetorch.int16
if dtype is torch.float16 or torch.bfloat16, and element typetorch.int32
if dtype is torch.int8.
For 2:4 sparse tensors, the metadata overhead is minor  just 2 bits per specified element.
Note
It’s important to note that torch.float32
is only supported for 1:2 sparsity. Therefore, it does not follow the same formula as above.
Here, we break down how to calculate the compression ratio ( size dense / size sparse) of a 2:4 sparse tensor.
Let (r, c) = tensor.shape and e = bitwidth(tensor.dtype), so e = 16 for torch.float16
and torch.bfloat16
and e = 8 for torch.int8
.
Using these calculations, we can determine the total memory footprint for both the original dense and the new sparse representation.
This gives us a simple formula for the compression ratio, which is dependent only on the bitwidth of the tensor datatype.
By using this formula, we find that the compression ratio is 56.25% for torch.float16
or torch.bfloat16
, and 62.5% for torch.int8
.
Constructing Sparse SemiStructured Tensors¶
You can transform a dense tensor into a sparse semistructured tensor by simply using the torch.to_sparse_semi_structured
function.
Please also note that we only support CUDA tensors since hardware compatibility for semistructured sparsity is limited to NVIDIA GPUs.
The following datatypes are supported for semistructured sparsity. Note that each datatype has its own shape constraints and compression factor.
PyTorch dtype 
Shape Constraints 
Compression Factor 
Sparsity Pattern 


