eigenpy breaks array reshaping

[rhaschke]

When using eigenpy, flattening of arrays fails:

import numpy
v = numpy.array([1,2,3,4]).reshape(-1,1)  # column vector
print v.flatten().shape  # should result in (4,), but yields (4,1) when using eigenpy

[jcarpent]

I will have a look at it. Thanks for raising this bug.

[jcarpent]

I've tried on my computer, and the bug you mentioned is not occurring. Can you provide me your NumPy version and the eigenpy version you're currently using?

[rhaschke]

numpy: 1.13.3
eigenpy: 1.5.1

But, I guess you were able to reproduce the problem?
I can only validate your fix tomorrow... Thanks for the quick response.

[rhaschke]

With your fix my particular problem still exists. I have created a minimal example to reproduce the issue:

from __future__ import print_function

import eigenpy
import numpy as np

quat = eigenpy.Quaternion(1,2,3,4)
a = quat.coeffs().T
print(a, a.flatten())

b = np.array([[1.,2.,3.,4.]])
print(b, b.flatten())

which yields:

[[ 2.  3.  4.  1.]] [[ 2.  3.  4.  1.]]
[[ 1.  2.  3.  4.]] [ 1.  2.  3.  4.]

You can see that the flatten() on the Eigen matrix doesn't yield a one-dimensional array, but it stays to be a 1xN matrix.

[rhaschke]

Using np.asarray(a).flatten() as a workaround, my issue is solved now.
When I was prototyping my code two days ago, I didn't notice that eigenpy objects cannot handle flatten() in the usual way and that drove me nearly crazy because I got mismatching dimensions all the time...

[jcarpent]

You should in fact use those lines:

from __future__ import print_function

import eigenpy
eigenpy.switchToNumpyArray() # Allows to convert Eigen::Matrix into np.array objects
import numpy as np

quat = eigenpy.Quaternion(1,2,3,4)
a = quat.coeffs().T
print(a, a.flatten())

b = np.array([[1.,2.,3.,4.]])
print(b, b.flatten())

The content summarizes the way it works. Hope it will help you.

[rhaschke]

This seems somewhat unintuitive: flatten() should always return a 1-dimensional np.array (because that's the purpose of flatten). The eigenpy array/matrix mode should only effect how generated matrix-like objects are handled, shouldn't it?
Anyway, thanks a lot for the clarification.

[jmirabel]

This seems somewhat unintuitive: flatten() should always return a 1-dimensional

This is a numpy issue:

import numpy
print (numpy.array([[1,2,3,4]]).flatten())
print (numpy.matrix([[1,2,3,4]]).flatten())

yields

[1 2 3 4]
[[1 2 3 4]]

[rhaschke]

Ahh, OK! Didn't know. I think, I always used arrays so far...