@ldes89150 thanks for the report! The code should never freeze, that's definitely a bug. However, to indicate masked values, you should use a NumPy masked array. The second part of the gallery example shows how to use this:

http://scikit-image.org/docs/dev/auto_examples/filters/plot_phase_unwrap.html

@scikit-image/core longer term, perhaps we should add a mask argument to the function that has the same semantics as the rest of the library?

@ldes89150 bug confirmed.
@jni agreed.

Quickly looking at the code, it seems there are some floating point comparisons that aren't NaN safe in the C code.

Not sure the best way to deal with this that won't completely kill performance.

Agreed. I think it is the floating point comparison that is tripping off the algorithm

https://github.com/scikit-image/scikit-image/blob/master/skimage/restoration/unwrap_2d_ljmu.c#L100

typedef enum {
  yes,
  no
} yes_no;

yes_no find_pivot(EDGE *left, EDGE *right, double *pivot_ptr) {

Really, bools were not sufficiently expressive here? 🤦‍♂️

@hmaarrfk a hang suggests a while loop is involved rather than a for-loop. I notice two whiles later on in that function...

@stefanv perhaps np.nan_to_num would be useful here? =)

We should look at how other NaN safe functions are implemented

@jni Thanks for the suggestion of using masked arrays!

Masking arrays unfortunately does not work for me. I have a 3000 x 3000 float array with NaNs masked out using MaskedArrays. Testing numpy.isnan(arr).any() returns False.
Maybe it is worth mentioning that the NaNs cover about 25% of the "outer" regions of the array including complete rows / columns. The rest is well connected, though. I did check value ranges.

@plu2k are you able to share your masked array with us for testing?

This is strange. Although using masked arrays works in some cases, when trying to corner the problem I found this minimal example:

import numpy as np
from skimage.restoration import unwrap_phase
wrapped = np.empty((100, 100), dtype=np.float64)
wrapped[:] = np.linspace(0, 4*np.pi, 100) % (2*np.pi)
wrapped[10, 10] = np.nan  #  <-- comment out to get working code
masked = np.ma.masked_array(wrapped, mask=np.isnan(wrapped))
print(np.isnan(masked).any())
unwrapped = unwrap_phase(masked)

which causes a freeze (skimage 0.15.0, numpy 1.16.2).

This is strange. Although using masked arrays works in some cases, when trying to corner the problem I found this minimal

Thanks, @plu2k, I can confirm that this is still the case on v0.19.2

See #7176 for an attempted fix.