Tensor must be 2D and (r, c) must both be a positive multiple of 64 
9/16 
2:4 

Tensor must be 2D and (r, c) must both be a positive multiple of 64 
9/16 
2:4 

Tensor must be 2D and (r, c) must both be a positive multiple of 128 
10/16 
2:4 
To construct a semistructured sparse tensor, start by creating a regular dense tensor that adheres to a 2:4 (or semistructured) sparse format.
To do this we tile a small 1x4 strip to create a 16x16 dense float16 tensor.
Afterwards, we can call to_sparse_semi_structured
function to compress it for accelerated inference.
>>> from torch.sparse import to_sparse_semi_structured
>>> A = torch.Tensor([0, 0, 1, 1]).tile((128, 32)).half().cuda()
tensor([[0., 0., 1., ..., 0., 1., 1.],
[0., 0., 1., ..., 0., 1., 1.],
[0., 0., 1., ..., 0., 1., 1.],
...,
[0., 0., 1., ..., 0., 1., 1.],
[0., 0., 1., ..., 0., 1., 1.],
[0., 0., 1., ..., 0., 1., 1.]], device='cuda:0', dtype=torch.float16)
>>> A_sparse = to_sparse_semi_structured(A)
SparseSemiStructuredTensor(shape=torch.Size([128, 128]), transposed=False, values=tensor([[1., 1., 1., ..., 1., 1., 1.],
[1., 1., 1., ..., 1., 1., 1.],
[1., 1., 1., ..., 1., 1., 1.],
...,
[1., 1., 1., ..., 1., 1., 1.],
[1., 1., 1., ..., 1., 1., 1.],
[1., 1., 1., ..., 1., 1., 1.]], device='cuda:0', dtype=torch.float16), metadata=tensor([[4370, 4370, 4370, ..., 4370, 4370, 4370],
[4370, 4370, 4370, ..., 4370, 4370, 4370],
[4370, 4370, 4370, ..., 4370, 4370, 4370],
...,
[4370, 4370, 4370, ..., 4370, 4370, 4370],
[4370, 4370, 4370, ..., 4370, 4370, 4370],
[4370, 4370, 4370, ..., 4370, 4370, 4370]], device='cuda:0',
dtype=torch.int16))
Sparse SemiStructured Tensor Operations¶
Currently, the following operations are supported for semistructured sparse tensors:
torch.addmm(bias, dense, sparse.t())
torch.mm(dense, sparse)
torch.mm(sparse, dense)
aten.linear.default(dense, sparse, bias)
aten.t.default(sparse)
aten.t.detach(sparse)
To use these ops, simply pass the output of to_sparse_semi_structured(tensor)
instead of using tensor
once your tensor has 0s in a semistructured sparse format, like this:
>>> a = torch.Tensor([0, 0, 1, 1]).tile((64, 16)).half().cuda()
>>> b = torch.rand(64, 64).half().cuda()
>>> c = torch.mm(a, b)
>>> a_sparse = to_sparse_semi_structured(a)
>>> torch.allclose(c, torch.mm(a_sparse, b))
True
Accelerating nn.Linear with semistructured sparsity¶
You can accelerate the linear layers in your model if the weights are already semistructured sparse with just a few lines of code:
>>> input = torch.rand(64, 64).half().cuda()
>>> mask = torch.Tensor([0, 0, 1, 1]).tile((64, 16)).cuda().bool()
>>> linear = nn.Linear(64, 64).half().cuda()
>>> linear.weight = nn.Parameter(to_sparse_semi_structured(linear.weight.masked_fill(~mask, 0)))
Sparse COO tensors¶
PyTorch implements the socalled Coordinate format, or COO format, as one of the storage formats for implementing sparse tensors. In COO format, the specified elements are stored as tuples of element indices and the corresponding values. In particular,
the indices of specified elements are collected in
indices
tensor of size(ndim, nse)
and with element typetorch.int64
,the corresponding values are collected in
values
tensor of size(nse,)
and with an arbitrary integer or floating point number element type,
where ndim
is the dimensionality of the tensor and nse
is the
number of specified elements.
Note
The memory consumption of a sparse COO tensor is at least (ndim *
8 + <size of element type in bytes>) * nse
bytes (plus a constant
overhead from storing other tensor data).
The memory consumption of a strided tensor is at least
product(<tensor shape>) * <size of element type in bytes>
.
For example, the memory consumption of a 10 000 x 10 000 tensor
with 100 000 nonzero 32bit floating point numbers is at least
(2 * 8 + 4) * 100 000 = 2 000 000
bytes when using COO tensor
layout and 10 000 * 10 000 * 4 = 400 000 000
bytes when using
the default strided tensor layout. Notice the 200 fold memory
saving from using the COO storage format.
Construction¶
A sparse COO tensor can be constructed by providing the two tensors of
indices and values, as well as the size of the sparse tensor (when it
cannot be inferred from the indices and values tensors) to a function
torch.sparse_coo_tensor()
.
Suppose we want to define a sparse tensor with the entry 3 at location (0, 2), entry 4 at location (1, 0), and entry 5 at location (1, 2). Unspecified elements are assumed to have the same value, fill value, which is zero by default. We would then write:
>>> i = [[0, 1, 1],
[2, 0, 2]]
>>> v = [3, 4, 5]
>>> s = torch.sparse_coo_tensor(i, v, (2, 3))
>>> s
tensor(indices=tensor([[0, 1, 1],
[2, 0, 2]]),
values=tensor([3, 4, 5]),
size=(2, 3), nnz=3, layout=torch.sparse_coo)
>>> s.to_dense()
tensor([[0, 0, 3],
[4, 0, 5]])
Note that the input i
is NOT a list of index tuples. If you want
to write your indices this way, you should transpose before passing them to
the sparse constructor:
>>> i = [[0, 2], [1, 0], [1, 2]]
>>> v = [3, 4, 5 ]
>>> s = torch.sparse_coo_tensor(list(zip(*i)), v, (2, 3))
