Optional Torch Multiprocessing in nnUNet for Improved Security and Compatibility
Opened this issue · 6 comments
Dear nnUNet team,
We are currently facing challenges when running nnUNet in Docker containers.
The requirement to use flags like --ipc=host or --shm-size for torch.multiprocessing (as reported and discussed in previous issues) makes it difficult to deploy models in our mhub.ai platform. We hesitate to suggest the use of --ipc=host, which is a simple solution but removes security restrictions and should therefore be used with caution. On the other hand, manually specifying the shm-size means an additional burden and makes the inference even more complicated from MHub's point of view.
We currently have some contributions to our portfolio on hold to discuss this topic. In our particular case, the inference is executed sequentially and therefore does not require multiprocessing per se. We propose to make the use of torch.multiprocessing optional during inference.
We welcome any comments and an open discussion on this topic.
Thank you very much!
Leo.
Hi Team,
I would also like to address this issue but from the perspective of user debugging. I think having an option to set num_workers=0 in the inference dataloader would allow users to better debug error traces that occur.
I see a few issues that would probably benefit from this as well:
#2509
#2514
#2182
Looking forward to hear what you think
Sounds like a reasonable proposal.
Currently, time is short on our side so if one of you, @surajpaib or @LennyN95, would be willing to create a PR with the feature in question we can certainly integrate this.
Setting num_workers=0 will probably be more work than just offering a new member function of nnUNetPredictor that covers this. This function would then not be supported via command line but if you are packing nnU-Net in a Docker you might not need that anyways. Would this be a sensible solution to your problem?
Best,
Fabian
I've come along similar issues when trying to deploy nnUNet models in docker containers, and although my multiprocessing issues might come from somewhere else, I also agree that having a sequential predictor might simplify a lot of things, especially when you don't need to do inference in big batches.
In my use-case, a quick fix that didn't require messing with your code is to simply use the predict_single_npy_array function while looping through my input arrays. This isn't a big issue for me as I already do all the dataloading manually, but might be a bit more cumbersome for other applications.
There is now a predict_from_files_sequential function in nnUNetPredictor that should do what you need :-)
@FabianIsensee Thanks Fabian! Did you have a look into the PR we opened and my comment on reproducibility? Is your predict_from_files_sequential used automatically when workers are set to 0?