Numpy vnl bridge unexpected behavior

1

Hi,
I tend to write too many things on the same line in Python and this caused an unexpected behavior for the numpy vnl bridge. It occurred in a different context but I have the following minimal example

import itk
                                                                                                                                                                                                                                                                                                                                                                                                                                         
m = itk.Matrix[itk.D,4,4]()
m.SetIdentity()
m2 = m+m
print(itk.GetArrayFromVnlMatrix((m+m).GetVnlMatrix().as_matrix()))
print(itk.GetArrayFromVnlMatrix(m2.GetVnlMatrix().as_matrix()))

which should show the same two matrices but actually shows

[[4.68850067e-310 0.00000000e+000 0.00000000e+000 0.00000000e+000]
 [0.00000000e+000 2.00000000e+000 0.00000000e+000 0.00000000e+000]
 [0.00000000e+000 0.00000000e+000 2.00000000e+000 0.00000000e+000]
 [0.00000000e+000 0.00000000e+000 0.00000000e+000 2.00000000e+000]]
[[2. 0. 0. 0.]
 [0. 2. 0. 0.]
 [0. 0. 2. 0.]
 [0. 0. 0. 2.]]

Any clue what’s going on and if this can and should be corrected? It seems to be some memory release before the deep copy of the numpy array but I could not see what was wrong in the bridge code.
Thanks in advance!
Simon

2 Likes
2

@fbudin any ideas?

3

No idea, but it is worth the look. Interestingly enough, since the implementation between images and VNL matrices are similar for the NumPy bridge, this problem could potentially arise with images…

4

As a follow up, trying:

In [7]: a=(m+m).GetVnlMatrix()

In [8]: print(itk.GetArrayFromVnlMatrix(a.as_matrix()))
[[4.65803687e-310 6.90632463e-310 6.90632461e-310 6.90632465e-310]
 [6.90632463e-310 6.90632463e-310 6.90632464e-310 6.90632464e-310]
 [6.90632463e-310 6.90632464e-310 6.90632463e-310 6.90632463e-310]
 [6.90632462e-310 6.90632462e-310 6.90632461e-310 6.90632464e-310]]

also returns garbage, which makes me think that the problem is in GetVnlMatrix(), not in the bridge part.

This is most likely due to the return by reference here.

1 Like
5

This PR solves the problem on my computer. @simon.rit: Could you test it on your machine? It is not the solution I wanted to implement (I wanted to use a SWIG typemap), but I couldn’t get the other solution to work. So I think that for now this could be a reasonable patch.

If this works for you, would you mind improving the PR with a test?

Thanks!

2 Likes
6

Great, thanks a lot for your efforts. Yes, I’ll test and provide a test. I now understand the reference issue and can provide a simpler example for the problem, without some one line operation:

>>> m = itk.Matrix[itk.D,3,3]()
>>> m.SetIdentity()
>>> v = m.GetVnlMatrix()
>>> del m
>>> print(itk.GetArrayFromVnlMatrix(v.as_matrix()))
[[4.65884433e-310 0.00000000e+000 0.00000000e+000]
 [0.00000000e+000 1.00000000e+000 0.00000000e+000]
 [0.00000000e+000 0.00000000e+000 1.00000000e+000]]

I’ll base my test on this, just the time to recompile the wrapping with your master branch and your PR.

2 Likes
7

I created another PR to add helper functions to convert NumPy array to ITK matrices (and back) directly, without having to worry about VNL matrices:

1 Like
8

See my PR


I wasn’t sure if a new commit was in order. Let me know if this is not how it should be done, I’m happy to correct the PR if you tell me how.

1 Like