>>> # Or another equivalent formulation to get s
>>> s = torch.sparse_coo_tensor(torch.tensor(i).t(), v, (2, 3))
>>> torch.sparse_coo_tensor(i.t(), v, torch.Size([2,3])).to_dense()
tensor([[0, 0, 3],
[4, 0, 5]])
An empty sparse COO tensor can be constructed by specifying its size only:
>>> torch.sparse_coo_tensor(size=(2, 3))
tensor(indices=tensor([], size=(2, 0)),
values=tensor([], size=(0,)),
size=(2, 3), nnz=0, layout=torch.sparse_coo)
Sparse hybrid COO tensors¶
PyTorch implements an extension of sparse tensors with scalar values to sparse tensors with (contiguous) tensor values. Such tensors are called hybrid tensors.
PyTorch hybrid COO tensor extends the sparse COO tensor by allowing
the values
tensor to be a multidimensional tensor so that we
have:
the indices of specified elements are collected in
indices
tensor of size(sparse_dims, nse)
and with element typetorch.int64
,the corresponding (tensor) values are collected in
values
tensor of size(nse, dense_dims)
and with an arbitrary integer or floating point number element type.
Note
We use (M + K)dimensional tensor to denote a Ndimensional sparse hybrid tensor, where M and K are the numbers of sparse and dense dimensions, respectively, such that M + K == N holds.
Suppose we want to create a (2 + 1)dimensional tensor with the entry [3, 4] at location (0, 2), entry [5, 6] at location (1, 0), and entry [7, 8] at location (1, 2). We would write
>>> i = [[0, 1, 1],
[2, 0, 2]]
>>> v = [[3, 4], [5, 6], [7, 8]]
>>> s = torch.sparse_coo_tensor(i, v, (2, 3, 2))
>>> s
tensor(indices=tensor([[0, 1, 1],
[2, 0, 2]]),
values=tensor([[3, 4],
[5, 6],
[7, 8]]),
size=(2, 3, 2), nnz=3, layout=torch.sparse_coo)
>>> s.to_dense()
tensor([[[0, 0],
[0, 0],
[3, 4]],
[[5, 6],
[0, 0],
[7, 8]]])
In general, if s
is a sparse COO tensor and M =
s.sparse_dim()
, K = s.dense_dim()
, then we have the following
invariants:
M + K == len(s.shape) == s.ndim
 dimensionality of a tensor is the sum of the number of sparse and dense dimensions,
s.indices().shape == (M, nse)
 sparse indices are stored explicitly,
s.values().shape == (nse,) + s.shape[M : M + K]
 the values of a hybrid tensor are Kdimensional tensors,
s.values().layout == torch.strided
 values are stored as strided tensors.
Note
Dense dimensions always follow sparse dimensions, that is, mixing of dense and sparse dimensions is not supported.
Note
To be sure that a constructed sparse tensor has consistent indices,
values, and size, the invariant checks can be enabled per tensor
creation via check_invariants=True
keyword argument, or
globally using torch.sparse.check_sparse_tensor_invariants
context manager instance. By default, the sparse tensor invariants
checks are disabled.
Uncoalesced sparse COO tensors¶
PyTorch sparse COO tensor format permits sparse uncoalesced tensors,
where there may be duplicate coordinates in the indices; in this case,
the interpretation is that the value at that index is the sum of all
duplicate value entries. For example, one can specify multiple values,
3
and 4
, for the same index 1
, that leads to an 1D
uncoalesced tensor:
>>> i = [[1, 1]]
>>> v = [3, 4]
>>> s=torch.sparse_coo_tensor(i, v, (3,))
>>> s
tensor(indices=tensor([[1, 1]]),
values=tensor( [3, 4]),
size=(3,), nnz=2, layout=torch.sparse_coo)
while the coalescing process will accumulate the multivalued elements into a single value using summation:
>>> s.coalesce()
tensor(indices=tensor([[1]]),
values=tensor([7]),
size=(3,), nnz=1, layout=torch.sparse_coo)
In general, the output of torch.Tensor.coalesce()
method is a
sparse tensor with the following properties:
the indices of specified tensor elements are unique,
the indices are sorted in lexicographical order,
torch.Tensor.is_coalesced()
returnsTrue
.
Note
For the most part, you shouldn’t have to care whether or not a sparse tensor is coalesced or not, as most operations will work identically given a sparse coalesced or uncoalesced tensor.
However, some operations can be implemented more efficiently on uncoalesced tensors, and some on coalesced tensors.
For instance, addition of sparse COO tensors is implemented by simply concatenating the indices and values tensors:
>>> a = torch.sparse_coo_tensor([[1, 1]], [5, 6], (2,))
>>> b = torch.sparse_coo_tensor([[0, 0]], [7, 8], (2,))
>>> a + b
tensor(indices=tensor([[0, 0, 1, 1]]),
values=tensor([7, 8, 5, 6]),
size=(2,), nnz=4, layout=torch.sparse_coo)
If you repeatedly perform an operation that can produce duplicate
entries (e.g., torch.Tensor.add()
), you should occasionally
coalesce your sparse tensors to prevent them from growing too large.
On the other hand, the lexicographical ordering of indices can be advantageous for implementing algorithms that involve many element selection operations, such as slicing or matrix products.
Working with sparse COO tensors¶
Let’s consider the following example:
>>> i = [[0, 1, 1],
[2, 0, 2]]
>>> v = [[3, 4], [5, 6], [7, 8]]
>>> s = torch.sparse_coo_tensor(i, v, (2, 3, 2))
As mentioned above, a sparse COO tensor is a torch.Tensor
instance and to distinguish it from the Tensor instances that use
