search

slub / ocrd_controller

Path to network implementation of OCR-D
MIT License
6 stars 3 forks source link

Changing default base version? #33

Closed markusweigelt closed 10 months ago

markusweigelt commented 10 months ago

Currently the default ocrd/all base version is maximum-git. Should we change this version to maximum-cuda-git

https://github.com/slub/ocrd_controller/blob/6c6eb2738a76fbec61272dcb9902e61965b9b2f7/Dockerfile#L3

Is it necessary for us to use the -git version? According to the Dockerfile, I don't see any need for it, and I would change the default version to maximum-cuda.

kba commented 10 months ago

The -git is not necessary, these are just aliases, all images do now contain the git repository data.

Whether you want CUDA support depends on whether you have the hardware (GPU) to make use of it. If you do, then this will improve performance of the tensorflow processors.

markusweigelt commented 10 months ago

Thank you for your response. Is there a reason why there are so many version designations then? I find this variety of aliases very confusing. Perhaps these could be reduced somewhat as well.

kba commented 10 months ago

-git is historical, because we did not ship the git repos in the containers for a long time to save space. But now that people do use it (as do you), we cannot get rid of it anymore without breaking deployments or rather users not noticing there is a new version.

The -cuda variant has a pretty complex setup and is even larger than the regular maximum image. It's also consistent with the core/core-cuda base images.

The {mini,medi,maxi}mum variants will eventually need to go though, since everybody is using maximum anyway and we plan to provide slim containers backed by the core network extension soon-ish.

bertsky commented 10 months ago

Currently the default ocrd/all base version is maximum-git.

That's just the default, though. It is still useful for building locally. On the CI we usually build for CUDA: https://github.com/slub/ocrd_controller/blob/6c6eb2738a76fbec61272dcb9902e61965b9b2f7/.github/workflows/publish.yml#L13

We could rename from maximum-git to maximum, but since (as @kba explained) this nowadays is just an alias anyway, we can as well keep it.