[Request] Add ARM/AArch64 whl for opencv-python-headless==4.4.0.46

[LukeIreland1]

Long story short, the numpy dependency for TF 2.3.x (<1.19.0) means that we have to use this version of opencv-python-headless to avoid it reinstalling numpy, and currently have to build this from source which takes a very long time.

Adding this particular whl would save us (ONCPUML @ Arm) a huge amount of time before we eventually move to TF 2.5.x. I'm happy to build and push myself, but I don't mind how it gets done.

[sergregory]

Hi @LukeIreland1 I'll take a look at what I can do with that. Probably we can set fewer restrictions on the numpy version as you've mentioned you were able to compile from sources with <1.19.0

[LukeIreland1]

The docker recipe is with Ubuntu 20.04, GCC 9 (TF doesn't support GCC10 yet), Python 3.8, Numpy 1.18.4 and of course opencv-python-headless 4.4.0.46.

Your idea would also solve the problem, with a bonus that we can use the latest version!

[LukeIreland1]

Any updates?

[JoseMoreville]

Hi! I'm having the same issue with MAC M1 chip and i was wondering if there is any update

Problem with the CMake installation, aborting build. CMake executable is cmake
----------------------------------------
ERROR: Failed building wheel for cmake
Building wheel for pip (PEP 517): started
Building wheel for pip (PEP 517): finished with status 'done'
Created wheel for pip: filename=pip-21.1.2-py3-none-any.whl size=1548016 sha256=0d4277ef88302462bd99da4a69736432e62d68ab9e6dd241f427d0a50318a5d9
Stored in directory: /Users/hades/Library/Caches/pip/wheels/a2/ce/31/b392aede8a115524ad9dff45e094fb3a48c15eef1d3338e2f8
Building wheel for packaging (PEP 517): started
Building wheel for packaging (PEP 517): finished with status 'done'
Created wheel for packaging: filename=packaging-20.9-py2.py3-none-any.whl size=40891 sha256=5e26f152646c6467a7a0cbbc702999a215f73d2a9548ec9940badc5ec0af4e26
Stored in directory: /Users/hades/Library/Caches/pip/wheels/65/ab/5b/83dbc6a48fc7cf74c19b52a8bbcc444f2a1a35c38235d83b66
Building wheel for setuptools (PEP 517): started
Building wheel for setuptools (PEP 517): finished with status 'done'
Created wheel for setuptools: filename=setuptools-57.0.0-py3-none-any.whl size=821665 sha256=c2a1407a3a3a8594fabf0010b9a57c6a0852a96c952890125e606c4fa07c9ef0
Stored in directory: /Users/hades/Library/Caches/pip/wheels/1f/6f/84/f79c03348267aa19978c7896b4d5e0400666a8c1e8c6fadcd5
Successfully built pip packaging setuptools
Failed to build numpy cmake
ERROR: Could not build wheels for numpy, cmake which use PEP 517 and cannot be installed directly

[LukeIreland1]

Hi! I'm having the same issue with MAC M1 chip and i was wondering if there is any update

It looks like your issue is actually that you can't build a cmake wheel. Do you definitely have cmake installed?

[JoseMoreville]

Hi! I'm having the same issue with MAC M1 chip and i was wondering if there is any update

It looks like your issue is actually that you can't build a cmake wheel. Do you definitely have cmake installed?

@LukeIreland1  not sure, i don't remember If it was installed by brew or pip

[LukeIreland1]

@LukeIreland1  not sure, i don't remember If it was installed by brew or pip

which cmake should tell you (probably). I have no experience with the cmake pip package, but I've had no issue (with cmake) using a normal package manager like brew, yum, or apt.

[LukeIreland1]

It looks like my team's docker image releases will include TF 2.5 sooner than the numpy requirement for the latest opencv-python-headless will change, or sooner than a whl for opencv-python-headless 4.4.0.46 will be added in any case.

[JoseMoreville]

It looks like my team's docker image releases will include TF 2.5 sooner than the numpy requirement for the latest opencv-python-headless will change, or sooner than a whl for opencv-python-headless 4.4.0.46 will be added in any case.

I managed to make it work with mini forge, conda also installing cmake cpython and cython.
For some reason that worked but i still need to check. It might be thanks to Rosseta 2.
It still took me a while to solve all the issues that cv2 gives when importing.
I'm not 100% of the steps but I'll try to reproduce the steps on other account or device to make a small "guide"

[LukeIreland1]

I managed to make it work with mini forge, conda also installing cmake cpython and cython.
For some reason that worked but i still need to check. It might be thanks to Rosseta 2.
It still took me a while to solve all the issues that cv2 gives when importing.
I'm not 100% of the steps but I'll try to reproduce the steps on other account or device to make a small "guide"

Yeah, this version installs fine, but then you obviously must build it from source which takes forever (probably the majority of the docker build time for us).