-
Hi, Could you please advise how to load the trained local models (at the 4 clients) and the trained global model in the prostate example and test them on a new testing dataset? I can see there are 2 folders "local_model.pt" and "best_local_model.pt" under the below rout for the MSD client and run_2 (fedavg experiment). Each of these 2 folders includes an "archive" folder that has "data" folder (contains 63 unknown files) and a "data.pkl" file. examples/prostate/workspace_prostate/client_MSD/run_2/app_client_MSD/ Could you please illustrate further about the content in the 2 aforementioned folders above and how to load the trained models with the updated/best weights? |
Beta Was this translation helpful? Give feedback.
Replies: 2 comments 1 reply
-
@atnassar, thanks for the question! As our other discussion, this question is more about Pytorch/MONAI, rather than NVFlare: you can load the .pt models into any pytorch script that is able to recognize and use the model (that's why I suggested to write a standalone MONAI trainer). "local_model.pt" is used to track the local model trained on the client, and "best_local_model.pt" is the record for the one with best validation score. |
Beta Was this translation helpful? Give feedback.
-
@atnassar, did this answer your question? |
Beta Was this translation helpful? Give feedback.
@atnassar, thanks for the question! As our other discussion, this question is more about Pytorch/MONAI, rather than NVFlare: you can load the .pt models into any pytorch script that is able to recognize and use the model (that's why I suggested to write a standalone MONAI trainer). "local_model.pt" is used to track the local model trained on the client, and "best_local_model.pt" is the record for the one with best validation score.
As for the "archive" folder, to be honest I did not see it in my experiments, @YuanTingHsieh do you have any idea?