torch.potrf fails on a positive definite matrix #199
Description
Hi, I could potentially be overlooking something simple here, but I think I'm hitting a rare edge case in torch.potrf. I've temporarily uploaded a 100x100 Tensor here that is positive definite. Numpy can compute the Cholesky on it, but Torch throws an error. As a hacky workaround I use torch.Tensor(np.linalg.cholesky(X.cpu().numpy())).type_as(X) if the factorization in Torch fails.
Source
#!/usr/bin/env python3
import numpy as np
import numpy.random as npr
import torch
with open('X.pt', 'rb') as f: X = torch.load(f)
X_np = X.numpy()
print('=== X is positive-definite, the minimum eigenvalue is:',
np.min(np.linalg.eigvals(X_np)))
print('\n\n=== Cholesky with np:')
print(np.linalg.cholesky(X_np))
print('\n\n=== Cholesky with Torch:')
print(torch.potrf(X))Output
=== X is positive-definite, the minimum eigenvalue is: 1.94934e-05
=== Cholesky with np:
[[ 8.96033478 0. 0. ..., 0. 0. 0. ]
[ 0.12424465 7.53945875 0. ..., 0. 0. 0. ]
[ 2.70835567 0.68308622 7.09788656 ..., 0. 0. 0. ]
...,
[ -5.82428885 7.08884764 -10.74139786 ..., 0.26134527 0. 0. ]
[ 0.87781757 1.51691735 -2.78839183 ..., 0.1106414 0.07375667
0. ]
[ 0.40712687 4.66132307 -4.44510603 ..., 0.2638523 0.01533402
0.04457394]]
=== Cholesky with Torch:
Traceback (most recent call last):
File "./t2.py", line 17, in <module>
print(torch.potrf(X))
RuntimeError: Lapack Error in potrf : the leading minor of order 100 is not positive definite at /home/bamos/src/pytorch/torch/lib/TH/generic/THTensorLapack.c:567