Feature/mesh motion #23
Conversation
* Applied the smartsim mesh motion solver to wing motion. * Reduce the number of epochs for time steps > 1, increased learning rate. * Cleaned up mesh-motion notebook.
Split trainer file and generate entities
|
@tmaric any reason why this has not been merged? I would have no objections to it. |
| @@ -3,6 +3,6 @@ cd "${0%/*}" || exit # Run from this directory | |||
| . ${WM_PROJECT_DIR:?}/bin/tools/RunFunctions # Tutorial run functions | |||
| #------------------------------------------------------------------------------ | |||
|
|
|||
| runApplication checkMesh -constant -writeAllFields | |||
| runApplication mpirun -np 4 checkMesh -constant -writeAllFields -parallel | |||
There was a problem hiding this comment.
Why not use runParallel?
There was a problem hiding this comment.
@tmaric I checked the notebook. I am not sure about the workaround using subprocess. The point of using SmartSim is that it manages resources and jobs. When using subprocess, we take resource management and exception handling into our own hands. Also, when running the notebook on a cluster, the preprocessing and cleanup would be executed on the login node. I'd prefer a solution employing the SmartSim API all the way through, which requires dealing with the subfolder issue. I am not yet sure how to deal with the new model-specific subfolders. One way would be to give all the models the same name and use overwrite=True. However, that workaround would create unnecessary copy operations, and I guess it does not work if an app modifies an existing dictionary (e.g., for setting an inflow profile). Maybe @al-rigazzi can give a hint on how to disable or work around the subfolder creation. Should we use the API at a lower level?
There is also a typo in the first print statement after running Allclean.
related to #29
There was a problem hiding this comment.
I checked the documentation of Experiment.create_model(), and the solution seems to be quite simple. One can use exp.create_model(..., path="path/to/simulation") to obtain the same behavior as before (not yet tested, though).
There was a problem hiding this comment.
The subprocess is not the workaround for the issue of copying config files into the model sub-directory introduced in smartsim 0.8.0.
The subprocess was used to separate pre-processing from the CFD+ML algorithm. The pre processing requires blocking mode, and CFD+ML non-blocking, this was the reason for introducing it originally. Additionally, more complex pre-processing may be necessary, which the user doesn't want to implement into a SmartSim workflow. For example, wingMotion requires simpleFoam to be run in a case before CFD+ML algorithm is started, and for that run there is no need to use smartredis at all.
That was the motivation, but I agree 100% that a consistent way would be to implement everything in a SmartSim workflow.
For that, I think we just need two experiments because of blocking/non blocking requirement: pre processing experiment and running experiment. We add meshing, initialization and decomposition as blocking steps in the preprocessing experiment. We run the CFD+ML experiment as is. This should be sufficient and consistent, without having to use any workarounds.
does this sound OK to you?
There was a problem hiding this comment.
Creating multiple Experiments would be one solution, but it's also possible to use multiple different models within one Experiment (example). You can decide to use blocking/nonblocking for each model individually.
There was a problem hiding this comment.
Then this is the best option IMO, I will try this first, alongside providing the folder argument to the model to avoid the 'to_copy' and 'generate' lines.
There was a problem hiding this comment.
path (Optional[str], default: None) – path to where the Model should be executed at runtime
This is not the same as of_model.attach_generator_files(to_copy="spinningDisk") copying files into the model sub-folder. FYI @AndreWeiner
https://www.craylabs.org/docs/api/smartsim_api.html#smartsim.experiment.Experiment.create_model
Working MLP mesh deformation for OpenFOAM's wing2D_pimpleFoam tutorial.