some other layout, one can use torch.Tensor.is_sparse
or
torch.Tensor.layout
properties:
>>> isinstance(s, torch.Tensor)
True
>>> s.is_sparse
True
>>> s.layout == torch.sparse_coo
True
The number of sparse and dense dimensions can be acquired using
methods torch.Tensor.sparse_dim()
and
torch.Tensor.dense_dim()
, respectively. For instance:
>>> s.sparse_dim(), s.dense_dim()
(2, 1)
If s
is a sparse COO tensor then its COO format data can be
acquired using methods torch.Tensor.indices()
and
torch.Tensor.values()
.
Note
Currently, one can acquire the COO format data only when the tensor instance is coalesced:
>>> s.indices()
RuntimeError: Cannot get indices on an uncoalesced tensor, please call .coalesce() first
For acquiring the COO format data of an uncoalesced tensor, use
torch.Tensor._values()
and torch.Tensor._indices()
:
>>> s._indices()
tensor([[0, 1, 1],
[2, 0, 2]])
Warning
Calling torch.Tensor._values()
will return a detached tensor.
To track gradients, torch.Tensor.coalesce().values()
must be
used instead.
Constructing a new sparse COO tensor results a tensor that is not coalesced:
>>> s.is_coalesced()
False
but one can construct a coalesced copy of a sparse COO tensor using
the torch.Tensor.coalesce()
method:
>>> s2 = s.coalesce()
>>> s2.indices()
tensor([[0, 1, 1],
[2, 0, 2]])
When working with uncoalesced sparse COO tensors, one must take into
an account the additive nature of uncoalesced data: the values of the
same indices are the terms of a sum that evaluation gives the value of
the corresponding tensor element. For example, the scalar
multiplication on a sparse uncoalesced tensor could be implemented by
multiplying all the uncoalesced values with the scalar because c *
(a + b) == c * a + c * b
holds. However, any nonlinear operation,
say, a square root, cannot be implemented by applying the operation to
uncoalesced data because sqrt(a + b) == sqrt(a) + sqrt(b)
does not
hold in general.
Slicing (with positive step) of a sparse COO tensor is supported only for dense dimensions. Indexing is supported for both sparse and dense dimensions:
>>> s[1]
tensor(indices=tensor([[0, 2]]),
values=tensor([[5, 6],
[7, 8]]),
size=(3, 2), nnz=2, layout=torch.sparse_coo)
>>> s[1, 0, 1]
tensor(6)
>>> s[1, 0, 1:]
tensor([6])
In PyTorch, the fill value of a sparse tensor cannot be specified
explicitly and is assumed to be zero in general. However, there exists
operations that may interpret the fill value differently. For
instance, torch.sparse.softmax()
computes the softmax with the
assumption that the fill value is negative infinity.
Sparse Compressed Tensors¶
Sparse Compressed Tensors represents a class of sparse tensors that have a common feature of compressing the indices of a certain dimension using an encoding that enables certain optimizations on linear algebra kernels of sparse compressed tensors. This encoding is based on the Compressed Sparse Row (CSR) format that PyTorch sparse compressed tensors extend with the support of sparse tensor batches, allowing multidimensional tensor values, and storing sparse tensor values in dense blocks.
Note
We use (B + M + K)dimensional tensor to denote a Ndimensional
sparse compressed hybrid tensor, where B, M, and K are the numbers
of batch, sparse, and dense dimensions, respectively, such that
B + M + K == N
holds. The number of sparse dimensions for
sparse compressed tensors is always two, M == 2
.
Note
We say that an indices tensor compressed_indices
uses CSR
encoding if the following invariants are satisfied:
compressed_indices
is a contiguous strided 32 or 64 bit integer tensorcompressed_indices
shape is(*batchsize, compressed_dim_size + 1)
wherecompressed_dim_size
is the number of compressed dimensions (e.g. rows or columns)compressed_indices[..., 0] == 0
where...
denotes batch indicescompressed_indices[..., compressed_dim_size] == nse
wherense
is the number of specified elements0 <= compressed_indices[..., i]  compressed_indices[..., i  1] <= plain_dim_size
fori=1, ..., compressed_dim_size
, whereplain_dim_size
is the number of plain dimensions (orthogonal to compressed dimensions, e.g. columns or rows).
To be sure that a constructed sparse tensor has consistent indices,
values, and size, the invariant checks can be enabled per tensor
creation via check_invariants=True
keyword argument, or
globally using torch.sparse.check_sparse_tensor_invariants
context manager instance. By default, the sparse tensor invariants
checks are disabled.
Note
The generalization of sparse compressed layouts to Ndimensional
tensors can lead to some confusion regarding the count of specified
elements. When a sparse compressed tensor contains batch dimensions
the number of specified elements will correspond to the number of such
elements perbatch. When a sparse compressed tensor has dense dimensions
the element considered is now the Kdimensional array. Also for block
sparse compressed layouts the 2D block is considered as the element
being specified. Take as an example a 3dimensional block sparse
tensor, with one batch dimension of length b
