6. Sparse Backend

TensorLy supports sparse tensors for some backends and algorithms.

When selecting a backend to be used for tensorly, this backend will also be applied to the sparse implementations. This is because many backends natively support both dense and sparse tensors as distinct objects. For example, TensorFlow and PyTorch both have (some) support for sparse tensors. Using TensorLy’s dense or sparse interfaces will give you appropriate objects for the selected backend.

6.1. Why a separate sparse backend?

Some algorithms need to be adapted to be sparse-safe.

In addition, not all backends have a native sparse representation. For instance, NumPy lacks a sparse array structure. In these cases, TensorLy makes opinionated decisions about how to handle sparse support, if at all. It is usually necessary for non-native sparse support to require additional dependencies. For the NumPy backend, the PyData/Sparse project is used as the sparse representation.

6.2. Algorithms

  • parafac

  • non_negative_parafac

  • partial_tucker

  • robust_pca

7. Usage

The sparse sub-package in tensorly.contrib.sparse contains a mirror of the interfaces in the usual, dense tensorly package. For example, unfolding a sparse tensor would use:

>>> import numpy as np
>>> import sparse
>>> from tensorly.contrib.sparse import tensor, unfold
>>> X = sparse.COO(np.arange(24).reshape((3, 4, 2)))
>>> X = sparse.COO(X.coords, X.data, shape=(100, 100, 100))
>>> X = tl.tensor(X)
>>> unfold(X, 0) # mode-1 unfolding
array([[ 0,  1,  2,  3,  4,  5,  6,  7],
       [ 8,  9, 10, 11, 12, 13, 14, 15],
       [16, 17, 18, 19, 20, 21, 22, 23]])

This separation makes it explicit whether a sparse or dense algorithm is being used.

Note that not all algorithms currently support sparse. This is because some algorithms may need to be rewritten to properly avoid densifying large sparse arrays, or creating arrays with large shapes. The algorithms listed above have been tested with sparse tensors. Other algorithms may work, but one should be careful to watch tensorly’s memory usage when using them.

With the PyData/Sparse backend, you can set the environment variable SPARSE_AUTO_DENSIFY=0 before importing sparse and it will cause it to raise a RuntimeError whenever a sparse tensor would be automatically densified. However, be aware that this does not protect against an algorithm attempting to create a new dense array with a large shape.

Another caveat to be aware of is that some algorithms may perform better with the dense variant, even when using sparse arrays. For example, the parafac algorithm has been written so as to avoid large memory usage for dense arrays. However, because the resulting decomposed vectors will generally be dense even if the input tensor is sparse. Thus, using the sparse parafac (tensorly.contrib.sparse.decomposition.parafac) may be slower and use more memory than the dense version (tensorly.decomposition.parafac) for a sparse input tensor.

7.1. Missing Values

The parafac function supports masks for missing values. A mask should be a boolean array of the same shape as the original tensor that is False/0 where the value is missing and True/1 where it is not. It is passed to the parafac() function via the mask parameter.

When using masks with sparse tensors, there are two important caveats:

  • The mask itself should be sparse. In the algorithm, the memory used will be proportional to the number of missing values. Values of the original sparse tensor that are 0 should be True/1 (i.e., non-missing) in the mask.

  • Sparse mask support is currently only supported with the sparse backend parafac (tensor.contrib.sparse.decomposition.parafac). The dense backend parafac will densify the array.

8. Example

In the following example, we construct a random sparse tensor that has a known rank (by construction), and decompose it with parafac.

>>> shape = (1000, 1001, 1002)
>>> rank = 5
>>> starting_factors = [sparse.random((i, rank)) for i in shape]
>>> starting_factors
[<COO: shape=(1000, 5), dtype=float64, nnz=50, fill_value=0.0>, <COO: shape=(1001, 5), dtype=float64, nnz=50, fill_value=0.0>, <COO: shape=(1002, 5), dtype=float64, nnz=50, fill_value=0.0>]

Here we construct a tensor from the random factors. Note that in general, a recomposed tensor will be dense, but for our constructed example it is sparse, so we can use kruskal_to_tensor without worrying about using too much memory.

>>> from tensorly.contrib.sparse.kruskal_tensor import kruskal_to_tensor
>>> tensor = kruskal_to_tensor(starting_factors)
>>> tensor
<COO: shape=(1000, 1001, 1002), dtype=float64, nnz=5044, fill_value=0.0>

This is how much memory the sparse array takes up, vs. how much it would take up if it were represented densly.

>>> tensor.nbytes / 1e9 # Actual memory usage in GB
>>> import numpy as np
>>> np.prod(tensor.shape) * 8 / 1e9 # Memory usage if array was dense, in GB

Now to decompose the tensor.

>>> from tensorly.decomposition import parafac # The dense version
>>> import time
>>> t = time.time(); factors = parafac(tensor, 5, init='random'); print(time.time() - t)

Note that the decomposition takes much longer when using the sparse variant.

>>> from tensorly.contrib.sparse.decomposition import parafac as sparse_parafac # The sparse version
>>> t = time.time(); sparse_factors = sparse_parafac(tensor, 5, init='random'); print(time.time() - t)

However, there can be advantages to using the sparse variant. It is currently required when using init='svd' to make TensorLy use the sparse SVD algorithm (from scipy.sparse). Choosing the sparse backend parafac also makes it use scipy.sparse.linalg.spsolve instead of numpy.linalg.solve, which can have advantages, for instance, when using the scikit-umfpack backend, it is more robust against nearly singular intermediate matrices.