Closed bertsky closed 1 year ago
is there a hard reason for Kraken to require scipy 1.11 and Python 3.8?
The thing is that scikit-image doesn't pin their scipy dependencies while at the same time not going out of their way to ensure compatibility. So they're now pinned to versions I know that work. I can go back to pinning scipy 1.10 + scikit-image 0.19.x but we'd be stuck there, as I can't define multiple valid combinations. IIRC 1.10/0.19.x also breaks python 3.11 compatibility.
Understood. Wow, what a(nother) nightmare.
Ok, so Flucht nach vorne is the best solution, I agree.
Next best alternative would be to maintain multiple branches of Kraken (each with their own pinned scipy/skimage versions) which would obviously be overkill.
So be it, thanks for clarifying!
Yeah, it's a bit annoying. I'm somewhat keen to get rid of shapely and scikit-image entirely because of these recurrent issues but that's probably only going to happen with the next iteration of the segmenter.
You just dropped 3.8 support, which broke when you required scipy 1.11, which in turn dropped Py38 due to Numpy's rather restrictive version support windows.
Since in OCR-D we rely on Py38 (as it is also the last release for which Nvidia provides Tensorflow 1.x wheels) and had a hard time recently deconflicting dependencies for that version already – is there a hard reason for Kraken to require scipy 1.11 and Python 3.8?