, and a block
shape of p, q
. If this tensor has n
specified elements, then
in fact we have n
blocks specified per batch. This tensor would
have values
with shape (b, n, p, q)
. This interpretation of the
number of specified elements comes from all sparse compressed layouts
being derived from the compression of a 2dimensional matrix. Batch
dimensions are treated as stacking of sparse matrices, dense dimensions
change the meaning of the element from a simple scalar value to an
array with its own dimensions.
Sparse CSR Tensor¶
The primary advantage of the CSR format over the COO format is better use of storage and much faster computation operations such as sparse matrixvector multiplication using MKL and MAGMA backends.
In the simplest case, a (0 + 2 + 0)dimensional sparse CSR tensor
consists of three 1D tensors: crow_indices
, col_indices
and
values
:
The
crow_indices
tensor consists of compressed row indices. This is a 1D tensor of sizenrows + 1
(the number of rows plus 1). The last element ofcrow_indices
is the number of specified elements,nse
. This tensor encodes the index invalues
andcol_indices
depending on where the given row starts. Each successive number in the tensor subtracted by the number before it denotes the number of elements in a given row.The
col_indices
tensor contains the column indices of each element. This is a 1D tensor of sizense
.The
values
tensor contains the values of the CSR tensor elements. This is a 1D tensor of sizense
.
Note
The index tensors crow_indices
and col_indices
should have
element type either torch.int64
(default) or
torch.int32
. If you want to use MKLenabled matrix operations,
use torch.int32
. This is as a result of the default linking of
pytorch being with MKL LP64, which uses 32 bit integer indexing.
In the general case, the (B + 2 + K)dimensional sparse CSR tensor
consists of two (B + 1)dimensional index tensors crow_indices
and
col_indices
, and of (1 + K)dimensional values
tensor such
that
crow_indices.shape == (*batchsize, nrows + 1)
col_indices.shape == (*batchsize, nse)
values.shape == (nse, *densesize)
while the shape of the sparse CSR tensor is (*batchsize, nrows,
ncols, *densesize)
where len(batchsize) == B
and
len(densesize) == K
.
Note
The batches of sparse CSR tensors are dependent: the number of specified elements in all batches must be the same. This somewhat artificial constraint allows efficient storage of the indices of different CSR batches.
Note
The number of sparse and dense dimensions can be acquired using
torch.Tensor.sparse_dim()
and torch.Tensor.dense_dim()
methods. The batch dimensions can be computed from the tensor
shape: batchsize = tensor.shape[:tensor.sparse_dim() 
tensor.dense_dim()]
.
Note
The memory consumption of a sparse CSR tensor is at least
(nrows * 8 + (8 + <size of element type in bytes> *
prod(densesize)) * nse) * prod(batchsize)
bytes (plus a constant
overhead from storing other tensor data).
With the same example data of the note in sparse COO format
introduction, the memory consumption of a 10 000
x 10 000 tensor with 100 000 nonzero 32bit floating point numbers
is at least (10000 * 8 + (8 + 4 * 1) * 100 000) * 1 = 1 280 000
bytes when using CSR tensor layout. Notice the 1.6 and 310 fold
savings from using CSR storage format compared to using the COO and
strided formats, respectively.
Construction of CSR tensors¶
Sparse CSR tensors can be directly constructed by using the
torch.sparse_csr_tensor()
function. The user must supply the row
and column indices and values tensors separately where the row indices
must be specified using the CSR compression encoding. The size
argument is optional and will be deduced from the crow_indices
and
col_indices
if it is not present.
>>> crow_indices = torch.tensor([0, 2, 4])
>>> col_indices = torch.tensor([0, 1, 0, 1])
>>> values = torch.tensor([1, 2, 3, 4])
>>> csr = torch.sparse_csr_tensor(crow_indices, col_indices, values, dtype=torch.float64)
>>> csr
tensor(crow_indices=tensor([0, 2, 4]),
col_indices=tensor([0, 1, 0, 1]),
values=tensor([1., 2., 3., 4.]), size=(2, 2), nnz=4,
dtype=torch.float64)
>>> csr.to_dense()
tensor([[1., 2.],
[3., 4.]], dtype=torch.float64)
Note
The values of sparse dimensions in deduced size
is computed
from the size of crow_indices
and the maximal index value in
col_indices
. If the number of columns needs to be larger than
in the deduced size
then the size
argument must be
specified explicitly.
The simplest way of constructing a 2D sparse CSR tensor from a
strided or sparse COO tensor is to use
torch.Tensor.to_sparse_csr()
method. Any zeros in the (strided)
tensor will be interpreted as missing values in the sparse tensor:
>>> a = torch.tensor([[0, 0, 1, 0], [1, 2, 0, 0], [0, 0, 0, 0]], dtype=torch.float64)
>>> sp = a.to_sparse_csr()
>>> sp
tensor(crow_indices=tensor([0, 1, 3, 3]),
col_indices=tensor([2, 0, 1]),
values=tensor([1., 1., 2.]), size=(3, 4), nnz=3, dtype=torch.float64)
CSR Tensor Operations¶
The sparse matrixvector multiplication can be performed with the
tensor.matmul()
method. This is currently the only math operation
