Further testing for preserving tensor context with operations

[AtomicCactus]

Describe the bug

Decomposing a 2D tensor along both modes, while specifying two ranks results in an error because internally there's a tensor created on the CPU as part of the process, which cannot be concatenated with the GPU.

Works fine when the rank is specified as an integer, but not as a list:

• rank=16 works
• rank=[16,16] crashes

Works fine on the CPU, but performance is not the same.

Steps or Code to Reproduce

import torch
import tensorly as tl
tl.set_backend('pytorch')

x = torch.randn((72, 27)).cuda()
result = tl.decomposition.partial_tucker(x, rank=[16, 32], modes=[0, 1])

Expected behavior

Tucker decomposition should not fail when ranks are provided as an array of values.

Actual result

Traceback (most recent call last):
  File "/home/yuri/Source/tensorly_bug.py", line 6, in <module>
    result = tl.decomposition.partial_tucker(x, rank=[16, 32], modes=[0, 1])
  File "/home/yuri/.local/lib/python3.10/site-packages/tensorly/decomposition/_tucker.py", line 166, in partial_tucker
    core, factors = initialize_tucker(
  File "/home/yuri/.local/lib/python3.10/site-packages/tensorly/decomposition/_tucker.py", line 64, in initialize_tucker
    U, _, _ = svd_interface(
  File "/home/yuri/.local/lib/python3.10/site-packages/tensorly/tenalg/svd.py", line 426, in svd_interface
    U, V = svd_flip(U, V, u_based_decision=u_based_flip_sign)
  File "/home/yuri/.local/lib/python3.10/site-packages/tensorly/tenalg/svd.py", line 42, in svd_flip
    signs = tl.concatenate((signs, tl.ones(tl.shape(V)[0] - tl.shape(U)[1])))
  File "/home/yuri/.local/lib/python3.10/site-packages/tensorly/backend/__init__.py", line 206, in wrapped_backend_method
    return getattr(
  File "/home/yuri/.local/lib/python3.10/site-packages/tensorly/backend/pytorch_backend.py", line 153, in concatenate
    return torch.cat(tensors, dim=axis)
RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cuda:0 and cpu! (when checking argument for argument tensors in method wrapper_CUDA_cat)

Versions

Linux-5.19.0-41-generic-x86_64-with-glibc2.35
Python 3.10.6 (main, Mar 10 2023, 10:55:28) [GCC 11.3.0]
NumPy 1.22.4
SciPy 1.7.3
TensorLy 0.8.1

[AtomicCactus]

I think we just need to make sure that on line #42 in svd.py the tl.ones(tl.shape(V)[0] - tl.shape(U)[1]) tensor is moved to the same device as the signs tensor:

[JeanKossaifi]

Good catch thanks @AtomicCactus! Could you open a small PR to fix the issue? Perhaps we could test this kind of issue, at least by changing the dtype of the input tensor in the future.

[AtomicCactus]

Sure thing! https://github.com/tensorly/tensorly/pull/504 I didn't add a unit test for this, since I'm not sure if any CI/CD pipeline or other system that runs those has a GPU.

[JeanKossaifi]

Thanks @AtomicCactus! I reviewed the PR -- for test we could look at dtype rather than device. Our current CI pipeline doesn't have GPU support.