sbb_binarization
sbb_binarization copied to clipboard
Update Tensorflow
Minimum Tensorflow version sticks with 2.4 but should be 2.12.1 to be in line with eynollah.
Due to being stuck on CentOS 7 with rather outdated nvidia drivers, we can only currently test up to Tensorflow <=2.11.x and CuDNN <=8.1.x here (applies for sbb_binarization
, eynollah
and sbb_pixelwise_segmentation
). I will make the updates to READMEs and requirements.txt asap.
@cneud your requirement change in eynollah pinned that to 2.12. That in turn excludes using it together (in the same venv, as is most useful for ocrd_all) with sbb-binarization, which with your change requires <2.12
.
This difference involves a significant switch from protobuf 3 to 4.
I have tried with the CI tests (running the OCR-D wrapper on kant_aufklaerung_1784-page-region
): TF 2.11 and 2.12 produce binary-identical results.
So can we please update to 2.12 here?
(Also, why pin eynollah to 2.12.1
rather than 2.12.*
?)