supported on CSR tensors.
>>> vec = torch.randn(4, 1, dtype=torch.float64)
>>> sp.matmul(vec)
tensor([[0.9078],
[1.3180],
[0.0000]], dtype=torch.float64)
Sparse CSC Tensor¶
The sparse CSC (Compressed Sparse Column) tensor format implements the CSC format for storage of 2 dimensional tensors with an extension to supporting batches of sparse CSC tensors and values being multidimensional tensors.
Note
Sparse CSC tensor is essentially a transpose of the sparse CSR tensor when the transposition is about swapping the sparse dimensions.
Similarly to sparse CSR tensors, a sparse CSC
tensor consists of three tensors: ccol_indices
, row_indices
and values
:
The
ccol_indices
tensor consists of compressed column indices. This is a (B + 1)D tensor of shape(*batchsize, ncols + 1)
. The last element is the number of specified elements,nse
. This tensor encodes the index invalues
androw_indices
depending on where the given column starts. Each successive number in the tensor subtracted by the number before it denotes the number of elements in a given column.The
row_indices
tensor contains the row indices of each element. This is a (B + 1)D tensor of shape(*batchsize, nse)
.The
values
tensor contains the values of the CSC tensor elements. This is a (1 + K)D tensor of shape(nse, *densesize)
.
Construction of CSC tensors¶
Sparse CSC tensors can be directly constructed by using the
torch.sparse_csc_tensor()
function. The user must supply the row
and column indices and values tensors separately where the column indices
must be specified using the CSR compression encoding. The size
argument is optional and will be deduced from the row_indices
and
ccol_indices
tensors if it is not present.
>>> ccol_indices = torch.tensor([0, 2, 4])
>>> row_indices = torch.tensor([0, 1, 0, 1])
>>> values = torch.tensor([1, 2, 3, 4])
>>> csc = torch.sparse_csc_tensor(ccol_indices, row_indices, values, dtype=torch.float64)
>>> csc
tensor(ccol_indices=tensor([0, 2, 4]),
row_indices=tensor([0, 1, 0, 1]),
values=tensor([1., 2., 3., 4.]), size=(2, 2), nnz=4,
dtype=torch.float64, layout=torch.sparse_csc)
>>> csc.to_dense()
tensor([[1., 3.],
[2., 4.]], dtype=torch.float64)
Note
The sparse CSC tensor constructor function has the compressed column indices argument before the row indices argument.
The (0 + 2 + 0)dimensional sparse CSC tensors can be constructed from
any twodimensional tensor using torch.Tensor.to_sparse_csc()
method. Any zeros in the (strided) tensor will be interpreted as
missing values in the sparse tensor:
>>> a = torch.tensor([[0, 0, 1, 0], [1, 2, 0, 0], [0, 0, 0, 0]], dtype=torch.float64)
>>> sp = a.to_sparse_csc()
>>> sp
tensor(ccol_indices=tensor([0, 1, 2, 3, 3]),
row_indices=tensor([1, 1, 0]),
values=tensor([1., 2., 1.]), size=(3, 4), nnz=3, dtype=torch.float64,
layout=torch.sparse_csc)
Sparse BSR Tensor¶
The sparse BSR (Block compressed Sparse Row) tensor format implements the BSR format for storage of twodimensional tensors with an extension to supporting batches of sparse BSR tensors and values being blocks of multidimensional tensors.
A sparse BSR tensor consists of three tensors: crow_indices
,
col_indices
and values
:
The
crow_indices
tensor consists of compressed row indices. This is a (B + 1)D tensor of shape(*batchsize, nrowblocks + 1)
. The last element is the number of specified blocks,nse
. This tensor encodes the index invalues
andcol_indices
depending on where the given column block starts. Each successive number in the tensor subtracted by the number before it denotes the number of blocks in a given row.The
col_indices
tensor contains the column block indices of each element. This is a (B + 1)D tensor of shape(*batchsize, nse)
.The
values
tensor contains the values of the sparse BSR tensor elements collected into twodimensional blocks. This is a (1 + 2 + K)D tensor of shape(nse, nrowblocks, ncolblocks, *densesize)
.
Construction of BSR tensors¶
Sparse BSR tensors can be directly constructed by using the
torch.sparse_bsr_tensor()
function. The user must supply the row
and column block indices and values tensors separately where the row block indices
must be specified using the CSR compression encoding.
The size
argument is optional and will be deduced from the crow_indices
and
col_indices
tensors if it is not present.
>>> crow_indices = torch.tensor([0, 2, 4])
>>> col_indices = torch.tensor([0, 1, 0, 1])
>>> values = torch.tensor([[[0, 1, 2], [6, 7, 8]],
... [[3, 4, 5], [9, 10, 11]],
... [[12, 13, 14], [18, 19, 20]],
... [[15, 16, 17], [21, 22, 23]]])
>>> bsr = torch.sparse_bsr_tensor(crow_indices, col_indices, values, dtype=torch.float64)
>>> bsr
tensor(crow_indices=tensor([0, 2, 4]),
col_indices=tensor([0, 1, 0, 1]),
values=tensor([[[ 0., 1., 2.],
[ 6., 7., 8.]],
[[ 3., 4., 5.],
[ 9., 10., 11.]],
[[12., 13., 14.],
[18., 19., 20.]],
[[15., 16., 17.],
[21., 22., 23.]]]),
size=(4, 6), nnz=4, dtype=torch.float64, layout=torch.sparse_bsr)
>>> bsr.to_dense()
tensor([[ 0., 1., 2., 3., 4., 5.],
[ 6., 7., 8., 9., 10., 11.],
[12., 13., 14., 15., 16., 17.],
[18., 19., 20., 21., 22., 23.]], dtype=torch.float64)
