Bug: Scikit-image 0.14.1 broken with numpy 1.16

[abhiTronix]

Bug Description

Latest Numpy commit numpy/numpy#11966 (comment) removed _validate_lengths and other functions from numpy.lib.arraypad hence brokes the latest scikit-image binaries.

System Information:

• Operating system: Ubuntu 18.04 LTS
• Python version: Python 3.6.7
• scikit-image version : Latest(0.14.1)

Way to reproduce

Python 3.6.7 (default, Oct 22 2018, 11:32:17) 
[GCC 8.2.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import skimage
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/usr/local/lib/python3.6/dist-packages/skimage/__init__.py", line 167, in <module>
    from .util.dtype import (img_as_float32,
  File "/usr/local/lib/python3.6/dist-packages/skimage/util/__init__.py", line 8, in <module>
    from .arraycrop import crop
  File "/usr/local/lib/python3.6/dist-packages/skimage/util/arraycrop.py", line 8, in <module>
    from numpy.lib.arraypad import _validate_lengths
ImportError: cannot import name '_validate_lengths'
>>> from numpy.lib.arraypad import _validate_lengths
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
ImportError: cannot import name '_validate_lengths'
>>> 

[hmaarrfk]

Thanks for the report. I think this was fixed recently in master

#3556

And backported to the 0.14 branch.

[hmaarrfk]

I don’t know how others feel about this, but I think issues should really be closed when the problem is solved for the end user.

People installing scikit image will continue to have a broken install until .14.2 is release.

[stefanv]

Users can probably find closed issues as well, and see that it has been fixed in development? It's probably about time to release .14.2 though, you're right.

[tw255]

how can i fix this issue? Or do we just wait for the new patch?

[hmaarrfk]

what distribution are you using? who is packaging 1.16 today?

[hmaarrfk]

@thisisit1212 In either case, if you must be on numpy 1.16 before 0.14.2 gets released on pypi, conda-forge actually has the package with the relevant patches applied.

See recipe here:
https://github.com/conda-forge/scikit-image-feedstock/blob/master/recipe/meta.yaml#L10

You can take the patches you need from there and patch 0.14.1 if you need to cut your own release with minimal differences with 0.14.1.

[tw255]

How would I go about doing what you mentioned? sorry im new to programming

[hmaarrfk]

@thisisit1212 thats great that you are getting into programming with python + scikit-image!!! I think they are great libraries.

As there are many ways to install python, explain to you how to install "non-standard" software isn't a trivial task without more details.

Can you please explain to me:

1. What operating system you are using.
2. How did you install python
3. How did you install scikit-image
4. How did you install numpy?

Finally, if you can, using python, run the following two lines of code and give me the output:

import numpy as np
print(np.__version__)

[tw255]

Im so sorry for the trouble I just realized I was using numpy 1.16 for whatever reason. I just downloaded 1.15.4 and it works fine now. Thanks!

[hmaarrfk]

Right, today, you have to try and download numpy 1.16. In a few weeks, you might not have to. That is why I was surprised that it happened to you.

[sk1p]

FYI numpy 1.16 was just released, so right now if you create a new virtualenv and install via pip install scikit-image, you get incompatible versions.

[hmaarrfk]

I wish there was a sad face. Unfortunately, many libraries (other than scikit image too!!!) will need time to get tested and released with numpy 1.16. I suggest that you wait a little but before jumping on.

[stefanv]

@jni It sounds like it might be worth cutting a 0.14.2 for this. Does anyone have the bandwidth to handle this right now?

[Borda]

Could it be as a fast hotfix for 0.14.Xr? otherwise, it breaks most automatic builds without a fixed package version and pushes users to ban 0.14.1 (not sure if it appears in other image versions)

[stefanv]

I'll trigger the process today.

Update: release is underway.

[himmelmaus]

Any update on when this will be released on PyPi? I have scikit-image .14.1 and numpy 1.16.0 but I still receive the same error (running macOS high sierra and python 3.6.3)

[jni]

@findoslice Sorry, we are seeing build errors on Travis that we are still getting to the bottom of...

As a temporary workaround, you can monkey-patch your version of scikit-image. The import error points to a file called arraycrop.py in your filesystem. You can replace it with this one:
https://github.com/sciunto/scikit-image/blob/eb8c1966eba5639dd4917491aa1b02c74bca56d7/skimage/util/arraycrop.py

[Borda]

or force being numpy<1.16.0

[himmelmaus]

Thanks guys, I handled it using numpy == 1.15.0 I was just wondering as it took me a while to find it

[hmaarrfk]

i wouldn't pin to 1.15.0 it was pretty buggy. 1.15.3 or 1.15.4 are pretty good

[mpapadem]

Hi all! I have been having the same problem as described above and using numpy 1.16.0

Although a data modelling geek, my computer science skills are still developing......How do I go back to a previous version of numpy , say 1.15.5 as suggested above?

I am using Windows (I know.......still on windows...., switching into linux soon), and have Anaconda

[soupault]

@mpapadem try to execute conda install numpy=1.15.4 in your conda environment.

[mpapadem]

@soupault thanks! it s done. But now I get a different error....:

ModuleNotFoundError: No module named 'numpy.core._multiarray_umath'

[soupault]

@mpapadem maybe try alpacahq/pylivetrader#73 (comment)? If it doesn't help, I'd suggest you to Google a bit more to see if there are any alternative solutions. From what I see, the issue isn't connected to scikit-image anymore.

[hmaarrfk]

@mpapadem, this original issue was quite specific to a bug we had in the code base. I suspect you are facing a different issue if you are still hitting issues.

Please open a new issue so we can give you more targeted help.

Since you have quite a unique setup, please share the outputs of

conda list
pip list

when you open that new issue.

[claudefalbriard]

Current pip upgrade of skimage updates Numpy to version 1.16.1. and Scikit-Image 0.14.2 is still broken with error message: ImportError: cannot import name 'img_as_float32'
Thanks for checking.

[stefanv]

@claudefalbriard You may have a different problem from the rest; please submit a new issue, filling out the template provided.

[hmaarrfk]

@scikit-image/core should we close this issue to encourage others to open a new one?

[claudefalbriard]

Sorry, my fault! Its OK. This error was caused by a missing restart of the Anaconda runtime environment. The skimage upgrade via pip worked perfectly and the samples are working OK now. Thanks.

[hmaarrfk]

i wouldn't recommend installing scikit-image via pip on anaconda.'
Why do you need to install it that way?