AttributeError: module 'torch.linalg' has no attribute 'lu_factor'

[yufei0726]

I encountered this problem when calling QPfunction, but I printed and saw that the eigenvalues ​​of the Q matrix were all positive and there were no diagonal zero elements. I hope someone can answer this question.

Traceback (most recent call last):
File "C:\ProgramData\anaconda3\envs\dQP\lib\site-packages\qpth\solvers\pdipm\batch.py", line 380, in pre_factor_kkt
Q_LU = lu_hack(Q)
File "C:\ProgramData\anaconda3\envs\dQP\lib\site-packages\qpth\solvers\pdipm\batch.py", line 9, in lu_hack
data, pivots = torch.linalg.lu_factor(x, pivot=not x.is_cuda)
AttributeError: module 'torch.linalg' has no attribute 'lu_factor'

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
File "runcarsim1.py", line 114, in
sys.exit(main())
File "runcarsim1.py", line 55, in main
control = controller(traj_ref, current_state)
File "C:\ProgramData\anaconda3\envs\dQP\lib\site-packages\torch\nn\modules\module.py", line 1102, in _call_impl
return forward_call(*input, **kwargs)
File "D:\syf\differentiable-control\QPnet.py", line 44, in forward
x = self.dQP(traj_ref, current_state)
File "D:\syf\differentiable-control\QPnet.py", line 231, in dQP
deltU = QPFunction(verbose = 0)(Q_obj.double(), P_obj.double(), G.double(), h.double(), e, e)
File "C:\ProgramData\anaconda3\envs\dQP\lib\site-packages\qpth\qp.py", line 93, in forward
ctx.Q_LU, ctx.S_LU, ctx.R = pdipm_b.pre_factor_kkt(Q, G, A)
File "C:\ProgramData\anaconda3\envs\dQP\lib\site-packages\qpth\solvers\pdipm\batch.py", line 382, in pre_factor_kkt
raise RuntimeError("""
RuntimeError:
qpth Error: Cannot perform LU factorization on Q.
Please make sure that your Q matrix is PSD and has
a non-zero diagonal.

[Weixy21]

This may be the problem of a new version Pytorch. You may try older version of Pytorch (e.g., '1.10.1+cu113', I received a similar warning for this problem that this will be unavailable in the future version).

[yufei0726]

I think I solved this problem. I checked the pytorch documentation and found that the functions with problems are only available in pytorch versions after 1.13. So I installed pytorch 1.13.1+cu116 and the corresponding torchvision and torchaudio to solve the problem. I hope my solution can help others. Thank you very much for your reply!

[yufei0726]

But I often get an error that my Q matrix is ​​not semi-positive definite. I don't know if there is any experience to deal with this situation.

[Weixy21]

Q is usually defined as an identity matrix. If there is such an issue with an identity Q, then it might also be the pytorch version problem. '1.10.1+cu113' has no such a problem.

[yufei0726]

I construct the optimization as a path tracking problem. The Q matrix is ​​currently derived from the dynamic parameters and the artificially set weight matrix.

[Weixy21]

Then you need to ensure that the artificially set weight matrix should always make Q positive definite, e.g., with non-negative diagonal items for diagonal Q.