The (0 + 2 + 0)dimensional sparse BSR tensors can be constructed from
any twodimensional tensor using torch.Tensor.to_sparse_bsr()
method that also requires the specification of the values block size:
>>> dense = torch.tensor([[0, 1, 2, 3, 4, 5],
... [6, 7, 8, 9, 10, 11],
... [12, 13, 14, 15, 16, 17],
... [18, 19, 20, 21, 22, 23]])
>>> bsr = dense.to_sparse_bsr(blocksize=(2, 3))
>>> bsr
tensor(crow_indices=tensor([0, 2, 4]),
col_indices=tensor([0, 1, 0, 1]),
values=tensor([[[ 0, 1, 2],
[ 6, 7, 8]],
[[ 3, 4, 5],
[ 9, 10, 11]],
[[12, 13, 14],
[18, 19, 20]],
[[15, 16, 17],
[21, 22, 23]]]), size=(4, 6), nnz=4,
layout=torch.sparse_bsr)
Sparse BSC Tensor¶
The sparse BSC (Block compressed Sparse Column) tensor format implements the BSC format for storage of twodimensional tensors with an extension to supporting batches of sparse BSC tensors and values being blocks of multidimensional tensors.
A sparse BSC tensor consists of three tensors: ccol_indices
,
row_indices
and values
:
The
ccol_indices
tensor consists of compressed column indices. This is a (B + 1)D tensor of shape(*batchsize, ncolblocks + 1)
. The last element is the number of specified blocks,nse
. This tensor encodes the index invalues
androw_indices
depending on where the given row block starts. Each successive number in the tensor subtracted by the number before it denotes the number of blocks in a given column.The
row_indices
tensor contains the row block indices of each element. This is a (B + 1)D tensor of shape(*batchsize, nse)
.The
values
tensor contains the values of the sparse BSC tensor elements collected into twodimensional blocks. This is a (1 + 2 + K)D tensor of shape(nse, nrowblocks, ncolblocks, *densesize)
.
Construction of BSC tensors¶
Sparse BSC tensors can be directly constructed by using the
torch.sparse_bsc_tensor()
function. The user must supply the row
and column block indices and values tensors separately where the column block indices
must be specified using the CSR compression encoding.
The size
argument is optional and will be deduced from the ccol_indices
and
row_indices
tensors if it is not present.
>>> ccol_indices = torch.tensor([0, 2, 4])
>>> row_indices = torch.tensor([0, 1, 0, 1])
>>> values = torch.tensor([[[0, 1, 2], [6, 7, 8]],
... [[3, 4, 5], [9, 10, 11]],
... [[12, 13, 14], [18, 19, 20]],
... [[15, 16, 17], [21, 22, 23]]])
>>> bsc = torch.sparse_bsc_tensor(ccol_indices, row_indices, values, dtype=torch.float64)
>>> bsc
tensor(ccol_indices=tensor([0, 2, 4]),
row_indices=tensor([0, 1, 0, 1]),
values=tensor([[[ 0., 1., 2.],
[ 6., 7., 8.]],
[[ 3., 4., 5.],
[ 9., 10., 11.]],
[[12., 13., 14.],
[18., 19., 20.]],
[[15., 16., 17.],
[21., 22., 23.]]]), size=(4, 6), nnz=4,
dtype=torch.float64, layout=torch.sparse_bsc)
Tools for working with sparse compressed tensors¶
All sparse compressed tensors — CSR, CSC, BSR, and BSC tensors — are conceptionally very similar in that their indices data is split into two parts: socalled compressed indices that use the CSR encoding, and socalled plain indices that are orthogonal to the compressed indices. This allows various tools on these tensors to share the same implementations that are parameterized by tensor layout.
Construction of sparse compressed tensors¶
Sparse CSR, CSC, BSR, and CSC tensors can be constructed by using
torch.sparse_compressed_tensor()
function that have the same
interface as the above discussed constructor functions
torch.sparse_csr_tensor()
, torch.sparse_csc_tensor()
,
torch.sparse_bsr_tensor()
, and torch.sparse_bsc_tensor()
,
respectively, but with an extra required layout
argument. The
following example illustrates a method of constructing CSR and CSC
tensors using the same input data by specifying the corresponding
layout parameter to the torch.sparse_compressed_tensor()
function:
>>> compressed_indices = torch.tensor([0, 2, 4])
>>> plain_indices = torch.tensor([0, 1, 0, 1])
>>> values = torch.tensor([1, 2, 3, 4])
>>> csr = torch.sparse_compressed_tensor(compressed_indices, plain_indices, values, layout=torch.sparse_csr)
>>> csr
tensor(crow_indices=tensor([0, 2, 4]),
col_indices=tensor([0, 1, 0, 1]),
values=tensor([1, 2, 3, 4]), size=(2, 2), nnz=4,
layout=torch.sparse_csr)
>>> csc = torch.sparse_compressed_tensor(compressed_indices, plain_indices, values, layout=torch.sparse_csc)
>>> csc
tensor(ccol_indices=tensor([0, 2, 4]),
row_indices=tensor([0, 1, 0, 1]),
values=tensor([1, 2, 3, 4]), size=(2, 2), nnz=4,
layout=torch.sparse_csc)
>>> (csr.transpose(0, 1).to_dense() == csc.to_dense()).all()
tensor(True)
Supported operations¶
Linear Algebra operations¶
The following table summarizes supported Linear Algebra operations on
sparse matrices where the operands layouts may vary. Here
T[layout]
denotes a tensor with a given layout. Similarly,
M[layout]
denotes a matrix (2D PyTorch tensor), and V[layout]
denotes a vector (1D PyTorch tensor). In addition, f
denotes a
scalar (float or 0D PyTorch tensor), *
is elementwise
multiplication, and @
is matrix multiplication.
PyTorch operation 
Sparse grad? 
Layout signature 

