Support multiple locations for models about nuke-ml-server HOT 5 OPEN

from nuke-ml-server.

Hi, thanks!

Do you have an impression whether the people using Docker containers are using in the context of VFX pipelines or are single users? I have a VFX pipeline in mind and hence the need for more versatility on the configuration..

I think that the os.walk() call will not be affected by any changes in sys.append.path. So the only solution is to build on-the-fly the "models" dir with symlinks to all models this particular server is supposed to see. That may be adequate for individuals playing with some machine learning models, but to integrate in a vfx pipeline it just doesn't seem to scale well. It's hard to debug, and maintain.

About the baseModel.py location. I understand it makes easier for a single user to just test and modify on top of it. But in order to rely on this software in a pipeline we should be able to install it as is, without local changes.
In that spirit, it would greatly help to separate what is API from what are model examples that one may not need installed.

Does that help motivate implementing the changes?

And if not so much, would you be open then, to consider merge requests to improve the package in those directions?

from nuke-ml-server.

@ringdk unfortunately we can't use docker for this.

@ldmoser We may just need Dmytro patch it in our branch.

from nuke-ml-server.

but yes, following up Lucio's comment, it would be nice to make this more flexible for both docker and local installs. Docker is great for services and we use it all the time for that but for what is essentially going to be just another process on a render farm using docker adds complexity which is not necessarily a great solution for us and I suspect other mid/large facilities. It also makes it very hard for us to burst to the cloud.

from nuke-ml-server.

From a practical point of view, for use not only as a playground but also in production, it would be good to separate the base Nuke MLserver code and API from the model code.
Then the base MLserver code will have a minimum of dependencies. And the base code of the server will not become outdated with every update of the machine learning packages.

from nuke-ml-server.