Update scipy / opencv / matplotlib / numpy versions

jasper-tms commented 6 months ago

In d8f4b88 we relaxed the scipy version requirement a bit, but scipy==1.4.1 still only has wheels for python3.8 and below (see available wheels at https://pypi.org/project/scipy/1.4.1/#files). The latest version of scipy currently is 1.13.0, which has wheels for python3.9 to python3.12. We should try just unpinning the scipy version to let pip install the latest version that's compatible with the version of python the user is trying to use. If unpinning breaks things, there might be some work to do to get everything to run with a more modern python version. Maybe aim for 3.10 or so.

This will also fix https://github.com/NeLy-EPFL/DeepFly3D/security/dependabot/2

I can get around to doing this sometime, though @azmaite if you're testing/using this code soon feel free to give the updated scipy a try.

jasper-tms commented 6 months ago

Oh the JB_stable branch already has scipy unpinned. It has been a while but iirc I was using that branch and didn't have any problems, so probably we can merge it into master and call it a day.

jasper-tms commented 5 months ago

I'm trying to get everything running with python3.10 and recent versions of a bunch of standard packages:

scipy 1.13.1 (the latest as of today)
matplotlib 3.9.0 (the latest as of today)
numpy 1.26.4 (the latest as of today)
opencv-python-headless 4.9.0.80 (not quite the latest, which is 4.10.0.82, but probably at some point I'll try the latest)

I've pushed some commits to the jsp-bugfixes branch that were necessary for this. With these commits, things overall work -- I'm getting 2D and 3D pose estimations that look to be correct on new data and on the data in the sample/test/ folder.

Remaining things to figure out:

[x] ax3d = Axes3D(fig) needed to be changed to ax3d = fig.add_subplot(111, projection='3d') for the 3d pose video rendering to work for me. (Not sure why the Axes3D thing didn't work, but GPT tells me that it is not a recommended way to do 3D plotting anymore so it seems fine to me to try to phase it out in favor of the now-recommended method of add_subplot(projection='3d').) The video output by the updated code almost identical to the expected old rendering but not quite -- looks like the new code produces slightly more zoomed out renderings. I guess I should see if I can change some zoom setting to produce nearly identical rendering. Solved in 1f0bef4296cd5b9fb6dc2be29fde8621a89f7cd1
[x] I should compare the output when df3d is run on the test images with both old versions of packages and the new versions. I would imagine I'll see the video zoom difference thing, but also potentially notice some more things that need to be addressed. Yep they're otherwise identical.

jasper-tms commented 5 months ago

numpy 2.0.0 was just released (https://pypi.org/project/numpy/), so as part of this issue I should also see what needs to be done to get deepfly3d to run with that

jasper-tms commented 2 months ago

There's a dependabot warning that there's a vulnerability in opencv-python versions below 4.8.1.78. That will be addressed when this issue is closed since I've tested with opencv-python 4.9.x

NeLy-EPFL / DeepFly3D

Update scipy / opencv / matplotlib / numpy versions #43