no 


no 


no 


no 


no 


no 


no 


no 


no 


yes 


no 


no 


no 


no 


no 


no 


yes 


no 


no 


yes 


yes 

where “Sparse grad?” column indicates if the PyTorch operation supports
backward with respect to sparse matrix argument. All PyTorch operations,
except torch.smm()
, support backward with respect to strided
matrix arguments.
Note
Currently, PyTorch does not support matrix multiplication with the
layout signature M[strided] @ M[sparse_coo]
. However,
applications can still compute this using the matrix relation D @
S == (S.t() @ D.t()).t()
.
Tensor methods and sparse¶
The following Tensor methods are related to sparse tensors:
Is 

Is 

Return the number of dense dimensions in a sparse tensor 

Return the number of sparse dimensions in a sparse tensor 

Returns a new sparse tensor with values from a strided tensor 

Returns a sparse copy of the tensor. 

Convert a tensor to coordinate format. 

Convert a tensor to compressed row storage format (CSR). 

Convert a tensor to compressed column storage (CSC) format. 

Convert a tensor to a block sparse row (BSR) storage format of given blocksize. 

Convert a tensor to a block sparse column (BSC) storage format of given blocksize. 

Creates a strided copy of 

Return the values tensor of a sparse COO tensor. 
The following Tensor methods are specific to sparse COO tensors:
Returns a coalesced copy of 

Resizes 

Removes all specified elements from a sparse tensor 

Returns 

Return the indices tensor of a sparse COO tensor. 
The following methods are specific to sparse CSR tensors and sparse BSR tensors:
Returns the tensor containing the compressed row indices of the 

Returns the tensor containing the column indices of the 
The following methods are specific to sparse CSC tensors and sparse BSC tensors:
The following Tensor methods support sparse COO tensors:
add()
add_()
addmm()
addmm_()
any()
asin()
asin_()
arcsin()
arcsin_()
bmm()
clone()
deg2rad()
deg2rad_()
detach()
detach_()
dim()
div()
div_()
floor_divide()
floor_divide_()
get_device()
index_select()
isnan()
log1p()
log1p_()
mm()
mul()
mul_()
mv()
narrow_copy()
neg()
neg_()
negative()
negative_()
numel()
rad2deg()
rad2deg_()
resize_as_()
size()
pow()
sqrt()
square()
smm()
sspaddmm()
sub()
sub_()
t()
t_()
transpose()
transpose_()
zero_()
Torch functions specific to sparse Tensors¶
Constructs a sparse tensor in COO(rdinate) format with specified values at the given 

Constructs a sparse tensor in CSR (Compressed Sparse Row) with specified values at the given 

Constructs a sparse tensor in CSC (Compressed Sparse Column) with specified values at the given 

Constructs a sparse tensor in BSR (Block Compressed Sparse Row)) with specified 2dimensional blocks at the given 

Constructs a sparse tensor in BSC (Block Compressed Sparse Column)) with specified 2dimensional blocks at the given 

Constructs a sparse tensor in Compressed Sparse format  CSR, CSC, BSR, or BSC  with specified values at the given 

Return the sum of each row of the given sparse tensor. 

This function does exact same thing as 

Performs a matrix multiplication of the dense matrices 

Performs a matrix multiplication of the sparse matrix 

Matrix multiplies a sparse tensor 

Performs a matrix multiplication of a sparse COO matrix 

Performs a matrix multiplication of the sparse matrix 

Applies a softmax function. 

Applies a softmax function followed by logarithm. 

Creates a sparse 2D tensor by placing the values from rows of 
Other functions¶
The following torch
functions support sparse tensors:
cat()
dstack()
empty()
empty_like()
hstack()
index_select()
is_complex()
is_floating_point()
is_nonzero()
is_same_size()
is_signed()
is_tensor()
lobpcg()
mm()
native_norm()
pca_lowrank()
select()
stack()
svd_lowrank()
unsqueeze()
vstack()
zeros()
zeros_like()
To manage checking sparse tensor invariants, see:
A tool to control checking sparse tensor invariants. 
To use sparse tensors with gradcheck()
function,
see:
Decorate function, to extend gradcheck for sparse tensors. 
Unary functions¶
We aim to support all zeropreserving unary functions.
If you find that we are missing a zeropreserving unary function that you need, please feel encouraged to open an issue for a feature request. As always please kindly try the search function first before opening an issue.
The following operators currently support sparse COO/CSR/CSC/BSR/CSR tensor inputs.
abs()
asin()
asinh()
atan()
atanh()
ceil()
conj_physical()
floor()
log1p()
neg()
round()
sin()
sinh()
sign()
sgn()
signbit()
tan()
tanh()
trunc()
expm1()
sqrt()
angle()
isinf()
isposinf()
isneginf()
isnan()
erf()
erfinv()