Error when running sparse Robust PCA

[dawe]

Describe the bug

Computation of sparse robust pca crashes as it calls svd with unexpected keyword (full_matrices)

Steps or Code to Reproduce

Mostly copied from tensorly docs

import tensorly.contrib.sparse as stl; import sparse
shape = (1000, 1001, 1002)
rank = 5
starting_weights = stl.ones((rank))
starting_factors = [sparse.random((i, rank)) for i in shape]
from tensorly.contrib.sparse.cp_tensor import cp_to_tensor
tensor = cp_to_tensor((starting_weights, starting_factors))
from tensorly.decomposition import robust_pca
from tensorly.contrib.sparse.decomposition import robust_pca as  sparse_robust_pca
dense_tensor = tensor.todense()
d_RPCA = robust_pca(dense_tensor)
s_RPCA = sparse_robust_pca(tensor)

Expected behavior

Should compute s_RPCA (equal to d_RPCA)

Actual result

Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/home/cittaro.davide/miniforge3/envs/sep22/lib/python3.9/site-packages/tensorly/contrib/sparse/core.py", line 12, in inner
    return func(*args, **kwargs)
  File "/home/cittaro.davide/miniforge3/envs/sep22/lib/python3.9/site-packages/tensorly/decomposition/robust_decomposition.py", line 112, in robust_pca
    svd_thresholding(unfold(D, i) + unfold(L[i], i) / mu, reg_J / mu),
  File "/home/cittaro.davide/miniforge3/envs/sep22/lib/python3.9/site-packages/tensorly/tenalg/proximal.py", line 845, in svd_thresholding
    U, s, V = tl.truncated_svd(matrix, n_eigenvecs=min(matrix.shape))
  File "/home/cittaro.davide/miniforge3/envs/sep22/lib/python3.9/site-packages/tensorly/tenalg/svd.py", line 226, in truncated_svd
    U, S, V = tl.svd(matrix, full_matrices=full_matrices)
  File "/home/cittaro.davide/miniforge3/envs/sep22/lib/python3.9/site-packages/tensorly/backend/__init__.py", line 206, in wrapped_backend_method
    return getattr(
TypeError: svd() got an unexpected keyword argument 'full_matrices'

Versions

>>> import platform; print(platform.platform())
Linux-3.10.0-1127.el7.x86_64-x86_64-with-glibc2.17
>>> import sys; print("Python", sys.version)
Python 3.9.13 | packaged by conda-forge | (main, May 27 2022, 16:56:21) 
[GCC 10.3.0]
>>> import numpy; print("NumPy", numpy.__version__)
NumPy 1.23.5
>>> import scipy; print("SciPy", scipy.__version__)
SciPy 1.9.1
>>> import tensorly; print("TensorLy", tensorly.__version__)
TensorLy 0.8.1

[aarmey]

Thanks for reporting this @dawe.

@JeanKossaifi, because this skip line is not indented within the if statement above, the sparse tests are always skipped. I assume this should be moved within the if statement?

[aarmey]

Some information I found:

• This is not actually an error with the full_matrices option. The actual error when you remove dynamic dispatch is that no svd method exists for COO matrices.
• The error appears this way because we need to add the full_matrices option to the stub in backend/core.py to match the real function signatures.
• One broader issue here is that svd_thresholding() needs to avoid calling tl.truncated_svd() in the first place, because it is not setup for sparse matrices.
• If this will be an option within svd_thresholding(), then we would need to expose it in robust_pca(). Adding an option for the SVD method is probably best so that the user can adjust this for performance, too.
• There still exists the issue with testing, I believe.

[JeanKossaifi]

Thanks for opening @dawe and thanks for digging into it @aarmey!

Regarding the testing, it should only be skipped if sparse is not installed - that should never be an issue for us (we can double check we indeed install pydata-sparse) but avoids test failures if users run the test suite locally without it installed.

pytest.importorskip only skips if import fails:

Import and return the requested module modname, or skip the current test if the module cannot be imported.

We haven't updated the sparse backend for a while and indeed quite likely the interfaces no longer match and should be updated.

[JeanKossaifi]

@dawe would you be interested in opening a small PR to fix the issue?

[ghbrown]

The same error seems to plague tensorly.contrib.sparse.decomposition.parafac.