The combination of pytorch and NVFlare result in training output (trained models, prediction csvs) at different locations (checkpoints for the local models are saved in the scratch dir; FL_global_model.pt and best_FL_global_model.pt are saved in the project dir). Output to be preserved should not be located in a scratch directory (danger of "temporary" location being cleaned up) and ideally all output should be at one location.
- Decide on single location.
- Move all output there as part of local/swarm training run.
- Make sure all downstream scripts (collecting results, evaluation, …) use the updated location.
The combination of pytorch and NVFlare result in training output (trained models, prediction csvs) at different locations (checkpoints for the local models are saved in the scratch dir;
FL_global_model.ptandbest_FL_global_model.ptare saved in the project dir). Output to be preserved should not be located in a scratch directory (danger of "temporary" location being cleaned up) and ideally all output should be at one location.